Renewable and Sustainable Energy Reviews 64 (2016) 91–105

Contents lists available at ScienceDirect

Renewable and Sustainable Energy Reviews

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

Structural health monitoring of offshore wind turbines: A review

through the Statistical Pattern Recognition Paradigm
Maria Martinez-Luengo, Athanasios Kolios n, Lin Wang
Cranfield Energy, Cranfield University, Bedford MK43 0AL, UK

art ic l e i nf o a b s t r a c t

Article history: Offshore Wind has become the most profitable renewable energy source due to the remarkable devel-
Received 6 February 2016 opment it has experienced in Europe over the last decade. In this paper, a review of Structural Health
Received in revised form Monitoring Systems (SHMS) for offshore wind turbines (OWT) has been carried out considering the topic
3 May 2016
as a Statistical Pattern Recognition problem. Therefore, each one of the stages of this paradigm has been
Accepted 23 May 2016
Available online 9 June 2016
reviewed focusing on OWT application. These stages are: Operational Evaluation; Data Acquisition,
Normalization and Cleansing; Feature Extraction and Information Condensation; and Statistical Model
Keywords: Development. It is expected that optimizing each stage, SHMS can contribute to the development of
Offshore wind turbines efficient Condition-Based Maintenance Strategies. Optimizing this strategy will help reduce labor costs of
Structural health monitoring OWTs' inspection, avoid unnecessary maintenance, identify design weaknesses before failure, improve
Statistical Pattern Recognition Paradigm
the availability of power production while preventing wind turbines' overloading, therefore, maximizing
Sensors
the investments' return. In the forthcoming years, a growing interest in SHM technologies for OWT is
Statistical model development
expected, enhancing the potential of offshore wind farm deployments further offshore. Increasing effi-
ciency in operational management will contribute towards achieving UK's 2020 and 2050 targets,
through ultimately reducing the Levelised Cost of Energy (LCOE).
& 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).

Contents

1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
2. Structural health monitoring systems for offshore wind turbines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
2.1. History and evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
2.2. SHM technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.2.1. Acoustic emission monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.2.2. Thermal imaging method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.2.3. Ultrasonic methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.2.4. Fatigue and modal properties monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
2.2.5. Strain monitoring. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3. Operational evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.1. Offshore wind turbines damage definition and detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.2. Variation in environmental and operational conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4. Data acquisition, normalization and cleansing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.1. Sensors types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96
4.2. Data collection and storage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
4.3. Data normalization and cleansing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

Abbreviations: AE, Acoustic Emission; OM, Operational Management; CB, Carbon Fiber; OMA, Operational Modal Analysis; CM, Condition Monitoring; OWF, Offshore Wind
Farm; EOC, Environmental and Operational Conditions; OWT, Offshore Wind Turbine; EU, European Union; O&M, Operations and Maintenance; FBG, Fiber Bragg Grating;
RSA, Response Surface Analysis; FEA, Finite Element Analysis; SHMS, Structural Health Monitoring Systems; FMECA, Failure Mode, Effects and Criticality Analysis; SVM,
Support Vector Machines; LCOE, Levelised Cost of Energy; WF, Wind Farm; MEMS, micro-electromechanical system; WSN, Wireless Sensor Network; NN, Neural Networks;
WT, Wind Turbine
n
Correspondence to: Tel.: þ 44 1234 754631.
E-mail address: a.kolios@cranfield.ac.uk (A. Kolios).

http://dx.doi.org/10.1016/j.rser.2016.05.085
1364-0321/& 2016 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
92 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105

5. Feature extraction and information condensation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98

6. Statistical model development . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
6.1. Supervised learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1.1. Response surface analysis (RSA) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1.2. Fisher's discriminant . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1.3. Neural networks (NN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1.4. Genetic algorithms. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.1.5. Support vector machine (SVM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.2. Unsupervised learning. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.2.1. Control chart analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
6.2.2. Outlier detection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.2.3. Neural networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
7. Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
8. Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

1. Introduction machines, SHM for OWT remains a research topic which is slowly
getting into the field deployment stage. This is due to the early
Over the past 15 years, Wind Energy has experienced a stage of the technology's deployment, the additional challenge
remarkable growth in the European Union (EU). While in 2000 that offshore environments pose to these technologies, and asso-
wind energy contributed 2.4% of the EU's electricity demand, by ciated costs to operators for hardware installation and data
2015 this percentage raised to 11.4%, or in absolute numbers, processing.
12.9 GW of installed capacity became 141.6 GW. This rapid Farrar and Sohn [17] were the first to introduce the Statistical
development is not only due to the targets set by the EU in 2006 Pattern Recognition Paradigm in the SHM field. This methodology
for all Member States [1], but also due to the scalability of wind follows four stages:
energy with units of larger capacity been deployed in larger
farms, further offshore [2]. According to Renewable UK, Offshore 1) Operational evaluation: This stage tries to set the boundaries of
Wind (OW) has officially become the most profitable renewable the problem by replying to four questions concerning the
energy source since, it can produce more renewable energy than implementation of the Damage Identification Facility. Questions
all of the other sources combined [3]. In Europe, including sites are related to: the motivation and economic justification for
under construction, there are 84 Offshore Wind Farms (OWF) in implementing the SHMS, the different Systems’ damage defi-
11 countries as of the end of 2015. Furthermore, 3,230 turbines nitions, the Environmental and Operational Conditions (EOC) in
are now installed and operational, reaching a cumulative instal- which the SHMS are used, and the data acquisition limitations
led capacity of 11,027 MW. In 2015 only, a grid-connected capa- in the operational environment.
city of 3,019 MW, was added, accounting for almost double of the 2) Data acquisition, normalization and cleansing: Data Acquisition
capacity added in 2014 [4]. Moreover, due to the increased refers to the selection of the excitation methods, type, quantity
deployment of 4–6 MW turbines in 2015, the average Offshore and location of sensors, and the Data Acquisition/Storage/
Wind Turbine (OWT) size became 4.2 MW, constituting a 13% Transmittal Hardware [18]. Data Normalization is another
increase over 2014. crucial aspect for the Damage Identification Process, as there
Considering wind energy as a mature technology, allows are numerous conditions in which measurements can be taken
developers and operators to gain confidence to include this [19]. Therefore, this Normalization constitutes the procedure of
energy technology within their mainstream portfolios. Increasing separating variations in sensor readings produced by damage,
availability of farms and reliability of units, decreasing from those produced by the variation in EOC. Data Cleansing is
unscheduled maintenance and eliminating unexpected cata- the procedure of selecting data which is passing on to or
strophic failures, are the targets that attract focus towards rejecting from the Feature Selection procedure [11]. Two exam-
deploying the next generation of wind farms. Structural Health ples of Data Cleansing processes are filtering and resampling,
Monitoring Systems (SHMS) can contribute significantly towards which constitute Signal Processing Techniques [20].
enhancing OWT's profitability, reliability and sustainability 3) Feature extraction and information condensation: This is the
through more systematic operational management approaches. aspect of the SHMS that attracts most attention, as these fea-
SHM represents the procedure of implementing a damage tures allow the distinction between damaged and non-damaged
detection strategy for engineering infrastructures related to structures [21,22]. Data Condensation is essential when analo-
aerospace, civil and mechanical engineering [5], being damage gue feature sets acquired along the structure's lifetime are
referring to the variations in material and/or geometric proper- envisioned. Due to the extraction of data from a structure
ties of these systems [6]. Some of the most known structural during long periods of time, robust data reduction techniques
damage roots are: moisture absorption, fatigue, wind gusts [7], have to be developed to preserve feature sensitivity to the
thermal stress, corrosion [8], fire and lightning strikes [9]. changes of interest.
Usually, there are two critical aspects that influence SHMS 4) Statistical model development: It is related with the imple-
development: the sensing technology (and the associated signal mentation of the algorithms that work with the extracted fea-
analysis), and the interpretation algorithm [10]. tures and calculate the extent of the damage to the structure.
Damage identification is performed through five similar dis- These algorithms can be divided into the two categories that are
ciplines [11]: SHM, Condition Monitoring (CM) [12], Non- shown in Fig. 1 [23–25]. All of these algorithms assess statistical
Destructive Evaluation [13], Statistical Process Control [14], and distributions of the measured or derived features, to enhance
Damage Prognosis [15,16]. Apart from the CM of rotating the damage identification process.
 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105 93

Fig. 1. Algorithms classification for Statistical Model Development.

This paper constitutes a comprehensive review of SHMS of OWT conjunction with the development of the space shuttle programs
following the process of the Statistical Pattern Recognition Paradigm. [11]. That effort carried out on other applications that are being
The paper has been divided in eight sections. In Section 2, a com- investigated for the National Aeronautics and Space Administra-
prehensive review of the SHMS for OWTs is carried out, presenting tion's Space Station [5]. Some of the most widely used technolo-
the history and evolution of SHMS and the different technologies that gies in this field are: fastener monitoring [30], blade tip clearance
can be employed to OW. Each one of the framework's stages [31], and fatigue monitoring [32]. The Civil Engineering commu-
abovementioned has been reviewed in greater detail focusing on nity has researched on vibration-based Damage Identification of
OWTs applications (Sections 3–6). Discussions of the capabilities and bridges and buildings since the 80 s [33,34]. This research is cur-
limitations of SHMS, the most used methods in the OW Industry and rently being applied to offshore structures due to the similarities
current technology gaps are presented in Section 7, followed by across industries.
conclusions in Section 8. During the 70 s and 80 s, the Oil and Gas industry carried out
extensive research to develop vibration-based Damage Identifica-
tion Techniques for offshore platforms [35]. Related to this tech-
2. Structural health monitoring systems for offshore wind nique, one of the research objectives was the detection of near-
turbines failing drilling equipment and the prevention of expensive oil
pumps from becoming inoperable [36]. Unfortunately, most efforts
2.1. History and evolution were not successful, as this problem is fundamentally different to
that of the rotating machines due to the impossibility of predicting
The identification of changes in the dynamic response of sys- where damage will occur and the structure's inaccessibility for
tems has been carried out qualitatively, since practice has intro- data acquisition purposes. Besides, numerous practical issues were
duced tools by employing acoustic techniques [26,27]. Lately the found apart from measurement acquisition difficulties, occasioned
emergence of SHM techniques has come together with the evo- by platform noise, instrumentation difficulties in hostile environ-
lution, miniaturization and cost reduction of digital computing ments, changing mass caused by marine growth, varying fluid
hardware [11]. Depending on the sector, this evolution took place storage levels, temporal variability of foundation conditions and
sooner or later. For example, CM Systems for rotating machines the inability of wave motion to excite higher vibration modes [11].
constituted one of the first developed systems, whilst SHMS for However, different applications could finally overcome these
the OW industry are currently an emerging research topic. issues and be implemented. For example, fatigue gauges are
Nowadays, CM of rotating machines constitutes the most commonly employed for fatigue monitoring by measuring the
prosperous application of SHM Technology in terms of profit- crack-growth proportional to the cumulative fatigue damage for
ability, reliability and level of development. The failure identifi- welded joints [37]. This approach was reviewed by [38] and [39].
cation process is based on pattern recognition related to dis- Another example is presented in [40], where different techniques
placement, velocity or acceleration time histories, commonly for corrosion monitoring are introduced and the application of
located at one point on the housing or shafts of the machinery flexible ultrasonic thin-film piezoelectric transducer arrays is
during standard-operating conditions and start-up or shut-down described. Lastly, [41] suggests a methodology to enhance the
transients [28]. Commercial software integrated with measure- reliability of SHM for flexible risers, which are widely used in
ment hardware is marketed to help the user systematically apply offshore oil exploration facilities and are essentially composite
this technology to the operating equipment [11]. These facts structures consisting of several metal armours and polymeric
mentioned above, supposed the transition of SHMS in this field layers.
from a research topic to a common industry practice. A good Most of the Wind Farms (WF) are either at the beginning or in the
example of this is the US Navy's Integrated Condition Assessment middle of their service life and, currently, the trend is to build these
System [29]. WFs much further offshore. Therefore, special consideration due to
The aerospace sector started studying the use of vibration- extreme weather conditions and complex dynamics, such as sensor
based Damage Identification during the late 70 s and early 80 s in tolerance and endurance, data acquisition and transmission, among
94 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105

