Mechanical Systems and Signal Processing 226 (2025) 112320

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Invited for the special issue in honor of professor John Mottershead

A general probabilistic framework for impact localisation based on

flexural wave propagation
∗
Dong Xiao , Francisco de Sá Rodrigues , Zahra Sharif-Khodaei , M.H. Aliabadi
Department of Aeronautics, Imperial College London, South Kensington, London SW7 2AZ, United Kingdom

ARTICLE INFO ABSTRACT

Communicated by Z. Mao Flexural wave propagation-based impact localisation methods for composite structures face
challenges in accurately estimating the group velocity profile (GVP) and quantifying localisation
Keywords:
uncertainties due to wave dispersion. This paper introduces a two-step probabilistic framework
Structural health monitoring
Impact localisation
to address these challenges. In the first step, a probabilistic model for GVP estimation is
Flexural wave propagation developed using Bayesian inference, leveraging wave dispersion relations from classical laminate
Group velocity profile theory to define the GVP probability space. Three GVP estimation methods depending on
Bayesian inference structural complexity and available knowledge are explored: physics-based, data-driven, and
hybrid physics-data method. The estimated GVP facilitates multi-frequency probabilistic impact
localisation in the second step, wherein the maximum likelihood impact location is identi-
fied solely using fast gradient-based optimisation. Experimental impact testing on laminated
composite flat, stiffened, and sandwich panels validates the framework, demonstrating its
efficiency, accuracy, and scalability in GVP estimation, impact localisation, and uncertainty
quantification across diverse composite structures. The data-driven method notably requires
only sparse reference impacts and a limited sensor network (each with no more than four) to
accurately estimate GVPs for these structures.

1. Introduction

Composite aerostructures are prone to a variety of impact events, such as dropped tools during maintenance, flying debris during
takeoff and landing, hail strikes, and collisions with ground vehicles [1]. These impacts can cause barely visible impact damage that
may compromise structural integrity. Real-time impact localisation is crucial, as it provides spatial information on potential damage
for further assessment, enabling more effective maintenance strategies and reducing repair costs.
Impact localisation in anisotropic composite aerostructures poses an inversion problem, aiming to determine the impact location
based on available structural local response data measured by structural health monitoring (SHM) passive sensing [2]. This inversion
problem is commonly addressed through impact-induced wave propagation analysis [3,4], leveraging the fundamental physical
relationship that the distance travelled by a wave equals the product of its velocity and time. The wave propagation time difference
can be estimated from the impact sensor signals, also known as the Time Difference of Arrival (TDOA). Based on the estimated
TDOA, impact location can be pinpointed using triangulation [5,6], provided the wave group velocity profile (GVP) is known, or
through a data-driven approach utilising TDOA as training feature [7]. To ensure accurate and reliable impact localisation, it is
crucial to focus on analysing the impact-induced wave propagation in laminated composites and precisely estimating the TDOA of
these waves from the sensor signals.

∗ Corresponding author.
E-mail addresses: [email protected] (D. Xiao), [email protected] (F.d.S. Rodrigues), [email protected] (Z. Sharif-Khodaei),
[email protected] (M.H. Aliabadi).

https://doi.org/10.1016/j.ymssp.2025.112320
Received 6 June 2024; Received in revised form 11 December 2024; Accepted 3 January 2025
Available online 16 January 2025
0888-3270/© 2025 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

The impact dynamic analysis reveals that impacts induce both in-plane waves and out-of-plane flexural waves in laminated
structures [8,9]. These waves exhibit multiple modes and display dispersive properties [10,11], wherein the wave velocities are
functions of the wave frequency. These dispersive properties emphasise the significance of wave frequency in both GVP and TDOA
estimation. Based on the type of waves utilised, the methods for impact localisation can be generally categorised into in-plane
extensional wave-based and out-of-plane flexural wave-based approaches.
Significant attention has been devoted to employing these two types of waves for impact localisation. Due to the high-
frequency nature of extensional waves in impact dynamics, observing these waves typically requires a high sampling frequency [12].
Capineri [13] and Bulletti [14] typically designed an analogue front-end to capture these high-frequency extensional waves for
impact localisation. The sampling frequency required to capture extensional waves typically exceeds several megahertz in their
designs. However, the impact experiments [13,15] demonstrated that in-plane extensional waves are not detectable for low-energy,
low-velocity impacts, including oblique 45-degree impacts [16]. The low-amplitude characteristics of the extensional waves limit
their applicability for impact localisation in low-velocity scenarios.
Flexural waves are suitable for impact localisation due to their effectiveness across various impact scenarios, regardless of
the impact velocity. The popularity of data-driven methods for impact localisation, utilising TDOA extracted from flexural waves
as training feature, has grown, as they do not require information regarding structural configurations and material properties.
These data-driven models include constructing neural networks [17–20], Gaussian process regression [21,22] based on training
impact data. However, the accuracy and generalisability of data-driven approaches heavily depend on the size and diversity of
the training data. Collecting experimental or numerical training impacts across various impact scenarios can be costly, although
multi-fidelity modelling allows for savings in high-fidelity experimental training impacts [23,24]. Additionally, surrogate-based
data-driven approaches struggle to localise impacts outside the coverage area of the training impacts due to their limited exploration
capabilities [25]. In contrast, impact localisation using triangulation with TDOA and GVP of flexural waves demonstrates superior
generalisability.
To make the triangulation solvable, two main strategies have been explored: sensor clustering, which assumes a uniform group
velocity from the impact point to a cluster of sensors [26,27], and pre-determined GVP, which is obtained in advance [28–30]. While
the sensor clustering approach is resource-intensive, requiring a significant number of sensors and extensive impact data acquisition,
pre-measuring the GVP offers a resource-efficient alternative. Since GVP is structure-dependent but impact-independent, it can be
applied universally across different impacts, regardless of variations in impact mass or velocity, further enhancing its practicality.
For impact localisation using triangulation with pre-determined GVP, the GVP of flexural waves can be obtained through various
methods, including analytical derivation [10,31,32], numerical simulation [33,34], experimental measurement [28,35–37], and
calibration based on parameterised GVP models [25,30]. Notably, Chow [31] extended Timoshenko’s beam theory [38] to cross-ply
laminated plates by incorporating a single shear deformation displacement, enabling the analytical derivation of dispersion relations
governing flexural wave propagation. To account for the different rotations of normals to the midplane about the 𝑥 and y axes,
Abrate [10] and Jeong [32] built on Mindlin plate theory [39] to derive more comprehensive dispersion relations for angle-ply
laminated plates. However, these studies did not extend their analysis to group velocity derivation from the spectrum relations.
In contrast, numerical approaches, such as finite element methods [33] and spectral finite element simulations [34], allow for the
accurate measurement of GVP in more complex structures, including stiffened configurations. Deng [33] conducted a numerical
comparison between the GVPs of stiffened and flat laminated plates, demonstrating that the maximum velocity deviations induced
by stiffeners in all directions are less than 10%. These numerical methods are particularly advantageous when analytical solutions
become inaccurate and impractical due to geometric or material complexities. Experimental techniques offer another versatile,
though often costly, approach for measuring GVP. Meo [35] applied wavelet transforms to extract TDOA data from impact sensor
signals and estimated GVPs for flexural waves at frequencies below 10 kHz. For higher frequency measurements, Kirkby [36]
highlighted the importance of distinguishing between in-plane and flexural waves for accurate measurements. Calibration methods
offer a more efficient approach by combining the physical principles of analytical derivation with the precision of experimental
measurements, significantly reducing the amount of required experimental data. For example, Zhu [30] calibrated GVPs using
a parameterised laminate model with minimal reference impacts. Similarly, Xiao [25] parameterised GVPs based on dispersion
relations derived from first-order shear deformation theory (FSDT) [40], and refined them through fast gradient-based optimisation
with limited impact data. These GVP estimation approaches contribute to advancing the accuracy and reliability of flexural
wave-based impact localisation methods.
However, it can be concluded that, aside from experimental measurements, other methods rely on reasonably accurate knowledge
of structural configurations and material properties—information that is often difficult to obtain for complex composite structures.
Furthermore, experimental approaches generally require a large number of precisely placed sensors and strategically positioned
impact excitations. For example, tens of sensors may need to be arranged along the circumference of a circle, with the impact
excitation located at its centre [24,28,37]. Alternatively, sensors can be evenly distributed and multiple impact excitations can be
carefully aligned with any two passive sensors to estimate the wave velocity along the line [35]. These demands make experimental
setups both time-consuming and resource-intensive. Therefore, developing efficient GVP estimation methods that minimise the
number of required experimental excitations and sensors, regardless of their placements, is essential to improve the practicality
of GVP measurements.
Moreover, achieving scalability in GVP estimation methods across different structural levels — from flat panels to single-stiffener
panels, and further to large, metre-scale stiffened panels — can further improve the GVP estimation efficiency. For this scalability
to be successful, the material properties or GVP of lower-level structures must be seamlessly integrated into the estimation of more
complex, higher-level structures that incorporate these components. This necessitates a GVP updating or calibration method that

2
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

transfers and integrates knowledge from simpler, lower-level structures to more complex, higher-level ones, ensuring accurate GVP
estimation as structural complexity increases while simultaneously reducing costs.
It is also noteworthy that most of the existing literature on wave-based impact localisation follows a deterministic approach [18–
20,25–27,30,33–37,41–43], often neglecting the role of uncertainties. Quantifying these uncertainties is crucial for providing more
reliable predictions, especially when using sparse passive sensor networks. While methods like Gaussian process regression [21,22]
have been applied in a data-driven manner, they often lack generalisability for impact localisation outside the reference impact
area. The Kalman filter [28,29] is another widely used technique for uncertainty quantification in wave-based impact localisation.
However, Kalman filter assumes that impact location, TDOA, and GVP are independent Gaussian variables, which contradicts
physical principles since these factors are correlated. These limitations highlight the need for a more robust impact localisation
method that accounts for uncertainties in both GVP and TDOA while adhering to the physical principles.
This study focuses on optimising resource usage in sensor deployment and data acquisition for impact localisation by employing
a sparse sensor network. To address challenges associated with sparse sensor networks, such as accurate GVP estimation and
uncertainty quantification, a two-step probabilistic framework based on flexural wave propagation is proposed. In the first step, a
Bayesian model for GVP estimation is developed, leveraging wave dispersion relations from classical laminate theory (CLT) [40,44]
to define the GVP probability space. Three GVP estimation methods are incorporated, tailored to structural complexity and available
data: a physics-based (PB) method that analytically derives the GVP based on FSDT; a data-driven (DD) method that estimates GVP
using sparse reference impact data; and a hybrid physics-data (HPD) approach, which enhances scalability by enabling rapid GVP
estimation for complex structures using reference GVP data from simpler components.
In the second step, the estimated GVP is used in multi-frequency probabilistic impact localisation. Within the Bayesian framework,
the posterior distribution of impact locations is efficiently determined using fast gradient-based optimisation and Markov Chain
Monte Carlo (MCMC) [45,46], providing reliable localisation while quantifying uncertainties. Experimental impact tests on a flat
composite panel, a stiffened composite panel, and a sandwich composite panel validate the framework, showing it to be accurate
and efficient in GVP estimation and impact localisation, even for impacts outside sensor-covered regions. For stiffened and sandwich
panels, where GVP cannot be analytically derived, the framework efficiently enabled probabilistic GVP estimation and impact
localisation with minimal sensor and reference data, underscoring its versatility across complex structures.
Key contributions of this study include:

• A general probabilistic framework for impact localisation on composite structures, utilising Bayesian inference based on
flexural wave propagation.
• Development of three GVP estimation methods tailored to structural complexity and available information, where the data-
driven method effectively estimates GVP with sparse reference impacts and a minimal sensor network and the hybrid
physics-data method increases the scalability of GVP estimation.
• Multi-frequency analysis allowing for the detection and exclusion of TDOA outliers in impact localisation, thereby improving
localisation reliability.
• Validation through impact testing on laminated composite flat, stiffened, and sandwich panels, demonstrating the framework’s
accuracy, scalability, and efficiency in GVP estimation, impact localisation, and uncertainty quantification across diverse
composite structures.

The remainder of this paper is organised as follows: Section 2 derives the dispersion relation from CLT used for defining the
probability space of the GVP. In Section 3, the relation between impact location, TDOA, and GVP is presented, followed by the
introduction of the probabilistic framework for GVP identification and impact localisation. As GVP is critical in impact localisation,
Section 4 develops methods for GVP estimation based on three practical scenarios of structural complexity and available knowledge.
Using the estimated GVP, Section 5 introduces the probabilistic impact localisation method using Bayesian inference. Section 6
outlines the experimental setup for impact testing on three different composite panels and evaluates three TDOA extraction methods
based on sensor signals. The results of the impact localisation for each composite panel are presented and analysed in Section 7.
Finally, Section 8 summarises the findings and outlook of this study.

2. Impact-induced dispersive flexural waves

This section begins with deriving the dispersion relations of impact-induced flexural waves in laminated composites using
classical laminate theory (CLT). The dispersion relations subsequently assist in defining the parameter space of the group velocity
profile (GVP) of flexural waves.

2.1. Dispersion relations of the impact-induced flexural waves

To model the out-of-plane responses of laminated composites subjected to low-velocity impacts, this study assumes a linear
structural response, and a perpendicular point impact on the plate surface. Additionally, perfect bonding between lamina layers is
assumed, with each lamina behaving as a linear elastic orthotropic material [44,47]. Since symmetrical laminates are widely used
in practical applications [10], the focus is on symmetrically laminated composites, which does not impose a significant limitation.
In a symmetrically laminated plate, in-plane and out-of-plane motions are decoupled. The governing equation for out-of-plane

3
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 1. GVP basis functions and GVP realisations: (a) positive basis functions 𝜙1 (𝜃), 𝜙3 (𝜃) and 𝜙5 (𝜃), (b) basis functions 𝜙2 (𝜃) and 𝜙4 (𝜃), (c) GVP realisations by
setting wave coefficient 𝐜. Basis functions 𝜙1 (𝜃) ≥ 0, 𝜙3 (𝜃) ≥ 0 and 𝜙5 (𝜃) ≥ 0, resulting GVP 𝑣𝑔 (𝜃) monotonically increases with velocity coefficients 𝑐1 , 𝑐3 , and 𝑐5 .
In contrast, velocity coefficients 𝑐2 and 𝑐4 affect the shape of the GVP, since 𝜙2 (𝜃) and 𝜙4 (𝜃) are positive in quadrants 1 and 3, but negative in quadrants 2 and
4.

displacement under a point load can thus be derived using CLT (assuming zero strains 𝜖𝑧𝑧 , 𝜖𝑥𝑧 , and 𝜖𝑦𝑧 ), resulting in a fourth-order
partial differential equation (PDE) [10,44,48]:
2
( ) ( )
𝜕2 2 2 2 2 2 2 2
𝜌ℎ 𝜕𝜕 𝑡𝑤
2 + 𝜕 𝑥2 𝐷11 𝜕𝜕 𝑥𝑤2 + 𝐷12 𝜕𝜕𝑦𝑤2 + 2𝐷16 𝜕𝜕𝑥𝜕𝑤𝑦 + 𝜕𝜕𝑦2 𝐷12 𝜕𝜕 𝑥𝑤2 + 𝐷22 𝜕𝜕𝑦𝑤2 + 2𝐷26 𝜕𝜕𝑥𝜕𝑤𝑦
( ) (1)
𝜕2 2 2 2
+2 𝜕 𝑥𝜕 𝑦
𝐷16 𝜕𝜕 𝑥𝑤2 + 𝐷26 𝜕𝜕 𝑦𝑤2 + 2𝐷66 𝜕𝜕𝑥𝜕𝑤𝑦 = 𝑝(𝑡)𝛿𝑘 (𝑥 − 𝑥∗ , 𝑦 − 𝑦∗ ).

In the governing equation, 𝛿𝑘 represent Kronecker delta, 𝑝(𝑡) denotes the force applied at the impact point located at 𝑀 = (𝑥∗ , 𝑦∗ ),
while 𝜌 and ℎ represent density and plate thickness, respectively. 𝐷𝑖𝑗 refers to the elements of the bending stiffness matrix 𝐃, and
𝑤 indicates the out-of-plane displacement. When no external force is applied (𝑝(𝑡) = 0), this PDE yields a wave solution of the form
𝑤(𝑟, 𝑡) = 𝑅𝑒[𝐴𝑒𝑖𝑘(⃗𝑛⋅⃗𝑟−𝑣𝑝 𝑡) ], where 𝑛⃗ = cos 𝜃 ⃗𝑖 + sin 𝜃 𝑗,
⃗ and 𝑟⃗ = 𝑥⃗𝑖 + 𝑦𝑗.
⃗ Here, 𝜃 is the phase velocity direction, 𝑖 denotes the imaginary
unit, 𝑘 and 𝑣𝑝 represent the wave number and wave phase velocity, respectively, and the wave frequency 𝜔 is the product of wave
number and wave phase velocity (𝜔 = 𝑘𝑣𝑝 ). The dispersion relations of this flexural wave in the direction of 𝜃 can be derived by
substituting this waveform into the governing PDE Eq. (1) as follows [10]:
𝑣4𝑝 (𝜃)
𝜌ℎ 𝜔2
= 𝐷11 cos4 𝜃 + 4𝐷16 sin 𝜃 cos3 𝜃 + 2(𝐷12 + 2𝐷66 ) sin2 𝜃 cos2 𝜃 + 4𝐷26 sin3 𝜃 cos 𝜃 + 𝐷22 sin4 𝜃 . (2)

It is noted that the wave phase velocity profile 𝑣𝑝 (𝜃) is dependent on the structural bending stiffness 𝐷𝑖𝑗 and wave frequency 𝜔.
Since the wave phase velocity 𝑣𝑝 is directly proportional to the square root of the frequency 𝜔, this flexural wave demonstrates
dispersive characteristics. Specifically, flexural waves of higher frequencies travel faster than those of lower frequencies.
For flat elements, the phase velocity direction aligns with the group velocity direction [49]. This allows the group velocity 𝑣𝑔 (𝜃)
of the flexural waves to be further derived as (consistent with the group velocity derivation of orthotropic plates on pages 65, 66,
116 in [10]):
𝜕𝜔 𝜕[𝑘𝑣𝑝 (𝜃)] 𝑑 𝑣𝑝 (𝜃)
𝑣𝑔 (𝜃) = = = 𝑣𝑝 (𝜃) + 𝑘 = 2𝑣𝑝 (𝜃), (3)
𝜕𝑘 𝜕𝑘 𝑑𝑘

which is equal to two times of the phase velocity. Since the TDOA extracted from sensor signals measures the arrival time of the
waves corresponding to their group velocity, the group velocity is the primary focus of this study.

2.2. Parameterisation of the flexural wave group velocity profile based on dispersion

Based on relation Eq. (3), the dispersion relations depicted in Eq. (2), serve as a parameterisation method for the group velocity
profile (GVP), which represents the GVP 𝑣𝑔 (𝜃) with the five basis functions 𝜙𝑗 (𝜃) = sin(𝑗−1) 𝜃 cos(5−𝑗) 𝜃 , 𝑗 = 1, 2, … , 5 and their
corresponding coefficient 𝑐𝑗 :
∑
𝑣4𝑔 (𝜃) = 16𝑣4𝑝 (𝜃) = 5𝑗=1 𝑐𝑗 𝜙𝑗 (𝜃), 𝜙𝑗 (𝜃) = sin(𝑗−1) 𝜃 cos(5−𝑗) 𝜃 ,
(4)
16𝐷11 𝜔2 64𝐷16 𝜔2 32(𝐷12 +2𝐷66 )𝜔2 64𝐷26 𝜔2 16𝐷22 𝜔2
𝑐1 = 𝜌ℎ , 𝑐2 = 𝜌ℎ , 𝑐3 = 𝜌ℎ
, 𝑐4 = 𝜌ℎ , 𝑐5 = 𝜌ℎ .

The GVP basis functions are plotted in Fig. 1(a) and (b). Notably, 𝜙1 (𝜃) ≥ 0, 𝜙3 (𝜃) ≥ 0 and 𝜙5 (𝜃) ≥ 0 in all wave direction 𝜃,
which results in a monotonic increase in the GVP 𝑣𝑔 (𝜃) with respect to the coefficients 𝑐1 , 𝑐3 and 𝑐5 . In contrast, the coefficients
𝑐2 and 𝑐4 primarily affect the shape of the GVP without altering the area under it. The area, represented by the velocity integral
2𝜋
𝐴𝑣 = ∫0 𝑣𝑔 (𝜃)𝑑 𝜃 across all angles 𝜃, remains constant regardless of variations in 𝑐2 and 𝑐4 . This constancy is illustrated in Fig. 1(b),
where the positive wave velocity of 𝜙2 (𝜃) and 𝜙4 (𝜃) in quadrants 1 and 3 offsets the negative velocity in quadrants 2 and 4, keeping
𝐴𝑣 unchanged.
By setting 𝑐3 = 2𝑐1 = 2𝑐5 , and 𝑐2 = 𝑐4 = 0, a circular GVP is generated, as shown in Fig. 1(c), where the ‘Ref’ circular GVP
meets this condition. Doubling 𝑐1 and 𝑐5 in ‘Ref’ (indicated by ‘2𝑐1 ’ and ‘2𝑐5 ’ in the figure legend) correspondingly increases 𝑣𝑔 (0)

4
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

and 𝑣𝑔 (𝜋∕2). Additionally, raising 𝑐2 and 𝑐4 from zero to the value of 𝑐1 in ‘Ref’ (represented as ‘𝑐2 , 𝑐4 ’ in the figure legend) results
in increased wave velocity in quadrants 1 and 3 and decreased velocity in quadrants 2 and 4.
Every realisation of velocity coefficients corresponds to a GVP. Given that wave velocity is greater than zero, the parameter
space of the velocity coefficients is defined by positive GVP:
∑
𝑣4𝑔 (𝜃) = 5𝑗=1 𝑐𝑗 𝜙𝑗 (𝜃) > 0,
(5)
𝜙1 (𝜃) ≥ 0, 𝜙3 (𝜃) ≥ 0, 𝜙5 (𝜃) ≥ 0, 𝑐1 > 0, 𝑐3 > 0, 𝑐5 > 0, 𝑐2 ≥ 0, 𝑐4 ≥ 0.
The velocity coefficients 𝑐2 and 𝑐4 may be negative for certain principal material directions. However, by rotating the principal
material direction and coordinate axes transformation during the calculation of GVP, these coefficients can achieve non-negative
values [44]. This study advocates for selecting principal material directions that guarantee non-negative velocity coefficients for
parameterising and computing the GVP. Detailed derivations of the parameter space for velocity coefficients corresponding to
positive GVPs are provided in Appendix A.
It is essential to recognise that the GVP 𝑣𝑔 (𝜃) within a structure is fundamentally governed by the structure’s inherent properties
and remains unaffected by external impacts as long as no damage occurs. This highlights the high generalisability of GVP for impact
localisation across varying impact conditions. Upon a damaging impact, wave propagation initiates at the point of contact, before
any damage is present. Consequently, the initial segment of the impact-induced wave reflects the characteristics of wave propagation
in an undamaged structure. However, considering the viscoelastic properties of laminated composites, this segment may also include
non-propagating wave modes with complex wave numbers that decay with distance. By analysing this initial segment, it is possible
to extract the TDOA of the propagating flexural wave mode, which can then be used to locate the impact with the known GVP.

