Process Safety and Environmental Protection 171 (2023) 71–86

Contents lists available at ScienceDirect

Process Safety and Environmental Protection

journal homepage: www.journals.elsevier.com/process-safety-and-environmental-protection

Advances in corrosion growth modeling for oil and gas pipelines: A review
Haonan Ma a, Weidong Zhang a, Yao Wang b, Yibo Ai a, *, Wenyue Zheng a, *
a
National Center for Materials Service Safety, University of Science and Technology Beijing, Beijing 100083, China
b
South China Branch of National Petroleum and Natural Gas Pipe Network Group, Guangzhou 510620, China

A R T I C L E I N F O A B S T R A C T

Keywords: To quantify the progress of corrosion damage and develop pipeline integrity management strategies, it is
Corrosion necessary to establish a reliable corrosion growth model. Due to the complexity of the corrosion process, the
Probabilistic models availability of data, and the limitations of various models in their applicability, there is currently no generally
Machine learning
accepted optimal corrosion growth prediction methodology. Corrosion data used for modeling, in-line inspection
Hybrid approach models
Oil and gas pipeline
techniques for detecting defects, and sources of uncertainty in the modeling process are briefly described. This
Corrosion rate paper focuses on reviewing the concepts, the performance, and the application of existing pipeline corrosion
growth models. The deterministic and probabilistic models are analyzed in detail according to the core methods
involved, and the latest applications of machine learning and deep learning in corrosion growth modeling are
also introduced. To leverage the strengths of various models, this paper presents hybrid approach models based
on the combinations of the aforementioned models, which have greater performance and interpretability than
single models and should be given more attention in the future development of corrosion growth prediction.
Finally, some suggestions for future development are put forward in light of the challenges and deficiencies
present in the current modeling process.

1. Introduction strategies, as it provides scientific guidance for reliability analysis and

determination of inspection intervals as well as prioritization of repairs.
Owing to their advantages of safety, transmission efficiency, and low An overly conservative corrosion growth model can result in excessively
cost, pipelines are the primary means for the long-distance trans­ frequent inspections and even unnecessary excavations and repairs,
portation and distribution of oil and gas (Li et al., 2021a). However, as increasing the costs to operators, while an underestimation of the
the pipeline ages, corrosion defects will inevitably develop on the in­ corrosion growth may leave critical defects unmitigated, which can later
ternal and external surfaces. Corrosion is one of the major causes of grow to failure, resulting in serious accidents. Therefore, in the imple­
pipeline failure, seriously affecting pipelines operation (Bhandari et al., mentation of pipeline integrity management strategies, it is imperative
2015). According to 11th Report of the European Gas Pipeline Incident to establish an advanced and accurate corrosion growth model.
Data Group ((EGIG), 2020), 26.63% of gas pipeline incident were caused The corrosion model is also, more importantly, a way of under­
by corrosion, as shown in Fig. 1. Accidents caused by corrosion will not standing all the accumulated information and a way of making the in­
only cause serious economic losses, but also have catastrophic conse­ formation fit together (Nešić, 2007). Specifically, a perfect corrosion
quences for human safety and the environment (Keshtegar et al., 2019; growth model should have the following characteristics: 1) the ability to
Ma et al., 2022). For example, in 2013, a corrosion leak and subsequent represent data diversity and accuracy; 2) a comprehensive physical
explosion at a gas pipeline in Qingdao, China, killed 62 people and basis, which can greatly reduce the uncertainty of the model; 3)
caused a loss of 118 million USD (Li et al., 2020). Although corrosion is robustness and computational validity (Guo et al., 2020). As a result, the
inevitable, it can be significantly reduced through effective safety and main objective of this study is to review and categorize existing
integrity management (Amaya-Gómez et al., 2019). knowledge on corrosion growth modeling, highlight the strengths and
The severity of the corrosion damage can be assessed by the corro­ limitations of existing methods, and provide suggestions for future
sion growth model. The corrosion growth model is critical for future directions.
pipeline condition assessment and pipeline integrity management Modeling and predicting the time evolution of corrosion growth is a

* Corresponding authors.
E-mail addresses: [email protected] (Y. Ai), [email protected] (W. Zheng).

https://doi.org/10.1016/j.psep.2022.12.054
Received 12 October 2022; Received in revised form 23 November 2022; Accepted 18 December 2022
Available online 27 December 2022
0957-5820/© 2022 Published by Elsevier Ltd on behalf of Institution of Chemical Engineers.
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