others, have to be accounted for [42,43] and [44]. As this sector [5], Acoustic Emission (AE) is a very effective technique that
grows, business economics currently demands management of OPEX detects failure mechanisms up to the microscale. However, this
and CAPEX costs [45]. For example, considering a 750 kW turbine technique is less effective when it comes to damage character-
with an expected 20-year service life, the operations and main- ization and assessment in the case that an appropriate algorithm is
tenance (O&M) costs account for between 25% and 30% of the overall not available. Complex damage mechanisms in WT blades have
energy generation cost or 75–90% of the investment costs [46]. been better understood since AE monitoring was carried out in a
Some of the technologies employed in OW differ from those blade during loading, enhancing the ability to assess damage
employed by the Civil or the Oil and Gas Industries in aspects of during testing [55–57]. Also during a certification test, the
accessibility, severity of the environmental conditions and more damaged area due to cracking in the blade was located due to the
complex loading due to excessive operational loads. Due to these sound of the cracking mechanism [56]. Fatigue tests can also be
differences, further development and research in these technolo- monitored, as [58] presents; such as the sound produced due to
gies has been prioritized in the past years aiming to adapt them to stress released waves or energy dissipation using piezoelectric
OW applications. In order to make wind power competitive with sensors [59,60].
other sources of energy, some of the performance indicators, such AE signals are defined by their amplitude and energy [61]. As
as availability, reliability, efficiency and integrity of turbines, still [59] explains, AE events will occur around a particular point, at a
have to be improved [45]. In the following section, a review of the structure under certain loading. That particular point will be the
current SHM techniques used in the OW Industry is presented. one at which the structure will fail at some point, being the feature
extremely useful in locating the failure. Even though most of the
2.2. SHM technologies relevant literature is related to tests in WT blades, a few cases
explain how the technique has been applied to a WT blade during
SHMS of OWTs are becoming very much in demand now that operation [62], by using a broadband radio to send the AE data
machines are growing in size and OWF are being developed fur- from the rotating frame to the ground with no signal resolution
ther from the coasts. In order to decrease the power generation loss. Even though the previously collected data had acceptable
costs and, therefore, the Levelized Cost of Energy, WT mass of levels of noise in low to moderate wind speeds, verification of the
components need to be optimized without compromising the fact that the noise does not increase with wind speeds and the
structure's integrity. This can be achieved through making the feasibility of those signals to be filtered, has to be assessed.
turbines more structurally flexible, which directly affects their
modal parameters, i.e. the resonance frequency [47]. Another 2.2.2. Thermal imaging method
important aspect is that OWTs' inspection and maintenance is This method aims to detect defects or anomalies in the material
considerably more expensive than onshore turbines’. Therefore, beneath the surface and it is based on the subsurface's tempera-
SHMS which are able to predict structural changes are becoming ture gradients. Thermal imaging can be applied to a WT blade by
crucial to diminish operation and maintenance (O&M) costs and to installing infrared cameras [63]. Irregularity of or damage to
assess the remaining lifetime of these structures. An example of a material is detected due to a change in the thermal diffusivity.
good application of SHMS to an onshore WT is presented in [48], Moreover, this technique can be divided in two categories
where a life-cycle management framework for online monitoring depending on the thermal excitation method used: active or
and performance assessment is applied to WT. passive. The passive thermal imaging method aims to investigate
SHMS have become a useful method to enhance Operational materials at different temperatures, other than the ambient, and
Management (OM) and optimize maintenance activities of modern therefore, it is not normally used in SHMS of OWTs; the active
infrastructure [35], as the information gathered can be employed approach has an external stimulus source (i.e. optical flash lamps,
in the development of a tailored, condition-based maintenance or heat lamps).
program [49]. This program aims to reduce the necessary down- A particular type of active thermal imaging method is called
time due to components inspection, prevent unnecessary repla- the thermoelastic stress method and it is based on the thermo-
cements and failures, and increase availability. Furthermore, due elastic effect, which consists of the change in temperature of an
to the capacity of monitoring the structure's integrity, design elastic solid produced by a change of stress [64]. As explained in
improvements can be implemented such as selection of lighter [65], in the damaged or abnormal region, different heat conduc-
blades that will enhance performance with less conservative tion, higher acoustical damping, and stress concentration take
margins of safety [50] and which will adapt quicker to the wind's place. This technique has been proven to be useful in WT blades
variability, capturing more energy [51]. fatigue tests [61,66], as stress concentrations during the test can be
General reviews of SHM can be found in [52] and [26,53] where observed before damage in the surface can be appreciated. A
assessment of the different methodologies was carried out. SHM promising variation of this methodology involves applying high
techniques for WT were reviewed by [7], however, the majority of power ultrasounds [67], or oscillating stresses with a mechanical
that review was based on bridges and civil infrastructures. A wide shaker, to the surface that is being tested [65]. This technique is
overview of how the EOC affects SHM techniques and the nor- called vibro-thermographic and is able to locate and assess crack
malization of the data that needs to be carried out for compen- dimensions, as [67] states. Furthermore, it can be used for asses-
sating these variations is given in [24]. A discussion between SHM sing voids and stress concentration in composites. Nevertheless,
and CM costs can be found in [54]. this method has the potential to become a promising SHM tech-
Within this section, the different SHM techniques and espe- nique for WTs, and more research needs to be carried out in order
cially those suitable for OWT blades, tower and foundation, are to reduce the sensitivity to temperature variations [61].
explained.
2.2.3. Ultrasonic methods
2.2.1. Acoustic emission monitoring Ultrasound is a method commonly used for assessing the inner
Failure mechanisms such as cracking, excessive deformation, structures of solid objects [68]. It has also turned out to be very
debonding, delamination, impacts, crushing, among others, all useful with composite structures. The basic principle of this
provoke transient changes in stored elastic energy in particular technique is that ultrasonic waves, emitted by a transmitter, pass
points of a structure. This energy release can be effectively used to through the tested material and are reflected and/or mode con-
monitor WTs and, particularly, their blades. As Ciang mentioned in verted by a flaw or anomaly. This modified signal is picked up by a
 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105 95

receiver once it has passed through the material (if not reflected). been proven to be effective in different types of structures,
In the simplest arrangement, transmitter and receiver are placed including composite structures [84,85].
on opposite surfaces of the material [69]. The aim of this technique
is to reveal planar cracks that take place perpendicularly to the 2.2.5. Strain monitoring
sound wave propagation direction [70]. An advantage of this Strain monitoring is the technique that detects microscopic
method is that it can detect cracks of just a few millimeters. length variations in a component at pre-established locations,
which does not necessarily mean damage detection. However,
2.2.4. Fatigue and modal properties monitoring these length variations are known to be directly related to stresses
Fatigue and Modal Properties Monitoring are among the most and loads applied to the material [86]. Due to the fact that total
important SHM techniques for OWT structures, as the con- deformations of large components, i.e. WT blades, are large
sequences of structural damage may be catastrophic. These because they are the sum of all the local deformations, they give
methods are very simple to implement on structures of any size as no indication of local damage. For that reason, strain sensors have
they are based in another CM technique, which is the most mature to be positioned at points of particular interest, where large
and successful methodology for rotating machinery monitoring; deformations are expected. This limits their applicability to overall
the vibration-based inspection method [35,71]. component damage sensing applications [87]. Strain monitoring
Modal Properties Monitoring is based on the principle that has been proven to be useful in continuous operational WT
modal parameters, such as resonance frequency, damping coeffi- monitoring as it was successfully employed in a 4.5 MW turbine
cient and modal curvatures, among others, experience certain [88]. However, in order to predict WT failures in blades, tower and
variations due to a change in different physical properties (i.e. foundation, prior knowledge of their component's stress field is
reduction in mass or stiffness) [72,73]. Due to these changes, the required so that sensors can be mounted in critical areas.
structure is considered to be damaged; that damage being iden- Another SHM technology for strain monitoring is the Strain
tifiable by comparison between the structure's modal parameters Memory Alloys Method, which relies on an irreversible crystal-
before and after an event. In other words, due to the fact that lographic transformation for their smart properties. The transfor-
modal properties changes are considered as damage indicators, mation consists of the change, due to the strain, from one crystal
this SHM technique is categorised as a pattern recognition pro- state to another. The parent austenitic crystal structure is para-
blem [53]. magnetic, while the product martensitic phase is ferromagnetic.
In order to be able to analyze the structure's dynamic response Any SHMS related to this group of smart materials is considered as
by studying its mode shapes, several accelerometers must be a passive system, as both full-time power supply and data storage
installed. Other analyses that can be carried out are curvature facilities are not necessary. Instead, power is only needed during
mode shapes and wavelet maps. These analyses are particularly the sensor's interrogation, being the actual reading stored within
relevant when they are carried out in service conditions [74]. the sensor element itself [89].
However, performing these analyses accurately to a full scale OWT
during operation is extremely difficult due to the high number of
uncertainties which the offshore environment presents [47] and, 3. Operational evaluation
therefore, special effort has been given to solve this issue in the
past years [75]. One reason that makes this analysis difficult is the 3.1. Offshore wind turbines damage definition and detection
fact that wind and wave loading applied to the structure cannot be
measured accurately in a continuous manner. This introduces the Damage definition constitutes a very important stage of the
difficulty of having to employ Operational Modal Analysis (OMA) Statistical Pattern Recognition Paradigm as the boundaries of the
for calculating the modal parameters based on the assumption problem are defined within it. Moreover, damage features have
that the structure is subjected to unknown random loads [76–79]. high variability among fields and structures. Therefore, identifi-
OMA methods are based on the principle that in the analyzed time cation of damage causes, consequences and features must be
interval, the system is linear and does not vary with time. One carried out at the beginning of any SHMS design phase. Several
issue pointed out in [80], is that most of the research regarding risk analysis techniques can be employed, Failure Mode, Effects
data variability due to changes in EOC was carried out in labora- and Criticality Analysis (FMECA) being considered one of the most
tories, where basic signal processing techniques were enough to relevant for this particular case [90] [91]. Different reviews of OWT
solve the damage detection problem [81]. Unfortunately, these failure modes have been made in [5,46,92].
techniques are not considered enough to be employed in an OWT One of the main concerns regarding OWT damage detection is to
during operation. identify the best way to detect structural damage. Usually the
Scour effect on the natural frequency of OWT was studied in change in modal properties is used for this purpose [35]. However
[82], where it was proved that while scour increases, the natural determining the best methodology constitutes a much broader field
frequencies of the support structure, and therefore the WT, than what can be expected at first sight, as numerous different
decreases. This phenomenon represents a threat for the turbine as choices are available. Proof of this is the review of damage detection
the natural frequency gets closer to the rotor's frequency of rota- methods through the change in modal properties presented in [26].
tion [47]. Therefore, continuous monitoring of WTs’ dynamics The relevant method are: Natural Frequency Based Methods [93],
variations due to scour is recommended as it is expected to be a Mode Shape Based Methods [94–98], Mode Shape Curvature Based
useful tool for developing maintenance plans regarding scour Methods [99–103], Strain Mode Shape Based Methods [71,104–
protection [82]. 108], Dynamically Measured Flexibility Based Methods [109–111],
Another type of modal monitoring called resistance-based and Neural Network Based Methods [112–117].
damage detection method has been found to be revolutionary Damage definition in OWT blades is closely related to the one
due to the fact that it has the capability of detecting local damage. that anisotropic reinforced laminated composites have. Delamina-
It uses piezoelectric materials which, by monitoring their electrical tion is one of the most common failure modes in composites [118],
impedance, can detect the presence of structural damage. which is responsible for causing stiffness reduction, variation in
According to [83], only local response of the structure will be resonant frequency, and decrease in buckling capacity. Such defects
transmitted to the sensor in case the excitation frequency is big might be caused by poor process control during manufacturing,
enough. Damage detection using this monitoring technique has impact loading, or other hazardous service environments [119].
96 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105