3. Wave propagation-based impact localisation

3.1. Impact location relation to TDOA and wave velocity

Based on wave propagation analysis, the physical relation between the impact location (𝑥, 𝑦), sensor locations (𝑥𝑗 , 𝑦𝑗 ), 𝑗 =
1, 2, … , 𝑁𝑠 , and the GVP 𝑣𝑔 (𝜃), along with the TDOA, is given as [6,25]:
√ √
(𝑥−𝑥𝑗 )2 +(𝑦−𝑦𝑗 )2 (𝑥−𝑥1 )2 +(𝑦−𝑦1 )2
𝛥𝑡1𝑗 = 𝑡𝑗 − 𝑡1 = − , (6)
𝑣𝑔 (𝜃𝑗 ) 𝑣𝑔 (𝜃1 )

where 𝛥𝑡1𝑗 represents the TDOA between the 𝑗th sensor and the 1-st sensor, and 𝑣𝑔 (𝜃𝑗 ) and 𝑣𝑔 (𝜃1 ) denote the wave propagation group
velocity from the impact point to the 𝑗th sensor and the 1-st sensor, respectively. The TDOA for all the passive sensors, numbering
𝑁𝑠 , is based on anchor sensor 1. Therefore, the TDOA for an impact on composite structures is represented as a vector of size 𝑁𝑠 :
𝛥𝐭 = [0, 𝛥𝑡12 , 𝛥𝑡13 , ⋯ , 𝛥𝑡1𝑁𝑠 ]𝑇 . (7)

In real-world applications, sensor locations (𝑥𝑗 , 𝑦𝑗 ), 𝑗 = 1, 2, … , 𝑁𝑠 are known from the instalment process, and TDOA values
are directly estimated from the impact sensor signals. For anisotropic structures, this results in 𝑁𝑠 − 1 nonlinear equations Eq. (6),
but 𝑁𝑠 + 2 unknowns remain: 2 impact location coordinates 𝑥, 𝑦 and 𝑁𝑠 group velocities 𝑣𝑔 (𝜃𝑗 ), 𝑗 = 1, 2, … , 𝑁𝑠 . Therefore, the
triangulation system is underdetermined and not directly solvable. For a sparse sensor network, the pre-determined GVP of flexural
waves makes the triangulation system determined (three sensors used) [28], even overdetermined (number of sensors greater than
3).

3.2. Probabilistic impact localisation framework based on wave propagation

In practice, the TDOA estimated from impact sensor signals inherently contains uncertainties arising from measurement errors
(such as sensor noise and environmental factors) and simulation inaccuracies (due to limitations in the TDOA extraction methods).
The distribution of TDOA is influenced by the kernel in the extraction method and how it handles the measurement errors.
Assuming that sensor noise, a primary source of measurement error, predominantly affects high-frequency signals, extracting TDOA
for low-frequency waves for impact localisation using techniques like continuous wavelet transform (CWT) [29] and Butterworth
filtering [20] significantly mitigates the impact of measurement errors. Additionally, Gaussian kernels are often assumed in the TDOA
extraction methods. For instance, Dehghan-Niri [29] modelled TDOAs as mutually independent Gaussian variables, accounting for
the Heisenberg uncertainty introduced by CWT during TDOA extraction. Other studies [24,28] also modelled TDOAs as independent
Gaussian variables, achieving successful impact localisation without explicitly discussing the uncertainty sources.
To investigate TDOA distribution experimentally, this study conducted 200 impacts at the same location on a composite plate
using a guided drop mass, with a setup similar to that shown in Fig. 4(a). The drop height was varied randomly between 1 and 4 mm
through hand-lifting, and four PZT sensors recorded the resulting impact signals at a sampling frequency of 2 MHz. TDOAs were
extracted from these signals using the Normalised Smoothed Envelope Threshold (NSET) method [20,23], which has demonstrated
high accuracy and consistency in TDOA extraction (see Section 6.2) and was used for impact localisation throughout this study.
Fig. 2 illustrates the normal probability plot of the TDOA data (shown as markers), where TDOAs for each sensor align closely
along a straight line, suggesting a high likelihood of normal distribution. The similar slopes of these lines indicate comparable
TDOA variances across sensors. This similarity in TDOA variance is mainly due to the TDOA being extracted by a common method
from sensor signals with a common sampling frequency and similar noise levels for each impact.

5
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 2. Normal probability plot of TDOAs extracted by the NSET method at frequencies: (a) 2 kHz, (b) 5 kHz. The TDOAs for each sensor align closely along
a straight line, suggesting a high likelihood of normal distribution. The similar slopes indicate comparable TDOA variance across the sensors.

Table 1
Group velocity profile (GVP) estimation methods and their applications and features.
Method Structural complexity Available knowledge Characteristics
PB Flat plates Structural configuration & material Using plate theories (CLT, FSDT)
properties
DD Complex structures Reference impact data GVP parameterised by dispersion relation from CLT;
Hybrid GA for optimisation
HPD Complex structures with a simpler Reference impact data & GVP of simpler component can be from PB method;
component GVP of simpler component Gradient-based optimisation

It is important to note that this study assumes the independence of TDOA data, as the TDOAs were estimated for independent
impacts with stationary sensors. However, in scenarios where this assumption does not hold — such as with moving receivers [50]
or dynamic tracking targets — alternative approaches, such as the Autoregressive Integrated Moving Average (ARIMA) model and its
variants [51], may be required. These methods effectively account for temporal correlation structures, ensuring the model accurately
captures the dependencies introduced by dynamic conditions.
Based on these assumptions and experimental validation, TDOAs in this study are modelled as mutually independent Gaussian
variables. Accordingly, Eq. (6) is reformulated to incorporate TDOA uncertainties modelled as Gaussian variables as follows:
√ √
𝑑 𝑑 (𝑥−𝑥𝑗 )2 +(𝑦−𝑦𝑗 )2 (𝑥−𝑥1 )2 +(𝑦−𝑦1 )2
𝛥𝑡1𝑗 = 𝑡𝑗 − 𝑡1 + 𝛿 = 𝑗 − 1 + 𝛿 = − + 𝛿, (8)
𝑣𝑔 (𝜃𝑗 ) 𝑣𝑔 (𝜃1 ) 𝑣𝑔 (𝜃𝑗 ) 𝑣𝑔 (𝜃1 )

where 𝛿 is a zero-mean Gaussian variable with variance 𝜎𝛿2 . Consistent with finds in Fig. 2, the variance of 𝛥𝑡1𝑗 , where 𝑗 = 2, … , 𝑁𝑠 ,
is assumed to be the same across sensors.
Additionally, GVP estimations, particularly those measured by experiments, also contain uncertainties due to material variability
and measurement limitations. This study focuses on quantifying the uncertainties in both GVP and TDOA and localising impacts
while accounting for these uncertainties. Fig. 3 presents impact localisation problem formulation and the developed two-step
probabilistic framework for wave propagation-based impact localisation, incorporating the uncertainties in both GVP and TDOA.
The first step focuses on estimating the GVP and quantifying its uncertainties. Three different cases are considered for GVP
estimation based on the structural complexity and the available knowledge: (1) plate structures with known material properties
(physics-based), (2) complex structures with only experimental or numerical reference impact data (data-driven), and (3) complex
structures with reference impact data as well as GVP information from a lower-level structure that is part of the complex structure
(hybrid physics-data). Using GVP parameterisation Eq. (5) and Bayesian inference, three corresponding GVP estimation methods
are developed for these cases: the physics-based (PB), data-driven (DD), and hybrid physics-data (HPD) approaches.
The second step involves probabilistic impact localisation using Bayesian inference, based on the previously estimated GVP.
To enhance robustness, the localisation is conducted in a multi-frequency manner. Additionally, the GVP parameterisation enables
efficient maximum likelihood localisation through gradient-based optimisation. The following sections, Sections 4 and 5, focus on
detailing the GVP estimation methods and the probabilistic impact localisation process within this two-step framework, respectively.

4. Group velocity profile (GVP) estimations

Three GVP estimation methods are developed to address varying levels of structural complexity and available information: the
physics-based (PB), data-driven (DD), and hybrid physics-data (HPD) approaches. A summary of the applicable structures, required
information, and key characteristics for each method is provided in Table 1.

6
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 3. Probabilistic impact localisation framework based on flexural wave propagation. This probabilistic framework accounts for the uncertainties in TDOA
and GVP, and consists of two main steps: probabilistic GVP estimation and probabilistic impact localisation.

These three methods exhibit a high degree of interaction. The PB method, based on plate theories, applies solely to flat plate
structures. As outlined in Section 2.2, the GVP for the DD method is parameterised by the dispersion relation derived from the CLT,
a plate theory also foundational to the PB approach, particularly for thin plates. The HPD method integrates both the PB and DD
approaches: the PB method provides a reference GVP derived from a simpler component of the target complex structure, enhancing
the efficiency of the DD estimates. The following subsections present detailed derivations for each of these methods.

4.1. Physics-based (PB) approach

For plate-like structures with known material properties, the GVP of flexural waves can be derived analytically using plate
theories or calculated numerically through semi-analytical FEA, such as the SAFEDC freeware [52], for specific wave frequencies.
Experimental studies have demonstrated [53–55] that CLT provides accurate GVP estimation at very low frequencies in thin
plates, while shear deformation theory is effective at relatively higher frequencies for moderately thick plates. This highlights the
importance of selecting an appropriate wave frequency to ensure accurate GVP estimates with CLT or FSDT.
In practical applications, impact dynamics often involve loading and unloading phases with finite contact duration and simple,
positive force histories (e.g., half-sine) [56,57]. By the Paley-Wiener-Schwartz theorem [58,59], the Fourier–Laplace transform
amplitudes of such finite-duration impact forces are bounded and decay with increasing frequency. As a result, the dominant
impact-induced waves are at low frequencies where plate theories can accurately estimate wave velocities.

4.1.1. GVP estimation based on CLT

Eqs. (2) and (3) illustrate the dispersion relations derived from the CLT for symmetrically laminated composite plates. However,
CLT tends to overestimate the wave velocity [10,25]. Both the phase and group velocities are unbounded as the frequency approaches
infinity. Consequently, the accuracy of the GVP diminishes as the frequency increases. Moreover, achieving a desired level of GVP

7
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

accuracy necessitates a decrease in wave frequency with an increase in plate thickness. Therefore, if GVPs derived from CLT are
utilised for impact localisation, it is advisable to base them on lower frequencies.

4.1.2. GVP estimation based on FSDT

In contrast, the first-order shear deformation theory (FSDT) or even higher-order plate theories, which consider shear deformation
and rotational inertia, result in more accurate and bounded estimation of phase and group velocities. For symmetrically laminated
composite plates, the spectrum relations are given by an equation with respect to the frequency 𝜔, wavenumber 𝑘 and phase velocity
direction 𝜃 using FSDT [10]:

det [𝐆](𝜔, 𝑘, 𝜃) = 0, (9)

where det is the determinant operator and the 3-by-3 matrix 𝐆 is symmetric. The elements in symmetric matrix 𝐆 are polynomials in
terms of frequency 𝜔, wavenumber 𝑘 and phase velocity direction 𝜃, with detailed expressions provided in the literature [10,25,32].
Applying Laplace expansion to det [𝐆] yields a polynomial in terms of 𝜔 and 𝑘, which is of degree 6 at most for both parameters.
In flat orthotropic plates, the direction of phase velocity aligns with that of group velocity [49], allowing the group velocity to
be calculated as 𝑣𝑔 (𝜃) = 𝜕 𝜔∕𝜕 𝑘. For a given frequency 𝜔 at direction 𝜃, its corresponding wavenumber 𝑘 can be obtained by solving
Eq. (9). The phase velocity 𝑣𝑝 (𝜃) then can then be calculated as 𝜔∕𝑘, allowing for the generation of the dispersion curve, represented
by 𝜔 = 𝑠𝑟(𝑘; 𝜃). To calculate the group velocity, denoted as 𝑣𝑔 (𝜃) = 𝜕 𝜔∕𝜕 𝑘, one method involves utilising the difference operation
𝑣𝑔 (𝜃) ≈ 𝛥𝜔∕𝛥𝑘 with dense sampling points on the spectrum relation curve 𝜔 = 𝑠𝑟(𝑘; 𝜃). However, this method is computationally
costly, and its accuracy depends on the difference step size. Previous studies [10,25,32] do not analytically derive the group velocity
from the spectrum relations. In this study, instead of relying on numerical approximations, the group velocity is analytically derived
by implicit differentiation, which takes the partial derivative of Eq. (9) with respect to the wavenumber 𝑘 for a given wave direction
𝜃:
𝜕 𝜕 det [𝐆] 𝜕 det [𝐆] 𝜕 𝜔
𝜕𝑘
det [𝐆] = 𝜕𝑘
+ 𝜕𝜔 𝜕𝑘
= 0. (10)

Simplifying Eq. (10) yields analytical group velocity 𝑣𝑔 :

[ ]
𝜕 det [𝐆]∕𝜕 𝑘 t r adj(𝐆) 𝜕𝐆
𝜕𝜔 𝜕𝑘
𝑣𝑔 = = − 𝜕 det [𝐆]∕𝜕 𝜔 = − [ ], (11)
𝜕𝑘 t r adj(𝐆) 𝜕𝐆
𝜕𝜔

𝜕𝐆 𝜕𝐆
where t r () denotes the trace operator, adj(𝐆) is the adjugate of the matrix 𝐆, 𝜕𝑘
and 𝜕𝜔
are the element-wise derivatives of the
matrix 𝐆 with respect to 𝑘 and 𝜔, respectively. The derivation of these derivatives is provided in Appendix B.

4.1.3. Comparison of dispersion curves: plate theories vs. SAFEDC

This study explores the use of plate theories for GVP estimation and impact localisation, avoiding the need for FEA. To assess
the accuracy of plate theories in modelling wave propagation in composite laminate structures, a detailed comparison of dispersion
curves was conducted. Since flexural waves dominate low-velocity impacts and the A0 wave mode, particularly in the low-frequency
range, carries the majority of the impact energy, the analysis focused specifically on the A0 mode. Dispersion curves for the A0 mode
of a composite laminated plate were calculated using CLT, FSDT, and the semi-analytical finite element-based software SAFEDC [52],
as detailed in Appendix C.
The comparison reveals that while CLT significantly overestimates both phase and group velocities in the higher frequency range,
it demonstrates reasonable agreement with FSDT and SAFEDC results in the low-frequency range. FSDT, however, shows superior
accuracy across the frequency spectrum, closely matching the results of SAFEDC, particularly for group velocity. These findings
underscore the reliability of plate theories, with FSDT standing out as a robust and computationally efficient alternative to more
complex numerical methods for modelling the A0 mode in impact localisation applications.
It is important to note that known material properties can be represented by probability distributions to account for material
uncertainties [60,61]. Consequently, the GVP estimated based on plate theories becomes probabilistic rather than deterministic.
A 97.5% confidence range of the GVP is defined based on the material uncertainty parameters, and the final 95% confidence
localisation area for the probabilistic GVP can subsequently be assessed using the probabilistic methodology developed in Section 5.
A comparative analysis of the GVPs estimated by CLT and FSDT for impact localisation at different frequencies is presented in
Section 7.2.

4.2. Data-driven (DD) approach

For complex structures without prior knowledge of material properties, the GVP can be probabilistically estimated solely using
experimental or numerical reference impact data. This reference dataset comprises independent reference impact locations 𝑥𝑖 , 𝑦𝑖 ,
𝑖 = 1, 2, … , 𝑁𝑖 and their associated TDOA 𝛥𝐭 𝑖 , 𝑖 = 1, 2, … , 𝑁𝑖 , extracted from impact sensor signals.

8
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

4.2.1. Probabilistic estimation of GVP solely using the reference impact data
By employing Bayesian inference [62,63] in combination with the GVP parameterisation method outlined in Eq. (4), the posterior
distribution of the GVP velocity coefficients 𝐜 ∈ R5×1 , given independent observations 𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 , 𝑖 = 1, 2, … , 𝑁𝑖 , is derived as:
𝑃 (𝑥𝑖 ,𝑦𝑖 ,𝛥𝐭 𝑖 ,𝑖=1,2,…,𝑁𝑖 |𝐜,𝜎𝛿 ) ∏𝑁 𝑃 (𝑥𝑖 ,𝑦𝑖 ,𝛥𝐭 𝑖 |𝐜,𝜎 )
𝑃 (𝐜, 𝜎𝛿 |𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 , 𝑖 = 1, 2, … , 𝑁𝑖 ) = 𝑃 (𝑥𝑖 ,𝑦𝑖 ,𝛥𝐭 𝑖 ,𝑖=1,2,…,𝑁 )
𝑃 (𝐜, 𝜎𝛿 ) = 𝑖=1𝑖 𝑃 (𝑥𝑖 ,𝑦𝑖 ,𝛥𝐭 𝑖 ) 𝛿 𝑃 (𝐜)𝑃 (𝜎𝛿 ), (12)
𝑖

where 𝑃 (𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 |𝐜, 𝜎

𝛿 ), 𝑃 (𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 ),
and 𝑃 (𝐜, 𝜎𝛿 ) represents the likelihood, marginal likelihood, and prior, respectively. The prior
𝑟𝑐 = 𝑃 (𝐜, 𝜎𝛿 ) = 𝑃 (𝐜)𝑃 (𝜎𝛿 ) reflects the independence between the GVP (parameterised by 𝐜) and TDOA standard deviation (SD) 𝜎𝛿 ,
where TDOA uncertainties arise from sensor measurement errors and limitations in TDOA extraction accuracy.
The prior component 𝑃 (𝐜) represents the distribution of the GVP velocity coefficients absent specific observations 𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 .
Because the GVP depends on structural properties, and its velocity coefficients are linearly related to bending stiffness, it is typically
assumed that the prior distribution of these coefficients follows a Gaussian distribution with a mean 𝐜̂ that accounts for material
property uncertainties [64,65]. The standard deviation of the prior depends on the variances of material properties. Assuming the
standard deviation ratio of the nominal value of material properties is denoted by 𝜂, the prior distribution of velocity coefficients
can be expressed as:
( )
∏ (𝑐 −𝑐̂ )2
𝑃 (𝐜) = 5𝑘=1 √ 1 2 𝑒𝑥𝑝 − 𝑘 2𝑘 ̂
, 𝜎̂ 𝑐 = 𝜂 𝐜, (13)
2𝜎̂
2𝜋 𝜎̂ 𝑐 𝑘 𝑐𝑘

where 𝑐̂𝑘 and 𝜎̂ 𝑐2𝑘 represent the mean and variance of the 𝑘th wave velocity coefficient.
The parameter 𝜎𝛿 in the prior can be handled in several ways: as a fixed parameter estimated by its maximum likelihood [66]
(constant 𝜎𝛿 ), as a probabilistic parameter using Jeffrey’s non-informative (JNI) prior 𝑃 (𝜎𝛿 ) ∝ 1∕𝜎𝛿 [67], or by modelling it as
an inverse chi-squared distribution (ICS) [62,68]. The JNI prior is widely regarded as a standard choice when there is no strong
prior information about the parameters [67], as it is invariant under reparameterisation and allows the data to primarily inform
the parameter estimation. This non-informative prior ensures minimal bias and reflects our lack of prior knowledge about the scale
of 𝜎𝛿 . In contrast, the ICS distribution is particularly well-suited for modelling variance in Bayesian inference due to its conjugacy
with the normal likelihood [69]. This conjugacy simplifies posterior computation and ensures analytical tractability. Moreover, the
ICS prior provides a more intuitive and interpretable framework for expressing prior uncertainty about the variance.
In the absence of strong prior information about 𝜎𝛿 and to facilitate faster computation, this study employs the maximum
likelihood approach (treating 𝜎𝛿 as a fixed parameter). This treatment simplifies the prior and computation, as 𝜎𝛿 is determined
directly by maximising likelihood based on observations 𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 . The likelihood 𝑃 (𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 |𝐜, 𝜎𝛿 ) denotes the probability of the
𝑖th observation given the GVP velocity coefficients 𝐜 and TDOA variance, which follows to a multivariate Gaussian distribution,
expressed as:
( 𝑖 2)
∏𝑁𝑠 ∏𝑁𝑠 (𝛿 )
𝑖𝑐 = 𝑃 (𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 |𝐜, 𝜎𝛿 ) = 𝑗=2 𝑃 (𝑥𝑖 , 𝑦𝑖 , 𝛥𝑡𝑖𝑗 |𝐜, 𝜎𝛿 ) = 𝑗=2 √ 1 exp − 𝑗 2 ,
2 2𝜋 𝜎𝛿 2𝜎𝛿
√ √ (14)
𝑑𝑗𝑖 𝑑1𝑖 (𝑥𝑖 −𝑥𝑗 )2 +(𝑦𝑖 −𝑦𝑗 )2 (𝑥𝑖 −𝑥1 )2 +(𝑦𝑖 −𝑦1 )2
𝛿𝑗𝑖 = 𝑡𝑖𝑗 − 𝑡𝑖1 − 𝛥𝑡𝑖𝑗 = − − 𝛥𝑡𝑖𝑗 = − − 𝛥𝑡𝑖𝑗 ,
𝑣𝑔 (𝜃𝑗𝑖 ) 𝑣𝑔 (𝜃1𝑖 ) 𝑣𝑔 (𝜃𝑗𝑖 ) 𝑣𝑔 (𝜃1𝑖 )

based on physical relation of wave propagation Eq. (8). In the expression, 𝛿𝑗𝑖 represents the TDOA residual of the 𝑗th sensor for the
𝑖th reference impact, calculated as the difference between the estimation from the GVP 𝑡𝑖𝑗 − 𝑡𝑖1 and the TDOA extracted from the
reference sensor signals 𝛥𝑡𝑖𝑗 . Additionally, 𝑑𝑗𝑖 and 𝜃𝑗𝑖 represent the distance and wave propagation direction from the 𝑖th reference
impact to 𝑗th sensor, respectively. 𝜎𝛿2 denotes the TDOA variance across the reference impacts. The likelihood of all independent
reference impacts is the product of the likelihood of each reference impact:
( 𝑖 2)
∏𝑁 ∏𝑁 ∏𝑁𝑠 ∏𝑁 ∏𝑁𝑠 (𝛿 )
𝑐 = 𝑖=1𝑖 𝑖𝑐 = 𝑖=1𝑖 𝑗=2
𝑃 (𝑥𝑖 , 𝑦𝑖 , 𝛥𝑡𝑖𝑗 |𝐜) = 𝑖=1𝑖 𝑗=2
√ 1 exp − 𝑗 2 . (15)
2 2𝜋 𝜎𝛿 2𝜎𝛿