are built from datasets and can learn the complex relationship between
influencing factors and corrosion rate or depth to solve nonlinear
problems in the corrosion process. In recent years, methodologies like as
probability theory, stochastic processes, machine learning, and deep
learning have evolved into an unceasing stream of research on corrosion
growth models. However, no review exists in the open literature that can
systematically summarize the ideas, benefits and drawbacks, and ap­
plications of the above modeling methods. More critically, the direction
of corrosion growth modeling in the future needs to debated. Thus, this
paper emphasizes on addressing the following questions: (i) What data
sources are used in corrosion growth modeling, and what uncertainties
exist in the modeling process? (ii) What deterministic methods for
calculating corrosion growth are available in industry? How to use
Fig. 1. Distribution of gas pipeline incidents from 2010 till 2019.
probabilistic models based on stochastic processes to address corrosion
growth issues? (iii) What are the applications of artificial intelligence
complex task. A significant challenge is that the growth of corrosion algorithms using machine learning and deep learning in corrosion
defects is stochastic, with both temporal and spatial variability. In the growth modeling (iv) By summarizing existing methods and knowledge,
time domain, the corrosion process is considered to be a dynamic and what are the future trends for corrosion growth prediction?
stable evolutionary process containing three stages: nucleation, exten­ The rest of this paper is organized as follows: Section 2 presents the
sion, and repassivation. In the spatial domain, due to the large length of corrosion data, In-line inspection technologies, and sources of uncer­
oil and gas pipelines, the environment along the pipeline may vary tainty in the modeling process; Section 3 introduces different types of
greatly, causing defects to grow at different rates in different parts of the empirical models, which are divided into deterministic models and
pipeline. In addition, corrosion damage can take numerous forms, probabilistic models; Section 4 describes the concepts and applications
including uniform corrosion, pitting corrosion, erosion corrosion, stress- of different data-driven models; Section 5 proposes three types of hybrid
corrosion, stray-current corrosion, and microbiologically influenced approach models, which can be a future research direction for corrosion
corrosion (MIC) (Enning and Garrelfs, 2014; Spark et al., 2020) (Fig. 2), growth modeling; Section 6 discusses some existing challenges and de­
and there are multiple sources of uncertainty in the process of corrosion ficiencies in the research and gives some suggestions for dealing with
deterioration and detection. All of these will increase the complexity of them in future work; Finally, the conclusion of this paper are given in
corrosion modeling and affect the accuracy of corrosion predictions. Section 7.
The corrosion growth models of oil and gas pipelines can be divided
into three types, mechanism-based models, empirical models, and data- 2. Preparation for modeling
driven models. Mechanism-based models are derived on the basis of the
mechanism of corrosion reactions, most of which use electrochemical Corrosion is a deterioration process caused by electrochemical or
models to describe initiation and growth of corrosion pit. Mechanism- chemical reactions between a metal and its environment (Kim et al.,
based models have been reviewed in depth by Papavinasam et al. 2021). The pipeline steel used for long-distance transmission of oil and
(2006) and Nešić (2007) and the results predicted by mechanism-based gas is typically API (American Petroleum Institute) 5 L X-grade series
models are too conservative, so mechanism-based models will not be steel with low-carbon microalloy chemical compositions. When the soil
discussed further in this paper. Empirical models are obtained by using is considered as the corrosion environment, the environmental variables
probability theory or stochastic processes, and utilize prior knowledge of that affect corrosion include temperature, humidity, pH, soil type, ionic
known degradation mechanisms to a certain extent. Data-driven models concentration, organic content, and etc. When the conveyed medium is

Fig. 2. Different forms of corrosion on oil and gas pipelines: (a) uniform corrosion, (b) pitting corrosion, (c) erosion-corrosion, (d) stray current corrosion, (e) MIC.