There are many other failure mechanisms for carbon–fiber com- 4. Data acquisition, normalization and cleansing
posites, such as fiber breakage, matrix cracking, fiber splitting, and
delamination, as listed in [120]. 4.1. Sensors types
The most probable failure mechanisms that an OWT's tower
and foundation can experience are corrosion and fatigue due to As previously mentioned, SHMS for OWT can be used to detect
combined wind and wave loading [121,122]. Failure of these damage in blades, tower and support structure. This section aims
structures due to the accelerated fatigue produced by the increase to introduce the different types of sensors and technologies which
are used and in which subsystem. From the top to the base of the
of stresses, when natural frequencies are found to be similar to the
OWT, blades constitute a difficult element to integrate SHMS due
rotor's frequency, can lead to catastrophic consequences which
to the high variety of failure modes that can develop, the high
must be avoided. This phenomenon, known as resonance, has to
strains they experience, the fact that they are rotating compo-
be dealt early in the design stage of these structures taking into
nents, and the high variability in their operating conditions [80].
account all operation stages through a structure's service life [123]. Different sensors can be used in blades, as confirmed by different
In the particular case of pile-foundations, scouring and reduction reviews [131]. Two approaches are followed: active and passive
in the foundation's integrity over time can be problematic. Scour sensing technologies, whereby active sensing, but not passive,
reduces the fundamental structural resonances of the support needs an external excitation [80].
structure. Therefore, it can be considered a damage indicator as it Tower and Foundation constitute two key elements of OWT as
can be correlated to a change in the natural frequency of the tower they are not replaceable unless a significant cost is assumed. These
and an increase in the fatigue damage [79,82,124]. are components that, once the turbine is installed, should sustain
associated loads and their partial failure would carry catastrophic
consequences. Therefore, early in the design stage, the intended
3.2. Variation in environmental and operational conditions turbine's service life and the possibility of extending it or repow-
ering it with a new nacelle, must be taken into account [132].
According to [125], the system's integrity state is a stochastic Furthermore, due to the difficulty and sensitivity of fatigue ana-
function of the initial system's integrity (quality), influenced by lysis, SHMS should be installed in order to be able to verify the
the acting loads (e.g. extreme loads, cyclic loads, environmental accuracy of the design calculations and implement an optimal
conditions). Even though a structure is considered to be damaged Operational Management Strategy. These SHMS will mainly con-
when at least one of its physical properties (mass, stiffness, etc.) sist of fatigue and modal properties monitoring (such as resonance
varies, changes in EOC might induce variations in these properties frequency or modal curvatures), corrosion and scour monitoring. It
without necessarily meaning that damage exists [80]. In fact, in should be noted that regarding SHMS for operating WTs, not much
the majority of the situations, it is extremely difficult to assess progress has been made in developing robust applications, espe-
whether or not EOC cause sensitive variations in the SHMS mea- cially for OWT blades [133].
surements [126]. For this reason, this topic has been recognized as Some of the methods that were introduced in Section 2
[125,133–136] include vibration monitoring-based methods
an important issue in SHMS and has been identified as a key
(accelerometers, piezo or micro-electromechanical systems
concern across the research community [17].
(MEMS)), strain (strain gauge or fiber optic cables), ultrasonic
SHMS for OWT are particularly relevant in the design phase,
waves which are widely applied in composite structures (piezo-
during shipping, installation and operation. The application of
electric transducer), acoustic emissions (usually barrel sensors),
SHMS in harsh environments is a particularly challenging task. The
impedance techniques, laser vibrometry, impedance tomography,
reasons are not only because these systems need to be prepared to thermography (infrared cameras), laser ultrasound, nanosensors,
withstand the severity of the environment for a long period, and and buckling health monitoring. The necessary sensors for
the ease of installation, ruggedness and reliability of equipment is implementation of these techniques are described below.
essential in providing key information about the tower's structural Structural dynamic responses are usually monitored by
integrity, but also because OWFs are being developed further than embedded strain gauges, piezoceramics or accelerometers [137].
ever before from coasts, which is making their Operational Man- Accelerometers are relatively simple devices whereby the oper-
agement critical [127]. ating principle is the comparison of the acceleration they experi-
The consequences that the variations in the EOC have on the ence with the acceleration due to gravity. They are commonly
dynamic behavior of structures were assessed in different studies provided as MEMS which are very small devices with computing
[119,128]. For example, a statistical methodology that propagates capability. These devices are commonly used for modal para-
variability in measured Frequency Response Function data and meters and vibration monitoring of blades, tower and foundation
calculates the level of uncertainty of the modal properties is of the WT. There are various types of accelerometers available,
explained in [129]. A good example of the effect of the variation in such as piezoelectric, optical, laser, capacitive, and servo. The
the EOC is presented in [80], where the turbulence suffered by the selection of an accelerometer for a specific application depends on
a number of factors, such as amplitude and frequency range of the
rotor affected the operational WT Control System. Other important
response, sensitivity, resolution, etc. [37]. The SHM of civil engi-
factors that strongly influence SHMS signals are extreme events,
neering structures using plastic optical-fiber based accelerometers
such as earthquakes. SHM technologies are known to have an
for estimating the natural frequencies by measuring the dynamic
accurate characterization of input excitations. Seismic excitations
response was carried out in [138]. Other types of sensors that can
are transient in nature, constituting an issue that limits the per- be used to analyze modal parameters are piezoelectric patches,
formance of most SHMS due to the fact that these technologies are which were used in [51] at critical locations with the aim of
based on the stationary stochastic-excitation assumption [27]. comparing their natural frequency. Velocimeters, on the other
Further information regarding this issue can be found in [130]. To hand, operate based on a principle similar to interferometry. In
conclude, any methodology employed has to be able to distinguish SHM these devices are primarily used to measure displacement by
between EOC that affect signals and damage features in order to integrating acceleration or velocity measurements of the struc-
allow the SHMS to detect only damages in the structure. tural members they are attached to [139].
 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105 97

Two popular sensor groups exist for the purpose of strain increases, and drastically once the crack density in a composite
measurement: traditional electrical and relatively modern fiber laminate specimen increases [151].
optic [86]. Electrical strain gauges have become so widely applied 2) FBG is made by illuminating the core of an optical fiber with a
that they dominate the entire field except for special applications. spatially varying pattern of intense Ultraviolet laser lights that
They are, along with electrical resistances, the most popular types have sufficient energy to break the highly stable silicon–oxygen
of sensors [140], closely followed by Fiber Bragg Grating (FBG) bonds, which will raise, to some degree, the refractive index
sensors, which recently have experienced considerable improve- [152]. Although the main use of FBG consists of measuring
ments [141]. strains crack evolution [151], impact damage can be detected by
Several electrical sensor types exist including capacitance, distributing FBG over the structure [137,138,153,154].
inductance, semiconductor and resistance. Each is sensitive to a 3) Optical fuses transversally positioned in laminated composites
differing electrical property [140]. Resistance strain gauges record have been proven to be useful in damage detection [155]. For
the resistance variation of an electrically conductive wire relative to example, if short length optical fibers are embedded through
displacement. This resistance variation occurs due to a change in the thickness of a graphite/epoxy laminate during the manu-
the cross sectional area and length of the wire as the specimen is facturing process, the fibers act as optical fuses, which will
elongated. Electrical resistances are generally used for identifying break in areas of low energy impact damage [156].
cracks in composite materials and joints. The most suitable material
for monitoring using this method is carbon fiber (CB) polymer- 4.2. Data collection and storage
matrix composites as their electrical properties are affected by
structural damage. This material is commonly used due to its It is widely recognized that dynamic data acquisition is a
strong, super-elastic, and piezoresistive properties [142]. These complex, tedious and costly process [157]. The recent develop-
sensors can also be used for identifying failures in conductive bolted ment of wireless monitoring has brought a big advance in SHM
and Infrastructure Asset Management [34] as it integrates wireless
joints. A novel method for analyzing the structural health of alu-
communications and mobile computing with sensors. The result is
mina nanocomposites, by the change in electrical conductivities
a more economic sensor platform that has three aims: acquisition
after indentation, is proposed in [143]. The utility of the electrical
of structural response data, local interrogation of collected mea-
resistance method for locating barely visible impact damage in
surement data, and wireless transmission of that data or analysis
carbon fiber composite structures was explained in [144].
results to a Wireless Sensor Network (WSN), which comprises
Piezoelectric materials, when subjected to stress, produce an
other wireless sensing units [34]. As explained in [158], a WSN is
electric field and vice versa when subjected to an electric field.
composed of four mean stages: communication, data acquisition,
Furthermore, changes in the fundamental properties of the
processing, and fusion stages. Moreover, WSNs encompass many
structure, such as mass, stiffness and damping, directly make the
fields: wireless communication, network technology, integrated
mechanical impedance vary, this variation being a clear damage
circuits, sensor technology, MEMS, among many others. WSNs are
indicator [145]. Damage detection using changes in the electro-
composed of data acquisition systems which have numerous
mechanical impedance of piezoelectric wafer active sensors can
design parameters: a number of channels, a maximum sampling
easily be done by attaching them to the structure [146].
rate, and resolution, among others; a computational core, where
Even though piezoelectric materials are the most common
all the data acquired are stored and which possess processing
sensor type for stress monitoring, there are many other sensors capabilities; and the wireless communication channel.
that can also be applied to this aim, such as: thin film sensors, A real WSN application is presented in [157], where three WTs
piezoelectric composite materials, rolling sensors, and optic-based in operation, instrumented with WSN, proved their efficacy in
sensors [69]. However, an important drawback this technology operational conditions. While in the first turbine instrumented,
has, is temperature and ambient vibrations effects in the piezo- the aim was to prove the accuracy in the collection and trans-
electric sensors’ performance in composites, as explained in [27]. mission of vibrational data from the turbine's tower, in the second
Temperature effect in blades must be compensated in the results, turbine instrumented, several strain gauges were also included at
as they are made from this material. In fact, [147] explained how a its base. In both, wireless communication channels, performance
rise in temperature and vibrations can jeopardise the detection of was assessed and their data used for offline output-only towers
the delamination caused by impacts. Other common drawbacks modal analysis.
that strain gauges might experience are described in [86]: non- The acquired data from WTs contain key features for future
linearity, hysteresis and zero shift due to cold work [140]. developments in the Wind Energy Industry. For that reason, opera-
Cracks and displacements can also be monitored by fiber-optic tors are starting to appreciate the importance of investing in SHMS
sensors which usually are: spectrometric, interferometric or [159]. However, even though monitoring has many proven advan-
intensity-modulated. An optical fiber is a glass or plastic fiber tages, it is expensive and its costs are the cause why only a few
designed to guide light along its length. Moreover, FBGs were also operational turbines have extensive sensor instrumentation [157]. An
proved to be useful as a corrosion transducer and temperature assumption usually made, is that traditional cable based monitoring
sensor simply by adding a metal coating to one segment of the systems are cheaper and easier to install. Nevertheless, this tech-
fiber [148]; as a pH-sensitive corrosion detector [149] and good at nology is not only more costly to install, but also introduces diffi-
delamination identification [150]. Furthermore, fiber-optic sensors culties in the installation process due to the cables. On the contrary,
are employed in SHMS for OWT in various forms: wireless sensors are substantially cheaper and easier to install than
traditional cable-based systems [160]. In the case of turbine blades,
1) Plastic fiber-optics can be attached, for example, to the blade of wireless communication eliminates the necessity of moving data
a WT to measure loads it bears. This measurement is carried out through a slip ring interface, which is difficult and costly.
by the reduction of the light source's power that takes place in Wireless sensors are not, exactly, cable-based sensor replace-
the plastic fiber-optic depending on the strain to which it is ments; without wires, wireless sensors usually depend on intern-
subjected [151]. This concept is used to sense strains in a ally stored power for operation. Inefficient use of wireless sensors
structure. When loads increase, the measured optical power is will deplete this precious energy source rapidly, making frequent
reduced being damage detectable due to the fact that the battery replacement necessary. Among the three different types of
normalized optical power decreases linearly as the strain WSN topologies (Star, Cluster tree, and Mesh [158]) there are
98 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105