The log likelihood is further derived as:

𝑁 (𝑁 −1) ∑𝑁 ∑𝑁𝑠 𝑖 2
ln 𝑐 = − 𝑖 2𝑠 ln(2𝜋 𝜎𝛿2 ) − 1 2 𝑖=1𝑖 𝑗=2 (𝛿𝑗 ) . (16)
2𝜎𝛿

The prior mean 𝐜̂ is unknown; thus, it is assumed to be equal to its maximum likelihood estimation. The Gaussian prior reaches
its maximum value when 𝐜 is equal to the maximum likelihood estimation 𝐜. ̂ Therefore, the maximum likelihood estimation 𝐜̂ serves
as the maximum posterior estimation, as the posterior is proportional to the likelihood multiplied by the prior. The maximisation
of the log likelihood results in the estimation of the TDOA variance across the reference impacts:
𝜕 ln 𝑐 𝑁 (𝑁 −1) ∑𝑁 ∑𝑁𝑠 𝑖 2
= − 𝑖 𝑠2 + 1 4 𝑖=1𝑖 𝑗=2 (𝛿𝑗 ) = 0,
𝜕 𝜎𝛿2 2𝜎𝛿 2𝜎𝛿
∑ ∑ (17)
𝑁 𝑁𝑠
𝜎̂ 𝛿2 = 𝑁 (𝑁1 −1) 𝑖=1𝑖 𝑗=2 (𝛿𝑗𝑖 )2 .
𝑖 𝑠

Hence, the GVP velocity coefficients with the maximised likelihood (posterior), denoted as 𝐜,
̂ are estimated by maximising the
log likelihood, which is equivalent to minimising the TDOA variance across the reference impacts:
∑ 𝑁 ∑ 𝑁𝑠 𝑖 2
𝐜̂ = ar gmax ln 𝑐 = ar gmin 𝜎̂ 𝛿2 = ar gmin 𝑁 (𝑁1 −1) 𝑖=1𝑖 𝑗=2 (𝛿𝑗 ) . (18)
𝐜 𝐜 𝑖 𝑠 𝐜
9
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

4.2.2. Maximising likelihood: hybrid GA (HGA) with gradient-based local search

Genetic algorithm (GA) [70] can be employed to estimate the GVP with the highest probability by maximising the log likelihood
(minimising the TDOA variance) as illustrated in Eq. (18), with the five GVP velocity coefficients treated as design variables. To
expedite the optimisation process, gradient information can be incorporated to facilitate a faster and more efficient local search.
These gradients include the first and second-order partial derivatives of the TDOA variance with respect to the five GVP velocity
coefficients 𝑐𝑚 , 𝑐𝑛 , 𝑚, 𝑛 = 1, 2, … , 5, which are derived as:
𝜕 𝜎̂ 𝛿2 ∑𝑁 ∑𝑁𝑠 𝑖 𝜕𝛿𝑗𝑖 𝜕 2 𝜎̂ 2 ∑𝑁 ∑𝑁𝑠 𝜕 𝛿𝑗𝑖 𝜕𝛿𝑗𝑖 𝜕2 𝛿𝑖
𝜕𝑐
= 𝑁 (𝑁2 −1) 𝑖=1𝑖 𝑗=2 𝛿𝑗 𝜕 𝑐 , 𝜕 𝑐 𝜕𝛿𝑐 = 𝑁 (𝑁2 −1) 𝑖=1𝑖 𝑗=2 𝜕𝑐 𝜕𝑐
+ 𝛿𝑗𝑖 𝜕𝑐 𝜕𝑗𝑐 . (19)
𝑚 𝑖 𝑠 𝑚 𝑚 𝑛 𝑖 𝑠 𝑚 𝑛 𝑚 𝑛

Based on the expression of the TDOA residual 𝛿𝑗𝑖 defined in Eq. (14) and the GVP parameterised by velocity coefficients
𝜕 𝛿𝑗𝑖 𝜕 2 𝛿𝑗𝑖
𝑐𝑚 , 𝑐𝑛 , 𝑚, 𝑛 = 1, 2, … , 5 as illustrated in Eq. (4), the intermediate derivatives 𝜕 𝑐𝑚
and 𝜕 𝑐𝑚 𝜕 𝑐𝑛
are further derived using the chain
rule:
𝜕 𝛿𝑗𝑖 𝜕 𝑡𝑖𝑗 𝜕 𝑡𝑖1 𝜕 𝑡𝑖𝑗 𝑑𝑗𝑖 𝜕 𝑣𝑔 (𝜃𝑗𝑖 ) 𝜕 𝑣𝑔 (𝜃𝑗𝑖 ) 𝜙𝑚 (𝜃𝑗𝑖 )
= − , =− , = ,
𝜕 𝑐𝑚 𝜕 𝑐𝑚 𝜕 𝑐𝑚 𝜕 𝑐𝑚 𝑣2𝑔 (𝜃𝑗𝑖 ) 𝜕 𝑐𝑚 𝜕 𝑐𝑚 4𝑣3𝑔 (𝜃𝑗𝑖 )
𝜕 2 𝛿𝑗𝑖 𝜕 2 𝑡𝑖𝑗 𝜕 2 𝑡𝑖1 𝜕 2 𝑡𝑖𝑗 2𝑑𝑗𝑖 𝜕 𝑣𝑔 (𝜃𝑗𝑖 ) 𝜕 𝑣𝑔 (𝜃𝑗𝑖 ) 𝑑 𝑖 𝜕 2 𝑣𝑔 (𝜃 𝑖 ) 𝜕 2 𝑣𝑔 (𝜃 𝑖 ) −3𝜙𝑚 (𝜃𝑗𝑖 )𝜙𝑛 (𝜃𝑗𝑖 )
(20)
= − , = − 2 𝑗 𝑖 𝜕 𝑐 𝜕 𝑐𝑗 , 𝜕 𝑐 𝜕 𝑐𝑗 = .
𝜕 𝑐𝑚 𝜕 𝑐𝑛 𝜕 𝑐𝑚 𝜕 𝑐𝑛 𝜕 𝑐𝑚 𝜕 𝑐𝑛 𝜕 𝑐𝑚 𝜕 𝑐𝑛 𝑣3𝑔 (𝜃𝑗𝑖 ) 𝜕 𝑐𝑚 𝜕 𝑐𝑛 𝑣𝑔 (𝜃𝑗 ) 𝑚 𝑛 𝑚 𝑛 16𝑣7𝑔 (𝜃𝑗𝑖 )

Incorporating the gradients illustrated in Eq. (19) into the GA for efficient local search, the GVP with the maximised likelihood
can be efficiently identified.

4.2.3. Bounding velocity coefficients by reference impact data

The maximisation of the log likelihood (equivalent to the minimisation of the TDOA variance), as depicted in Eq. (18), is
significantly influenced by the bounds of these five design variables, namely, the velocity coefficients. The upper and lower bounds
of the five velocity coefficients can be determined using the reference impact data 𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 , 𝑖 = 1, 2, … , 𝑁𝑖 . Based on the physical
relation of wave propagation shown in Eq. (8), the following inequality can be derived, and the equivalent velocity 𝑣(𝜃 𝑖 ) is defined:
̄ 1𝑗
𝑑𝑗𝑖 −𝑑1𝑖 𝑑𝑗𝑖 −𝑑1𝑖
min[𝑣𝑔 (𝜃𝑗𝑖 ), 𝑣𝑔 (𝜃1𝑖 )] ≤ ≤ max[𝑣𝑔 (𝜃𝑗𝑖 ), 𝑣𝑔 (𝜃1𝑖 )], 𝑣(𝜃
̄ 1𝑗𝑖 )= . (21)
𝛥𝑡𝑖𝑗 +𝛿 𝛥𝑡𝑖𝑗

When excluding TDOA uncertainties (𝛿=0), the equivalent velocity 𝑣(𝜃 𝑖 ) lies midway between the actual velocities 𝑣 (𝜃 𝑖 ) and 𝑣 (𝜃 𝑖 )
̄ 1𝑗 𝑔 𝑗 𝑔 1
of two separate wave propagation paths. However, due to inevitable TDOA uncertainties (𝛿 ≠ 0), the equivalent velocity 𝑣(𝜃 𝑖 ) is
̄ 1𝑗
not bounded by actual velocities 𝑣𝑔 (𝜃𝑗𝑖 ) and 𝑣𝑔 (𝜃1𝑖 ). A large negative TDOA uncertainty (𝛿 < 0) may leads the equivalent velocity
to be greater than max[𝑣𝑔 (𝜃𝑗𝑖 ), 𝑣𝑔 (𝜃1𝑖 )] (excluding negative equivalent velocity). A positive TDOA uncertainty (𝛿 > 0) may leads the
equivalent velocity to be smaller than min[𝑣𝑔 (𝜃𝑗𝑖 ), 𝑣𝑔 (𝜃1𝑖 )]. Across all wave propagation paths in the reference impacts, the equivalent
velocity forms a velocity set ̄ = {𝑣(𝜃 ̄ 1𝑗𝑖 ), 𝑖 = 1, 2, … , 𝑁 , 𝑗 = 2, 3, … , 𝑁 }. Considering the randomness of the TDOA uncertainty, the
𝑖 𝑠
maximum and minimum of this equivalent velocity set, max() ̄ and min(), ̄ respectively, bound the actual maximal wave group
velocity max[𝑣𝑔 (𝜃)] and minimal wave group velocity min[𝑣𝑔 (𝜃)].
To mitigate the influence of TDOA uncertainty on equivalent velocity, only the equivalent velocities corresponding to a large
value of the measured TDOA 𝛥𝑡𝑖𝑗 should be included in the velocity set. In this study, the TDOA threshold for calculating the
equivalent velocity is set to half of the maximal TDOA for all reference impacts. In this way, the actual maximal wave velocity
max[𝑣𝑔 (𝜃)] and minimal wave velocity min[𝑣𝑔 (𝜃)] are bounded as:
̄ ≤ min[𝑣𝑔 (𝜃)] ≤ max[𝑣𝑔 (𝜃)] ≤ max().
min() ̄ (22)

With the bounded wave velocity, the velocity coefficients 𝑐1 and 𝑐5 can be directly bounded since they are only associated with
the wave velocity in the 𝜃 = 0 and 𝜃 = 𝜋∕2 directions, respectively:
̄ 4 ≤ 𝑐1 = 𝑣4 (𝜃 = 0) ≤ [max()]
[min()] ̄ 4,
𝑔
̄ 4 4 ̄ 4. (23)
[min()] ≤ 𝑐5 = 𝑣 (𝜃 = 𝜋∕2) ≤ [max()] 𝑔

To generate a circular wavefront (𝑣𝑔 (𝜃) is constant), the conditions 𝑐3 = 2𝑐1 = 2𝑐5 , and 𝑐2 = 𝑐4 = 0 should be satisfied. These
conditions define the bounds for the velocity coefficient 𝑐3 . Additionally, since the lamina placed in the 𝜃 = 0 and 𝜃 = 𝜋∕2 directions
does not contribute to the bending stiffness 𝐷16 (𝑐2 ) and 𝐷26 (𝑐4 ), the laminated composite with a layup of [0◦ ∕45◦ ]𝑠 achieves the
largest ratio of 𝐷16 ∕𝐷11 among symmetrically laminated composites [44], which provides an upper bound for 𝑐2 and 𝑐4 . Assuming
a lamina modulus ratio of 𝐸1 ∕𝐸2 = 𝛽 ≥ 1, this largest ratio 𝑐2 ∕𝑐1 can be estimated based on CLT:
𝑐2 4𝐷16 4(𝛽−1) 4
𝑐1
= 𝐷11
≤ 29.5𝛽+1.5
≤ 29.5
. (24)

The same analysis can be applied to 𝑐4 ∕𝑐5 . Therefore, the velocity coefficients 𝑐2 , 𝑐3 , and 𝑐4 are bounded by:
̄ 4 ≤ 𝑐3 ≤ 2[max()]
2[min()] ̄ 4,
̄ 4 ̄ 4 ∕7.375,
−[max()] ∕7.375 ≤ 𝑐2 ≤ [max()] (25)
̄ 4 ∕7.375 ≤ 𝑐4 ≤ [max()]
−[max()] ̄ 4 ∕7.375.

The bounds of 𝑐2 and 𝑐4 are symmetric about 0 since their signs are influenced by the selected principal material direction. By
rotating the principal material direction when negative values are present, 𝑐2 and 𝑐4 can be reassessed as positive.

10
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Algorithm 1: Data-driven GVP estimation with reference impact data.

Input: 𝑥𝑖 , 𝑦𝑖 , 𝑠𝑖 ; /* Reference impact data */
Input: 𝑥𝑗 , 𝑦𝑗 , 𝑁𝑠 , 𝜆 ; /* Sensors’ locations, confidence level */
Output: 𝑣̂ 𝑔 (𝜃), 𝐜,̂ 𝛼̂ ; /* GVP and velocity coefficients with maximum posterior, velocity confidence ratio */
Step 1: TDOA extraction from reference impact sensor signals:
𝛥𝐭 𝑖 ← 𝑠𝑖 ; /* TDOA extraction */
Step 2: Velocity coefficient bounds estimation:
̄ 𝑣(𝜃
, ̄ 1𝑗 𝑖 ) ← 𝑥𝑖 , 𝑦𝑖 , 𝑥 , 𝑦 , 𝛥𝐭 𝑖 ;
𝑗 𝑗 /* Equivalent velocity set */
𝐜𝑙𝑏 , 𝐜𝑢𝑏 ← ̄ ; /* Bounds of velocity coefficients */
Step 3: Maximum likelihood estimations of GVP and TDOA variance by HGA:
𝛿𝑗𝑖 ← 𝜙𝑚 , 𝑥𝑖 , 𝑦𝑖 , 𝑥𝑗 , 𝑦𝑗 , 𝐜 ; /* TDOA residual function */
∑𝑁 ∑𝑁𝑠 𝑖 2
𝜎𝛿2 ← 𝑁 (𝑁1 −1) 𝑖=1𝑖 𝑗=2 (𝛿𝑗 ) , 𝐜 ; /* TDOA variance function */
𝑖 𝑠
𝜕 𝜎̂ 𝛿2
𝜕 𝑐𝑚
← 𝛿𝑗𝑖 , 𝜙𝑚 , 𝐜 ; /* Gradient function */
𝜕 𝜎̂ 2
̂ 𝜎̂ 𝛿2 , 𝑣̂ 𝑔 (𝜃)
𝐜, ← 𝜎𝛿2 , 𝜕𝑐 𝛿 , 𝐜𝑢𝑏 , 𝐜𝑙𝑏 ; /* Maximum likelihood GVP and TDOA variance by HGA */
𝑚
Step 4: Velocity confidence ratio estimation by MCMC:
𝐋𝑐 , 𝐏𝑟𝑐 ← 𝜎̂ 𝛿2 , 𝐜 ; /* Likelihood and prior based on identified TDOA variance */
𝐂 ← 𝐋𝑐 , 𝐏𝑟𝑐 ; /* MCMC samples of velocity coefficients */
𝛼̂ ← 𝐂, 𝐜,̂ 𝜆; /* Velocity confidence ratio */

4.2.4. Identifying 97.5% confidence range of the GVP

The gradient information derived in Section 4.2.2 and the velocity coefficient bounds identified in Section 4.2.3 facilitate the
rapid determination of the TDOA variance 𝜎̂ 𝛿2 , the velocity coefficients 𝐜, ̂ and the corresponding GVP 𝑣̂ 𝑔 (𝜃) with maximum likelihood.
This identified TDOA variance 𝜎̂ 𝛿2 refines the posterior distribution from 𝑃 (𝐜, 𝜎𝛿 |𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 , 𝑖 = 1, 2, … , 𝑁𝑖 ) to 𝑃 (𝐜|𝜎̂ 𝛿 , 𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 , 𝑖 =
1, 2, … , 𝑁𝑖 ), enabling efficient posterior estimation and calculation of the confidence range [𝐜𝑙𝑏 , 𝐜𝑢𝑏 ] for velocity coefficients at a
given confidence level 𝜆.
The confidence range can be defined by the velocity confidence ratio 𝛼 and the maximum likelihood velocity coefficients 𝐜̂ as
̂ 4 , 𝛼 4 𝐜].
[𝐜∕𝛼 ̂ With the identified velocity coefficients 𝐜, ̂ 𝛼 can be calculated by integrating the posterior distribution over the probability
space 𝛺𝑐 , as detailed in Appendix A.
̂ 4 , 𝛼 4 𝐜],
[𝐜𝑙𝑏 , 𝐜𝑢𝑏 ] = [𝐜∕𝛼 ̂ 𝛼 ≥ 1,
𝛼 4 𝐜̂ (26)
𝛼̂ = ar gmin ∫𝐜∕𝛼̂ 4 𝑃 (𝐜|𝜎̂ 𝛿 , 𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 , 𝑖 = 1, 2, … , 𝑁𝑖 )𝑑 𝛼 ≥ 𝜆.
𝛼

Both Markov Chain Monte Carlo (MCMC) [45,46] and importance sampling [71] are effective methods for estimating this
posterior distribution. Here, an efficient adaptive MCMC method, known as Delayed Rejection Adaptive Metropolis (DRAM) [68],
is employed, which combines adaptive Metropolis sampling with delayed rejection to enhance sample accuracy and robustness. The
initial proposal covariance for the MCMC samples is set as diag 𝜂 2 𝐜̂2 , consistent with the Gaussian prior. Using these MCMC samples,
the velocity confidence ratio 𝛼 is efficiently determined through the bisection method. The upper bound for the derived 𝛼̂ is based
on the prior, with a maximum value of max(1∕(1 − 3𝜂), 1 + 3𝜂), according to the 3𝜎 principle.
Algorithm 1 provides the pseudo-algorithm for data-driven GVP estimation, which involves four main steps: TDOA extraction,
estimation of velocity coefficient bounds, maximum likelihood estimation of velocity coefficients and TDOA variance by HGA, and
velocity confidence ratio estimation via MCMC. In the maximum likelihood estimation step, the fixed parameter for TDOA variance
is identified and subsequently incorporated as known information in the MCMC sampling. This approach reduces the number of
probabilistic parameters and enhances computational efficiency and sampling convergence.

4.3. Hybrid physics-data (HPD) approach

In scenarios where limited reference impact data 𝑥𝑖 , 𝑦𝑖 , 𝛥𝐭 𝑖 , 𝑖 = 1, 2, … , 𝑁𝑖 , is available for a target complex structure, and the
GVP from a lower-level structure incorporated within the target is known, the GVP of the complex structure can be probabilistically
estimated by integrating these two sources of information. This approach demonstrates the scalability of GVP estimation, enabling
rapid GVP identification for more complex, higher-level structures by leveraging the reference GVP from their simpler components.
For example, complex stiffened plate structures consist of both flat plate and stiffener components. The GVP of the lower-level flat
plates can be estimated using the PB method with known material properties, and this estimation can assist in the GVP identification
of the overall complex stiffened structure.

4.3.1. Gaussian prior and maximum likelihood estimation of the GVP

Assuming that the available reference GVP corresponds to the reference velocity coefficients 𝐜, ̄ these coefficients cannot
accurately serve as the prior mean. This is because the reference GVP may not be accurate enough to mimic the wave propagation in

11
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

the target structure. If the reference GVP is derived from a lower-level structure, it may differ significantly in velocity amplitude from
the actual GVP of the target structure, although it may have a similar shape. Therefore, the prior mean of the velocity coefficient
again adopts its maximum likelihood estimation. However, the reference GVP enables the representation of the prior mean, as the
summation of a scaled reference velocity coefficients and small deviations:
𝐜̂ = (1∕𝛾)4 𝐜̄ + 𝛿𝑐 , (27)

where 𝛾 and 𝛿𝑐 represent the scaling factor and deviations, respectively. In this representation, the scaling factor 𝛾 controls the
amplitude of the prior mean, and the deviations fine-tune the prior mean. This representation makes the difference between the
hybrid approach present in this subsection and data-driven approach present in Section 4.2. With such representation, the maximum
likelihood estimation of the velocity coefficients 𝐜 can be identified solely based on gradient-based optimisation instead of using GA.
The subsequent subsections detail the gradient-based optimisation approach for the maximum likelihood estimation of the velocity
coefficients.

4.3.2. Maximising likelihood: a two-stage gradient-based approach

The maximum likelihood estimation of the velocity coefficients with the highest probability 𝐜̂ can be implemented using HGA
with gradient-based local search, as illustrated in Section 4.2.2. However, given the availability of reference velocity coefficients 𝐜,
̄
the representation illustrated in Eq. (27) suggests a two-stage gradient-based approach to optimise 𝐜,̂ involving the identification of
the scaling factor 𝛾 and the small deviations 𝛿𝑐 subsequently.
Instituting 𝐜 = (1∕𝛾)4 𝐜̄ into the maximum likelihood estimation shown in Eq. (18), the maximum likelihood alters to:
1 ∑ 𝑁𝑖 ∑ 𝑁𝑠 𝑖 2 1 ∑𝑁𝑖 ∑𝑁𝑠 ̄𝑖 2
𝐜̂ = ar gmin 𝜎̂ 2 = ar gmin
𝛿
(𝛿 ) = ar gmin
𝑁𝑖 (𝑁𝑠 −1) 𝑖=1 𝑗=2 𝑗 (𝛿 ) , 𝑁𝑖 (𝑁𝑠 −1) 𝑖=1 𝑗=2 𝑗
𝐜 𝐜 𝐜=(1∕𝛾)4 𝐜̄
𝑑𝑗𝑖 𝑑1𝑖
(28)
𝛿̄𝑗𝑖 = (𝑡̄𝑖𝑗 − 𝑡̄𝑖1 )𝛾 − 𝛥𝑖1𝑗 = ( − )𝛾 − 𝛥𝑡𝑖1𝑗 ,
̄ 𝑗𝑖 )
𝑣(𝜃 ̄ 1𝑖 )
𝑣(𝜃

where 𝛿̄𝑗𝑖 represents the TDOA residual of the 𝑖th reference impact and 𝑗th sensor given the reference velocity coefficients 𝐜̄ and
scaling factor 𝛾. The first and second-order derivatives of the TDOA variance 𝜎̂ 𝛿2 with respect to the scaling factor 𝛾 can be further
derived as follows:
𝑑 𝜎̂ 𝛿2 ∑𝑁 ∑𝑁𝑠 ̄𝑖 𝑖 𝑑 2 𝜎̂ 2 ∑𝑁 ∑𝑁𝑠 𝑖
𝑑𝛾
= 𝑁 (𝑁2 −1) 𝑖=1𝑖 𝑗=2 𝛿𝑗 (𝑡̄𝑗 − 𝑡̄𝑖1 ), 𝑑 𝛾 2𝛿 = 𝑁 (𝑁2 −1) 𝑖=1𝑖 𝑗=2 (𝑡̄𝑗 − 𝑡̄𝑖1 )2 . (29)
𝑖 𝑠 𝑖 𝑠

It is evident that since the second-order derivative is non-negative, this convex optimisation problem can be effectively addressed
using gradient descent algorithms. In this study, the Newton’s method [72] is adopted for this purpose to identify the optimised 𝛾̂
which maximises the likelihood.
The optimised scaled velocity coefficients (1∕̂𝛾 )4 𝐜̄ closely approximate the mean velocity coefficients 𝐜̂ in terms of amplitude.
In the second stage optimisation, the scaled velocity coefficients (1∕̂𝛾 )4 𝐜̄ obtained from the first stage optimisation are utilised as
the initial point. The objective is to further minimise the TDOA variance (maximise the log likelihood) with respect to the velocity
coefficients, as illustrated in Eq. (18), employing the trust-region-reflective algorithm [73] based on the gradient information derived
in Eq. (19). This two-stage approach relies solely on rapid gradient-based optimisation, thereby achieving high computational
efficiency.
However, since gradient-based optimisation methods tend to converge to local optima, this two-stage process is repeated until
the scaling factor satisfies |𝛾 − 1| < 0.01 to approach the global optimum. Similar to the data-driven method, the velocity coefficients
are optimised within the bounds developed in Section 4.2.3. This finite optimisation space allows for efficient implementation of
the optimisation process.