72
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

considered as the external environment, the environmental variables process of corrosion, but the amount of data available is limited due to
that affect corrosion include temperature, pressure, flow rate, ionic the high cost of collection. Experimental and simulation data are simple
content, CO2 and H2S content, and so on. In order to slow down pipeline to gather, but owing to the simplification of the corrosion process, they
corrosion, some anti-corrosion measures such as coating and cathodic have certain deviations and can be used as a supplement to the field data
protection are also used. Cathodic protection can lower the electro­ when necessary.
chemical potential of the corrosion process and thus reducing the anodic
dissolution of the steel. Therefore, in an operating pipeline that have 2.2. In-line inspection technologies
corrosion-protection measures involved in the system, information and
data of all these key parameters should be prepared prior to modeling. ILI is performed by special high-accuracy detection tools, which uses
one of the non-destructive testing methods to detect, locate, and size
2.1. Pipeline corrosion data defects as well as to measure the geometry of the pipeline and its specific
parameters (Timashev and Bushinskaya, 2015). Technological ad­
An excellent model requires a large amount of high-quality data. In vancements in this field have led to the availability of multiple ILI
the modeling process, more than 80% of the time can be spent on data- technologies for different kinds of pipe sizes, product types, internal
related tasks such as data acquisition, data transmission and storage, and constraints, and pipeline defect forms that can more accurately classify
data preprocessing (Zhang et al., 2021). Adequate, well-distributed, and defect types and predict defect sizes. Such instruments are magnetic flux
complete dataset is the key to successful modeling. There are three main leakage (MFL), electromagnetic acoustic transducers (EMAT), ultra­
sources for corrosion data of oil and gas pipelines: sound tools (UT), eddy current testing (ET), and calipers, as shown in
1) Field data: the data is collected by measuring the in-service Fig. 3.
pipeline and its surrounding environment. For example, In-line inspec­ Timashev and Anna (2016) introduced in detail the operating prin­
tion (ILI) data (Desjardins, 2001; Worthingham et al., 2000), Alternating ciple, characteristics, the advantages and disadvantages, and applica­
Current Voltage Gradient survey, Close Interval Potential Survey, soil bility of the above ILI tools. Due to discrepancies in magnetic strengths,
data, pipeline materials data and operational parameters data (El-Ab­ defect size calculation methods, and defect size models between
basy et al., 2016). Due to the use of a large number of different types of different ILI tools, the data measured by different tools is difficult to
sensors, the data generated has the characteristics of large volume and compare directly(Al-Amin et al., 2014). The significant advantage of ILI
diversity. is that it can provide information on the size and growth rate of the
2) Experimental data: by simulating the same environment as the entire defect population to understand the corrosion mechanism and
actual pipeline, the data is measured in the laboratory by the weight-loss what is happening along the entire pipeline. At the same time, the ILI
or immersion test method. For example, researchers have carried out tools have several limitations, including (1) the probability of detection
corrosion studies on steel (X52 (Cunha Lins et al., 2012; Li and Casta­ of defects, (2) the probability of correct identification of detected de­
neda, 2014), X60 (Quej-Ake et al., 2018), X65 (Qin et al., 2022), X70 fects, (3) the accuracy of measurement of defects, and (4) the minimum
(Barbalat et al., 2012) and X80 (Arzaghi et al., 2020; Wang et al., detectable defect size. These limitations depend largely on the tech­
2015d)) widely used in underground pipelines under different envi­ nology employed (McNealy et al., 2010; Wang et al., 2015c; Xie and
ronmental conditions. Tian, 2018).
3) Simulation data: the data is generated through software simula­
tion or the mathematical models. Layouni et al. (2017) used MATLAB to 2.3. Sources of uncertainty
simulate the ILI signals of defects. Masi et al. (2015) used OLGA software
to provide information on temperature profiles, pressure profiles, and There are many sources of uncertainty in modeling corrosion growth.
velocity profiles along the pipeline. It is critical to identify them in order to reduce these uncertainties, so
By collecting data from relevant sources, corrosion datasets are that accurate predictions of corrosion depth or rate can be achieved.
established to provide support for subsequent modeling and prediction These sources of uncertainty are mainly divided into aleatory uncer­
(Yang et al., 2022). Field data can accurately reflect the deterioration tainty and epistemic uncertainty.

Fig. 3. The detection probes of the main ILI tools: (a) MFL, (b) EMAT, (c) UT, (d) Caliper.

73
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

• Aleatory uncertainty stems from the inherent variability of corrosion underground corrosion tests of bare steel pipe coupons without
(Melo et al., 2019), such as uncertainty in basic variables (initial cathodic protection in a variety of soils (2010).
state estimation, material properties, etc.). This uncertainty cannot • ASME B 31.8 S: It contains guidance on estimating corrosion rates
be eliminated but can be represented by probability (Kiureghian and based on soil resistivity measurements and site observations, which
Ditlevsen, 2009). indicates that the maximum corrosion rate of 0.3 mm/year may be
• Epistemic uncertainty stems from a lack of knowledge or data. This appropriate for worst-case external corrosion conditions (2010).
uncertainty can be reduced by improving data collection methods • GRI-00/0230: This report provides a typical worst-case external
and increasing the amount of available data for analysis, and is corrosion rate. The pitting corrosion rate is 0.56 mm/year, and the
represented by possibility theory, interval analysis, evidence theory, general corrosion rate is 0.3 mm/year (Dawson and Ganim, 2016;
and probability theory (Helton et al., 2010). Leis and Bubenik, 2001).