several important issues for WSN use in SHMS. These were sum- jeopardise the accuracy of the data set. For this reason, the whole
marized in [157]: compatibility issues between different types of set might be discarded from the feature selection process. Signal
sensors, their sampling frequencies, the problem of transmission processing techniques, such as filtering and resampling, can also
bandwidth and real-time ability variance, the selection of a wireless be thought of as data cleansing procedures [24,34,169,170].
transmission frequency, topology choice, data fusion method, and
the contrast between the energy consumption requirements of
different applications to that of each different device. 5. Feature extraction and information condensation
Even though WSN have been proven to be applicable to OWFs
[80], they have the disadvantage of the high amount of power Feature extraction constitutes the methodology that refers to
needed by the sensors, which had been tried to be diminished the identification of the damage sensitive physical characteristics.
with an increased interest in data telemetry with energy har- It is usually determined by the data obtained from the structure
vesting [161,162]. In order to provide enough power to the sensors and is application specific [171]. Many methods can be used for
without using batteries, piezoelectric, thermoelectric and photo- damage feature identification, the most basic one being compar-
voltaic energy harvesting techniques were assessed in [163], on a ison of SHMS output data with similar data obtained when the
cross section of a CX-100 WT blade. The aim was to determine the same structure has experienced a damaging event. This metho-
feasibility of powering individual nodes that would compose the dology is based on the fact that damaging events have already
sensor network. In another study [164], a 4-channel AE wireless occurred. Another process for feature identification is the numeric
node was powered by structural vibration and wind energy har- simulation of the damaged system's response to postulated inputs,
vesting modules. which is currently the most used technique in several industries,
e.g. the automotive industry. Another option for recognizing these
4.3. Data normalization and cleansing sensitive features would be testing the structure or a representa-
tive specimen in a laboratory, introducing the expected damage.
The ability to normalize the measured data with respect to Damage-accumulation testing, during which structural compo-
varying EOC is a key aspect of a SHMS in order to avoid false nents of the system under study are subjected to a realistic load-
positive indications of damage [19]. One example of the normal- ing, can be used also to identify appropriate features [119]. As
ization process is carried out to the measured inputs when modal Farrar and Worden explain in [11], this methodology might involve
parameters are being extracted. Two strategies can be employed induced-damage testing, fatigue testing, corrosion growth or
for normalizing these data: when the EOC are available and are not temperature cycling to accumulate certain types of damage. As
available. such, numerous articles in this theme issue are devoted to the
The most important aspect regarding accuracy of data nor- feature extraction portion of SHM [18,20–22].
malization comes with the damage sensitive features that must be Data Condensation constitutes an inherent part of the Feature
extracted from these data. Those damage sensitive features must Extraction procedure. The different types and quantity of sensors
not be lost or diluted by the normalization process. There are needed to make any SHMS work efficiently and accurately usually
different data normalization techniques. Some examples are: the produce huge amounts of data. Therefore, data condensation is,
subtraction of the mean value of a measured time history for direct most of the time, a necessary stage occurring before the analysis of
current off-sets removal from the signal, the division by the the extracted data through the statistical models. One possibility
standard deviation of the signal for normalizing varying ampli- of data condensation is to summarize all the damage sensitive
tudes in the signal, curve fitting of analytical forms of the fre- features in their adimensional form into feature vectors of small
quency response function to measured frequency response func- dimension. This constitutes an accurate way of estimating the
tions in experimental modal analysis, among others. If the struc- feature's statistical distribution [119]. Moreover, data condensation
ture is linear, this normalization procedure removes the influence is not only beneficial due to the savings in computational power,
of the input from the parameter estimation procedure. but also necessary in case of comparisons of many data sets over
Data normalization constitutes a very important part of the the lifetime of the structure. Even though the more data con-
damage identification process as it affects significantly Neural densation is achieved, the more computational power is saved; the
Network (NN) performance [165]. Even though not all sources of sensitivity of the chosen features to the structural changes under a
variability in the data acquisition mechanism can be eliminated, certain level of variability in the EOC has to be ensured by the
they need to be identified and minimized as much as possible employment of robust data reduction techniques (such as, Prin-
[166]. Therefore, appropriate measurements need to be carried out cipal Components Analysis [172], Discriminant Analysis [173],
in order that such sources of variability can be statistically quan- Regression Analysis [174], etc. [175]).
tified [19]. An example of data normalization in OWT is explained Another option for data condensation in AE is proposed in
in [167], where a non-linear regression model to perform data several studies [5,176]. This technique is based on the use of
normalization was used in real-life data obtained from the Structural Neural Systems, a highly distributed sensor concept that
monopile of an OWT. Further research on this topic will be carried mimics the signal processing in the biological neural system [50].
out in the future because, in order to achieve successful SHM goals, This methodology is employed in situations when a great level of
data normalization procedures able to discriminate whether accuracy in the damage evaluation is needed, as both the number
measurement variations are motivated by damage in the structure, of sensors and the amount of power needed for condensing and
or by changes in the EOCs [168]. processing the data increase considerably. Moreover, an
Data cleansing is the procedure of selectively choosing data to improvement in this technology is presented in [177] by the
pass on or to reject from the feature selection process or, in other connection in series or array pattern of multiple piezoceramic
words, is the procedure of selectively discarding data that might patches. This connection decreases the amount of channels
not represent the system's behavior [11]. Data cleansing is a dif- necessary for data collection of AEs or high strains.
ficult process due to the fact that it is commonly based on experts’
knowledge gained in previous data acquisition processes. An 6. Statistical model development
example of data cleansing could be when a sensor is discovered to
be loose and, therefore, based on the judgment of the experts; the Statistical Model Development is the Pattern Recognition sec-
measurements carried out by that sensor are not accurate and can tion that addresses the applicability of the algorithms that operate
 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105 99

on the extracted features, identifying and quantifying damage in a as the auto-associative NN employed in [185] or in [186] where NN
structure. There are two main types of algorithms: supervised and were trained with FEA data of the bridge.
unsupervised learning [23–25]. These categories of algorithms NNs were also used for structural damage detection in plate
correspond to SHMS that do contain and do not contain data from truss structures, where damage was assessed by evaluating dif-
the damaged structure, respectively. Supervised learning approa- ferent learning rates, network types, reduction techniques of
ches are preferable, as by their application, damage can be clas- network topologies, and dimension analysis [187]. Different
sified and quantified, while damage identification is the further reports [188–190] assess the benefits and drawbacks of using
level of damage, according to Rytter's Damage States of a System sensors and NN to detect impact in composite materials, which
[178], that unsupervised learning algorithms allow [119]. could be a possibility for SHM of OWT blades.

6.1. Supervised learning 6.1.4. Genetic algorithms

Ruotolo and Surace did most of the research related to this field
When supervised learning approaches are employed, very high between 1996 and 2001 [191–194]. In 1997 they formulated a
demand of data is associated with them, as data from every con- problem for choosing the location and depth of cracks in beams
ceivable damage situation must be available [179]. The two pos- employing measured modal parameters, which afterwards will be
sible sources of damage data come from: physics-based modelling optimized by a genetic algorithm [192,195] and [191]. Never-
(i.e. from Finite Element Analysis (FEA)), and experiments. Diffi- theless, there are some practical issues because, as the structure's
culty in obtaining these data in some fields jeopardises the complexity increases either size or geometry, the optimization
applicability of this approach (e.g. aviation). Moreover, to accu- becomes prohibitive [27]. The same authors carried out a similar
mulate enough training data, copies of the system of interest that study in 1998 where genetic algorithms, simulated annealing, and
can be intentionally damaged in different ways, might be neces- eigensensitivity analyses were compared in order to identify sev-
sary. The different analyses that can be categorised as supervised eral damage scenarios in a FEA of a frame structure [196]. Similar
learning algorithms are introduced below. studies were carried out in [197] for detecting damage in a
composite beam.
6.1.1. Response surface analysis (RSA)
The RSA obtains the approximation relationship between the 6.1.5. Support vector machine (SVM)
resonance frequencies and other damage parameters (i.e. damage The SVM constitutes a powerful framework for general classi-
location, and size). An example of this technique is explained in fication and regression problems; as many different types of dis-
[180], where damages were satisfactorily identified in beams and criminant functions, such as linear, nonlinear, neural network, and
plates made of CB reinforced plastic. The technique was applied to radial-basis discriminant functions, can be put in this tool with no
data simulated in analytical models. Nevertheless, the applicability real modifications [198]. While in [199], a SVM is applied to
of this technique, experimentally, is low as numerous data from damage classification problems in ball bearings and truss struc-
various damage conditions are required. tures, in [168], nonlinear principal component analysis based on
the unsupervised support vector machine is introduced and
6.1.2. Fisher's discriminant incorporated for data normalization.
This method introduces a linear transformation of the original
multivariate distributions into univariate distributions whose 6.2. Unsupervised learning
means are as far apart as possible, while the variances of those
transformed distributions are as small as possible [119]. It was Unsupervised Learning constitutes an alternative to Supervised
satisfactorily applied in [6] where linear and quadratic dis- Learning when no damage state data are available. However, the
crimination methodologies were implemented to measurements drawback Unsupervised Learning algorithms have, is that they can
taken from a concrete bridge column subjected to static and only be used for detection and possibly locating the damage [200].
dynamic testing. No relevant applications of this methodology For that reason, they have perhaps received less attention than
have been found for OW; however in [181] a new co-training Supervised Learning approaches. A common type of Unsupervised
algorithm based on modified Fisher's Linear Discriminant Analysis Learning algorithms is known as novelty detection or anomaly
was proposed for semi-supervised learning, which is meant to be detection method [201–203]. The idea of novelty detection is that
very useful in applications such as brain-computer interface only training data from the normal EOC of the structure or system
design. are used to establish the diagnostic. To do so, a model of the
normal EOC is created with the aim of comparing it with the one
6.1.3. Neural networks (NN) made with the newly acquired data. When significant deviations
NN are commonly used in SHMS for identifying, locating, and are detected, the algorithm indicates novelty, which means that
quantifying damage in structures. This methodology is nowadays the system has departed from the normal condition and, therefore,
very well known as substantial textbooks and monograph acquired damage. Unsupervised Learning algorithms can be
accounts exist [182]. NN are the group of statistical learning roughly categorized into three groups, i.e. Control Chart Analysis,
models inspired by biological NN. The reason NN are extremely Outlier Detection, and Neural Networks.
useful in SHM applications is the fact that they are used to esti-
mate or approximate functions that can depend on a large number 6.2.1. Control chart analysis
of inputs and are generally unknown [179]. This methodology continuously monitors the features extracted
Some of the studies that have employed NN in the past for from the measurements, for anomalies. When the observations
assessing structural damage include: the evaluation of two NNs for fluctuate outside the control limits, the monitoring system alarms
damage assessment, namely the Multilayer Perceptron Network the abnormality of the system's condition [27]. In [6], Control
and the Radial Basis Function Network [183]; and the damage Chart Analysis for monitoring a reinforced concrete bridge column
detection and location in a numerical simulation of a two- was used. It has also been frequently used for process control of
dimensional truss structure by using a feed-forward NN [184]. chemical plants, manufacturing facilities, and nuclear power
Other studies employed NN for assessing the integrity of bridges plants.
100 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105

6.2.2. Outlier detection various EOC of a structure, without any damage state. Singular
Outlier, or novelty detection, is the primary class of algorithms value decomposition is used to estimate the rank of this matrix.
applied in unsupervised learning applications. These algorithms After that, the same matrix is increased by adding an additional
assess statistical distributions of the measured or derived features column containing a new feature vector, this time corresponding
to enhance the damage identification process [204]. When applied to a potential damage state of the structure. In case this new
in an Unsupervised Learning mode, statistical models are typically feature vector corresponds to a damaged structure, it will be
used to answer questions regarding the existence and location of independent from the previously measured vectors and, therefore,
damage. When applied in a Supervised Learning mode and cou- the rank of the matrix will increase [205].
pled with analytical models, the statistical procedures can be used The basic principle of novelty detection is that a model of the
to better determine the type of damage, the extent of damage and system is built using training data only acquired from normal EOC
remaining useful life of the structure. The statistical models are of the structure. While the monitoring of the structure takes place,
also used to minimize false indications of damage (both false- newly acquired data are compared with the model. In the case that
positive and false-negative), as these are undesirable. significant deviations are found, the algorithm indicates novelty,
Outlier Detection Methodologies use changes in the rank of a which means that the system has deviated from the normal con-
matrix as a damage indicator [196]. Firstly, a matrix is composed dition and, therefore, is damaged [206]. Three different novelty
by putting the feature vectors in columns, measured during indices to detect damage in composite plates where introduced in

Table 1
Technology assessment: capabilities and limitations.

Technology Capabilities Limitations

Acoustic emission monitoring  Very effective detecting failure mechanisms up to microscale.  Limited application offshore
Type of sensors:  Allows a simple, rapid and cost-effective inspection or monitoring  Variable damage characterization and assessment
- Piezoelectric transducers of a structure. effectiveness depending on the algorithm.
 Good response at low frequencies.  Optimization of data processing needed as it still
 Multifunctional character of piezoelectric sensors. takes up much time and computational effort.
 High sensitivity to background noise.
 AE systems can only qualitatively gauge how much
damage is contained in a structure.
 Determining acoustic signature of the structure is
very difficult

Thermal imaging method  Fast.  Limited implementation in offshore structures.

Type of sensors:  Cost effective.  Camera resolution for detecting cracks
- Impedance tomography  Trials using drones are currently being conducted, which will  Laborious Image processing
- Thermography (infrared cameras) detect cracks up to 0.3 mm based on technology limitations, avoid  Cracks detection needs more automation from
the necessity of having personnel inside the turbine and be even footage.
more cost effective. Moreover, time required would be less than
traditional sensors.