5. Probabilistic impact localisation

5.1. Probabilistic impact localisation: Bayesian inference

Given the observations, including the TDOA 𝛥𝐭 for an impact estimated from its corresponding sensor signals and the known
deterministic GVP 𝑣𝑔 (𝜃), the impact location can be determined in a probabilistic model using Bayesian inference [63]:
𝑃 (𝛥𝐭,𝑣𝑔 |𝑥,𝑦,𝜎𝛿 ) 𝑃 (𝛥𝐭,𝑣𝑔 |𝑥,𝑦,𝜎𝛿 )
𝑃 (𝑥, 𝑦, 𝜎𝛿 |𝛥𝐭, 𝑣𝑔 ) = 𝑃 (𝛥𝐭,𝑣𝑔 )
𝑃 (𝑥, 𝑦, 𝜎𝛿 ) = 𝑃 (𝛥𝐭,𝑣𝑔 )
𝑃 (𝑥, 𝑦)𝑃 (𝜎𝛿 ), (30)

where 𝑃 (𝑥, 𝑦, 𝜎𝛿 |𝛥𝐭, 𝑣𝑔 ), 𝑃 (𝛥𝐭, 𝑣𝑔 |𝑥, 𝑦, 𝜎𝛿 ), 𝑃 (𝑥, 𝑦, 𝜎𝛿 ) and 𝑃 (𝛥𝐭, 𝑣𝑔 ) represent the posterior, likelihood, prior, and marginal likelihood
(evidence), respectively. The prior 𝑥𝑦 = 𝑃 (𝑥, 𝑦, 𝜎𝛿 ) = 𝑃 (𝑥, 𝑦)𝑃 (𝜎𝛿 ) indicates the independence between impact location and TDOA
variance, as TDOA uncertainties arise primarily from sensor measurement errors and the inherent limitations in TDOA extraction
methods.
Similar to the data-driven GVP estimation approach in Section 4.2, TDOA SD 𝜎𝛿 is treated as a fixed parameter and estimated
using its maximum likelihood [66] (constant 𝜎𝛿 ), rather than using Jeffrey’s non-informative prior 𝑃 (𝜎𝛿 ) ∝ 1∕𝜎𝛿 [67] or modelling
it as an inverse chi-squared distribution [62,68]. This approach simplifies both the prior and the Bayesian inference process, as 𝜎𝛿2 is
automatically determined by the observations (Known TDOA 𝛥𝐭 and GVP 𝑣𝑔 (𝜃)). Additionally, Section 7.2.2 presents a comparison
of these three approaches to TDOA variance in impact localisation.

12
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

The likelihood 𝑃 (𝛥𝐭, 𝑣𝑔 |𝑥, 𝑦, 𝜎𝛿 ) can be derived as the probability of a multivariate Gaussian distribution based on Eq. (8):
( 2 )
∏𝑁𝑠 ∏𝑁𝑠 𝛿
𝑥𝑦 = 𝑃 (𝛥𝐭, 𝑣|𝑥, 𝑦, 𝜎𝛿 ) = 𝑗=2 𝑃 (𝛥𝑡1𝑗 , 𝑣|𝑥, 𝑦, 𝜎𝛿 ) = 𝑗=2 √ 1 exp − 𝑗2 ,
2 2𝜋 𝜎𝛿 2𝜎𝛿
√ √ (31)
𝑑𝑗 𝑑1 (𝑥−𝑥𝑗 )2 +(𝑦−𝑦𝑗 )2 (𝑥−𝑥1 )2 +(𝑦−𝑦1 )2
𝛿𝑗 = (𝑡𝑗 − 𝑡1 ) − 𝛥𝑡1𝑗 = 𝑣𝑔 (𝜃𝑗 )
− 𝑣𝑔 (𝜃1 )
− 𝛥𝑡1𝑗 = 𝑣𝑔 (𝜃𝑗 )
− 𝑣𝑔 (𝜃1 )
− 𝛥𝑡1𝑗 ,

where 𝑡𝑗 − 𝑡1 signifies the time difference estimated from the GVP 𝑣𝑔 (𝜃), while 𝛥𝑡1𝑗 denotes the time difference extracted from the
impact sensor signals. Consequently, 𝛿𝑗 represents the residual of the predicted and measured TDOA.

5.2. Maximum likelihood estimations of impact location and TDOA variance

Based on the Bayesian inference described in Eq. (30), the posterior probability of the impact occurring at any location (𝑥, 𝑦)
can be estimated. Given the uniform distribution of the prior 𝑃 (𝑥, 𝑦), the impact location with the highest probability, denoted as
𝑥, ̂ can be determined by maximising the log likelihood:
̂ 𝑦,
(𝑁 −1) ∑ 𝑁𝑠 2
̂ 𝑦̂ = ar gmax ln 𝑥𝑦 = ar gmax − 𝑠2 ln(2𝜋 𝜎𝛿2 ) − 1 2 𝑗=2
𝑥, 𝛿𝑗 . (32)
𝑥,𝑦 𝑥,𝑦 2𝜎𝛿

𝜕 ln 
The maximum of the log likelihood is achieved when the first-order partial derivative with respect to 𝜎𝛿2 is zero. Setting equal
𝜕 𝜎𝛿2
to zero yields the estimation of the TDOA variance:
𝜕 ln 𝑥𝑦 (𝑁 −1) ∑ 𝑁𝑠 2
2 = − 𝑠 2 + 1 4 𝑗=2 𝛿𝑗 = 0,
𝜕 𝜎𝛿 2𝜎𝛿 2𝜎𝛿
∑ 𝑁𝑠 (33)
1
𝜎̂ 𝛿2 = 𝑁𝑠 −1
𝛿2 .
𝑗=2 𝑗

The TDOA variance is estimated as the average residual of sum of the square error of the extracted TODA from sensor signals and
the estimated TODA from GVP. Incorporating the TDOA variance estimation derived in Eq. (33) into Eq. (32), the maximisation of
log likelihood is equivalent to minimisation of the TDOA variance:
(𝑁 −1) ∑ 𝑁𝑠 2
̂ 𝑦̂ = ar gmin 𝑠2 [ln(2𝜋 𝜎̂ 𝛿2 ) + 1] = ar gmin 𝜎̂ 𝛿2 = ar gmin 𝑁 1−1 𝑗=2
𝑥, 𝛿𝑗 . (34)
𝑥,𝑦 𝑥,𝑦 𝑥,𝑦 𝑠

The maximum likelihood estimation of TDOA variance 𝜎̂ 𝛿2 (from Eq. (34)) refines the posterior from 𝑃 (𝑥, 𝑦, 𝜎𝛿 |𝛥𝐭, 𝑣𝑔 ) to
𝑃 (𝑥, 𝑦|𝛥𝐭, 𝑣𝑔 , 𝜎̂ 𝛿 ). This transformation enables a more focused assessment of impact localisation reliability, which is quantified using
a confidence area. This area is defined by a confidence radius, ensuring that the cumulative posterior probability within the
overlapping area of circular region centred at the posterior mean and structural domain exceeds a confidence level 𝜆. The posterior
mean (𝑥, ̄ and the confidence radius 𝑟𝜆 are thus defined by definite integration as follows:
̄ 𝑦)
(𝑥, ̄ = ∫𝛺 (𝑥, 𝑦)𝑃 (𝑥, 𝑦|𝛥𝐭, 𝑣𝑔 , 𝜎𝛿 )𝑑(𝑥, 𝑦),
̄ 𝑦)
𝑥𝑦
𝑟 2𝜋 (35)
𝑟𝜆 = ar gmin ∫0 ∫0 𝑃 (𝑥, 𝑦|𝛥𝐭, 𝑣𝑔 , 𝜎𝛿 )𝑑 𝜙𝑑 𝑟 > 𝜆, 𝑥 = 𝑟 cos 𝜙 + 𝑥,
̄ 𝑦 = 𝑟 sin 𝜙 + 𝑦.
̄
𝑟

To approximate the posterior distribution 𝑃 (𝑥, 𝑦|𝛥𝐭, 𝑣𝑔 , 𝜎̂ 𝛿 ), an efficient adaptive MCMC method, namely Delayed Rejection
Adaptive Metropolis (DRAM) [68], is adopted, which utilises adaptive Metropolis sampling with delayed rejection. The initial
proposal distribution for the MCMC samples is set to a two-dimensional Gaussian distribution with mean (𝑥, ̂ and covariance
̂ 𝑦)
max[𝑣𝑔 (𝜃)]2 𝜎̂ 𝛿2 𝐼. Through these MCMC samples, the posterior mean (𝑥, ̄ and confidence radius 𝑟𝜆 are efficiently computed.
̄ 𝑦)
The choice of the initial proposal distribution considers the relationship among the confidence radius, TDOA variance, and
wave velocity. Since the likelihood follows a Gaussian form, increased TDOA variance and higher wave velocity both result in
a more gradual decrease in likelihood as the impact location diverges from its maximum likelihood estimate. Consequently, the
confidence radius exhibits a monotonic increase with both TDOA variance and wave velocity. This adaptive covariance, denoted
by max[𝑣𝑔 (𝜃)]2 𝜎̂ 𝛿2 𝐼 based on the maximum wave velocity max[𝑣𝑔 (𝜃)], ensures the proposal distribution maintains adequately heavy
tails.
A 95% confidence localisation area is utilised in this study to denote the localisation reliability. In scenarios with a deterministic
GVP, the 95% confidence radius directly specifies the localisation area with 95% confidence. However, for a probabilistic GVP,
the 97.5% confidence radius is employed. The multiplication of the 97.5% confidence range of the probabilistic GVP by a 97.5%
confidence localisation radius results in a localisation area with 95% confidence.

5.3. Maximising likelihood: gradient-based optimisation

The process of identifying the impact location with the highest likelihood involves solving a minimisation problem outlined
in Eq. (34). Traditionally, this problem has been tackled using heuristic optimisation algorithms, such as GA. However, GA-based
optimisation incurs significant computational costs and requires numerous evaluations of the objective function (specifically, the
TDOA variance 𝜎̂ 𝛿2 ) to thoroughly explore the design space. Enhancing computational efficiency is essential for enabling real-time
impact localisation directly at the sensor node of a passive impact monitoring system [74], rather than relying on a central computing

13
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Algorithm 2: Probabilistic impact localisation with known GVP using Bayesian inference.
Input: 𝑣𝑔 (𝜃), 𝜆, 𝑥𝑗 , 𝑦𝑗 , 𝑠𝑗 , 𝑁𝑠 ; /* GVP, confidence level, sensors’ locations, sensor signals */
Output: 𝑥, ̂ 𝜎̂ 𝛿 ;
̂ 𝑦, /* Impact location and TDOA variance with maximum likelihood */
Output: 𝑥, ̄ 𝑟𝜆 ;
̄ 𝑦, /* Posterior mean impact location and confidence radius */
Step 1: TDOA extraction from impact sensor signals:
𝛥𝐭 ← 𝑠𝑗 ; /* TDOA extraction */
Step 2: Maximum likelihood estimation of impact location and TDOA variance:
𝑑𝑗 , 𝜃𝑗 ← 𝑥, 𝑦, 𝑥𝑗 , 𝑦𝑗 ; /* Impact-to-sensor distances and directions */
𝑑𝑗 𝑑
𝛿𝑗 ← ( 𝑣 − 𝑣 (𝜃1 ) ) − 𝛥𝑡1𝑗 ; /* TDOA residual */
𝑔 (𝜃𝑗 ) 𝑔 1
∑𝑁𝑠 2
𝜎𝛿2 ← 𝑁 1−1 𝑗=2 𝛿𝑗 ; /* TDOA variance */
𝑠
𝜕 𝛿𝑗 𝜕 𝛿𝑗
𝜕𝑥
, 𝜕𝑦
← 𝑥, 𝑦, 𝑑𝑗 , 𝜃𝑗 , 𝑣𝑔 (𝜃𝑗 ) ; /* Derivatives of TDOA residual to impact location */
𝜕 𝜎̂ 𝛿2 𝜕 𝜎̂ 𝛿2 𝜕 𝛿𝑗 𝜕 𝛿𝑗
𝜕𝑥
, 𝜕𝑦
← 𝜕𝑥
, 𝜕𝑦
, 𝛿𝑗 ; /* Derivatives of TDOA variance to impact location */
𝑥0 , 𝑦0 ← 𝛥𝑡1𝑗 ; /* Initial impact location */
𝜕 𝜎̂ 𝛿2 𝜕 𝛿𝑗
𝑥, ̂ 𝜎̂ 𝛿2 ← 𝑥0 , 𝑦0 , 𝜎𝛿2 ,
̂ 𝑦, 𝜕𝑥
, 𝜕𝑦
; /* Maximum likelihood estimation of impact location by TRF */
Step 3: Localisation mean and confidence radius estimations by MCMC:
𝑥𝑦 , 𝑥𝑦 ← 𝜎̂ 𝛿2 ; /* Likelihood and prior based on identified TDOA variance */
𝑋 , 𝑌 ← 𝑥𝑦 , 𝑥𝑦 , 𝑥, ̂ max[𝑣𝑔 (𝜃)]2 𝜎̂ 𝛿2 𝐼 ;
̂ 𝑦, /* MCMC samples */
𝑥, ̄ 𝑟𝜆 ← 𝜆, 𝑋 , 𝑌 ;
̄ 𝑦, /* Estimations of posterior mean impact location and confidence radius */

node. To expedite the optimisation process, this study adopts gradient-based approaches, leveraging the gradient information of the
objective function with respect to the impact location. These required gradients are derived using the chain rule:
𝜕 𝜎̂ 𝛿2 ∑ 𝑁𝑠 𝜕𝛿 𝜕 𝜎̂ 2 ∑𝑁𝑠 𝜕𝛿 𝜕 2 𝜎̂ 2 ∑𝑁𝑠 𝜕 𝛿𝑗 2 𝜕2 𝛿
𝜕𝑥
= 𝑁 2−1 𝑗=2 𝛿𝑗 𝜕𝑥𝑗 , 𝜕 𝑦𝛿 = 𝑁 2−1 𝑗=2 𝛿𝑗 𝜕𝑦𝑗 , 𝜕 𝑥2𝛿 = 𝑁 2−1 𝑗=2 ( 𝜕𝑥 ) + 𝛿𝑗 𝜕𝑥2𝑗 ,
𝑠 𝑠 𝑠 (36)
𝜕 2 𝜎̂ 𝛿2 ∑ 𝑁𝑠 𝜕 𝛿 𝑗 2 𝜕2 𝛿 𝜕 2 𝜎̂ 2 ∑𝑁𝑠 𝜕 𝛿𝑗 𝜕 𝛿𝑗 𝜕2 𝛿
𝜕 𝑦2
= 𝑁 2−1 𝑗=2 ( 𝜕 𝑦 ) + 𝛿𝑗 𝜕 𝑦2𝑗 , 𝜕𝑥𝜕𝛿𝑦 = 𝑁 2−1 𝑗=2 𝜕𝑥 𝜕𝑦
+ 𝛿𝑗 𝜕 𝑥𝜕𝑗𝑦 .
𝑠 𝑠

𝜕 𝛿𝑗 𝜕 𝛿𝑗 𝜕 2 𝛿𝑗 𝜕 2 𝛿𝑗 𝜕 2 𝛿𝑗
The intermediate partial derivatives 𝜕𝑥
, 𝜕𝑦
, 𝜕 𝑥2
, 𝜕 𝑦2
, and 𝜕 𝑥𝜕 𝑦
are further calculated based on the TDOA residual 𝛿𝑗 defined
in Eq. (31). Their detailed derivation is presented in Appendix D. Utilising these gradients, the trust-region-reflective (TRF)
algorithm [73], a gradient-based optimisation method well suited to bounded or constrained optimisation problems, efficiently
localises impacts by solving Eq. (34) in a bounded and finite structural space.
The utilisation of gradient-based optimisation stems from the property that when wavefronts with a convex interior (most often
in real structures), such as circles or ellipses, the objective function (𝜎𝛿2 ) typically exhibit a simple shape with only one minimum
within the sensor coverage area. By initialising the optimisation process with two initial points, one inside and one outside the
sensor coverage area, and both closest to the sensor with minimal TDOA, gradient-based optimisation can efficiently locate this
minimum and identify the impact localisation with maximised likelihood.
Algorithm 2 presents the pseudo-algorithm for probabilistic impact localisation with a known GVP, leveraging Bayesian inference.
This algorithm comprises three main steps: TDOA extraction, maximum likelihood estimation of the impact location and TDOA
variance, and localisation confidence radius estimation via MCMC. During the maximum likelihood estimation process, the fixed
parameter for TDOA variance is identified. This parameter is then incorporated as known information within the MCMC, thereby
reducing the number of probabilistic parameters and enhancing computational efficiency.

5.4. Probabilistic impact localisation with probabilistic GVP

Instead of considering a deterministic GVP, the GVP can be regarded as probabilistic, with deterministic values assigned varying
probabilities. This probabilistic GVP can be estimated solely from the physical structural material properties, accounting for material
uncertainties, as present in Section 4.1. Alternatively, it can be estimated based on reference impact data, as demonstrated in
Section 4.2, or through a hybrid approach that combines reference impact data with physical material properties, as outlined in
Section 4.3.
In accordance with the GVP parameterisation method depicted in Eq. (4), the probabilistic GVP is defined within the probability
space spanned by the five velocity coefficients 𝐜. Each observation of the velocity coefficients corresponds to a realisation of the
GVP. Assuming the velocity coefficients with the highest probability are given by 𝐜, ̂ corresponding to the GVP 𝑣̂ 𝑔 (𝜃) with the highest
probability, the impact localisation with the highest probability can be identified using this probabilistic impact localisation method
by inputting this GVP 𝑣̂ 𝑔 (𝜃).
To estimate the 95% confidence localisation area for the probabilistic GVP, the 97.5% confidence interval of the random velocity
coefficients 𝐜 is defined as [𝐜𝑙𝑏 , 𝐜𝑢𝑏 ], which can be characterised by the velocity bounding ratio 𝛼 from Eq. (26). It is observed that
the velocity coefficients with the highest probability correspond to the geometric mean of the confidence bounds. Performing Monte

14
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Carlo simulations of the velocity coefficients within this 97.5% confidence range results in estimations of the impact location and
the corresponding 97.5% confidence localisation area. The union of these areas, constitutes the final 95% confidence localisation
area for the probabilistic GVP.

5.5. Multi-frequency probabilistic impact localisation

Due to the dispersive nature of flexural waves and the availability of TDOAs and GVPs across multiple frequencies, impact
localisation can be performed using a multi-frequency approach. Using Bayesian inference again, the posterior impact location 𝑥, 𝑦,
given the observed data (TDOAs and GVPs at different frequencies 𝛥𝑡𝑖 , 𝑣𝑖𝑔 , 𝑖 = 1, 2, … , 𝑁𝑓 ), is expressed as:
𝑃 (𝛥𝐭 𝑖 ,𝑣𝑖𝑔 ,𝑖=1,2,…,𝑁𝑓 |𝑥,𝑦,𝜎𝛿 )𝑃 (𝑥,𝑦,𝜎𝛿 )
𝑃 (𝑥, 𝑦, 𝜎𝛿 |𝛥𝐭 𝑖 , 𝑣𝑖𝑔 , 𝑖 = 1, 2, … , 𝑁𝑓 ) = . (37)
𝑃 (𝛥𝐭 𝑖 ,𝑣𝑖𝑔 ,𝑖=1,2,…,𝑁𝑓 )

However, the likelihood 𝑃 (𝛥𝐭 𝑖 , 𝑣𝑖𝑔 , 𝑖 = 1, 2, … , 𝑁𝑓 |𝑥, 𝑦, 𝜎𝛿 ) is intractable due to the unknown explicit correlations between TDOAs and
GVPs at each frequency, even though they are highly correlated. In contrast, the likelihood at a single frequency, 𝑃 (𝛥𝐭, 𝑣𝑔 |𝑥, 𝑦, 𝜎𝛿 )
is explicitly known, allowing for posterior estimation of the impact location at a single frequency (as shown in Section 5.1).
When the likelihood is intractable, two methods can be employed to estimate the multi-frequency impact location: Approximate
Bayesian Computation (ABC) [75] and arithmetic averaging. ABC approximates Bayesian updating by simulating TDOAs and GVPs
at multiple frequencies, given an assumed impact location, and comparing the simulated data to the observed data. The generative
models required for simulation include the dispersion relations (for GVP simulation across frequencies, but not available) and
the impact location relations to TDOAs and GVPs (for TDOA simulation with given impact location and GVPs). However, the
GVP estimations from the DD and HPD methods only provide the marginal distribution of GVPs at a single frequency, not their
joint distribution across frequencies. The application of ABC requires the definition of a covariance function (introducing new
hyperparameters) of GVPs at different frequencies. Additionally, ABC introduces an acceptance tolerance (another hyperparameter)
to act as a criterion for accepting or rejecting simulations. Determining the covariance and tolerance and can be computationally
intensive and complex.
In contrast, arithmetic averaging provides a more straightforward alternative by estimating the multi-frequency impact location
as the average of the uni-frequency estimations. Provided that the posterior mean and 95% localisation radius of the 𝑖th frequency
are denoted as (𝑥̄ 𝑖 , 𝑦̄𝑖 , 𝑟𝑖𝜆 ), arithmetic averaging merges these uni-frequency estimates into a multi-frequency estimate (𝑥, ̄ 𝑟𝜆 ) as:
̄ 𝑦,
1 ∑𝑁𝑓 𝑖 𝑖 𝑖
(𝑥, ̄ 𝑟𝜆 ) = 𝑁
̄ 𝑦, 𝑖=1
(𝑥̄ , 𝑦̄ , 𝑟𝜆 ). (38)
𝑓

This method assumes perfect correlation (a value of 1) between GVPs at different frequencies, which is physically plausible, as GVPs
are strongly correlated by the dispersion relations.
Although simple and intuitive, arithmetic averaging is highly sensitive to outliers. Outliers in probabilistic impact localisation
may arise from inaccurate GVP estimation and imprecise TDOA extraction from the sensor signals. The GVP can be analytically
obtained with known material properties for plate-like structures (present in Section 4.1), through a data-driven approach based on
reference impact data (present in Section 4.2), or via a hybrid approach based on reference impact data and reference GVP from a
lower-level structure (present in Section 4.3). Therefore, inaccurate TDOA extraction from the sensor signals is the main source of the
localisation outlier. The detection and elimination of TDOA outliers at the TDOA extraction stage are illustrated in Section 6.2.2.
This procedure significantly decreases the possibility of the presence of uni-frequency impact localisation outlier. Consequently,
arithmetic averaging is employed to integrate these uni-frequency impact localisation estimates, where the corresponding TDOAs
are not outliers, into a multi-frequency estimation. As the frequency whose TDOA is an outlier is not included in impact localisation,
computational costs are reduced as well.