Since epistemic uncertainty can be lowered, it should be given more 3.1.2. Linear corrosion rate model
attention in modeling. Epistemic uncertainty can be classified as model The linear corrosion rate model is a deterministic model that assumes
uncertainty and data uncertainty (Guo et al., 2020). Model uncertainty that the corrosion rate grows linearly with time (Li et al., 2022b). Its
includes parameter uncertainty, model form uncertainty, and solution form is relatively simple, as shown in Eq. (1).
approximation error (Mullins et al., 2016). Data uncertainty includes
D(t) = D0 + Ḋt (1)
into measurement uncertainty and data sparsity. Model uncertainty and
data uncertainties are coupled because data is also used to determine
Where, D(t)is the corrosion depth over time, t is the time interval, D0 is
model parameters or change the form of the model. In order to better
the corrosion depth at the initial time, and Ḋ is the corrosion rate.
describe epistemic uncertainty, we here provide an illustrating example:
When there is only one set of ILI data, the initial time can be regarded
Assuming that the corrosion growth model is represented by f(X1 ,…,
as zero or the moment of third-party interference or coating peeling off
Xi ,…,Xn ,), the predicted corrosion depth D can be obtained by D = f(X1 ,
(Vanaei et al., 2017). When a pipeline has at least two ILI data collected
…, Xi , …, Xn , ). Where, the random variable X1 can be pH, temperature
at different times, the corrosion rate can be calculated with Eq. (2).
and organic content in the soil, or flow rate, pressure and H2S in the
medium, or different kinds of coatings applied. By statistically analyzing dT2 − dT1
Ḋ = (2)
the observed data Xi = (x1 , x2 , …, xm ) of any random variable Xi , the t2 − t1
distribution gi (x) that the random variable Xi follows can be obtained.
The distribution is expressed as the parameter uncertainty. Therefore, Where, dT1 and dT2 are the defect depths of the first and second in­
parameter uncertainty can be affected by the quality and quantity of the spections, and t2 − t1 is the time interval between the two inspections.
datasets. The quality of the dataset refers to the accuracy of the data. The
quantity of datasets refers to the size of the available samples. The model 3.1.3. Power function corrosion rate model
form uncertainty stems from the form f(X1 , …, Xi , …, Xn ) of corrosion The power function corrosion rate model is one of the most suc­
growth model selected. This uncertainty is mainly caused by the cessful empirical models, which is used in various forms of corrosion,
simplification of the complex corrosion process, which can be assumed such as uniform corrosion, pitting corrosion, corrosion fatigue, and even
that the corrosion growth process is a linear growth process or other fracture toughness evaluation (Velázquez et al., 2017). A common
form of math functions. Solution approximation error stems from the application of this power function in corrosion shown in the “Standard
approximate solution of the model by different solvers. Measurement Guide for Examination and Evaluation of Pitting Corrosion” published
uncertainty mainly includes measurement error, probability of detec­ by ASTM is that it guides engineers to use inspection data to predict the
tion, and reportability error. Therefore, in order to improve the accuracy remaining life of metal structures subject to pitting corrosion (2005).
of prediction results, it is very important to choose an appropriate Romanoff (1957) first applied power functions to calculate the
modeling method and improve the quantity and quality of datasets. maximum pitting corrosion depth D over time t, as illustrated in Eq.(3):
D = ktn (3)
3. Prediction method based on empirical models
This model can also be converted into an approximate linear formula
In corrosion growth models, there are two types of empirical models: by taking logarithm, as shown in Eq. (4).
deterministic models and probabilistic models. Deterministic models log D = log k + n log t (4)
provide a specific corrosion rate for each corrosion defect under given
conditions, while probabilistic models take into account the temporal Taking into account the initial time t0 of corrosion (Li and Mah­
and spatial variability of the corrosion process and produce a probability moodian, 2013), the corrosion depth can be expressed by Eq. (5):
distribution of corrosion depth (Peng et al., 2020). The concept and D = k(t − t0 )n (5)
application of each method are summarized below.
Where, kand n are parameters related to the material-environment
system and can be fitted by statistical methods using field or labora­
3.1. Deterministic models tory corrosion data. The value of n is less than 1. As long as the above
parameters are known, the corrosion depth at any given location at the
3.1.1. Single-value corrosion rate model future time can be calculated.
The single-value corrosion rate (SVCR) model is one of the most
widely used deterministic models in the industry. The model uses a
constant corrosion rate with a conservative value, and the corrosion rate 3.2. Probabilistic models
is independent of the age and depth of corrosion defects. Some general
standards are as follows: It is well known that corrosion is a dynamic and complex destructive
phenomenon (Ahammed and Melchers, 1997), so measurable parame­
• NACE SP0502: When any other guiding data is lacking, it is recom­ ters such as corrosion rate and maximum corrosion depth exhibits a
mended to use the pitting rate of 0.4 mm/year to determine re- large dispersion. However, deterministic models cannot take into ac­
inspection intervals. This rate represents the upper 80% confidence count the effects of measurement errors and the randomness of corrosion
level of maximum pitting rates for long-term (up to 17-year duration) (Foorginezhad et al., 2021) and can only give point estimates at specific