Ultrasonic methods  It is sensitive to both surface and subsurface discontinuities.  Surface must be accessible to transmit ultrasound.
Type of sensors:  The depth of penetration for flaw detection or measurement is  Skill and training required is more extensive than
- Piezoelectric transducers superior to other NDT methods. other methods.
 Only single-sided access is needed when the pulse-echo technique  Coupling medium to promote the transfer of sound
is used. energy into the test specimen is required.
 It is highly accurate in determining reflector position and esti-  Difficulty of inspection of rough, irregular, very
mating size and shape. small, exceptionally thin or not homogeneous
 Minimal preparation is required. materials.
 Electronic equipment provides instantaneous results.  Difficulty of inspection of cast iron and other coarse
 Detailed images can be produced with automated systems. grained materials.
 It has other uses, such as thickness measurement, in addition to  Linear defects oriented parallel to the sound beam
flaw detection. may go undetected.
 Reference standards are required for both equip-
ment calibration and the characterization of flaws.

Fatigue and modal properties  High reliability, mature technology  Difficult analysis in operating conditions.
monitoring  Easy installation.  High number of uncertainties when applied in the
Type of sensors:  There are many different techniques available for this purpose. offshore environment.
- Accelerometers.  Recent developments in Operational Modal Analysis solve some  Environmental and Operational Conditions changes
- MEMS. limitations. have to be accounted in the results.
- Plastic optical-fiber based  Stable performance.  Difficulties in wind and wave loads measuring.
accelerometers.
- Velocimeters.

Strain monitoring  Easy installation process once appropriate training has been  Not very robust system.
Type of sensors: undertaken.  The installation is very sensitive to misalignments.
- Strain gauge (capacitance, inductance,  Mature technology.  Reduced service life.
semiconductor and resistance).  Optical fiber might be the future of strain monitoring as it is less  Distance between the sensor and the Data Acqui-
- Fiber optic cables. prone to fatigue, eliminates wiring issues and allows more points to sition System influences accuracy and limits sensor
- Fiber Bragg Grating (FBG). be monitored with the same cable. location.
 Mechanical properties limitations
 Can be affected by EMI noise.
 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105 101

[207]. Thus, a stochastic subspace approach to determine damage 8. Conclusions

existence in a structure was used in [208].
In this paper, a review of the Statistical Pattern Recognition
6.2.3. Neural networks Paradigm for SHMS for OWT has been carried out. It is expected
NN in the Unsupervised Learning mode work in the same way that by the assessment of each one of the stages present in this
as in the Supervised, apart from the fact that no data from paradigm, SHMS can contribute in the development of an appro-
damaging events are available. A good example is the adaptive NN priate Condition Based Maintenance Strategy. The optimization of
model proposed in [209]. In the model, data obtained from FEA
this strategy will lead to reducing labor costs of WT inspection,
simulations are used to train the NN; being the modal parameters
preventing unnecessary replacement of components, discovering
from the FEA simulations used as inputs. The NN output will
design weaknesses before failure, improving the availability of
consist of structural parameters. Once modal parameters from the
power while preventing WTs overloading, and maximizing return
actual structure become available, the NN is used to calculate the
associated structural parameters. Finally, the FEA model is updated in wind farm investments [50]. Increasing efficiency in operational
using these new structural parameters, calculating the associated management will contribute towards achieving UK's 2020 and
modal parameters. Training will stop when the measured modal 2050 targets, through ultimately reducing the Levelised Cost of
parameters are acceptably not so different from those calculated Energy (LCOE) [211].
from the FEA model. In [210] a discussion of delamination detec-
tion within composites applying a similar methodology can be
found. Good agreement between experimental and analytical Acknowledgments
results was achieved. In [48], synthetic damage patterns are
introduced in the FEA models. These models’ structural responses This work was supported by grant EP/L016303/1 for Cranfield
to the damage patterns are calculated, analyzed, and archived in a University, Centre for Doctoral Training in Renewable Energy
“damage catalogue” which was used for posterior deteriorations Marine Structures (REMS) (http://www.rems-cdt.ac.uk/) from the
and damage assessment of the WT structure, in near real time. The UK Engineering and Physical Sciences Research Council (EPSRC).
most recent NN application to WT blades is explained in [81],
where the different NN types that can be used are identified.

References
7. Discussion
[1] European Commission. Communication from the Commission to the Eur-
Previous sections have reviewed the different SHM technolo- opean Parliament, the Council, the European Economic and Social Commit-
gies that could be employed for OWTs. A summary of related tee and the Committee of the Regions. 〈http://ec.europa.eu/europe2020/pdf/
europe2020stocktaking_en.pdf〉; 2014 [accessed 03.8.15].
critical aspects, such as cost/effectiveness, capabilities and limita- [2] EWEA Business Intelligence. Aiming high: rewarding ambition in wind
tions can be found in Table 1, based on a structured survey aiming energy; November 2015. http://dx.doi.org/10.1021/nn700402d.
to map current practice within the industry. Responses have [3] RenewableUK. Wind Energy in the UK. State of the Industry Report Summary
2015 2015:1–43.
indicated high levels of interest and engagement in this topic [4] Ho A, Mbistrova A, Corbetta G. EWEA. The European offshore wind industry:
obtaining responses from multiple industrial stakeholders key trends and statistics 2015. European Wind Energy Asociation; 2016.
including sensor providers, equipment providers, consultancies [5] Ciang CC, Lee J-R, Bang H-J. Structural health monitoring for a wind turbine
system: a review of damage detection methods. Meas Sci Technol
and designers, and developers/operators. 2008;19:122001. http://dx.doi.org/10.1088/0957-0233/19/12/122001.
Some of the conclusions derived from the data collection pro- [6] Farrar CR. Sohn H. Pattern recognition for structural health monitoring. 2000.
cess can be summarized below: [7] Ghoshal A, Sundaresan MJ, Schulz MJ, Pai PF. Structural health monitoring
techniques for wind turbine blades. J Wind Eng Ind Aerodyn 2000;85:309–
24. http://dx.doi.org/10.1016/S0167-6105(99)00132-4.
 First generation of wind farms were equipped with sensors after [8] Adedipe O, Brennan F, Kolios AJ. Review of Corrosion fatigue in offshore
their deployment. The technologies mostly used were strain and structures: present status and challenges in the offshore wind sector. Renew
Sustain Energy Rev 2016;61:141–54.
fatigue and modal properties monitoring.
[9] Larsen FMF, Sorensen T. New lightning qualification test procedure for large
 The percentage of instrumented turbines within a wind farm is wind turbine blades. In: International Conference on Lightning 2003;36:1–10.
between 3% and 12%, showing a wide discrepancy in best [10] Kim H, Melhem H. Damage detection of structures under earthquake exci-
practice. tation using discrete wavelet analysis. Eng Struct 2004;26:347–62.
[11] Farrar CR, Worden K. An introduction to structural health monitoring. Philos
 Strain gauges, accelerometers and inclinometers are the tech- Trans A: Math Phys Eng Sci 2007;365:303–15. http://dx.doi.org/10.1098/
nologies mostly used for Structural Health Monitoring. LVDTs rsta.2006.1928.
(Linear Variable Differential Transducer) are of interest as they [12] Bently DE, Hatch CT. Fundamentals of rotating machinery diagnostics. New
York: ASME Press; 2003.
can measure displacements with high reliability and accuracy, [13] Shull PJ. Nondestructive evaluation theory, techniques, and applications.
however at the expense of cost. New York: Marcel Dekker, Inc.; 2002.
 To date, natural frequency analysis is considered to be the most [14] Montgomery DC. Introduction to statistical quality control. New York: John
Wiley & Sons Inc.; 1997.
commonly applied practice for detecting deviations from the [15] Farrar CR, Lieven NAJ. Damage prognosis: the future of structural health
design intent in wind turbine foundations, as other techniques monitoring. Philos Trans A: Math Phys Eng Sci 2007;365:623–32. http://dx.
are either very expensive, low maturity or of low accuracy. doi.org/10.1098/rsta.2006.1927.
 Fiber optics technology for strain measurement is considered a [16] Farrar CR, Sohn H, Hemez FM, Anderson MC, Bement MT, Cornwell PJ, et al.
Damage prognosis: current status and future needs 2003.
promising future technology due to the numerous benefits it [17] Farrar CR, Doebling SW, Nix DA. Vibration-based structural damage identi-
brings. However, this option has not been costed or imple- fication. Philos Trans A: Math Phys Eng Sci 2001:359.
[18] Fassois SD, Sakellariou JS. Time-series methods for fault detection and
mented in any case and installation needs to be carefully
identification in vibrating structures. Philos Trans A: Math Phys Eng Sci
assessed. 2007;365:411–48. http://dx.doi.org/10.1098/rsta.2006.1929.
 A necessity of the development of new methodologies to col- [19] Farrar CR, Sohn H, Worden K. Data normalization: a key for structural health
lect, review, purify and analyze the data collected by CM and monitoring 2001.
[20] Friswell MI. Damage identification using inverse methods. Philos Trans A:
SHM solutions has been pointed out by most of the Math Phys Eng Sci 2007;365:393–410. http://dx.doi.org/10.1098/
interviewees. rsta.2006.1930.
102 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105