6. Experimental impact testing and TDOA extraction

In this section, the experimental setup for impact testing on three different types of composite panels is detailed. Three TDOA
extraction methods are investigated based on the experimental impact sensor signals to achieve accurate and reliable TDOA
extraction.

6.1. Experimental setup

To validate the proposed probabilistic impact localisation methods and demonstrate their applicability across different structural
configurations, experimental impact testing was performed on three composite panels: a flat panel (FP), a stringer-stiffened panel
(SP), and a sandwich panel (SWP) with composite face sheets and an aluminium honeycomb core, as shown in Fig. 4. Both the FP
and SP panels were fabricated using carbon fibre-reinforced laminates (M21/T800s). The FP panel has dimensions of 290 mm by
200 mm, while the SP panel measures 475 mm by 250 mm. The FP panel features a stacking sequence of [0∕45∕ − 45∕90]2𝑠 , with
each ply having a thickness of 0.25 mm. The SP panel, comprising both the stringer and the single plate region, has a stacking
sequence of [−45∕45∕0∕0∕90∕0]𝑠 and a ply thickness of 0.125 mm. In contrast, the SWP panel, measuring 300 mm by 300 mm,
includes upper and lower composite face sheets (both 2 mm thick) and a 10 mm thick aluminium honeycomb core. However, its
layup and material composition of the laminate face sheets are unknown. During impact testing, the FP and SWP panels were fully

15
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 4. Experimental setup for impact testing: (a) laminated composite flat panel, (b) laminated composite stringer-stiffened panel, (c) sandwich panel with
composite face sheets and an aluminium honeycomb core. Impacts were generated by dropping a mass along a vertical rail. PZT sensors bonded to the panels
were connected to the NI PXI-5105 oscilloscope for impact sensor data acquisition.

Table 2
Mechanical properties of a unidirectional lamina (M21/T800s).
Property 𝐸1 [MPa] 𝐸2 [MPa] 𝐺12 [MPa] 𝐺13 [MPa] 𝐺23 [MPa] 𝜈12 𝜌 [k g∕m3 ]
Value 171 000 11 470 4830 4830 3720 0.33 1600

𝐸1 : elastic modulus in the fibre direction, 𝐸2 : elastic modulus perpendicular to the fibre direction, 𝐺12 , 𝐺13 , 𝐺23 : shear modulus, 𝜈12 :
Poisson’s ratio, 𝜌: density.

clamped along two opposite longitudinal edges, while the remaining two edges were left free. The SP panel, however, was simply
supported at its four corners, with a support area measuring 20 mm by 20 mm.
Two types of PZT sensors and bonding methods were used across the panels. Four PZT sensors (PIC 255) were bonded to the
FP panel using super glue, while four PZT sensors (PIC DuraAct) were bonded to the SP and SWP panels using Hexcel Reduc 312
epoxy-based film adhesive. These sensors and bonding techniques have been widely used in studies involving guided wave SHM [76].
Across all panels, the four sensors formed a rectangular sensor coverage area. The sensors were connected to a NI PXI-5105 8 channel
oscilloscope and the impact sensor signals were recorded using NI SignalExpress software at a sampling frequency of 200 kHz. A
drop mass released from a height of 2 cm generated impacts on the panels, producing an impact energy of approximately 20 mJ. A
total of 35 impact tests, labelled 1–35, were conducted separately on the FP and SP panels. Meanwhile, 49 impacts, labelled 1–49,
were performed on the SWP panel, as illustrated in Fig. 5. The impacts on the FP and SWP panels were evenly distributed within
their sensor coverage areas, with grid spacing of 20 mm by 20 mm and 22 mm by 22 mm, respectively. In contrast, the 35 impacts
on the SP panel were more sparsely distributed, with grid spacing of 40/45 mm by 40 mm. Additionally, some impacts on the SP
panel, specifically those numbered 1–10 and 26–35, were located slightly outside the sensor coverage area.
The implementation of the PB method necessitates knowledge of material properties. Table 2 presents the mechanical properties
of the unidirectional lamina used (M21/T800s). With these mechanical properties and known laminate stacking sequence, the GVP
of simple flat laminated composite panels can be analytically computed based on plate theories.
It is noteworthy that the drop mass impacts occurred on the opposite side of the sensor for the SP panel and SWP panel, whereas
impacts occurred on the same side as the sensor for the FP panel. Accurately localising impacts on these panels demonstrates the
effectiveness of the proposed probabilistic impact localisation framework for real composite structures, regardless of the sensor
attachment side.

16
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 5. Experimental composite panel layup: (a) flat panel (FP) (modified based on Figure 4(a) in [25]), (b) stringer-stiffened panel (SP), (c) sandwich composite
panel with laminate face sheets (SWP). Four PZT sensors placed on the panel were used to capture the impact responses. The 35 and 49 impacts on the respective
FP and SWP were all within the sensor network, while some of the 35 impacts on the SP were outside the sensor coverage.

17
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 6. Spearman’s correlation (SC) and normalised Euclidean distance (NED) between analytical TDOA and extracted TDOA for 35 impacts on the experimental
flat panel: (a)–(c) Spearman’s correlation of NSET, CWTT, and CWTMA, (d)–(f) normalised Euclidean distance of NSET, CWTT, and CWTMA. The NSET method
yields the most consistent TDOA extraction in both rank and value, with 34 impacts having a Spearman’s correlation greater than 0.8 and most of impacts
having a normalised Euclidean distance less than 0.4.

6.2. TDOA extraction from impact sensor signals

Given the dispersive nature of impact-induced flexural waves, it is essential that the TDOA extraction method accurately estimates
TDOAs across a range of frequencies [77,78]. To achieve this, the TDOA extraction methods must include a tunable frequency
parameter. In this study, three such methods are compared: Normalised Smoothed Envelope Threshold (NSET) [20,23], Continuous
Wavelet Transform Threshold (CWTT) [33], and Continuous Wavelet Transform Maximum Amplitude (CWTMA) [29,35].
In the NSET method, the tunable frequency parameter is the cutoff frequency 𝑓𝑐 of the low-pass Butterworth filter. Conversely,
for CWTT and CWTMA, the central frequency of the wavelets serves as the tunable frequency, exploiting the superior time–frequency
resolution provided by wavelet transforms. NSET and CWTT are threshold-based methods, which determine the TDOA when the
signal envelopes (obtained via a low-pass Butterworth filter in NSET) or the CWT amplitude curves (obtained from the CWT in
CWTT) exceed a predetermined threshold. By contrast, CWTMA identifies the TDOA as the time corresponding to the maximum
amplitude in the CWT curves. In both NSET and CWTT, the threshold is set to 0.025 of the maximum amplitude, consistent with the
literature [20,23]. For CWTT and CWTMA, Morlet wavelets are employed to achieve optimal time-frequency resolution, following
established practices in previous studies [29,33,35].

6.2.1. Comparison of three TDOA extraction methods

To evaluate the quality of TDOA estimation using the three methods, 35 impacts at various locations on the FP panel were
analysed. For these impacts, the corresponding analytical TDOA 𝛥𝐭 𝑎 was computed using the analytical GVP 𝑣𝑔 (𝜃) derived from
FSDT (as shown in Section 4.1.2). The analytical TDOA and the experimentally estimated TDOA from sensor signals are represented
as vectors of equal length, as described in Eq. (7). To assess the performance of these TDOA estimation methods, the two vectors
are compared using two key metrics: ranking consistency and numerical distance, measured by Spearman’s correlation (SC) and
Euclidean distance (ED), respectively.
Spearman’s correlation (SC), ranging from −1 to 1, evaluates the consistency in ranking between analytical and estimated TDOAs.
A higher correlation value suggests better agreement in the rank order of TDOA elements, indicating strong shape similarity between
the actual and estimated GVPs. Conversely, the Euclidean distance (ED), normalised by the norm of the analytical TDOA, measures
the numerical discrepancy between the two vectors. A smaller value here indicates greater amplitude similarity between the actual
and estimated GVPs. An effective TDOA extraction method should yield TDOAs with high Spearman’s correlation and low normalised
Euclidean distance (NED) when compared to the analytical TDOA. Fig. 6 presents histograms showing the Spearman’s correlation
and normalised Euclidean distance between the analytical and extracted TDOAs at four different frequencies (5:5:20 kHz) for each
of the three methods.
It is observed that the CWTMA method performs poorly, yielding over 10 impacts with Spearman’s correlation values below
0.4, indicating low TDOA ranking consistency. In contrast, the NSET method consistently achieves high-ranking TDOA extraction,
with only one impact (number 18) showing a negative Spearman’s correlation. This anomaly occurs because the analytical TDOA

18
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 7. Comparison of the normalised CWT amplitude curve (black) at 5 kHz from the CWTT method and the normalised filtered envelope (red) at the same
cutoff frequency from the NSET method, both derived from impact sensor signals of FP18. The CWTT method exhibits TDOA overestimation, attributed to the
low time resolution and averaging effects inherent in the wavelets used for the CWT. In contrast, the NSET method, with its low-pass Butterworth filter, preserves
the signal’s temporal characteristics, resulting in more accurate TDOA estimation.

values for this impact satisfy 𝛥𝑡𝑎12 = 𝛥𝑡𝑎14 and 𝛥𝑡𝑎13 = 0, where minor deviations significantly affect the correlation. Similarly, impact
18 also shows low Spearman’s correlation for the CWTT method. Overall, NSET demonstrates robust performance across different
frequencies, while CWTT shows slight variations in Spearman’s correlation consistency across frequencies.
Regarding normalised Euclidean distance, NSET has the smallest mean and variance among the 35 impacts, with over four-fifths
exhibiting distances under 0.5. A normalised Euclidean distance of 0.5 and 1 indicates that the extracted TDOA is approximately 1.5
and 2 times the analytical TDOA, respectively. Though NSET maintains the ranking, it slightly scales the analytical TDOA values.
Compared to NSET, CWTT overestimates the TDOA to a greater degree. The CWTMA method shows the largest Euclidean distances,
with values up to 20 times those of NSET and CWTT, indicating poor ranking and distance consistency. Overall, NSET achieves the
best ranking and distance accuracy, making it the preferred method for TDOA extraction in impact localisation.
The performance differences between the NSET and CWTT methods can be attributed to their respective signal processing
approaches. The CWTT method, which utilises the Continuous Wavelet Transform (CWT) to extract TDOA from the amplitude curve
at a specific frequency, is sensitive to the time resolution and averaging effects of wavelets. Low-frequency wavelets inherently have
poorer time resolution, leading to broader peaks and slower transitions in the amplitude curve. As a result, TDOA values estimated
by the CWTT method may be overestimated, as shown in Fig. 7, where the normalised CWT amplitude at 5 kHz exhibits a gradual
rise to its maximum, inflating the calculated TDOA 𝛥𝑇14 .
In contrast, the NSET method leverages a low-pass Butterworth filter to extract the signal envelope. The Butterworth filter’s
maximally flat amplitude response allows it to retain the essential temporal features of the signal while suppressing irrelevant high-
frequency noise and low-frequency distortions. This ensures that the envelope of the filtered signal closely follows the sinusoidal
waveform at the cutoff frequency, preserving the fidelity of TDOA estimation. Consequently, the NSET method achieves higher
accuracy, with TDOA values more closely matching the analytical results, as evidenced by the Euclidean distance metrics presented
in Fig. 6. This robust performance makes the NSET method a more reliable choice for TDOA extraction in scenarios requiring high
precision.

6.2.2. TDOA outlier detection by multi-frequency analysis

The precision of TDOA extraction from sensor signals significantly influences the accuracy of impact localisation. In real-world
scenarios where the impact location is unknown, the absence of analytical TDOA poses challenges in evaluating TDOA extraction
accuracy. However, since high-frequency flexural waves propagate faster than low-frequency waves, TDOA extraction from high-
frequency components yields smaller TDOA values. Leveraging multi-frequency TDOA extractions enables the detection of TDOA
outliers and facilitates the evaluation of TDOA accuracy, as demonstrated in Fig. 8, which showcases the maximum TDOA for the
35 impacts on the flat panel across various frequencies using NSET and CWTT methods.
It is evident that with NSET, the lowest frequency of 5 kHz corresponds to the highest maximum TDOA values for all 35 impacts.
However, CWTT does not consistently demonstrate this frequency-TDOA relationship across all 35 impacts. For example, impact 11
and 35 yield the highest maximum TDOA values at the higher frequencies of 20 kHz and 15 kHz using CWTT, respectively. If CWTT
is used for TDOA extraction to locate impacts, these two TDOA extractions are identified as outliers and should not be utilised.
Consistent with the Spearman’s correlation and Euclidean distance analysis based on analytical TDOA, as shown in Fig. 6, NSET
exhibits a better frequency-TDOA relationship than CWTT. Therefore, NSET is employed for TDOA extraction to localise impacts in
the subsequent study.

7. Impact localisation results and analysis

This section mainly presents the impact localisation results and analysis for three experimental composite panels using the
developed probabilistic framework with three GVP estimation methods (PB, DD, and HPD). For the FP panel, the GVP is estimated by

19
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 8. Maximum TDOA across various frequencies: (a) NSET, (b) CWTT. Higher frequency, higher velocity flexural waves give a smaller value of maximum
TDOA. TDOA extraction that violates this rule is identified as an outlier and excluded from the impact localisation.

Fig. 9. Frequency selection for multi-frequency impact localisation. Multi-frequency maximum TDOA analysis allows for the selection of frequencies for impact
localisation. Both reference impact data (if available) and target impact data assist in this process.

the PB and DD methods, since analytical GVP is available using plate theories. For the SP panel, due to the presence of stiffener, DD
and HPD are compared in estimating GVP and impact localisation. The impacts on the SWP panel, without knowing the structural
configurations and material properties, are identified using the DD method.

7.1. Frequency selection for multi-frequency impact localisation

The identification of TDOA outliers also facilitates the selection of frequencies for multi-frequency impact localisation. In the PB
method, where no reference impacts are available, the frequencies used for impact localisation are solely determined by the target
impact itself. In contrast, in the DD and HPD methods, the available reference impacts aid in determining the frequency candidates
for impact localisation, and each target impact to be localised further refines these frequency candidates. Fig. 9 illustrates the
maximum TDOA as the frequency increases for the three experimental composite panels. Across panels, four experimental impacts
were employed as reference impacts for impact localisation using the DD and HPD methods. The reference impacts for the FP and
SP panels are numbered 12, 14, 22, and 24, while the reference impacts for the SWP panel are numbered 17, 19, 31, and 33.

20
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 10. Impact localisation on the FP panel based on analytical GVP and maximum likelihood estimation from: (a)–(c) CLT with frequencies 5 kHz, 10 kHz and
20 kHz, (d)–(f) FSDT with frequencies 5 kHz, 10 kHz and 20 kHz. Localisation accuracy decreases with increasing frequency for CLT, while FSDT consistently
provides more accurate impact localisation than CLT.

Fig. 9(a)–(c) shows that, across the three panels, reference impacts exhibit a decreasing maximum TDOA. This observation aligns
with the physical principle that higher frequencies correspond to higher velocities, resulting in smaller maximum TDOA values.
However, the critical frequency at which the maximum TDOA converges to a constant value varies across panels due to their different
structural configurations. To increase the robustness of impact localisation, the candidate frequencies for impact localisation are cut
off at their corresponding critical frequency. Consequently, seven frequencies ranging from 5 kHz to 20 kHz with a step of 2.5 kHz
were selected for the FP panel, eight frequencies (5:1:10, 12, 15 kHz) were chosen for the SP panel, and six frequencies ranging
from 3 kHz to 8 kHz with a step of 1 kHz were identified for the SWP panel as candidate frequencies for impact localisation.
When employing DD and HPD method for impact localisation, the frequencies used for impact localisation are filtered again
based on the maximum TDOA of each target impact to be localised. Fig. 9(d) illustrates the maximum TDOA of three target
impacts numbered 2, 3, and 9 on FP panel. The decreasing trend persists for impacts 3 and 9 only until 15 kHz and 12.5 kHz,
respectively. Therefore, the frequencies for impact 3 and 9 were selected as 5 kHz to 15 kHz and 5 kHz to 12.5 kHz, respectively.
These frequencies, characterised by a decreasing maximum TDOA, are chosen for probabilistic impact localisation. The localisation
estimations at these individual frequencies are then combined using arithmetic averaging.

7.2. Physics-based (PB) impact localisation on FP panel

Among the three experimental panels, only the FP panel allows for the analytical estimation of GVP based on plate theories with
known material properties.

7.2.1. Impact localisation using analytical GVPs obtained from CLT and FSDT
Fig. 10 illustrates the impact localisation on the flat panel based on the analytical GVPs calculated using both CLT and FSDT at
various frequencies, where the TDOA variance is fixed as its maximum likelihood estimate. In the figures, as indicated in the right-
hand legend, the red solid circles, blue solid squares, magenta solid hexagrams, and magenta hollow circles represent the sensors 𝑠𝑗 ,
actual impact locations 𝑥, 𝑦, the posterior mean location estimates 𝑥, ̄ and the posterior 95% localisation confidence area/radius
̄ 𝑦,
𝑟𝜆 , respectively. The localisation error is defined as the distance
√ between the actual impact locations 𝑥, 𝑦 and the posterior mean
location estimates 𝑥, ̄ calculated as root sum squared error (𝑥 − 𝑥)
̄ 𝑦, ̄ 2 + (𝑦 − 𝑦)
̄ 2.
The localisation results indicate that the accuracy of impact localisation using CLT decreases as frequency increases. This
significant decline in accuracy is primarily due to the fact that the GVP becomes less reliable at higher frequencies, as noted by
Abrate [10]. According to CLT, wave group velocities become unbounded as the frequency approaches infinity. Consequently, when
using GVPs derived from CLT for impact localisation, it is recommended to focus on lower frequencies. Fortunately, impacts with
simple force histories, such as half-sine pulses, predominantly excite structural modes at lower frequencies [79], and these low-
frequency waves dominate the sensor signals. Therefore, using GVPs from CLT at lower frequencies remains effective for impact
localisation.
Additionally, it can be observed that FSDT consistently provides more accurate impact localisation than CLT across various
frequencies. This advantage arises from FSDT’s ability to estimate a more precise and bounded GVP, particularly at higher

21
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 11. Impact localisation on the FP panel using JNI, ICS at frequency 10 kHz and the MCMC samples illustration for impact FP1: (a) impact localisation
with JNI at frequency 10 kHz, (b) impact localisation with ICS frequency 10 kHz, (c) 10 000 MCMC samples of three methods for impact FP1. Both JNI and
ICS result in larger localisation confidence radii compared to CST due to their variability in TDOA variance.

frequencies, as demonstrated in [25]. FSDT demonstrates overwhelming advantage over CLT for GVP estimation and impact
localisation for plate-like structures.
However, as the frequency increases from 10 kHz to 20 kHz, impact 9 shows a significantly larger 95% confidence radius and
large localisation error (31.9 mm). This increased confidence radius is attributed to errors in TDOA estimation for this impact. As
illustrated in Fig. 8, the maximum TDOA for impact 9 at 20 kHz is incorrectly larger than the values at lower frequencies (15 kHz and
10 kHz). This overestimation of maximum TDOA, combined with higher wave velocities, inflates the TDOA variance and results
in a larger localisation confidence radius. Therefore, it is crucial to detect and remove TDOA outliers in multi-frequency impact
localisation to improve accuracy.

7.2.2. Comparison of three TDOA variance treatments in Bayesian inference

Section 5.1 outlines three possible treatments of TDOA variance in Bayesian inference: as a fixed parameter estimated by
maximum likelihood (CST) [66], using Jeffrey’s non-informative prior (JNI) [67], or modelled as an inverse chi-squared distribution
(ICS) [62,68]. This study employed the CST approach due to its efficiency. In Fig. 11, impact localisation results on the FP panel
using JNI and ICS methods at a frequency of 10 kHz are illustrated, alongside 10 000 MCMC samples for impact FP1. Compared
to CST localisation shown in Fig. 10(e), the mean localisation accuracy remains similar across methods, but JNI and ICS exhibit
larger localisation confidence radii than CST. This is further supported by the MCMC sample distributions, where JNI and ICS show
increased variance in localisation predictions, primarily due to their handling of TDOA variance as a variable rather than fixing it
at its maximum likelihood estimation as in CST.
MCMC sampling convergence for the three methods was also analysed. Fig. 12 illustrates the localisation error of the posterior
mean, the standard deviation (SD) of the posterior distribution, and the distance between location estimates derived from the
posterior mean and maximum likelihood estimation (by gradient-based optimisation) for impact FP11, all measured with increasing
MCMC sample sizes. In the CST method, the localisation error and SD converge steadily at a sample size of 1000, displaying minimal
variation as the sample size increases. By contrast, the ICS and JNI methods show significant variability in localisation SD, requiring
larger sample sizes to ensure convergence. At a sample size of 10 000, both ICS and JNI methods converge, with the distance between
the posterior mean and MLE estimates remaining under 0.5 mm. Using MCMC to estimate the posterior distribution (posterior mean
and variance) of impact location yields reliable results, while gradient-based optimisation for maximum likelihood estimation of
TDOA variance and impact location efficiently provides initial spatial information on the impact.
Although all three methods deliver accurate localisation, with the posterior mean closely matching the actual impact location
and the true impact point falling within the 95% confidence radius for most impacts, this study selected the CST approach due to its
rapid convergence and the availability of an accurate MLE estimate of impact location without requiring MCMC. A sample size of
10 000 was chosen to balance convergence reliability with computational efficiency, costing approximately 5 s for all three TDOA
variance treatments.