74
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

times. In this case, probabilistic models based on stochastic processes are independent, reflecting the heterogeneity of the defects between the
more useful. Here are some of the more common probabilistic models for locations, while the shape parameter can address the temporal trend of
oil and gas pipelines: the average of a random variable that follows the Gamma process. By
designing the appropriate shape parameter, explanatory variables and
3.2.1. Generalized extreme value distribution model random effects can be incorporated into the model (Lawless and
The objective of extreme value statistics theory is to establish a Crowder, 2004).
probability model for the observed extreme value based on certain As an extension of the Gamma process model, Pandey et al. (2005)
sample sets. AZIZ (1956) was the first to apply extreme value statistics to considered the dependence between degradation and time. Zhou et al.
pitting problems and discussed some of the limitations of extreme value (2012) adopted the gamma distributed time-independent growth model
statistics. ELDREDGE (1957) was among the first to use extreme value and the gamma process-based growth model to predict the pipeline
statistics to obtain maximum pit depths on oil well pipes. Extreme value corrosion depth growth, and used the copula function to characterize the
statistics theory has now become the common method for predicting spatial correlation between the depth growth of different defects. Guida
local corrosion depth (Yarveisy et al., 2022). The core result of extreme and Pulcini (2013) proposed a state-dependent inverse Gamma process
value statistics is that the natural model of the maxima is the generalized for modeling pitting growth in nonlinear trends. This model was
extreme value distribution (GEVD) (Fougeres et al., 2006). The corro­ mathematically more tractable because it did not require discretization
sion rate distribution of the pipeline is described by GEVD as Eq. (6): of time and state. Fan et al. (2015) considered the dependence of the
{ [ (x − μ)− 1/ξ ]} degradation increment at the next moment on the current degradation
G(x) = exp − 1 + ξ (6) state. Xiang and Zhou (2018) used a uniform gamma process to char­
σ
acterize the growth of defect depth, taking into account the bias and
random errors associated with the ILI tool.
Where σ > 0, μ, ξ ∈ ℝ. The parametersμ, σ, ξ are the location, scale, and
shape factors.
3.2.3. Inverse Gaussian process-based corrosion growth model
The probability density function ft (x) of the corrosion depth at time t
Similar to the Gamma process, the Inverse Gaussian process differs
can be calculated from the corrosion rate distribution and the depth
only in the distribution of jump sizes. In this model, the increment of the
distribution at the initial time t0 using a convolution integral as Eq. (7):
corrosion growth is subjected to the Inverse Gaussian (IG) distribution,
∫∞ ( )
which means that it is nonnegative. {x(t), t ≥ 0} is considered as a
ft (x) = G(υ)ft0 x − υδt dυ (7)
0 random variable with an Inverse Gaussian distribution over time t. The
probabilistic density function can be given by
Where δt = t − t0 is the time interval. √̅̅̅̅̅ ( )
GEVD can express three distributions (Gumbel, Frechet, and Wei­ θ θ(t)(x(t) − μ(t))2
fIG (x(t)|μ(t), θ(t)) = x(t)− 3/2 exp − I(0,∞) (x(t)) (9)
bull) in a unified form (Eq. (6)). For relatively short exposure periods, 2π 2μ(t)2 x(t)
the Gumbel and Weibull distributions are the most likely to be suitable
for corrosion depth and rate (Bhandari et al., 2017). In the literature on where, I(0,∞) (x(t)) is an indicator function, μ(t) is the mean parameter (i.
pitting corrosion, the two-parameter Gumbel (Marsh and Taylor, 1988; e.,E[x(t)] = μ(t)) and θ(t) is the shape parameter. μ(t) is a monotonically
Strutt et al., 1985) and Weibull (Khan et al., 2021; Ossai et al., 2015) increasing function. Both parameters vary with time (Peng et al., 2020),
distributions are commonly used. This is because the mathematical and the relationship between them is shown in Eq. (10).
complexity of the three-parameter distribution is high, and in most
cases, two-parameter distributions can well fit the corrosion data. For θ(t) = ξ⋅μ(t)2 (10)
long-term exposure periods, considering the long-term stabilization of Due to the analytical tractability and physical interpretability of the
the diffusion-controlled pit growth process, the Frechet distribution may IG process, it is widely used. Zhang et al. (2013) established the IG
be the best fit. Caleyo et al. (2009b) used these three methods to fit the Process-based growth model of underground pipeline corrosion based
maximum pit depth data for buried pipelines respectively, and found on the ILI data affected by measurement uncertainties, using Markov
that the Frechet distribution was the best fit for the corrosion rate data Chain Monte Carlo to evaluate the posterior distribution of the model
and the Weibull distribution did not appropriately fit the heavy upper parameters. Zhang (2014) developed an IG process-based model for the
tail of the data. depth growth of corrosion defects in underground pipelines on a hier­
archical Bayesian framework incorporating measurement
3.2.2. Gamma process-based corrosion growth model uncertainty-influenced ILI data. Zhang et al. (2017) combined the in­
The corrosion process can be thought of as the accumulation of verse Gaussian process with the state space model to describe the
irreversible damage caused by a series of shocks, each arriving randomly corrosion process of oil and gas pipelines and predict their remaining
in time and causing a random amount of damage. Therefore, this phe­ life, of which the model parameters were estimated by combining the
nomenon can be modeled by a compound Poisson process defined by the experience maximization algorithm and the particle filter algorithm.
arrival of jumps described by a homogeneous Poisson process with the Zhou et al. (2017b) used Gamma and IG processes to model defect
arrival rate and random variables representing the jump size. The growth and characterized the dependence between different defect
gamma process can be viewed as the limit of a compound Poisson pro­ growths using the sum of Gaussian copula and stochastic processes. Guo
cess, with arrival rates and jump sizes approaching infinity and zero, et al. (2018) proposed an improved IG process that takes into account
respectively (Ye and Chen, 2014). From mathematics, the Gamma pro­ the dependencies between degradation increments and previous
cess is a continuous probability distribution defined by using only two degradation states. Wang et al. (2020) used the binary IG process to
parameters: the scale parameter and the shape parameter. The proba­ establish two performance degradation models of pipeline corrosion
bility density function of the Gamma process (Joshi, 2011; van Noort­ depth and residual strength.
wijk et al., 2007) can be illustrated with Eq. (8):
3.2.4. Markov process-based corrosion growth model
x(t)κ(t)− 1 e− x(t)/θ
f (x(t)|κ(t), θ) = κ(t)
∫∞ I(0,∞) (x(t)) (8) In this model, the pipeline corrosion growth process is regarded as a
θ 0
tκ(t)− 1 e− t dt Markov process, which means that the future state of the pipeline de­
pends only on the current state and has nothing to do with the previous
Where, κ(t) is the shape parameter, θ is the scale parameter, and x(t) is
states (Shekari et al., 2016). It is assumed that the thickness of the
the corrosion depth over time. The scale parameter is constant and time-
pipeline wall is divided intoN discrete states, and each discrete state can