[21] Mal A, Banerjee S, Ricci F. An automated damage identification technique Materials and Plasma Technology, IMRMPT 2015 and 1st Symposium on
based on vibration and wave propagation data. Philos Trans A: Math Phys Nanoscience and Nanotechnology; Dann Carlton Hotel Bucaramanga,
Eng Sci 2007;365:479–91. http://dx.doi.org/10.1098/rsta.2006.1933. Bucaramanga; Colombia; 4 May 2015 through 9 May 2015; Code 119613.
[22] Staszewski WJ, Robertson AN. Time-frequency and time-scale analyses for Journal of Physics: Conference Series Volume 687, Issue 1, 12 February 2016,
structural health monitoring. Philos Trans A: Math Phys Eng Sci Article number 012001.
2007;365:449–77. http://dx.doi.org/10.1098/rsta.2006.1936. [50] Schulz MJ, Sundaresan MJ. Smart sensor system for structural condition
[23] Hayton P, Utete S, King D, King S, Anuzis P, Tarassenko L. Static and dynamic monitoring of wind turbines. Denver, Colorado, USA: National Renewable
novelty detection methods for jet engine health monitoring. Philos Trans R Energy Laboratory; 2006.
Soc A: Math Phys Eng Sci 2007;365:493–514. http://dx.doi.org/10.1098/ [51] Sundaresan MJ, Schulz MJ, Ghoshal A. Structural health monitoring static test
rsta.2006.1931. of a wind turbine blade. Denver, Colorado, USA 2002.
[24] Sohn H. Effects of environmental and operational variability on structural [52] Chang PC, Flatau A, Lui SC. Review paper: health monitoring of civil infra-
health monitoring. Philos Trans A: Math Phys Eng Sci 2007;365:539–60. structure. Struct Health Monit 2003:341–58.
http://dx.doi.org/10.1098/rsta.2006.1935. [53] Doebling S, Farrar CR, Prime M. A summary review of vibration-based
[25] Worden K, Manson G. The application of machine learning to structural damage identification methods. Shock Vib Dig 1998:1–34.
health monitoring. Philos Trans A: Math Phys Eng Sci 2007;365:515–37. [54] Butterfield S, Sheng S, Oyague F. Wind energy's new role in supplying the
http://dx.doi.org/10.1098/rsta.2006.1938. world’s energy—what role will structural health monitoring play. In: Pro-
[26] Doebling S, Farrar C, Prime M, Shevitz D. Damage identification and health ceedings of the 8th International Workshop on Structural Health Monitoring.
monitoring of structural and mechanical systems from changes in their 7th International Workshop on Structural Health Monitoring, Stanford, CA,
vibration characteristics: a literature review 1996. doi:10.2172/249299. USA. DEStech Publications Inc. 2009.
[27] Sohn H, Farrar CR, Hemez FM. A review of structural health monitoring lit- [55] Sutherland H, Beattie A, Hansche B, Musial W, Allread JM, Johnson J. The
erature: 1996–2001. Masachusets, USA 2002. application of non-destructive techniques to the testing of a wind turbine
[28] Randall RB. State of the art in monitoring rotating machinery – part 1. J blade. USA: Sandia National Laboratories; 1994.
Sound Vib 2004;38:10–6. [56] Joosse PA, Blanch MJ, Dutton AG, Kouroussis DA, Philippidis TP, Vionis PS.
[29] Byington CS, MJR. Prognostic enhancements to diagnostic systems for Acoustic emission monitoring of small wind turbine blades. In: Proceedings
improved condition-based maintenance [military aircraft]. In: Aerospace of the 21st ASME wind energy symposium conjunction with 40th AIAA
conference proceedings, IEEE 2002;6:2815–24. aerospace sciences meeting, Reno, USA 2002:1–11.
[30] Bao P, Yuan M, Dong S, Song H, Xue J. Fiber Bragg grating sensor fatigue crack [57] Anastassopoulos AA, Kouroussis DA, Nikolaidis VN, Proust A, Dutton AG,
real-time monitoring based on spectrum cross-correlation analysis. J Sound Blanch MJ, et al. Structural integrity evaluation of wind turbine blades using
Vib 2013;332:43–57. http://dx.doi.org/10.1016/j.jsv.2012.07.049. pattern recognition analysis on acoustic emission data. In: Proceedings of the
[31] Woike M, Abdul-Aziz A, Oza N, Matthews B. New sensors and techniques for 25th european conference on acoustic emission testing EWGAE 2002, Pra-
the structural health monitoring of propulsion systems. Sci World J 2013: gue, Czech Republic 2002.
1–11. http://dx.doi.org/10.1155/2013/596506. [58] Beattie A. Acoustic emission monitoring of a wind turbine blade during a
[32] Ignatovich SR, Menou A, Karuskevich MV, Maruschak PO. Fatigue damage and fatigue test. In: Proceedings of the 35th aerospace sciences meeting and
sensor development for aircraft structural health monitoring. Theor Appl Fract exhibit, American Institute of Aeronautics and Astronautics 1997. doi:10.
Mech 2013;65:23–7. http://dx.doi.org/10.1016/j.tafmec.2013.05.004. 2514/6.1997-958.
[33] Brownjohn JMW. Structural health monitoring of civil infrastructure. Philos [59] Jørgensen ER, Borum KK, Mcgugan M, Thomsen CL, Jensen FM, Debel CP,
Trans: Math Phys Eng Sci 2007;365:589–622. et al. Full scale testing of wind turbine blade to failure - flapwise loading.
[34] Lynch JP. An overview of wireless structural health monitoring for civil Denmark 2004;1392.
structures. Philos Trans A: Math Phys Eng Sci 2007;365:345–72. http://dx. [60] Wells R, Hamstad MA, Mukherjee AK. On the origin of the first peak of
doi.org/10.1098/rsta.2006.1932. acoustic emission in 7075 aluminum alloy. J Mater Sci 1983;18:1015–20.
[35] Carden EP. Vibration based condition monitoring: a review. Struct Health [61] Dutton AG. Thermoelastic Stress Measurement and Acoustic Emission
Monit 2004;3:355–77. http://dx.doi.org/10.1177/1475921704047500. Monitoring in Wind. European wind energy conference (EWEC) 2004, Lon-
[36] Grafe H. Model updating of large structural dynamics models using mea- don 2004.
sured response functions. UK: Imperial College of Science, Technology and [62] Dutton AG. Acoustic emission condition monitoring of wind turbine rotor B
Medicine, University of London; 1998. lades: laboratory certification testing to large scale in-service deployment.
[37] Wang ML, Lynch JP, Sohn H. Sensor technologies for civil infrastructures. UK: European wind energy conference 2003:1–11.
Cambridge; 2014. [63] Avdelidis NP, Almond DP, Ibarra-Castanedo C, Bendada A, Kenny S, Maldague
[38] Montalvão D, Maia NMM, Ribeiro AMR. A review of vibration-based struc- X. Structural integrity assessment of materials by thermography. Conference
tural health monitoring with special emphasis on composite materials. on damage in composite materials (CDCM) 2006.
Shock Vib Dig 2006;38:295–324. [64] Krstulovic-Opara L, Klarin B, Neves P, Domazet Z. Thermal imaging and
[39] Kaewunruen S. Monitoring structural deterioration of railway turnout sys- Thermoelastic Stress Analysis of impact damage of composite materials. Eng
tems via dynamic wheel/rail interaction. Case Stud Nondestruct Test Eval Fail Anal 2011;18:713–9. http://dx.doi.org/10.1016/j.engfailanal.2010.11.010.
2014;1:19–24. http://dx.doi.org/10.1016/j.csndt.2014.03.004. [65] Rantala J, Wu D, Busse G. Amplitude modulated lock-in vibrothermography
[40] Eason TJ, Bond LJ, Lozev MG. Ultrasonic thickness structural health mon- for NDE of polymers and composites. J Res Nondestruct Eval 1996;7:215–28.
itoring photoelastic visualization and measurement accuracy for internal [66] Hahn F, Kensche CW, Paynter RJH, Dutton AG, Kildegaard J, Kosgaard C.
pipe corrosion. Proc. SPIE, Smart Materials and Nondestructive Evaluation for Design, fatigue test and NDE of a sectional wind turbine rotor blade. J
Energy Systems 2015:9439. Thermoplast Compos Mater 2002;15:267–77.
[41] Jacques R, Clarke T, Morikawa S, Strohaecker T. Monitoring the structural [67] Walle G, Abuhamad M, Toma E, Netzelmann U. Defect indications in sono-
integrity of a flexible riser during dynamic loading with a combination of thermography in relation to defect location and structure. In: Proceedings of
non-destructive testing methods. NDT E Int 2010;43:501–6. http://dx.doi. quantitative infrared thermography conference, Rhode Saint Genèse, Bel-
org/10.1016/j.ndteint.2010.05.005. gium 2004.
[42] Kolios AJ, Chahardehi A, Brennan FP. Experimental determination of the [68] Li S, Shi K, Yang K, Xu J. Research on the defect types judgment in wind
overturning moment and net force generated by a novel VAWT – experiment turbine blades using ultrasonic NDT. IOP Conference Series Materials Science
design under load uncertainty. Exp Tech 2013;37:7–14. and Engineering 2015;87:012056. doi:10.1088/1757-899X/87/1/012056.
[43] Borg M, Collu M, Kolios AJ. Offshore floating vertical axis wind turbines, [69] Sørensen BF, Lading L, Sendrup P, McGugan M, Debel CP, Kristensen OJDD,
dynamics modelling state of the art. Part II: Mooring Line & Structural et al. Fundamentals for remote structural health monitoring of wind turbine
Dynamics. Renew Sustain Energy Rev 2014;39:1226–34. blades - a preproject 2002;1336. doi:Risø-R-1336(EN).
[44] Brennan FP, Kolios AJ. Structural integrity considerations for the H2OCEAN [70] Park B, An Y-K, Sohn H. Visualization of hidden delamination and debonding
multi modal wind-wave platform. In: Scientific proceedings of european in composites through noncontact laser ultrasonic scanning. Compos Sci
wind energy association annual Event, Barcelona, Spain; 2014. Technol 2014;100:10–8.
[45] Liu WY, Tang BP, Han JG, Lu XN, Hu NN, He ZZ. The structure healthy con- [71] Farrar CR, Doebling SW. Damage detection and evaluation II: field applica-
dition monitoring and fault diagnosis methods in wind turbines: a review. tions to large structures. Modal Anal Test NATO Sci Ser 1999;363:345–78.
Renew Sustain Energy Rev 2015;44:466–72. http://dx.doi.org/10.1016/j. [72] Zhang H, Schulz MJ, Ferguson F, Pai PF. Structural health monitoring using
rser.2014.12.005. transmittance functions. Mech Syst Signal Process 1999;13:765–87.
[46] García Márquez FP, Tobias AM, Pinar Pérez JM, Papaelias M. Condition [73] Gross E, Simmermacher T, Rumsey M, Zadoks RI. Application of damage
monitoring of wind turbines: techniques and methods. Renew Energy detection techniques using wind turbine modal data. ASME Wind Energy
2012;46:169–78. http://dx.doi.org/10.1016/j.renene.2012.03.003. Symposium, Reno, NV, USA 1999.
[47] Devriendt C, Magalhaes F, Weijtjens W, Sitter G, De, Cunha A, Guillaume P. [74] Jang J-H, Yeo I, Shin S, Chang S-P. Experimental investigation of system
Structural Health Monitoring of offshore wind turbines using automated identification-based damage assessment on structures. J Struct Eng
operational modal analysis. Struct Health Monit 2014;13:644–59. http://dx. 2002;128:673–82.
doi.org/10.1177/1475921714556568. [75] Carne TG, James GH. The inception of OMA in the development of modal
[48] Smarsly K, Hartmann D, Law KH. A computational framework for life-cycle testing technology for wind turbines. Mech Syst Signal Process
management of wind turbines incorporating structural health monitoring. Struct 2010;24:1213–26. http://dx.doi.org/10.1016/j.ymssp.2010.03.006.
Health Monit 2013;12:359–76. http://dx.doi.org/10.1177/1475921713493344. [76] Hermans L, Van Der Auweraer H. Modal testing and analysis of structures
[49] Kolios A, Martinez-Luengo M. Operational management of offshore energy under operational conditions: industrial applications. Mech Syst Signal
assets. In: Proceedings of the 3rd International Meeting for Researchers in Process 1999;13:193–216. http://dx.doi.org/10.1006/mssp.1998.1211.
 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105 103