7.2.3. Multi-frequency impact localisation on FP panel with analytical GVPs

As identified in Section 7.1, a total of 7 frequencies ranging from 5 kHz to 20 kHz with increments of 2.5 kHz are considered for
impact localisation. For each impact on the flat panel, a multi-frequency analysis of the maximum TDOA is conducted to identify and
remove frequency outliers from the impact localisation process. Fig. 13 provides a comparison of multi-frequency impact localisation
with and without the removal of frequency outliers. The mean localisation errors across these 35 impacts are nearly identical, at
approximately 10 mm for both cases. However, the removal of frequency outliers results in a smaller mean confidence radius of
6.9 mm, compared to 7.7 mm without removing frequency outliers. The reason behind this lies in the 95% confidence radius of
impact 9. For impact 9, removing the frequency outliers at 15 kHz, 17.5 kHz, and 20 kHz leads to a reduction in the confidence
radius from 34.4 mm to 18.3 mm. This reduced confidence radius encompasses the actual impact location, indicating a robust
estimation. The multi-frequency impact localisation with the removal of frequency outliers demonstrates to be accurate and robust.

22
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 12. Localisation error of the posterior mean, the SD of the posterior distribution, and the distance between location estimates derived from the posterior
mean and maximum likelihood estimation for impact FP11 as the increase of MCMC samples. For CST, localisation error and SD converge steadily with minimal
variation, whereas the ICS and JNI methods show significant variability in localisation SD with increasing sample size, requiring larger sample sizes to ensure
convergence.

Fig. 13. Multi-frequency impact localisation on the flat panel with analytical GVP: (a) removal of frequency outliers, (b) without removing frequency outliers.
The removal of frequency outliers eliminates the presence of outliers in the localisation results, resulting in more accurate and robust localisation results.

7.3. Data-driven (DD) impact localisation on FP panel

In data-driven impact localisation, the GVP is probabilistically estimated solely based on reference impact data. Bounding the
GVP velocity coefficients is crucial as it determines the optimisation space for maximum likelihood estimation of GVP. The size
of reference impacts significantly influences both the bounding of velocity coefficients and the likelihood to be maximised. In the
following subsections, the effects of reference impact size on the bounding of the GVP velocity coefficients and estimation of the
GVP are investigated. Additionally, impact localisation based on GVPs estimated in a data-driven approach are conducted.

7.3.1. Bounding GVP velocity coefficients

A total of 9 impacts, numbered [12, 14, 24, 22, 13, 19, 23, 17, 18], located within a small rectangular area, as illustrated in
Fig. 5, are considered as reference impacts. The TDOA extracted by the NSET method is utilised to calculate the equivalent velocity
defined in Eq. (21). Its minimum and maximum values serve as the velocity bounds. Fig. 14 presents the lower (LB) and upper (UB)
bounds of equivalent velocity with increasing reference impact size and varying frequencies for TDOA extraction.
It is observed that involving only 1 reference impact is insufficient to accurately identify the velocity bounds. The velocity bounds
converge when the reference impact size is increased to 2. Further increasing the reference impact size does not alter the identified
velocity bounds. This information suggests that a minimum of 2 reference impacts is required in the developed data-driven GVP
estimation method.
Additionally, analysing the bounds across TDOA extraction frequencies, it is found that a higher frequency for TDOA extraction
corresponds to a higher upper bound of the velocity, while the lower bounds are close. This finding is consistent with the dispersion
of flexural waves, where higher-frequency waves have higher propagation velocities. Based on the identified lower and upper bounds
of equivalent velocity, the bounds for the GVP velocity coefficients 𝐜 can be directly estimated using Eqs. (23) and (25).

7.3.2. GVP estimations by data-driven method with HGA

The GVP with maximised likelihood is identified by maximising the log-likelihood within the bounds of the velocity coefficients.
This optimisation task is addressed using a HGA with gradient-based local search, as outlined in Section 4.2.2. To incorporate
additional randomness through GA’s crossover, mutation, and selection processes, the optimisation is iterated 10 times to identify the

23
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 14. Equivalent velocity bounds estimated with increasing reference impact size and varying frequencies for TDOA extraction. The velocity bounds converge
when the reference impact size is increased to 2 for different frequencies.

Fig. 15. Comparison of identified TDOA standard deviation between the hybrid GA and conventional GA. Both optimisation algorithms identified nearly identical
GVPs with maximised likelihood at frequencies of 5 kHz and 10 kHz.

global optimum. Leveraging gradient information, this HGA demonstrates approximately four times greater computational efficiency
than the conventional GA. On a PC with an Intel i7-10700 CPU, it takes on 4.42 s to identify the GVP with maximised likelihood using
the HGA, compared to 15.59 s using the conventional GA. Fig. 15 provides a comparison of the identified TDOA SD 𝜎𝛿 between the
hybrid GA and conventional GA when maximising the log-likelihood (minimising the TDOA variance) with increasing the reference
impact size. The TDOA of reference impacts is estimated using the NSET method from impact sensor signals at frequencies of 5 kHz
and 10 kHz, respectively. The evidence indicates that both the HGA and conventional GA identified nearly identical GVPs with
maximised likelihood for both frequencies. Additionally, as higher frequencies correspond to smaller TDOA values, the 10 kHz
frequency exhibits a smaller TDOA SD.
Fig. 16 illustrates the identified GVPs with maximised likelihood across various reference impact sizes, compared with the
analytical GVP calculated using known material properties and FSDT. As depicted in Fig. 10, the analytical GVPs from FSDT at
frequencies of 5 kHz and 10 kHz demonstrate high localisation accuracy and reliability (small confidence radius), serving as reference
GVPs. Variations in the GVPs are observed across different reference impact sizes, but they exhibit the same shape. With an increase
in the number of reference impacts, the GVP aligns more closely with the analytical GVP. Enhancing the size of the reference impacts
improves the reliability of the estimated GVP. It is noted that with a minimum of 2 reference impacts, a significantly accurate GVP
can be identified. This smaller number indicates a significant saving in the collection of reference impact data.

7.3.3. Data-driven impact localisation based on GVPs with maximised likelihood

The TDOAs of reference impacts can be extracted from sensor signals for any selected frequency, and the corresponding GVP
with maximised likelihood at this frequency can be estimated accordingly for impact localisation. Seven frequencies are considered
in this study, ranging from 5 kHz to 20 kHz with increments of 2.5 kHz. Fig. 17 illustrates the data-driven impact localisation of
the flat panel using four reference impacts and TDOAs extracted at single frequencies of 5 kHz and 10 kHz, as well as at multiple
frequencies. In the impact localisation illustration shown in Fig. 17, the black hollow diamonds represent the reference impacts,
denoted by 𝑥𝑖 , 𝑦𝑖 in the figure legend.
It is observed that with only four reference impacts, the localisation of impacts on the flat panel based on estimated GVPs,
including those outside the coverage area of reference impacts, is quite accurate across single and multiple frequencies. Benefitting
from averaging after the removal of frequency outliers, the multi-frequency impact localisation features a smaller confidence
radius covering the actual impact location compared to that of uni-frequency methods. This indicates that multi-frequency impact
localisation demonstrates higher robustness than uni-frequency methods.

24
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 16. Identified GVP with maximised likelihood of the FP panel based on various reference impact sizes at frequency: (a) 5 kHz, (b) 10 kHz. With a minimum
of 2 reference impacts, a significantly accurate GVP can be identified in a data-driven manner.

Fig. 17. Data-driven impact localisation of the FP panel using 4 reference impacts and TDOAs extracted at (a) uni-frequency 5 kHz, (b) uni-frequency 10 kHz,
(c) multi-frequency 5:2.5:20 kHz, respectively. Multi-frequency impact localisation provides more reliable results than that of uni-frequency.

Table 3
Multi-frequency impact localisation results of the FP panel using GVPs obtained by the PB and DD methods.
Method Mean error [mm] Max error [mm] Mean CR [mm] Max CR [mm]
PB 9.7 18.6 6.9 18.3
DD 8.9 16.0 9.1 29.7

To compare PB impact localisation with DD impact localisation, Table 3 presents a comparison of multi-frequency impact
localisation results for the FP panel using the PB and DD methods. The mean and maximum multi-frequency DD localisation errors
are 8.9 mm and 16.0 mm, respectively, slightly more accurate than the multi-frequency PB localisation. This enhanced accuracy
results from the inclusion of reference impact data, which improves the precision of the estimated GVPs. However, larger variations
in the estimated GVPs at different frequencies, as shown in Fig. 16, contribute to greater mean and maximum localisation confidence
radius in DD localisation. With only 4 reference impacts, the estimation of GVPs at different frequencies can be relatively accurate,
leading to precise impact localisation based on these estimated GVPs, even for the impacts occurring outside the reference coverage
area. For real-world flat plate structures, depending on the availability of structural material properties, both PB and DD are efficient
GVP estimation methods leading to accurate impact localisation.

7.4. Impact localisation on SP panel: DD vs. HPD

For complex stiffened composite panel, calculating the analytical GVP using PB method based on plate theories and known lamina
properties is not feasible. Instead, estimating the GVP requires FEM to acquire the simulated impact responses at predetermined
‘sensorised’ points around an impact point. This simulated impact can be regarded as the reference impacts for estimating the GVP
by the DD and HPD method. Therefore, this subsection mainly presents the maximum likelihood estimation of GVPs and impact
localisation by DD and HPD methods for the experimental SP panel. In the HPD method, the reference GVPs are derived from
analytical GVPs estimated by CLT, assuming the stiffened SP panel is flat but with the same size and laminate layup.

25
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 18. Minimised TDOA SD by the DD and HPD methods of the SP panel based on 4 reference impacts for the maximum likelihood estimation of GVP. The
use of HGA in the DD method identifies a smaller TDOA SD benefiting from its global search capability compared to the gradient-based HPD method. However,
the differences of minimised TDOA SD are small.

Fig. 19. Identified GVPs with maximised likelihood of the SP panel by the DD and HPD methods based on 4 reference impacts at frequencies 5 kHz and 15 kHz.
Although TDOA SD presents a maximum difference of 2 μs, as shown in Fig. 18, the GVPs with maximised likelihood by the DD and HPD methods are almost
identical, with a maximum velocity difference of 18 m/s in all directions.

7.4.1. GVP estimations of SP panel: DD vs. HPD

Fig. 18 illustrates the minimised TDOA SD for the maximum likelihood estimation of GVP by both DD and HPD methods across
the selected frequencies. It is evident that DD identified smaller TDOA SDs than the HPD method due to the utilisation of global
optimisation using HGA. However, the increasing and decreasing trends for both DD and HPD methods remain consistent. Increasing
the frequency from 5 kHz to 9 kHz decreases the TDOA SD, indicating good TDOA consistency at these frequencies. However, further
increases in frequency do not guarantee a decrease in TDOA SD, as seen at frequencies 10 kHz and 15 kHz. This suggests that at
these two frequencies, the TDOA values exhibit greater variability.
The slight differences in the TDOA SD between the DD and HPD methods are consistent with minor disparities in their
corresponding GVPs with maximised likelihood. Fig. 19 depicts the identified GVPs with maximised likelihood by both the DD
and HPD methods based on 4 reference impacts at frequencies 5 kHz and 15 kHz. The GVPs identified by both methods are nearly
identical. Even for the frequency 15 kHz, which corresponds to the greatest difference in TDOA SD, the maximum difference in
GVP is only 18 m/s. It is worth noting that the HPD method, which exclusively employs gradient-based optimisation, demonstrates
higher computational efficiency compared to the DD method.

7.4.2. Impact localisation with identified GVPs: DD vs. HPD

Fig. 20 illustrates the multi-frequency localisation of 35 impacts on the stiffened panel using 4 reference impacts with both the
DD and HPD methods. The confidence radius, represented by the magenta dashed circles in the figures, is scaled by a factor of 0.5
for clarity. Since the DD and HPD methods yield nearly identical GVPs, these impacts are localised at nearly the same locations
with almost identical confidence radii between the two methods. Table 4 summarises the mean and maximum localisation errors
and confidence radii. It is evident that both the DD and HPD methods demonstrate similar accuracy, with mean and maximum
localisation error of approximately 15 mm and 50 mm for all 35 impacts, respectively. The mean and maximum confidence radii
for all 35 impacts are approximately 43 mm and 110 mm, respectively.

26
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 20. Multi-frequency impact localisation of the SP panel based on 4 reference impacts using: (a) DD, (b) HPD. Since the DD and HPD methods yield nearly
identical GVPs, impacts are localised at nearly the same locations with almost identical confidence radii between the two methods. And it is observed that the
impacts inside the sensor network are more accurately located than those outside the sensor network.

Table 4
Multi-frequency impact localisation results of the stiffened SP panel.
Localisation method Mean error [mm] Max error [mm] Mean CR [mm] Max CR [mm]
DD (ALL) 15.3 54.1 42.2 106.8
DD (IN) 11.8 30.1 23.8 106.8
DD (OUT) 18.3 54.1 55.9 91.1
HPD (ALL) 14.5 49.1 44.8 114.1
HPD (IN) 10.7 35.0 27.6 114.1
HPD (OUT) 17.3 49.1 57.7 88.8

ALL: all impacts, IN: impacts inside the sensor network, OUT: impacts outside the sensor network.

Additionally, these 35 impacts can be divided into two groups: 15 impacts numbered 11–25 inside the sensor network, and
20 impacts numbered 1–10, 26–35 outside the sensor network. The impact localisation for impacts inside the sensor network
demonstrates higher accuracy and reliability compared to impacts outside the sensor network. As shown in Table 4, the DD method
localises impacts inside the sensor network with mean and maximum errors of 11.8 mm and 30.1 mm, respectively, compared to
higher values of 18.3 mm and 54.1 mm for impacts outside the sensor network. This is primarily because the four reference impacts
belong to the impact group inside the sensor network, and the identified GVPs with maximised likelihood exhibit higher generality
for impacts closer to the reference impacts.

7.5. Data-driven (DD) impact localisation on SWP panel

Due to the unknown material properties, the DD method was employed for impact localisation on the sandwich composite
panel. Four impacts numbered 17, 19, 31, 33 were utilised for the GVP maximum likelihood estimation and probabilistic impact
localisation.

7.5.1. GVP estimations of SWP panel by DD method

Fig. 21 depicts the maximum likelihood estimation of GVPs by the DD method based on 4 reference impacts at frequencies of
3 kHz, 5 kHz, and 8 kHz. As anticipated, higher frequencies correspond to GVPs with larger amplitudes. Furthermore, the GVPs at
different frequencies exhibit slight variations in shape, underscoring the importance of including multiple frequencies for impact
localisation to enhance reliability and robustness.

7.5.2. Impact localisation with maximised likelihood GVPs

Fig. 22 illustrates the data-driven impact localisation of the SWP using 4 reference impacts at single frequencies of 3 kHz and
8 kHz, as well as multi-frequency localisation spanning from 3 kHz to 8 kHz. Their corresponding mean and maximum localisation
errors and confidence radii are summarised in Table 5. It is evident that at 3 kHz frequency, the impact localisation achieved
higher accuracy and reliability (smaller confidence radii) than at 8 kHz. The multi-frequency impact localisation averages these
impact estimates across frequencies that are not outliers, resulting in more intermediate results. Specifically, the multi-frequency
impact localisation achieved a very high level of accuracy, with mean and maximum localisation errors of 6.7 mm and 16.9 mm,
respectively. Although impacts numbered 1:7:43 exhibited larger confidence radii than the remaining impacts, they were still
accurately localised. This indicates that the larger TDOA uncertainties (larger TDOA variance) of these impacts do not degrade
the accurate impact localisation (posterior mean) due to the randomness of the TDOA uncertainties.

27
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 21. Identified GVPs with maximised likelihood of the SWP panel by the DD method based on 4 reference impacts at frequencies 3 kHz, 5 kHz and 8 kHz.
Higher frequencies correspond to GVPs with larger amplitudes. Slight variations in GVP shape are exhibited at different frequencies.

Fig. 22. Data-driven impact localisation of the SWP panel using 4 reference impacts at (a) uni-frequency 3 kHz, (b) uni-frequency 8 kHz, (c) multi-frequency
3:1:8 kHz, respectively. The multi-frequency impact localisation averages the impact localisation across frequencies, resulting in more intermediate results.

Table 5
Impact localisation results of the SWP panel.
Localisation method Mean error [mm] Max error [mm] Mean CR [mm] Max CR [mm]
DD (3 kHz) 5.8 15.0 8.3 34.8
DD (8 kHz) 8.6 42.8 16.9 76.5
DD (MF) 6.7 16.9 10.0 48.4

7.6. Analysis of the developed probabilistic impact localisation framework

The localisation accuracy achieved with the DD method using four reference impacts on the FP, SP, and SWP panels, summarised
in Tables 3, 4, and 5, demonstrates the effectiveness of the GVP parameterisation in accurately modelling flexural wavefronts in
both the FP and SWP panels. For these panels, the mean localisation error remains below 10 mm, with a maximum error under
20 mm. However, the presence of a stiffener in the SP panel introduces slight deviations in the GVP [33], which the parameterisation
struggles to capture precisely, resulting in mean and maximum errors approximately 2 times greater than those in the FP and SWP
panels. Nonetheless, achieving a mean error below 16 mm for the SP panel with stiffeners is considered acceptable, given that more
than half of these impacts are outside the sensor network and only 4 reference impacts are used that are inside the sensor network.
In contrast to other data-driven impact localisation methods developed in the literature [17–22], which requires dense reference
impact data across the whole structures, the proposed data-driven method exhibited significant advantage in significant saving the
reference impacts. This proposed data-driven method accurately identified GVPs and located impacts across three different plate-like
composite structures using sparse reference impacts (no more than 4) and a sparse sensor network (no more than 4).
The GVP’s structure independence allows for significant scalability across various structures, irrespective of sensor number or
placement. This means that if sensor positions are altered between two identical structures, additional reference data collection
is unnecessary, unlike in traditional data-driven impact localisation methods [17–22]. Furthermore, the developed HPD method
enhances scalability by enabling rapid estimation of a structure’s GVP based on that of its lower-level components.
As shown in Fig. 12(c), the maximum likelihood estimation of the impact location (𝑥, ̂ obtained through gradient-based
̂ 𝑦)
optimisation is nearly identical to the posterior mean (𝑥, ̄ estimated via MCMC. Consequently, maximum likelihood estimation
̄ 𝑦)

28
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

alone (without conducting MCMC) enables highly efficient location estimation at (𝑥, ̂ with localisation uncertainty represented by
̂ 𝑦),
max[𝑣𝑔 (𝜃)]𝜎̂𝛿 . This approach is particularly advantageous for real-time impact monitoring in SHM systems with limited computational
resources at the sensor node level.
In summary, the probabilistic impact localisation framework based on GVP estimation is applicable to various plate-like
structures, including those with stiffeners. Depending on structural complexity and available information, three GVP estimation
methods can be selected for efficient and accurate impact localisation.

8. Discussion, conclusion and outlook

8.1. Discussion on flexural wave propagation-based impact localisation

This study presents a general probabilistic framework for impact localisation using flexural wave propagation, employing
Bayesian inference for estimating the GVP and localising impacts probabilistically with TDOA data extracted from sensor signals.
The success of this framework hinges on accurate GVP estimation and TDOA extraction.
Depending on the available knowledge and structural complexity, the GVP of a target structure can be identified using physics-
based (PB), data-driven (DD), or hybrid physics-data (HPD) methods. For laminated flat plate structures, the PB method leverages
known laminate properties to accurately estimate GVP. In contrast, for more complex structures, the DD and HPD methods offer
greater flexibility and improved performance. The DD method is particularly versatile, applicable to various structures as long as
reference impact data are available. Notably, the developed DD method allows for GVP estimation even with sparse data and sensor
arrays, reducing the cost of data collection. The HPD method enhances scalability by enabling rapid GVP identification for complex
structures based on simpler components with known GVPs. Each method excels in different scenarios, providing a comprehensive
and efficient approach to fast and accurate GVP estimation across a wide range of structural types.
However, the current GVP parameterisation relies on flexural wave dispersion relations derived from CLT, which is most effective
for thin plates. Its application to shell structures depends on the ratio of the shell’s radius of curvature to its thickness. For shallow
shell structures, like aerospace vehicle fuselages with a curvature-to-thickness ratio on the order of 500 [80], the proposed methods
are expected to accurately estimate GVP and localise impacts. In contrast, for thicker shells, such as compressed natural gas composite
tanks with a radius-to-thickness ratio below 20, the effectiveness diminishes due to significant curvature effects on wave speed and
mode interactions [81–83]. In such cases, shell theories or FEA are needed to model wave propagation and GVP for accurate impact
localisation.
Accurate and consistent TDOA extraction is essential for generalising this framework across diverse impact scenarios, which
typically involve unknown factors like impact mass, velocity, and impactor shape. Consistency in TDOA means that impacts at
the same location should yield the same TDOA, while accuracy ensures that TDOA follows the physical wave propagation relation
between location, velocity, and time. The NSET method has been experimentally validated to consistently extract TDOA, accounting
for uncertainties related to environmental conditions and impact variations [20,84]. The authors have also demonstrated TDOA
extraction consistency across small-mass (100 g) and large-mass (5500 g) near-damage impacts [85], ensuring reliable performance
even in near-damage scenarios. This consistency and accuracy in TDOA extraction underpin the general applicability of the wave
propagation-based impact localisation framework.
Both GVP and TDOA are temperature-dependent, and the DD and HPD methods estimate GVP at the temperature corresponding
to the reference impacts. To extend these methods across varying temperatures, temperature compensation techniques for wave
propagation [86,87] have been developed. By measuring the in-service temperature with sensors, GVP calibration through
temperature compensation can ensure accurate GVP estimation and TDOA extraction under real operating conditions, improving
impact localisation accuracy across different temperatures.

8.2. Conclusion and outlook

Depending on the structural complexity and available knowledge, the GVP of the target structure can be determined using
physics-based (PB), data-driven (DD), or hybrid physics-data (HPD) methods, and then applied for probabilistic impact localisation.
Experimental validations on a flat panel (FP), stiffened panel (SP), and sandwich panel (SWP) have led to the following key
conclusions:

• The GVP parameterisation method, based on flexural wave dispersion relations from CLT, effectively models the wavefront in
plate-like composite structures, particularly for flat laminated and sandwich panels.
• The data-driven GVP estimation method successfully identifies accurate GVPs for plate-like composite structures using sparse
reference impacts and a sparse sensor network (both no more than 4).
• The hybrid physics-data GVP estimation method allows fast GVP estimation of a complex structure based on the known GVP
of its simpler component.
• The NSET method demonstrates its ability to extract TDOA accurately and consistently from waves at different frequencies.
• Maximum TDOA analysis aids in detecting and excluding TDOA outliers, enhancing localisation accuracy. Multi-frequency
localisation offers more robust results.