75
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

be regarded as a Markov state. The maximum corrosion depth at any the time at which pm (t0 ) was observed or the time of the initial inspec­
point at time t is represented by a discrete random variable D(t), and the tion. The value of v varies with the type of soil structure and depends on
probability that it is at i state is expressed as P{D(t) = i} = pi (t),i = 1,2, characteristics such as pipe-soil potential, soil water content, soil bulk
…,N. λi (t)represents the rate of transition probability, so the probability density, and pipe coating type.
of transition that pitting depth at the i states promotes to the next state in Five kinds of corrosion growth models (NACE, Linear Growth, Time
unit time interval δt can be written asλi (t)δt + ο(δt)(Caleyo et al., Independent-GEVD, Time Dependent-GEVD, and Markov) were used to
2009a), which is affected by environmental conditions around the predict pitting corrosion depth distribution in the dataset selecting two
pipeline (Ossai et al., 2016). Thus, the growth of corrosion can be rep­ consecutive MFLs in the same pipeline (Valor et al., 2013b), as shown in
resented by the transition from state i to state j, where j ≥ i + 1. One of Fig. 4. The result could be concluded that compared with deterministic
the main problems in this model is to calculate the transition probabil­ models and other probabilistic models, the pitting depth distribution
ity. In the continuous-time, non-homogenous linear growth Markov after ten years, predicted by the Markov model, was more in line with
process, the transition probability satisfies the forward Kolmogorov the real results (Pipe85–2006).
equation, as shown in Eq. (11).
{ 3.2.5. Brownian motion Process-based corrosion growth model
dpi,j − λj (t)pi,j (t) + λj− 1 (t)pi,j− 1 (t), j ≥ i + 1
= (11) The widely used form of the Brownian motion process {W(t), t ≥ 0}
dt − λj (t)pi,j (t)
to characterize degradation process {x(t), t ≥ 0}is:
Where, pi,j is the transition probability from state i to state j, which is the x(t) = βt + σ W(t) (14)
element of the Markov process transition matrix.
Provan and Rodriguez (1989) first used the nonhomogeneous Mar­ Where, σ is the diffusion parameter, and β is the drift parameter that a
kov process to establish pit depth growth model. They proposed a positive non-decreasing deterministic function over time (Wang, 2010).
non-homogeneous transition probability rate relationship shown in Eq. Eq. (14) indicates that the mean value of the degradation trajectory is
(12), and used the experimental data of two different materials, stainless positive and increases monotonically over time, but it does not guar­
steel (Provan and Rodriguez, 1989) and aluminum (AZIZ, 1956)to fit antee that the specific realization of the degradation trajectory is posi­
pitting corrosion parameters λ and k. They applied an iterative method tive and monotonically increasing. Thus, this model cannot correctly
to calculate the transition probability (Caleyo et al., 2009a; Peng et al., characterize the positive and monotonically increasing nature of
2020). corrosion growth (Park and Padgett, 2005).
Elsayed and Liao (2004) developed the geometric Brownian motion
1 + λt
λj = jλ (12) process-based growth rate model, which can overcome perfectly the
1 + λtk
underlying drawbacks of the models, as follows:
Caleyo et al. (2009a) proposed a continuous-time, non-homogenous
ln(r(t)) = ln(r0 ) + βt + σ W(t) (15)
linear growth (pure birth) Markov process to model the external pitting
corrosion in underground pipelines. The binomial closed form solution
where r(t) represents degradation rate and r0 represents degradation
of the forward Kolmogorov equation is used to represent the transition
rate at the initial time. It follows thatln(r(t2 )) = ln(r(t1 )) +βt +σ W(Δt)
probability between pit depth (Markov) states in a time interval. The key
with Δt = t2 − t1 and t2 > t1 . Therefore, the degradation rate at the
to this model is to assume that the stochastic mean of the pitting
present time is dependent on the current state of the degradation.
corrosion derived from the Markov process equals the deterministic
Based on Elsayed’s method, Zhang and Zhou (2015) provided a
mean of the process, which allows the identification of the transition
state-dependent growth model. The model used a piecewise linear path
probability function only from the pitting initial time and the pitting
to approximate the actual growth path of corrosion defects and char­
exponent, as shown in Eq. (13). This model can predict the distribution
acterized the average growth rate between two consecutive inspections
of pitting corrosion depth and rate based on the observed soil properties
as a geometric Brownian motion process. Compared with the Markov
and pipeline coating types.
model, this method does not require discrete damage states or assess­
( )( )vm [ ( )v ]n− m
∑n
n− 1 t0 − tini t0 − tini ment of the possibility of transition. Compared with the model based on
pn (t) = pm (t0 ) 1− (13) the gamma process or the inverse Gaussian process, this model does not
m=1
n− m t − tini t − tini
involve specific assumptions about the mean growth path. But the
Where, tini and ν are parameters that describe the nonlinear time evolu­ disadvantage of this model is that it requires multiple sets of ILI data
tion of the corrosion depth in a buried pipeline, while t0 corresponds to with short inspection intervals.