[77] Cauberghe B. Applied frequency-domain system identification in the field of [107] Stubbs N, Kim J-T. Damage localization in structures without baseline modal
experimental and operational modal analysis. Vrije Universiteit Brussel, parameters. AIAA J 1996;34:1644–9.
Vakgroep Werktuigkunde, Faculteit Toegepaste Wetenschappen, 2004. [108] Park G. Self-monitoring and self-healing jointed structures. Key Eng Mater
[78] Devriendt C, Guillaume P. Identification of modal parameters from trans- 2001:75–84.
missibility measurements. J Sound Vib 2008;314:343–56. http://dx.doi.org/ [109] Li GQ, Hao KC, Lu Y, Chen SW. A Flexibility approach for damage identifica-
10.1016/j.jsv.2007.12.022. tion of cantilever-type structures with bending and shear deformation.
[79] Devriendt C, El-kafafy M, Sitter G De, Guillaume P. Damping estimation of Comput Struct 1999;73:565–72. http://dx.doi.org/10.1016/S0045-7949(98)
offshore wind turbines on a monopile foundation using state-of-the-art 00295-8.
operational modal analysis techniques. International Conference on Noise [110] Bernal D. Load vectors for damage localization. J Eng Mech 2002;128:7–14.
and Vibration Engineering, Leuven 2012:2647–62. [111] Zhao J, DeWolf JT. Sensitivity study for vibrational parameters used in
[80] Antoniadou I, Dervilis N, Papatheou E, Maguire AE, Worden K, Sheffield S. damage detection. J Struct Eng 1999;125:410–6.
Aspects of structural health and condition monitoring of offshore wind [112] Feng MQ, Bahng EY. Damage assessment of jacketed RC columns using
turbines. Philos Trans A: Math Phys Eng Sci 2015:373. vibration tests. J Struct Eng 1999;125:265–71.
[81] Dervilis N, Choi M, Taylor SG, Barthorpe RJ, Park G, Farrar CR, et al. On [113] Mangal L, Idichandy VG, Ganapathy C. ART-based multiple neural networks
damage diagnosis for a wind turbine blade using pattern recognition. J for monitoring offshore platforms. Appl Ocean Res 1996;18:137–43.
Sound Vib 2014;333:1833–50. http://dx.doi.org/10.1016/j.jsv.2013.11.015. http://dx.doi.org/10.1016/0141-1187(96)00024-7.
[82] Weijtjens W, Verbelen T, De Sitter G, Devriendt C. Foundation structural [114] Waszczyszyn Z, Ziemianski L. Neural networks in mechanics of structures
health monitoring of an offshore wind turbine: a full-scale case study. Struct and materials - new results and prospects of applications. Comput Struct
Health Monit 2015. http://dx.doi.org/10.1177/1475921715586624. 2001;79:2261–76. http://dx.doi.org/10.1016/S0045-7949(01)00083-9.
[83] Park G, Inman DJ. Impedance-based structural health monitoring damage [115] Zubaydi A, Haddara MR, Swamidas ASJ. Damage identification in a ship's
prognosis—for aerospace, civil and mechanical system. Chichester, UK: structure using neural networks. Ocean Eng 2002;29:1187–200. http://dx.
Wiley; 2005. p. 275–92. doi.org/10.1016/S0029-8018(01)00077-4.
[84] Peairs DM, Park G, Inman DJ. Improving accessibility of the impedance-based [116] Barai SV, Pandey PC. Time-delay neural networks in damage detection of
structural health monitoring method. J Intell Mater Syst Struct 2004;15:129–39. railway bridges. Adv Eng Softw 1997;28:1–10. http://dx.doi.org/10.1016/
[85] Pitchford C, Grisso BL, Inman DJ. Impedance-based structural health mon- S0965-9978(96)00038-5.
itoring of wind turbine blades. Proc SPIE 2007. http://dx.doi.org/10.1117/ [117] Marwala T, Hunt HEM. Fault identification using finite element. Mech Syst
12.715800. Signal Process 1999;13:475–90.
[86] Schubel PJ, Crossley RJ, Boateng EKG, Hutchinson JR. Review of structural [118] Zak A, Krawczuk M, Ostachowicz W. Vibration of a laminated composite plate
health and cure monitoring techniques for large wind turbine blades. Renew with closing delamination. J Intell Mater Syst Struct 2001;12:545–51.
Energy 2013;51:113–23. http://dx.doi.org/10.1016/j.renene.2012.08.072. http://dx.doi.org/10.1106/9PFK.-LXAD-9WLL-JXMG.
[87] Balageas D, Fritzen C-P, Guemes A. Structural health monitoring. London, [119] Sohn H, Farrar CR, Hemez F, Czarnecki J. A review of structural health mon-
UK: ISTE Ltd.; 2006. itoring literature 1996–2001 2002.
[88] Schroeder K, Ecke W, Apitz J, Lembke E, Lenschow G. A fibre Bragg grating [120] Marín JC, Barroso A, París F, Cañas J. Study of fatigue damage in wind turbine
blades. Eng Fail Anal 2009;16:656–68. http://dx.doi.org/10.1016/j.
sensor system monitors operational load in a wind turbine rotor blade. Meas
engfailanal.2008.02.005.
Sci Technol 2006;17:1167. http://dx.doi.org/10.1088/0957-0233/17/5/S39.
[121] Adedipe O, Brennan F, Kolios AJ. Corrosion fatigue load frequency sensitivity
[89] Li X, Levy C, Li M, Keshri AK, Agarwal A. A multifunctional MWCNT com-
analysis. Mar Struct 2015;42:115–36. http://dx.doi.org/10.1016/j.
posite: strain sensing, damping and application to structural vibration con-
marstruc.2015.03.005.
trol. ASME international mechanical engineering congress & exposition
[122] Adedipe O., Brennan F., Kolios A.J. Corrosion fatigue crack growth in offshore
proceedings 2009:323–34.
wind monopile steel HAZ material. MARSTRUCT 2015, 5th Int. Conf. Mar.
[90] Stamatis DH. Failure mode and effect analysis: FMEA from theory to
Struct., Southampton, United Kingdom; 2015.
execution. 2nd ed.. WI, USA: AS. Milwaukee; 2003.
[123] Van Der Tempel J. Design of support structures for offshore wind turbines.
[91] Sinha Y, Steel JA. A progressive study into offshore wind farm maintenance
Delft, 2006. TU Delft 2006.
optimisation using risk based failure analysis. Renew Sustain Energy Rev
[124] Gomez HC, Gur T, Dolan D. Structural condition assessment of offshore wind
2015;42:735–42.
turbine monopile foundations using vibration monitoring data. Proc SPIE
[92] Li D, Ho S-CM, Song G, Ren L, Li H. A review of damage detection methods for
2013;8694:86940B. http://dx.doi.org/10.1117/12.2018263.
wind turbine blades. Smart Mater Struct 2015;24:033001. http://dx.doi.org/
[125] Adams D, White J, Rumsey M, Farrar CR. Structural health monitoring of wind
10.1088/0964-1726/24/3/033001.
turbines: method and application to a HAWT. Wind Energy 2011;14:603–23.
[93] Salawu OS. Detection of structural damage through changes in frequency: a
http://dx.doi.org/10.1002/we.
review. Eng Struct 1997;19:718–23. http://dx.doi.org/10.1016/S0141-0296
[126] Cross EJ. On structural health monitoring in changing environmental and
(96)00149-6. operational conditions. University of Sheffield, 2012.
[94] Allemang RJ, Brown DL. Correlation coefficient for modal vector analysis. In: [127] Faulkner P, Cutter P, Owens A. Structural health monitoring systems in dif-
Proceedings of International Modal Analysis Conference & Exhibit, Union ficult environments - offshore wind turbines. In: Proceedings of the 6th
Coll 1982:110–16. European workshop on structural health monitoring 2012:1–7.
[95] Salawu OS, Williams C. Bridge assessment using forced-vibration testing. J [128] Hu W, Thöns S, Rohrmann RG, Said S, Rücker W. Vibration-based structural
Struct Eng 1995;121:161–73. http://dx.doi.org/10.1061/(ASCE)0733-9445 health monitoring of a wind turbine system Part II : environmental/opera-
(1995)121:2(161). tional effects on dynamic properties. Eng Struct 2015;89:273–90.
[96] Ren W-X, De Roeck G. Structural damage identification using modal data. I: [129] Doebling SW, Farrar CR. Using statistical analysis to enhance modal-based
simulation verification. J Struct Eng 2002;128:87–95. http://dx.doi.org/ damage identification in structural damage assessment using advanced sig-
10.1061/(ASCE)0733-9445(2002)128:1(87). nal processing procedures. In: Proceedings of DAMAS '97, University of
[97] Ren W-X, De Roeck G. Structural damage identification using modal data. II: Sheffield, UK 1997:199–210.
test verification. J Struct Eng 2002;128:96–104. http://dx.doi.org/10.1061/ [130] Sakellariou JS, Fassois SD. Parametric output error based identification and
(ASCE)0733-9445(2002)128:1(96). fault detection in structures under earthquake excitation Proceedings of
[98] Araújo dos Santos JV, Mota Soares CM, Mota Soares CA, Pina HLG. A damage European cost F3 conference on system identification and structural health
identification numerical model based on the sensitivity of orthogonality monitoring 2000:323–32.
conditions and least squares techniques. Comput Struct 2000;78:283–91. [131] Taylor SG, Farinholt K, Choi M, Jeong H, Jang J, Park G, et al. Incipient crack
[99] Ratcliffe CP, Bagaria WJ. Vibration technique for locating delamination in a detection in a composite wind turbine rotor blade. J Intell Mater Syst Struct
composite beam. AIAA J 1998;36:1074–7. 2014;25:613–20. http://dx.doi.org/10.1177/1045389X13510788.
[100] Abdel Wahab MM, De Roeck G. Damage detection in bridges using modal [132] Martinez Luengo M, Kolios A. Failure mode identification and end of life
curvatures: application to a real damage scenario. J Sound Vib 1999;226: scenarios of offshore wind turbines: a review. Energies 2015;8:8339–54.
217–35. http://dx.doi.org/10.1006/jsvi.1999.2295. http://dx.doi.org/10.3390/en8088339.
[101] Sampaio RPC, Maia NMM, Silva JMM. Damage detection using the frequency [133] Overgaard LCT, Lund E, Thomsen OT. Structural collapse of a wind turbine blade.
response function curvature method. J Sound Vib 1999;226:1029–42. Part A: static test and equivalent single layered models. Compos Part A Appl Sci
http://dx.doi.org/10.1006/jsvi.1999.2340. Manuf 2010;41:257–70. http://dx.doi.org/10.1016/j.compositesa.2009.10.011.
[102] Oh BH, Jung BS. Structural damage assessment with combined data of static [134] Kirikera G, Sundaresan M, Nkrumah F, Grandhi G, Ali B, Mullapudi S, et al.
and modal tests. J Struct Eng 1998;124:956–65. Wind turbines. Encycl. Struct. Health Monit 2009:1–23.
[103] Abdel Wahab MM. Effect of modal curvatures on damage detection using [135] Sørensen FB, Jørgensen E, Debel PC, Jensen MF, Jensen MH, Jacobsen KT, et al.
model updating. Mech Syst Signal Process 2001;15:439–45. http://dx.doi. Improved design of large wind turbine blade of fibre composites based on
org/10.1006/mssp.2000.1340. studies of scale effects (Phase 1) - Summary Report (Risø-R-1390(EN))
[104] Kim J-T, Stubbs N. Model-uncertainty impact and damage-detection accuracy 2004;1390. Doi: Risø-R-1390(EN).
in plate girder. J Struct Eng 1995;121:1409–17. http://dx.doi.org/10.1061/ [136] Zhou HF, Dou HY, Qin LZ, Chen Y, Ni YQ, Ko JM. A review of full-scale
(ASCE)0733-9445(1995)121:10(1409). structural testing of wind turbine blades. Renew Sustain Energy Rev
[105] Worden K. An experimental appraisal of the strain energy damage location 2014;33:177–87. http://dx.doi.org/10.1016/j.rser.2014.01.087.
method. Key Eng Mater 2001:35–46. [137] Kessler S. Damage detection in composite materials using frequency
[106] Kim J-T, Stubbs N. Improved damage identification method based. J Sound response methods. Compos Part B Eng 2002;33:87–95. http://dx.doi.org/
Vib 2002;252:223–38. http://dx.doi.org/10.1006/jsvi.2001.3749. 10.1016/S1359-8368(01)00050-6.
104 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105