29
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

• The JNI and CST approaches for modelling prior TDOA variance produce nearly identical posterior mean impact localisation
estimates, but with larger localisation variance compared to the constant prior TDOA variance (CST) estimated by maximum
likelihood. While all three approaches are effective, the CST method offers the advantages of faster computation and provides
a useful initial localisation estimate before inferring the posterior.

Looking ahead, future work will involve integrating GVP estimation and TDOA extraction methods into the Bayesian framework,
eliminating the need for manual selection. This can be achieved through Bayesian model class selection [88,89], allowing for
automatic method determination.
Additionally, while this study assumes a normal distribution for TDOA, a hypothesis experimentally validated in Section 3.2
with hundreds of impacts, this validation may not cover all impact scenarios, particularly those that cause damage. Future work
will therefore include comparative studies of various likelihood functions, such as Laplace, Gaussian–Laplace, and generalised normal
distributions [90], to assess their applicability in capturing a broader range of impact conditions.
Further research will also investigate wave propagation in laminated shells with a small radius-to-thickness ratio, where curvature
significantly affects wave speed and mode interactions [81–83], influencing localisation accuracy.
Finally, building on established impact localisation and force reconstruction techniques, efforts will be directed towards
developing a digital twin model for detecting and characterising impact damage using passive SHM sensing.

CRediT authorship contribution statement

Dong Xiao: Writing – review & editing, Writing – original draft, Software, Methodology, Data curation, Conceptualization.
Francisco de Sá Rodrigues: Writing – review & editing, Methodology, Data curation, Conceptualization. Zahra Sharif-Khodaei:
Writing – review & editing, Supervision, Methodology, Conceptualization. M.H. Aliabadi: Writing – review & editing, Supervision,
Methodology, Conceptualization.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared
to influence the work reported in this paper.

Acknowledgements

The first author acknowledges and expresses gratitude to the China Scholarship Council (CSC) for providing a scholarship (No.
[2021]339) to fund his Ph.D. studies.

Appendix A. Parameter space of GVP velocity coefficients

Each realisation of velocity coefficients within their parameter space correspond to a positive GVP in all wave propagation
direction 𝜃:
∑
𝑣4 (𝜃) = 5𝑗=1 𝑐𝑗 𝜙𝑗 (𝜃) = 𝑣4𝑝𝑜𝑠 (𝜃) + 𝑣4𝑛𝑒𝑔 (𝜃) > 0, 𝜙𝑗 (𝜃) = sin(𝑗−1) 𝜃 cos(5−𝑗) 𝜃 ,
(39)
𝑣4𝑝𝑜𝑠 (𝜃) = 𝑐1 cos4 𝜃 + 𝑐3 sin2 𝜃 cos2 𝜃 + 𝑐5 sin4 𝜃 , 𝑣4𝑛𝑒𝑔 (𝜃) = 𝑐2 sin 𝜃 cos3 𝜃 + 𝑐4 sin3 𝜃 cos 𝜃 .

The division of the positive and negative velocity functions 𝑣4𝑝𝑜𝑠 (𝜃) and 𝑣4𝑛𝑒𝑔 (𝜃) is determined by their velocity sign in 𝜃 ∈ [𝜋∕2, 𝜋]. The
positive GVP is ensured when the minimal positive velocity function min[𝑣4𝑝𝑜𝑠 (𝜃)] can compensate for the minimal negative velocity
function min[𝑣4𝑛𝑒𝑔 (𝜃)] in 𝜃 ∈ [0, 𝜋]:
min[𝑣4𝑝𝑜𝑠 (𝜃)] + min[𝑣4𝑛𝑒𝑔 (𝜃)] > 0, 𝜃 ∈ [0, 𝜋]. (40)

The minimal positive velocity 𝑣4𝑝𝑜𝑠 (𝜃)

can be easily calculated:
{
min(𝑐1 , 𝑐5 ) if 𝑐3 ≥ 2 min(𝑐1 , 𝑐5 ),
min[𝑣4𝑝𝑜𝑠 (𝜃)] = (41)
𝑐3 ∕2 if 𝑐3 < 2 min(𝑐1 , 𝑐5 ).

Meanwhile, the negative velocity function reaches it minimum when 𝜃 is the solution to:
𝜕 𝑣4𝑛𝑒𝑔 (𝜃)
𝜕𝜃
= 𝑐2 cos4 𝜃 − 𝑐4 sin4 𝜃 + 3(𝑐4 − 𝑐2 )𝑠𝑖𝑛2 𝜃 𝑐 𝑜𝑠2 𝜃 = 0. (42)

Letting 𝑐 𝑜𝑠2 𝜃 = 𝑞 ∈ [0, 1] and 𝑠𝑖𝑛2 𝜃 = 1 − 𝑞, this equality can be simplified into a quadratic equation, which only have a solution 𝑞𝑒
in the range of [0, 1]:
𝑔(𝑞) = 𝑐2 𝑞 2 − 𝑐4 (1 − 𝑞)2 + 3(𝑐4 − 𝑐2 )𝑞(1 − 𝑞) = 4(𝑐2 − 𝑐4 )𝑞 2 + (5𝑐4 − 3𝑐2 )𝑞 − 𝑐4 = 0,
(43)
𝑔(0)𝑔(1) = −𝑐4 𝑐2 ≤ 0.

30
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

The solution 𝑞𝑒 of this quadratic equation is derived as:

⎧
⎪1∕2 if 𝑐2 = 𝑐4 ,
𝑞𝑒 = ⎨ 3𝑐 −5𝑐 ±√9(𝑐 −𝑐 )2 +4𝑐 𝑐 (44)
2 4 2 4 2 4
⎪ if 𝑐2 ≠ 𝑐4 .
⎩ 8(𝑐2 −𝑐4 )

In the scenario where 𝑐2 ≠ 𝑐4 , the choice of the ± operation is determined by the bounds 𝑞𝑒 ∈ [0, 1]. Substituting 𝑐 𝑜𝑠2 𝜃 = 𝑞𝑒 into the
negative velocity function yields its minimum:
| √ |
min[𝑣4𝑛𝑒𝑔 (𝜃)] = − |[𝑐2 𝑞𝑒 + 𝑐4 (1 − 𝑞𝑒 )] 𝑞𝑒 (1 − 𝑞𝑒 )| . (45)
| |
Appendix B. Derivation of group velocity based on FSDT

The elements of symmetric matrix 𝐆 are given by [10,25,32]:

𝑔11 = −(𝐴44 sin2 𝜃 + 𝐴55 cos2 𝜃 + 2𝐴45 sin 𝜃 cos 𝜃)𝑘2 + 𝐼1 𝜔2 ,
𝑔12 = 𝑖(𝐴55 cos 𝜃 + 𝐴45 sin 𝜃)𝑘,
𝑔13 = 𝑖(𝐴45 cos 𝜃 + 𝐴44 sin 𝜃)𝑘,
(46)
𝑔22 = (𝐷11 cos2 𝜃 + 2𝐷16 sin 𝜃 cos 𝜃 + 𝐷66 sin2 𝜃)𝑘2 + 𝐴55 − 𝐼3 𝜔2 ,
𝑔23 = [(𝐷12 + 𝐷66 ) sin 𝜃 cos 𝜃 + 𝐷16 cos2 𝜃 + 𝐷26 sin2 𝜃]𝑘2 + 𝐴45 ,
𝑔33 = (𝐷22 sin2 𝜃 + 𝐷66 cos2 𝜃 + 2𝐷26 sin 𝜃 cos 𝜃)𝑘2 + 𝐴44 − 𝐼3 𝜔2 ,

In these elements, 𝑖 denotes the imaginary unit (𝑖2 = −1). 𝐷𝑖𝑗 where 𝑖, 𝑗 = 1, 2, 6, 𝐴𝑖𝑗 where 𝑖, 𝑗 = 1, 2, 6, represent laminate stiffness
ℎ∕2
detailed in [25]. 𝐼1 = 𝜌ℎ denotes the mass per unit length. 𝐼3 = ∫−ℎ∕2 𝜌𝑧2 𝑑 𝑧 is typically referred to as the rotational inertia. The
derivation of group velocity shown in Eq. (11) involves calculations of the adj(𝐆) (adjugate of the matrix 𝐆), 𝜕𝐆 𝜕𝑘
(element-wise
derivatives of the matrix 𝐆 with respect to 𝑘) and 𝜕𝐆
𝜕𝜔
(element-wise derivatives of the matrix 𝐆 with respect to 𝜔). The adjugate of
the matrix 𝐆 can be easily obtained by cofactor expansion. And these element-wise derivatives are also simply derived as follows:
[ 𝜕𝐆
𝜕𝑘
] = −2(𝐴44 sin2 𝜃 + 𝐴55 cos2 𝜃 + 2𝐴45 sin 𝜃 cos 𝜃)𝑘,
11
𝜕𝐆
[ 𝜕𝑘 ] = 𝑖(𝐴55 cos 𝜃 + 𝐴45 sin 𝜃), [ 𝜕𝐆 ] = 𝑖(𝐴45 cos 𝜃 + 𝐴44 sin 𝜃),
𝜕 𝑘 13
12
𝜕𝐆
[ 𝜕𝑘 ] = 2(𝐷11 cos2 𝜃 + 2𝐷16 sin 𝜃 cos 𝜃 + 𝐷66 sin2 𝜃)𝑘,
22
(47)
[ 𝜕𝐆 ]
𝜕 𝑘 23
= 2[(𝐷12 + 𝐷66 ) sin 𝜃 cos 𝜃 + 𝐷16 cos2 𝜃 + 𝐷26 sin2 𝜃]𝑘,
[ 𝜕𝐆 ] = 2(𝐷22 sin2 𝜃 + 𝐷66 cos2 𝜃 + 2𝐷26 sin 𝜃 cos 𝜃)𝑘,
𝜕 𝑘 33
𝜕𝐆
𝜕𝜔
= diag([2𝐼1 𝜔, −2𝐼3 𝜔, −2𝐼3 𝜔]).

where diag is the diagonal operator that turns a vector into a diagonal matrix.

Appendix C. Dispersion curve comparison

This appendix presents a comparison of dispersion curves for a composite laminated plate calculated using classical laminate
theory (CLT), first-order shear deformation theory (FSDT), and the semi-analytical finite element-based freeware SAFEDC [52].
SAFEDC serves as a numerical benchmark for calculating dispersion relations. Its semi-analytical finite element formulation accounts
for detailed material properties and complex laminate structures, offering high accuracy across a wide frequency range.
The laminated plate under consideration is composed of AS4M3502, with the stacking sequence [0∕90∕ − 45∕45]2𝑠 and a thickness
of 2 mm. The material properties of AS4M3502 are detailed in [52]. Given that flexural waves dominate low-velocity impacts and
the A0 wave mode (particularly in the low-frequency range) carries the majority of the impact energy, the dispersion curves of
the A0 mode were selected for analysis and comparison. Wave propagation along 30◦ of the laminate, a common orientation for
examining anisotropic effects, was used for consistency with previous studies (e.g., Figure 11 in [52]).
Fig. 23 illustrates the dispersion curves for the A0 mode, calculated using CLT, FSDT, and SAFEDC. It can be observed that as
expected, CLT significantly overestimates both phase and group velocities in the higher frequency range. However, in the low-
frequency range, its results are in reasonable agreement with those obtained using FSDT and SAFEDC. Despite its limitations
in accuracy, CLT captures the general shape of the group velocity profile (GVP), which can be useful for parameterising wave
propagation characteristics. In contrast, FSDT provides more accurate dispersion relations. Its predictions for both phase and group
velocities closely align with those of SAFEDC, especially in the critical group velocity relevant to impact localisation. The accuracy of
FSDT ensures it is a robust and computationally efficient alternative to more complex numerical methods, such as those implemented
in SAFEDC.
This comparison confirms the validity of plate theories, particularly FSDT, in modelling the A0 mode of guided waves. While
CLT exhibits limitations due to overestimation in the higher frequency range, it remains a useful tool for capturing the overall shape
of the GVP. FSDT, however, proves to be highly accurate and suitable for deriving dispersion relations, particularly in applications
such as impact localisation, where precision in group velocity estimation is critical.

31
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Fig. 23. Comparison of the A0 mode dispersion curves calculated using CLT, FSDT, and the semi-analytical finite element-based software SAFEDC [52] along 30◦
of the AS4M3502 laminate: (a) phase velocity, (b) group velocity. CLT significantly overestimates both phase and group velocities in the higher frequency range,
showing reasonable agreement only at lower frequencies. In contrast, FSDT provides more accurate dispersion relations, with its predictions closely aligning with
those of SAFEDC, particularly for group velocity.

Appendix D. Partial derivative of TDOA residual with respect to impact location

𝜕𝛿 𝜕𝛿 𝜕2 𝛿 𝜕2 𝛿 𝜕2 𝛿
The intermediate first and second order partial derivatives 𝜕𝑥𝑗 , 𝜕 𝑦𝑗 , 𝜕 𝑥2𝑗 , 𝜕𝑦2𝑗 , and 𝜕𝑥𝜕𝑗𝑦 are involved in the maximum likelihood
estimation of the impact location with highest probability by gradient-based optimisation approach. These first-order intermediate
partial derivatives can be further derived by chain rule:
𝜕 𝛿𝑗 𝜕 𝑡𝑗 𝜕 𝑡1 𝜕 𝑡𝑗 1 𝜕 𝑑𝑗 𝑑𝑗 𝜕 𝑣𝑔 (𝜃𝑗 )
= − , = (𝑣 − ),
𝜕𝑥 𝜕𝑥 𝜕𝑥 𝜕𝑥 𝑔 (𝜃𝑗 ) 𝜕𝑥 𝑣2𝑔 (𝜃𝑗 ) 𝜕 𝑥
𝜕 𝛿𝑗 𝜕 𝑡𝑗 𝜕 𝑡1 𝜕 𝑡𝑗 1 𝜕 𝑑𝑗 𝑑𝑗 𝜕 𝑣𝑔 (𝜃𝑗 )
= − , = ( 𝑣 (𝜃 − ), (48)
𝜕𝑦 𝜕𝑦 𝜕𝑦 𝜕𝑦 𝑔 𝑗 ) 𝜕𝑦 𝑣2𝑔 (𝜃𝑗 ) 𝜕𝑦
𝜕 𝑑𝑗 𝑥−𝑥𝑗 𝜕 𝑑𝑗 𝑦−𝑦𝑗
𝜕𝑥
= 𝑑𝑗
, 𝜕𝑦
= 𝑑𝑗
.

These second-order intermediate partial derivatives are further derived as:

𝜕 2 𝛿𝑗 𝜕 2 𝑡𝑗 𝜕2 𝑡
𝜕 𝑥2
= 𝜕 𝑥2
− 𝜕 𝑥21 ,
𝜕 2 𝑡𝑗 1 𝜕 𝑑𝑗
2
1 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝑑𝑗 1 𝜕 𝑑𝑗 −2𝑑𝑗 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝑑𝑗 𝜕 2 𝑣𝑔 (𝜃𝑗 )
𝜕 𝑥2
= ( 𝑣 (𝜃 2 − 𝑣2 (𝜃 ) ) − [( + ) 𝜕𝑥 + ],
𝑔 𝑗 ) 𝜕𝑥 𝑔 𝑗 𝜕𝑥 𝜕𝑥 𝑣2𝑔 (𝜃𝑗 ) 𝜕 𝑥 𝑣3𝑔 (𝜃𝑗 ) 𝜕𝑥 𝑣2𝑔 (𝜃𝑗 ) 𝜕 𝑥2
𝜕 2 𝛿𝑗 𝜕 2 𝑡𝑗
𝜕2 𝑡
𝜕 𝑦2
= 𝜕 𝑦2 − 𝜕𝑦21 ,
𝜕 2 𝑡𝑗 1 𝜕 𝑑𝑗
2
1 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝑑𝑗 1 𝜕 𝑑𝑗 −2𝑑𝑗 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝑑𝑗 𝜕 2 𝑣𝑔 (𝜃𝑗 )
𝜕 𝑦2
= ( 𝑣 (𝜃 2 − 𝑣2 (𝜃 ) ) − [( + ) + ], (49)
𝑔 𝑗 ) 𝜕𝑦 𝑔 𝑗 𝜕𝑦 𝜕𝑦 𝑣2𝑔 (𝜃𝑗 ) 𝜕 𝑦 𝑣3𝑔 (𝜃𝑗 ) 𝜕𝑦 𝜕𝑦 𝑣2𝑔 (𝜃𝑗 ) 𝜕 𝑦2
𝜕 2 𝛿𝑗 𝜕2 𝑡 𝑗 𝜕2 𝑡 1
𝜕 𝑥𝜕 𝑦
= 𝜕 𝑥𝜕 𝑦
− 𝜕 𝑥𝜕 𝑦
,
𝜕 2 𝑡𝑗 1 𝜕 2 𝑑𝑗 1 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝑑𝑗 1 𝜕 𝑑𝑗 −2𝑑𝑗 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝑑𝑗 𝜕 2 𝑣𝑔 (𝜃𝑗 )
= ( 𝑣 (𝜃 − ) − [( + ) + ],
𝜕 𝑥𝜕 𝑦 𝑔 𝑗 ) 𝜕 𝑥𝜕 𝑦 𝑣2𝑔 (𝜃𝑗 ) 𝜕 𝑦 𝜕𝑥 𝑣2𝑔 (𝜃𝑗 ) 𝜕 𝑦 𝑣3𝑔 (𝜃𝑗 ) 𝜕𝑦 𝜕𝑥 𝑣2𝑔 (𝜃𝑗 ) 𝜕 𝑥𝜕 𝑦
𝜕 2 𝑑𝑗 (𝑦−𝑦𝑗 )2 𝜕 2 𝑑𝑗 (𝑥−𝑥𝑗 )2 𝜕 2 𝑑𝑗 −(𝑥−𝑥𝑗 )(𝑦−𝑦𝑗 )
𝜕 𝑥2
= , = , = .
𝑑𝑗3 𝜕 𝑦2 𝑑𝑗3 𝜕 𝑥𝜕 𝑦 𝑑𝑗3

𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 2 𝑣𝑔 (𝜃𝑗 ) 𝜕 2 𝑣𝑔 (𝜃𝑗 ) 𝜕 2 𝑣𝑔 (𝜃𝑗 )

The intermediate derivatives, 𝜕𝑥
, 𝜕𝑦
, 𝜕 𝑥2
, 𝜕 𝑦2
, 𝜕 𝑥𝜕 𝑦
are non-zero only when the optimised GVP exhibits non-
circular wavefronts. This is because even small variations in the location coordinates 𝑥, 𝑦 result in changes in the wave propagation
direction 𝜃, thereby altering the wave propagation velocity 𝑣𝑔 (𝜃). These derivatives can be derived using the chain rule:
𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝜃𝑗 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝜃𝑗 𝜕 2 𝑣𝑔 (𝜃𝑗 ) 𝜕 2 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝜃𝑗 2 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 2 𝜃𝑗
= , = , 𝜕 𝑥2
= ( 𝜕𝑥 ) + ,
𝜕𝑥 𝜕 𝜃𝑗 𝜕𝑥 𝜕𝑦 𝜕 𝜃𝑗 𝜕𝑦 𝜕𝜃2 𝑗
𝜕 𝜃𝑗 𝜕 𝑥2
𝜕 2 𝑣𝑔 (𝜃𝑗 ) 𝜕 2 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝜃𝑗 2 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 2 𝜃𝑗 𝜕 2 𝑣𝑔 (𝜃𝑗 ) 𝜕 2 𝑣𝑔 (𝜃𝑗 ) 𝜕 𝜃𝑗 𝜕 𝜃𝑗 𝜕 𝑣𝑔 (𝜃𝑗 ) 𝜕 2 𝜃𝑗
(50)
𝜕 𝑦2
= ( 𝜕𝑦 ) + , = + .
𝜕𝜃2 𝑗
𝜕 𝜃𝑗 𝜕 𝑦2 𝜕 𝑥𝜕 𝑦 𝜕 𝜃𝑗2 𝜕𝑥 𝜕𝑦 𝜕 𝜃𝑗 𝜕 𝑥𝜕 𝑦

The computation of these intermediate derivatives involves the derivatives of the GVP 𝑣𝑔 (𝜃) with respect to the direction 𝜃 and
the derivatives of the direction 𝜃 with respect to the impact coordinates (𝑥, 𝑦). As the wave propagation direction 𝜃𝑗 is given by
cos(𝜃𝑗 ) = (𝑥𝑗 − 𝑥)∕𝑑𝑗 , the partial derivatives of 𝜃𝑗 with respect to impact location (𝑥, 𝑦) are derived as:
𝜕 𝜃𝑗 (𝑦𝑗 −𝑦) (𝑦𝑗 −𝑦) 𝜕 𝜃𝑗 −(𝑥𝑗 −𝑥) (𝑥−𝑥𝑗 )
= (𝑥𝑗 −𝑥)2 +(𝑦𝑗 −𝑦)2
= , = = ,
𝜕𝑥 𝑑𝑗2 𝜕𝑦 (𝑥𝑗 −𝑥)2 +(𝑦𝑗 −𝑦)2 𝑑𝑗2
𝜕 2 𝜃𝑗 2(𝑥−𝑥𝑗 )(𝑦−𝑦𝑗 ) 𝜕 2 𝜃𝑗 −2(𝑥−𝑥𝑗 )(𝑦−𝑦𝑗 ) 𝜕 2 𝜃𝑗 (𝑦−𝑦𝑗 )2 −(𝑥−𝑥𝑗 )2 (51)
𝜕 𝑥2
= , = , = .
𝑑𝑗4 𝜕 𝑦2 𝑑𝑗4 𝜕 𝑥𝜕 𝑦 𝑑𝑗4

32
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

Based on the dispersion relation that parameterises the GVP, as shown in Eq. (4), obtaining the partial derivative of the wave
velocity with respect to the wave direction is straightforward:
𝜕 𝑣𝑔 (𝜃) ∑5 𝑑 𝜙𝑗 (𝜃)
𝜕𝜃
= 31 𝑗=1 𝑐𝑗 𝑑 𝜃 ,
4𝑣𝑔 (𝜃)
𝜕 2 𝑣𝑔 (𝜃) ∑5 𝑑 2 𝜙𝑗 (𝜃) 𝜕 𝑣𝑔 (𝜃) 2 (52)
𝜕𝜃2
= 31 3
𝑗=1 𝑐𝑗 𝑑 𝜃 2 − 𝑣 (𝜃) ( 𝜕 𝜃 ) .
4𝑣𝑔 (𝜃) 𝑔

Data availability

Data will be made available on request.