Fig. 4. The pitting depth distribution predicted by five corrosion growth models (Valor et al., 2013b).

76
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

The Brownian bridge process, a conditional Brownian motion pro­ determining the optimal hyper-parameters is the process of solving the
cess, was proposed to model corrosion growth by Wang et al. (2015b). optimization problem. Heuristic algorithms are generally used to fit the
The geometric Brownian bridge process incorporates the advantages of hyper-parameters in the model. Common heuristic algorithms include
the geometric Brownian motion process and considers the passive states Particle Swarm Optimization (PSO), Firefly Algorithms, and Genetic
that metallic structures may undergo. The expression of this Brownian Algorithms (GA) (El Amine Ben Seghier et al., 2020). In evaluations, the
bridge process is shown in Eq. (16). performance of the established models is examined through perfor­
t − t0 ( ) ( ) mance indicators (Rachman et al., 2021). With the advancement of
ln(x(t)) = βt + W(t) − W tf , t ∈ t0 , tf (16) computer technology, a variety of machine learning models have been
tf − t0
developed to predict pipeline corrosion growth. Due to limited space,
Where, tf is the repassivation time and t0 is the initial time. ln(x(t))is a this paper only introduces Bayesian networks (BN) and artificial neural
geometric Brownian bridge process with given initial time and specified networks (ANN).
time.
This model takes into account that the evolution of localized corro­ 4.1.1. Bayesian network
sion is time-dependent, following the passive film breakdown stage, Bayesian network, also known as belief network or Bayesian belief
metastable pit steady state, and pit growth stage mechanism. Fig. 5 networks, is a classical probabilistic graphical model (Agarwal et al.,
shows the simulation of external corrosion rate and corrosion depth over 2021; Li et al., 2022d). It uses a directed acyclic graph to describe the
time in this model. Although Wang’s work used external corrosion data dependencies between attributes and a conditional probability table to
in his analysis, his method Eq. (16) can be applicable to internal describe the joint probability distribution of attributes (Li et al., 2022c).
corrosion processes. Bayesian network consists of two parts: network structure and param­
eters. The network structure is a directed acyclic graph, where each node
4. Prediction method based on data driven models corresponds to an attribute. If there is a direct dependency between two
attributes, they are connected by an edge, where parameters are used to
In the industrial 4.0 framework, multiple systems of oil and gas quantitatively describe this dependency (Kamil et al., 2021). Bayesian
pipelines are gradually being digitalized in operation monitoring, risk network is one of the most effective theoretical models for solving
management, maintenance, and inspection, which promotes the rapid complex uncertainty (Yeo et al., 2016) and correlation problems, and is
development and application of data-driven technology in pipeline widely used in quantifying the corrosion depth (Balekelayi and Tesfa­
corrosion damage prediction (Lee et al., 2019). Data-driven models are mariam, 2020; Zhang and Zhou, 2014) and failure probability (Guan
capable of learning linear or nonlinear relationships between influ­ et al., 2016; Tesfamariam et al., 2021) of oil and gas pipelines.
encing factors and corrosion rate/depth from rich datasets and have BN is able to integrate various types of available information
achieved some excellent outcomes in predicting corrosion growth. The (analytical models, expert knowledge, published literature, and histor­
prediction methods based on data driven models are divided into two ical data) into a flexible framework, then build causal networks and
categories: machine learning and deep learning. graphically represent them for better understanding and solving prob­
lems (Chen and Pollino, 2012; Cockburn and Tesfamariam, 2012), as
shown in Fig. 6. Koch et al. (2015) combined a BN model with geospatial
4.1. Machine learning
information to predict corrosion depth and leakage probability at spe­
cific locations in carbon steel oil pipelines. Jain et al. (2015) proposed a
Machine learning has been used in numerous data-intensive in­
BN model to quantitatively evaluate the external corrosion rate of oil
dustries (such as finance, engineering, and medicine) due to its appli­
and gas pipelines located in eastern China by using ILI data. Shabarchin
cability and computational convenience in solving complex and difficult
and Tesfamariam (2016) integrated general corrosion, pitting &
problems (Sambo et al., 2018). It relies on efficient artificial intelligence
erosion-corrosion, MIC, and failure pressure models into a flexible
algorithms, rich datasets, and superior computational performance to
network and used the BN model to estimate failure probabilities of
extract information from raw data to build relationships between in­
corrosion defects and analyze the influence of fluid composition and
dependent and dependent variables for prediction and classification.
operating conditions on corrosion rate. Yang et al. (2017) presented an
The machine learning process consists of three important steps: data
object-oriented BN-based corrosion failure model to assess the condition
preprocessing, training, and evaluation (Zhou et al., 2017a). Since the
of pipelines. Li et al. (2021c) developed a BN model used with Fuzzy Set
original field data collected by sensors is incomplete, inconsistent, un­
Theory to capture uncertainties of data and model. Yazdi et al. (2021)
structured, and noisy, it needs to be cleaned, de-noised, converted, and
used continuous BN with hierarchical Bayesian analysis to obtain the
fused before training. In the training phase, to reduce the complexity of
accurate failure probability and MIC rate. Dawuda et al. (2021) pro­
the model and improve the prediction accuracy and convergence speed,
posed a BN model to estimate the MIC rate, which takes into account the
it is necessary to select an appropriate machine learning algorithm and
variability and interdependence of corrosion parameters.
to optimize the hyper-parameters of the model. The process of