[138] Antunes PC, Dias JM, Varum H, André P. Dynamic structural health mon- [167] Weijtjens W, Verbelen T, De Sitter G, Devriendt C. Data normalization for
itoring of a civil engineering structure with a POF accelerometer. Sens Rev foundation SHM of an offshore wind turbine: a real-life case study. In: Pro-
2014;34:36–41. http://dx.doi.org/10.1108/SR-04-2013-656. ceedings of the 2nd European conference of the prognostics and health
[139] Ahlborn TM, Shuchman R, Sutter LL, Brooks CN, Harris DK, Burns JW, et al. management society, Acoustics and Vibrations Research Group (AVRG), Vrije
The state-of-the-practice of modern structural health monitoring for bridges: Universiteit Brussel, Brussel, Belgium 2014:788–95.
a comprehensive review. Burns 2010;48105:70. [168] Oh CK, Sohn H. Damage diagnosis under environmental and operational
[140] Dally JW, Riley WF. Experimental stress analysis. J Appl Mech 1996:33. variations using unsupervised support vector machine. J Sound Vib
http://dx.doi.org/10.1115/1.3625012. 2009;325:224–39. http://dx.doi.org/10.1016/j.jsv.2009.03.014.
[141] Bogue R. Uk start up poised to take the fibre optic sensor marked by storm. [169] Park G, Inman DJ. Structural health monitoring using piezoelectric impe-
Sens Rev 2005;25:24–7. dance measurements. Philos Trans A: Math Phys Eng Sci 2007;365:373–92.
[142] Wang S, Chung DDL, Chung JH. Impact damage of carbon fiber polymer- http://dx.doi.org/10.1098/rsta.2006.1934.
matrix composites, studied by electrical resistance measurement. Compos [170] Todd MD, Nichols JM, Trickey ST, Seaver M, Nichols CJ, Virgin LN. Bragg
Part A: Appl Sci Manuf 2005;36:1707–15. http://dx.doi.org/10.1016/j. grating-based fibre optic sensors in structural health monitoring. Philos
compositesa.2005.03.005. Trans A: Math Phys Eng Sci 2007;365:317–43. http://dx.doi.org/10.1098/
[143] Inam F, Bhat BR, Vo T, Daoush WM. Structural health monitoring capabilities rsta.2006.1937.
in ceramic-carbon nanocomposites. Ceram Int 2014;40:3793–8. http://dx. [171] Abe S. Feature selection and extraction. In: Support vector machines for
doi.org/10.1016/j.ceramint.2013.09.039. pattern classification. Advances in pattern recognition 2010:471. Doi:10.1007/
[144] Swait TJ, Jones FR, Hayes SA. A practical structural health monitoring system for 978-1-84996-098-4.
carbon fibre reinforced composite based on electrical resistance. Compos Sci [172] Wold S, Esbensen K, Geladi P. Principal component analysis. Chemom Intell
Technol 2012;72:1515–23. http://dx.doi.org/10.1016/j.compscitech.2012.05.022. Lab Syst 1987;2:37–52. http://dx.doi.org/10.1016/0169-7439(87)80084-9.
[145] Mascarenas DL, Todd MD, Park G, Farrar CR. Development of an impedance- [173] McLachlan GJ. Discriminant analysis and statistical pattern recognition.
based wireless sensor node for structural health monitoring. Smart Mater United States: Wiley; 2004.
Struct 2007;16:2137–45. http://dx.doi.org/10.1088/0964-1726/16/6/016. [174] Chatterjee S, Simonoff JS. Handbook of regression analysis. United States:
[146] Hamzeloo SR, Shamshirsaz M, Rezaei SM. Damage detection on hollow Wiley; 2012.
cylinders by electro-mechanical impedance method: experiments and finite [175] Namey E, Guest G, Thairu L, Johnson L. Data reduction techniques for large
element modeling. C R - Mec 2012;340:668–77. http://dx.doi.org/10.1016/j. qualitative data sets. In: Handbook for team-based qualitative research
crme.2012.07.001. 2007:137–63.
[147] Staszewski WJ, Biemans C, Boller C, Tomlinson GR. Impact damage detec- [176] Kirikera GR, Shinde V, Schulz MJ, Ghoshal A, Sundaresan M, Allemang R.
tionin composite structures - recent advances. In: Structural health mon- Damage localisation in composite and metallic structures using a structural
itoring, Standford University. Palo Alto, California 2000:754–63. neural system and simulated acoustic emissions. Mech Syst Signal Process
[148] Yu-Lung Lo, Fu-Yuan Xiao. Measurement of corrosion and temperature using 2007;21:280–97. http://dx.doi.org/10.1016/j.ymssp.2006.01.010.
a single-pitch bragg grating fiber sensor. J Intell Mater Syst Struct [177] Ghoshal A, Martin WN, Schulz MJ, Chattopadhyay A, Prosser WH, Kim HS.
1998;9:800–7. Health monitoring of composite plates using acoustic wave propagation,
[149] Panova AA, Pantano P, Walt DR. In situ fluorescence imaging of localized continuous sensors and wavelet analysis. J Reinf Plast Compos 2007;26:
corrosion with a pH-sensitive imaging fiber. Anal Chem 1997;69:1635–41. 95–112. http://dx.doi.org/10.1177/0731684407069965.
http://dx.doi.org/10.1021/ac961025c. [178] Rytter A. Vibration based inspection of civil engineering structures. Den-
[150] Ling H, Lau K, Cheng L, Jin W. Fibre optic sensors for delamination identifi- mark: Aalborg University; 1993.
cation in composite beams using a genetic algorithm. Smart Mater Struct [179] Farrar CR, Worden K. Chapter 11: Supervised learning – classification and
2005;14:287–95. http://dx.doi.org/10.1088/0964-1726/14/1/030. regression. Structural health monitoring: a machine learning perspective
[151] Takeda N. Characterization of microscopic damage in composite laminates 2013:361–401.
and real-time monitoring by embedded optical fiber sensors. Int J Fatigue [180] Inada T, Shimamura Y, Todoroki A, Kobayashi H, Nakamura H. Damage
2002;24:281–9. http://dx.doi.org/10.1016/S0142-1123(01)00083-4. identification method for smart composite cantilever beams with piezo-
[152] Doyle C. Fibre bragg grating sensors - an introduction to Bragg gratings and electric materials. In: Structural health monitoring 2000, Stanford University,
interrogation techniques. 2003. Palo Alto, California 1999:986–94.
[153] Tsuda H, Toyama N, Urabe K, Takatsubo J. Impact damage detection in CFRP [181] Tan X-M, Chen M-Y, Gan JQ. A co-training algorithm based on modified
using fiber Bragg gratings. Smart Mater Struct 2004;13:719–24. http://dx.doi. Fisher's linear discriminant analysis. Intell Data Anal 2015;19:279–92.
org/10.1088/0964-1726/13/4/009. http://dx.doi.org/10.3233/IDA-150717.
[154] Takeda N, Okabe Y, Kuwahara J, Kojima S, Ogisu T. Development of smart [182] Bishop CM. Neural Networks for Pattern Recognition. Cambridge, United
composite structures with small-diameter fiber Bragg grating sensors for Kingdom: Cambridge University Press; 1998.
damage detection: quantitative evaluation of delamination length in CFRP [183] Rytter A, Kirgegaard P. Vibration based inspection using neural networks. In:
laminates using Lamb wave sensing. Compos Sci Technol 2005;65:2575–87. Structural damage assessment using advanced signal processing procedures,
http://dx.doi.org/10.1016/j.compscitech.2005.07.014. proceeding of DAMAS '97, University of Sheffield, UK 1997:97–108.
[155] Deng L, Cai CS. Applications of fiber optic sensors in civil engineering. Struct [184] Faravelli L, Pisano AA. Damage assessment toward performance control. In:
Eng Mech 2007;25:577–96. http://dx.doi.org/10.12989/sem.2007.25.5.577. Structural damage assessment using advanced signal processing procedures,
[156] Kolle JJ. Transverse optical fuse for composite damage detection. SAMPE Quat proceeding of DAMAS '97, University of Sheffield, UK 1997:185–98.
1993;24:35–40. [185] Chan TH, Ni YQ, Ko JM. Neural network novelty filtering for anomaly
[157] Swartz R, Lynch J, Zerbst S, Sweetman B, Rolfes R. Structural monitoring of detection of tsing Ma bridge cables. Structural Health Monitoring 2000,
wind turbines using wireless sensor networks. Smart Struct Syst 2010;6:1–8. Stanford University, Palo Alto, California 1999:430–9.
[158] Wang P, Yan Y, Tian GY, Bouzid O, Ding Z. Investigation of wireless sensor [186] Liu PL, Sun SC. The Application of Artificial Neural Networks on the Health
networks for structural health monitoring. J Sens 2012:7. http://dx.doi.org/ Monitoring of Bridges. Structural health monitoring, current status and
10.1155/2012/156329. perspectives, Stanford University, Palo Alto, California 1997:103–10.
[159] Wang F, Liu J. Duty-cycle-aware broadcast in wireless sensor networks. Proc. - [187] Hermann H-G, Streng J. Problem-specific neural networks for detecting
IEEE INFOCOM 2009:468–76. http://dx.doi.org/10.1109/INFCOM.2009.5061952. structural damage. Structural health monitoring, current status and per-
[160] Lynch JP, Loh KJ. A summary review of wireless sensors and sensor networks spectives, Stanford University, Palo Alto, California 1997:267–78.
for structural health monitoring. Shock Vib Dig 2006;38:91–128. http://dx. [188] Feng M, Bahng E. Damage assessment of bridges with jacketed RC columns
doi.org/10.1177/0583102406061499. using vibration test. Smart Structures and Materials 1999: Smart Systems for
[161] Song G, Li H, Gajic B, Zhou W, Chen P, Gu H. Wind turbine blade health mon- Bridges, Structures, and Highways. Proc. SPIE 1999:316–27.
itoring with piezoceramic-based wireless sensor network. Int J Smart Nano [189] Choi MY, Kwon IB. Damage detection system of a real steel truss bridge by
Mater 2013;4:150–66. http://dx.doi.org/10.1080/19475411.2013.836577. neural networks. Smart Structures and Materials 2000: Smart Systems for
[162] Joyce BS, Farmer J, Inman DJ. Electromagnetic energy harvester for mon- Bridges, Structures, and Highways. Proceedings SPIE, Newport Beach, Cali-
itoring wind turbine blades. Wind Energy 2014;17:869–76. http://dx.doi.org/ fornia: 2000:295–306.
10.1002/we. [190] Maseras-Gutierrez M, Staszewski W, Found M, Worden K. Detection of
[163] Carlson CP, Schlichting AD, Ouellette S, Farinholt K, Park G. Energy Harvesting impacts in composite materials using piezoceramic sensors and neural net-
to Power Sensing Hardware Onboard Wind Turbine Blade. In: Proceedings of works. Smart Structures and Materials 1999: Smart Structures and Integrated
the IMAC, Jacksonville, Florida USA 2011;28. Systems. Proceedings SPIE 1998:491–7.
[164] Godinez-Azcuaga VF, Inman DJ, Ziehl PH, Giurgiutiu V, Nanni A. Recent [191] Ruotolo R, Surace C. Diagnosis of damage in a steel frame. In: Proceedings of
advances in the development of a self-powered for wireless sensor network the 16th International Modal Analysis Conference (IMAC). Part 1, Santa Bar-
for structural health prognosis. Proc. SPIE - Int. Soc. Opt. Eng. 2011. http://dx. bara, CA, USA 1998:609–15.
doi.org/10.1117/12.880249. [192] Ruotolo R, Surace C. Damage assessment of multiple cracked beams:
[165] Isik F, Ozden G, Kuntalp M. Importance of data preprocessing for neural numerical results and experimental validation. J Sound Vib 1997;206:
networks modeling: the case of estimating the compaction parameters of 567–88. http://dx.doi.org/10.1006/jsvi.1997.1109.
soils. Energy Educ Sci Technol Part A: Energy Sci Res 2012;29:463–74. [193] Ruotolo R, Surace C, Mares C. Damage identification using simulated
[166] Asdrubali F, Baldinelli G, D’Alessandro F, Scrucca F. Life cycle assessment of annealing. In: Proceeding of International Modal Analysis Conference - IMAC,
electricity production from renewable energies: review and results harmo- SEM 1997;1:954–60.
nization. Renew Sustain Energy Rev 2015;42:1113–22. http://dx.doi.org/ [194] Mares C, Ruotolo R, Surace C. Using transmissibility data to assess structural
10.1016/j.rser.2014.10.082. damage. Key Eng Mater 1999;167:236–45.
 M. Martinez-Luengo et al. / Renewable and Sustainable Energy Reviews 64 (2016) 91–105 105

[195] Mares C, Surace C. An application of genetic algorithms to identify damage in [204] Surace C, Worden K. Novelty detection in a changing environment: a nega-
elastic structures. J Sound Vib 1996;195:195–215. http://dx.doi.org/10.1006/ tive selection approach. Mech Syst Signal Process 2010;24:1114–28.
jsvi.1996.0416. http://dx.doi.org/10.1016/j.ymssp.2009.09.009.
[196] Ruotolo and Surace. Damage detection using singular value decomposition. [205] Basseville M, Abdelghani M, Benveniste A. Subspace-based fault detection
In: Structural damage assessment using advanced signal processings proce- algorithms for vibration monitoring. Automatica 2000;36:101–9. http://dx.
dures, proceeding of DAMAS '97, University of Sheffield, UK 1997:87–96. doi.org/10.1016/S0005-1098(99)00093-X.
[197] Krawczuk M, Ostachowicz W, Kawiecki G. Detection of Delaminations in [206] Farrar CR, Worden K. Chapter 9: machine learning and statistical pattern
Cantilevered Beams Using Soft Computing Methods. European cost F3 con- recognition. In: Structural health monitoring: a machine learning perspective
ference on system identification and structural health monitoring, Madrid, 2013:295–320.
[207] Worden K, Pierce SG, Manson G, Philp WR, Staszewski WJ, Culshaw B.
Spain 2000:243–52.
Detection of defects in composite plates using lamb waves and novelty
[198] Vapnik V. Statistical Learning Theory. New York, USA: John Wiley; 1998.
detection. Int J Syst Sci 2000;31:1397–409.
[199] Worden K, Lane AJ. Damage identification using support vector machines.
[208] Mevel L, Benveniste A, Basseville M, Goursat M. In: Operation structural
Smart Mater Struct 2001;10:540–7.
damage detection and diagnosis. European cost F3 conference on system
[200] Farrar CR, Worden K. Chapter 10. Unsupervised learning - novelty detection. In:
identification and structural health monitoring, Madrid, Spain: 2000:641–50.
Structural health monitoring: a machine learning perspective 2013:321–60.
[209] Chang TYP, Chang CC, Xu YG. Updating structural parameters: an adaptive
[201] Worden K. Structural fault detection using a novelty measure. J Sound Vib
neural network approach. In: Structural health montoring 2000, Stanford
1997;201:85–101. http://dx.doi.org/10.1006/jsvi.1996.0747. University, Palo Alto, California 1999:379–89.
[202] Markou M, Singh S. Novelty detection: a review - Part 1: statistical approa- [210] Sanders W, Akhavan F, Watkins S, Chandrashekhara K. Fiber optic vibration
ches. Signal Process 2003;83:2481–97. http://dx.doi.org/10.1016/j. sensing and neural networks methods for prediction of composite beam
sigpro.2003.07.018. delamination. Smart Struct Integr Syst Proc SPIE 1997:858–67.
[203] Markou M, Singh S. Novelty detection: a review - part 2: neural network [211] European Union Committee. The EU's target for renewable energy: 20% by
based approaches. Signal Process 2003:1–26. http://dx.doi.org/10.1016/j. 2020. Published by the Authority of the House of Lords 2008.
sigpro.2003.07.019.