References

[1] Faivre Vincent, Morteau Emilie, Damage tolerant composite fuselage sizing, characterization of accidental damage threat, Airbus Tech Mag (FAST) (48)
(2011) 10–16, https://aircraft.airbus.com/en/fast-magazine-articles.
[2] M.H. Ferri Aliabadi, Z. Sharif Khodaei, Structural health monitoring for advanced composite structures, Computational and Experimental Methods in
Structures, vol. 08, WORLD SCIENTIFIC (EUROPE), 2018, https://doi.org/10.1142/q0114.
[3] Mircea Calomfirescu, Axel S Herrmann, Christian König, Jan Müller, Investigation on wave propagation in composites as requirement for impact detection,
in: 25th International Congress of the Aeronautical Sciences, 2006, pp. 1–7, https://www.icas.org/icas_archive/ICAS2006/.
[4] S. Mahzan, W.J. Staszewski, Impact localisation of a smart composite panel using wave velocity propagation, Manuf. Eng. (2008) https://core.ac.uk/
download/pdf/12005636.pdf.
[5] A. Tobias, Acoustic-emission source location in two dimensions by an array of three sensors, Non-Destr. Test. 9 (1) (1976) 9–12, https://doi.org/10.1016/
0029-1021(76)90027-X.
[6] Tribikram Kundu, Acoustic source localization, Ultrasonics 54 (1) (2014) 25–38, https://doi.org/10.1016/j.ultras.2013.06.009.
[7] F. Dipietrangelo, F. Nicassio, G. Scarselli, Structural health monitoring for impact localisation via machine learning, Mech. Syst. Signal Process. 183 (2023)
109621, https://doi.org/10.1016/j.ymssp.2022.109621.
[8] F.C. Moon, A critical survey of wave propagation and impact in composite materials, Technical Report NASA-CR-121226, 1973, https://ntrs.nasa.gov/
citations/19730021203.
[9] Byoung Sung Kim, Francis Moon, Impact induced stress waves in an anisotropic plate, AIAA J. 17 (10) (1979) 1126–1133, https://doi.org/10.2514/3.61287.
[10] Serge Abrate, Impact on Composite Structures, Cambridge University Press, Cambridge, 1998, https://doi.org/10.1017/CBO9780511574504.
[11] F.C. Moon, Theoretical Analysis of Impact in Composite Plates, Technical Report NASA-AMS-1099, 1973, https://ntrs.nasa.gov/citations/19730019054.
[12] Kaige Zhu, Xinlin P. Qing, Bin Liu, Tiren He, A passive localization method for stiffened composite structures with a parameterized laminate model, J.
Sound Vib. 489 (2020) 115683, https://doi.org/10.1016/j.jsv.2020.115683.
[13] Lorenzo Capineri, Andrea Bulletti, A versatile analog electronic interface for piezoelectric sensors used for impacts detection and positioning in structural
health monitoring (SHM) systems, Electronics 10 (9) (2021) 1047, https://doi.org/10.3390/electronics10091047.
[14] Andrea Bulletti, Eugenio Marino Merlo, Lorenzo Capineri, Analysis of the accuracy in impact localization using piezoelectric sensors for structural health
monitoring with multichannel real-time electronics, in: 2020 IEEE 7th International Workshop on Metrology for AeroSpace (MetroAeroSpace), 2020, pp.
480–484, https://doi.org/10.1109/MetroAeroSpace48742.2020.9160275.
[15] T.M. Tan, C.T. Sun, Wave propagation in graphite/epoxy laminates due to impact, Technical Report NASA-CR-168057, 1982, https://ntrs.nasa.gov/citations/
19830014054.
[16] I.M. Daniel, T. Liber, R.H. LaBedz, Wave propagation in transversely impacted composite laminates, Exp. Mech. 19 (1) (1979) 9, https://doi.org/10.1007/
BF02327764.
[17] Llewellyn Morse, Zahra Sharif Khodaei, M.H. Aliabadi, Reliability based impact localization in composite panels using Bayesian updating and the Kalman
filter, Mech. Syst. Signal Process. 99 (2018) 107–128, https://doi.org/10.1016/j.ymssp.2017.05.047.
[18] Z. Sharif-Khodaei, M. Ghajari, M.H. Aliabadi, Determination of impact location on composite stiffened panels, Smart Mater. Struct. 21 (10) (2012) 105026,
https://doi.org/10.1088/0964-1726/21/10/105026.
[19] Prabakaran Balasubramanian, Vikram Kaushik, Sumaya Y Altamimi, Marco Amabili, Mohamed Alteneiji, Comparison of neural networks based on
accuracy and robustness in identifying impact location for structural health monitoring applications, Struct. Health Monit. 22 (1) (2023) 417–432,
https://doi.org/10.1177/14759217221098569.
[20] Aldyandra Hami Seno, Zahra Sharif Khodaei, M.H. Ferri Aliabadi, Passive sensing method for impact localisation in composite plates under simulated
environmental and operational conditions, Mech. Syst. Signal Process. 129 (2019) 20–36, https://doi.org/10.1016/j.ymssp.2019.04.023.
[21] Aldyandra Hami Seno, M.H. Ferri Aliabadi, Uncertainty quantification for impact location and force estimation in composite structures, Struct. Health
Monit. (2021) 147592172110202, https://doi.org/10.1177/14759217211020255.
[22] M.R. Jones, T.J. Rogers, K. Worden, E.J. Cross, A Bayesian methodology for localising acoustic emission sources in complex structures, Mech. Syst. Signal
Process. 163 (2022) 108143, https://doi.org/10.1016/j.ymssp.2021.108143.
[23] Aldyandra Hami Seno, M.H. Ferri Aliabadi, Multifidelity data augmentation for data driven passive impact location and force estimation in composite
structures under simulated environmental and operational conditions, Mech. Syst. Signal Process. 195 (2023) 110288, https://doi.org/10.1016/j.ymssp.
2023.110288.
[24] Nur M.M. Kalimullah, Amit Shelke, Anowarul Habib, A probabilistic framework for source localization in anisotropic composite using transfer learning based
multi-fidelity physics informed neural network (mfpinn), Mech. Syst. Signal Process. 197 (2023) 110360, https://doi.org/10.1016/j.ymssp.2023.110360.
[25] Dong Xiao, Zahra Sharif-Khodaei, M.H. Aliabadi, Hybrid physics-based and data-driven impact localisation for composite laminates, Int. J. Mech. Sci. 274
(2024) 109222, http://spiral.imperial.ac.uk/handle/10044/1/113093.
[26] T. Kundu, X. Yang, H. Nakatani, N. Takeda, A two-step hybrid technique for accurately localizing acoustic source in anisotropic structures without knowing
their material properties, Ultrasonics 56 (2015) 271–278, https://doi.org/10.1016/j.ultras.2014.08.009.
[27] Mario Emanuele De Simone, Francesco Ciampa, Salvatore Boccardi, Michele Meo, Impact source localisation in aerospace composite structures, Smart
Mater. Struct. 26 (12) (2017) 125026, https://doi.org/10.1088/1361-665X/aa973e.
[28] E. Dehghan Niri, A. Farhidzadeh, S. Salamone, Nonlinear Kalman filtering for acoustic emission source localization in anisotropic panels, Ultrasonics 54
(2) (2014) 486–501, https://doi.org/10.1016/j.ultras.2013.07.016.
[29] E. Dehghan Niri, S. Salamone, A probabilistic framework for acoustic emission source localization in plate-like structures, Smart Mater. Struct. 21 (3)
(2012) 035009, https://doi.org/10.1088/0964-1726/21/3/035009.
[30] Kaige Zhu, Xinlin P. Qing, Bin Liu, A two-step impact localization method for composite structures with a parameterized laminate model, Compos. Struct.
192 (2018) 500–506, https://doi.org/10.1016/j.compstruct.2018.03.052.

33
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

[31] T.S. Chow, On the propagation of flexural waves in an orthotropic laminated plate and its response to an impulsive load, J. Compos. Mater. 5 (3) (1971)
306–319, https://doi.org/10.1177/002199837100500302.
[32] Hyunjo Jeong, Young-Su Jang, Wavelet analysis of plate wave propagation in composite laminates, Compos. Struct. 49 (4) (2000) 443–450, https:
//doi.org/10.1016/S0263-8223(00)00079-9.
[33] Deshuang Deng, Xu Zeng, Zhengyan Yang, Yu Yang, Sheng Zhang, Shuyi Ma, Hao Xu, Lei Yang, Zhanjun Wu, Multi-frequency probabilistic imaging fusion
for impact localization on aircraft composite structures, Struct. Health Monit. (2024) 14759217241233181, https://doi.org/10.1177/14759217241233181.
[34] Francesco Ciampa, Michele Meo, Ettore Barbieri, Impact localization in composite structures of arbitrary cross section, Struct. Health Monit. 11 (6) (2012)
643–655, https://doi.org/10.1177/1475921712451951.
[35] M. Meo, G. Zumpano, M. Piggott, G. Marengo, Impact identification on a sandwich plate from wave propagation responses, Fifth International Conference
on Composite Science and Technology, Compos. Struct. 71 (3) (2005) 302–306, https://doi.org/10.1016/j.compstruct.2005.09.028.
[36] E. Kirkby, R. de Oliveira, V. Michaud, J.A. Månson, Impact localisation with FBG for a self-healing carbon fibre composite structure, Compos. Struct. 94
(1) (2011) 8–14, https://doi.org/10.1016/j.compstruct.2011.07.030.
[37] Tribikram Kundu, Samik Das, Kumar V. Jata, Detection of the point of impact on a stiffened plate by the acoustic emission technique, Smart Mater. Struct.
18 (3) (2009) 035006, https://doi.org/10.1088/0964-1726/18/3/035006.
[38] S.P. Timoshenko, On the correction for shear of the differential equation for transverse vibrations of prismatic bars, Lond. Edinb. Dublin Philos. Mag. J.
Sci. 41 (245) (1921) 744–746, https://doi.org/10.1080/14786442108636264.
[39] R.D. Mindlin, Influence of rotatory inertia and shear on flexural motions of isotropic, elastic plates, J. Appl. Mech. 18 (1) (1951) 31–38, https:
//doi.org/10.1115/1.4010217.
[40] Stephen Timoshenko, S. Woinowsky-Krieger, Theory of Plates and Shells, McGraw-Hill, 1959, https://archive.org/details/TheoryOfPlatesAndShells.
[41] Shirsendu Sikdar, Paresh Mirgal, Sauvik Banerjee, Low-velocity impact source localization in a composite sandwich structure using a broadband piezoelectric
sensor network, Compos. Struct. 291 (2022) 115619, https://doi.org/10.1016/j.compstruct.2022.115619.
[42] Byeong-Wook Jang, Chun-Gon Kim, Impact localization of composite stiffened panel with triangulation method using normalized magnitudes of fiber optic
sensor signals, Compos. Struct. 211 (2019) 522–529, https://doi.org/10.1016/j.compstruct.2019.01.028.
[43] F. Ciampa, M. Meo, Acoustic emission source localization and velocity determination of the fundamental mode A0 using wavelet analysis and a
Newton-based optimization technique, Smart Mater. Struct. 19 (4) (2010) 045027, https://doi.org/10.1088/0964-1726/19/4/045027.
[44] J.N. Reddy, Mechanics of Laminated Composite Plates and Shells: Theory and Analysis, CRC Press, 2003, https://doi.org/10.1201/b12409.
[45] Simon Jackman, Estimation and inference via Bayesian simulation: An introduction to Markov chain Monte Carlo, Am. J. Political Sci. 44 (2) (2000)
375–404, https://doi.org/10.2307/2669318.
[46] Bradley P. Carlin, Siddhartha Chib, Bayesian model choice via Markov chain Monte Carlo methods, J. R. Stat. Soc. Ser. B Stat. Methodol. 57 (3) (1995)
473–484, https://doi.org/10.1111/j.2517-6161.1995.tb02042.x.
[47] George Staab, Laminar Composites, Butterworth-Heinemann, 2015, https://doi.org/10.1016/B978-0-12-802400-3.00003-9.
[48] Robert Seydel, Fu-Kuo Chang, Impact identification of stiffened composite panels: I. system development, Smart Mater. Struct. 10 (2) (2001) 354–369,
https://doi.org/10.1088/0964-1726/10/2/323.
[49] R.S. Langley, Wave motion and energy flow in cylindrical shells, J. Sound Vib. 169 (1) (1994) 29–42, https://doi.org/10.1006/jsvi.1994.1004.
[50] Xin Chen, Yu Jade Morton, Wenxian Yu, Trieu-Kien Truong, GNSS spoofer localization with counterfeit clock bias observables on a mobile platform, IEEE
Sens. J. 24 (14) (2024) 21916–21930, https://doi.org/10.1109/JSEN.2023.3310976.
[51] Jing Wang, Yi Liu, Haiyan Wu, YangQuan Chen, Jairo Viola, Analysis and prediction of supermarket energy consumption time series with significant
fractional order characteristics, Authorea Prepr. (2024) https://doi.org/10.22541/au.170667004.49964987/v1.
[52] Menglong Liu, Wenyan Zhang, Xiao Chen, Lun Li, Kai Wang, Hong Wang, Fangsen Cui, Zhongqing Su, Modelling guided waves in acoustoelastic and
complex waveguides: From SAFE theory to an open-source tool, Ultrasonics 136 (2024) 107144, https://doi.org/10.1016/j.ultras.2023.107144.
[53] L. Maio, V. Memmolo, F. Ricci, N.D. Boffa, E. Monaco, R. Pecora, Ultrasonic wave propagation in composite laminates by numerical simulation, Compos.
Struct. 121 (2015) 64–74, https://doi.org/10.1016/j.compstruct.2014.10.014.
[54] L. Maio, V. Memmolo, F. Ricci, N.D. Boffa, E. Monaco, Investigation on fundamental modes of guided waves propagating in symmetric and nonsymmetric
composite laminates, Proc. Inst. Mech. Eng. C 231 (16) (2017) 2988–3000, https://doi.org/10.1177/0954406217698721.
[55] Fabrizio Ricci, E. Monaco, N.D. Boffa, L. Maio, V. Memmolo, Guided waves for structural health monitoring in composites: A review and implementation
strategies, Impact induced dynamic response and failure behavior of aircraft structures, Prog. Aerosp. Sci. 129 (2022) 100790, https://doi.org/10.1016/j.
paerosci.2021.100790.
[56] K.N. Shivakumar, W. Elber, W. Illg, Prediction of impact force and duration due to low-velocity impact on circular composite laminates, J. Appl. Mech.
52 (3) (1985) 674–680, https://doi.org/10.1115/1.3169120.
[57] Dong Xiao, Zahra Sharif Khodaei, M.H. Ferri Aliabadi, Impact identification based on surrogate-assisted efficient global optimisation, Fracture, Damage
and Structural Health Monitoring, Proc. Struct. Integr. 52 (2024) 667–678, http://spiral.imperial.ac.uk/handle/10044/1/113099.
[58] Stephen D. Casey, David F. Walnut, Systems of convolution equations, deconvolution, Shannon sampling, and the wavelet and gabor transforms, SIAM
Rev. 36 (4) (1994) 537–577, https://doi.org/10.1137/1036140.
[59] Robert S. Strichartz, A Guide To Distribution Theory And Fourier Transforms, World Scientific Publishing Company, 2003, https://doi.org/10.1142/5314.
[60] Loujaine Mehrez, Alireza Doostan, David Moens, Dirk Vandepitte, Stochastic identification of composite material properties from limited experimental
databases, part II: Uncertainty modelling, Mech. Syst. Signal Process. 27 (2012) 484–498, https://doi.org/10.1016/j.ymssp.2011.09.001.
[61] H.B. Motra, J. Hildebrand, F. Wuttke, The Monte Carlo method for evaluating measurement uncertainty: Applicationfor determining the properties of
materials, Probab. Eng. Mech. 45 (2016) 220–228, https://doi.org/10.1016/j.probengmech.2016.04.005.
[62] Andrew Gelman, John B. Carlin, Hal S. Stern, Donald B. Rubin, Bayesian Data Analysis, Chapman and Hall/CRC, New York, 1995, https://doi.org/10.
1201/9780429258411.
[63] George E.P. Box, George C. Tiao, Bayesian Inference in Statistical Analysis, John Wiley & Sons, 2011, https://doi.org/10.1002/9781118033197.
[64] V. Ajith, S. Gopalakrishnan, Spectral element approach to wave propagation in uncertain composite beam structures, J. Vib. Acoust. 133 (051006) (2011)
https://doi.org/10.1115/1.4003945.
[65] Faker Bouchoucha, Mohamed Najib Ichchou, Mohamed Haddar, Guided wave propagation in uncertain elastic media, Ultrasonics 53 (2) (2013) 303–312,
https://doi.org/10.1016/j.ultras.2012.07.001.
[66] Scott R. Eliason, Maximum Likelihood Estimation: Logic and Practice, SAGE, 1993, https://us.sagepub.com/en-us/nam/book/maximum-likelihood-
estimation.
[67] Andrew Gelman, Bayes, jeffreys, prior distributions and the philosophy of statistics, Statist. Sci. 24 (2) (2009) 176–178, https://www.jstor.org/stable/
25681293.
[68] Heikki Haario, Marko Laine, Antonietta Mira, Eero Saksman, DRAM: Efficient adaptive MCMC, Stat. Comput. 16 (4) (2006) 339–354, https://doi.org/10.
1007/s11222-006-9438-0.
[69] Pedro L. Ramos, Alex L. Mota, Paulo H. Ferreira, Eduardo Ramos, Vera L.D. Tomazella, Francisco Louzada, Bayesian analysis of the inverse generalized
gamma distribution using objective priors, J. Stat. Comput. Simul. 91 (4) (2021) 786–816, https://doi.org/10.1080/00949655.2020.1830991.
[70] John H. Holland, Genetic algorithms, Sci. Am. 267 (1) (1992) 66–73, https://www.jstor.org/stable/24939139.

34
D. Xiao et al. Mechanical Systems and Signal Processing 226 (2025) 112320

[71] Surya T. Tokdar, Robert E. Kass, Importance sampling: a review, WIREs Comput. Stat. 2 (1) (2010) 54–60, https://doi.org/10.1002/wics.56.
[72] Tjalling J. Ypma, Historical development of the Newton–raphson method, SIAM Rev. 37 (4) (1995) 531–551, https://doi.org/10.1137/1037125.
[73] Thomas F. Coleman, Yuying Li, An interior trust region approach for nonlinear minimization subject to bounds, SIAM J. Optim. 6 (2) (1996) 418–445,
https://doi.org/10.1137/0806023.
[74] Hailing Fu, Zahra Sharif Khodaei, M.H. Ferri Aliabadi, An event-triggered energy-efficient wireless structural health monitoring system for impact detection
in composite airframes, IEEE Internet Things J. 6 (1) (2019) 1183–1192, https://doi.org/10.1109/JIOT.2018.2867722.
[75] Jean-Michel Marin, Pierre Pudlo, Christian P. Robert, Robin J. Ryder, Approximate Bayesian computational methods, Stat. Comput. 22 (6) (2012)
1167–1180, https://doi.org/10.1007/s11222-011-9288-2.
[76] Nan Yue, Zahra Sharif Khodaei, M.H. Ferri Aliabadi, Uncertainty Quantification Of Guided Wave Structural Health Monitoring For Aeronautical Composite
Structures, World Scientific, 2024, https://doi.org/10.1142/9781800614703_0004.
[77] Dong Xiao, Zahra Sharif-Khodaei, M.H. Aliabadi, Time of arrival extraction for impact localisation on composite structures, e-J. Nondestruct. Test. 29 (07)
(2024) https://spiral.imperial.ac.uk/handle/10044/1/113191.
[78] Lukas Grasböck, Alexander Humer, Sven Herold, Investigation of different time of arrival estimation methods for impact induced guided waves, in:
Symposium for Smart Structures and Systems 2022, 2022, https://publica.fraunhofer.de/handle/publica/439329.
[79] Dong Xiao, Zahra Sharif-Khodaei, M.H. Aliabadi, Impact force identification for composite structures using adaptive wavelet-regularised deconvolution,
Mech. Syst. Signal Process. 220 (2024) 111608, http://spiral.imperial.ac.uk/handle/10044/1/113092.
[80] Sungmin Lee, Nickolas Vlahopoulos, Anthony M. Waas, Analysis of wave propagation in a thin composite cylinder with periodic axial and ring stiffeners
using periodic structure theory, J. Sound Vib. 329 (16) (2010) 3304–3318, https://doi.org/10.1016/j.jsv.2010.02.023.
[81] Kostas P. Soldatos, On the theories used for the wave propagation in laminated composite thin elastic shells, Z. Angew. Math. Phys. ZAMP 36 (1) (1985)
120–133, https://doi.org/10.1007/BF00949037.
[82] A.A. Khdeir, J.N. Reddy, D. Frederick, On the transient response of cross-ply laminated circular cylindrical shells, Int. J. Impact Eng. 9 (4) (1990) 475–484,
https://doi.org/10.1016/0734-743X(90)90036-U.
[83] J.N. Reddy, Exact solutions of moderately thick laminated shells, J. Eng. Mech. 110 (5) (1984) 794–809, https://doi.org/10.1061/(ASCE)0733-9399(1984)
110:5(794).
[84] Aldyandra Hami Seno, M.H. Ferri Aliabadi, Impact localisation in composite plates of different stiffness impactors under simulated environmental and
operational conditions, Sensors 19 (17) (2019) 3659, https://doi.org/10.3390/s19173659.
[85] Dong Xiao, Zahra Sharif Khodaei, M.H. Aliabadi, A generalisable data-driven approach to identifying impacts on composite structures, 2024, https:
//www.researchgate.net/publication/385587958_A_generalisable_data-driven_approach_to_identifying_impacts_on_composite_structures.
[86] Anthony J. Croxford, Jochen Moll, Paul D. Wilcox, Jennifer E. Michaels, Efficient temperature compensation strategies for guided wave structural health
monitoring, Ultrasonics 50 (4) (2010) 517–528, https://doi.org/10.1016/j.ultras.2009.11.002.
[87] Surajit Roy, Kuldeep Lonkar, Vishnuvardhan Janapati, Fu-Kuo Chang, A novel physics-based temperature compensation model for structural health
monitoring using ultrasonic guided waves, Struct. Health Monit. 13 (3) (2014) 321–342, https://doi.org/10.1177/1475921714522846.
[88] Jianye Ching, Yi-Chu Chen, Transitional Markov chain Monte Carlo method for Bayesian model updating, model class selection, and model averaging, J.
Eng. Mech. 133 (7) (2007) 816–832, https://doi.org/10.1061/(ASCE)0733-9399(2007)133:7(816).
[89] Ka-Veng Yuen, Recent developments of Bayesian model class selection and applications in civil engineering, Probabilistic Methods for Modeling,
Simulation and Optimization of Engineering Structures under Uncertainty in honor of Jim Beck’s 60th Birthday, Struct. Saf. 32 (5) (2010) 338–346,
https://doi.org/10.1016/j.strusafe.2010.03.011.
[90] Xunxue Cui, Kegen Yu, Songsheng Lu, Direction finding for transient acoustic source based on biased TDOA measurement, IEEE Trans. Instrum. Meas. 65
(11) (2016) 2442–2453, https://doi.org/10.1109/TIM.2016.2583224.

35