Fig. 5. The simulation of external corrosion rate and corrosion depth over time (Wang et al., 2019).

77
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

Fig. 6. The pitting corrosion depth Bayesian network model (Demissie et al., 2016).

To add temporal dimension to BN models, Dynamic Bayesian Net­ (Xia et al., 2016).
works (DBN) have been proposed, which can represent the dynamic Common artificial neural networks include Back-Propagation (BP)
behavior of the system with temporal features (Li et al., 2021b). In DBN, neural networks, Feed-Forward neural networks, Probabilistic neural
not only connections within time slices but also connections between networks, and Radial Basis Function (RBF) neural networks. Liao et al.
time slices are allowed (Mihajlovic and Petkovic, 2001). The arc con­ (2012) used BP neural networks, GA and BP neural networks, and PSO
necting two variables belonging to different time slices indicates the and BP neural network to predict the internal corrosion rate respec­
time probability correlation between these variables (Weber and Jouffe, tively. The results showed that GA and BP neural networks have the least
2003). Arzaghi et al. (2018) used the DBN method to represent the absolute error. El-Abbasy et al. (2014) developed an ANN model to
temporal nature and various pitting corrosion rates in the predict the corrosion rate using geometrical characterization of the
corrosion-fatigue degradation processes of subsea pipelines. Xiang and pipeline, multiphase flow parameters, and deterministic models. The
Zhou (2019) integrated ILI measurement error characterization, growth results indicated that the ANN model outperforms current deterministic
model update, and reliability analysis into a single DBN model to predict methods in terms of prediction accuracy. Shaik et al. (2020) developed
the growth of corrosion depth. Yazdi et al. (2022b) proposed a an intelligent model based on a feed-forward BP neural network using
DBN-based method to evaluate uncertain and time-dependent charac­ historical fields detection data. The results showed that the accuracy of
teristics of subsea pipelines and verified the efficiency and effectiveness the feed-forward BP neural network is slightly better than of ANN, fuzzy
of the method. logic, and regression.
As ANNs are used to analyze various types of corrosion data (Abbas
4.1.2. Artificial neural networks et al., 2018), more and more researchers are trying to determine the
The artificial neural network is a parallel processing technology that relationship between different variables and corrosion growth. Table 1
is similar in structure to the human brain and operates by imitating the shows the ANN models built with different variables. In general, the
way a human brain works. It is a highly self-organizing, adaptive, and more input variables to consider, the better the model. Ben Seghier et al.
self-trained approximator with high associative memory and strong (2021) used eight different combinations of the input variables to esti­
nonlinearity, using nonlinear mapping to extract relations hidden in the mate the maximum pitting depth in oil and gas pipelines. The results of
data (Ren et al., 2012). A neural network consists of multiple neuron the study showed that the highest prediction accuracy of maximum
nodes, including the input layer, hidden layer, and output layer, as pitting depth was obtained when using the largest number of input
shown in Fig. 7. In the hidden layer, the input layer and the output layer variables.
are connected by neurons using a specific nonlinear function. By
training on sample data, neural networks can predict accurate solutions
4.2. Deep learning

Deep learning, as a rapidly developing branch of machine learning, is

essentially a deep neural network with multiple hidden layers. It differs
from the traditional neural network in that its network is more complex
and larger, and the prediction effect is significantly improved. The
typical methods are convolutional neural networks (CNN), recurrent
neural networks (RNN), and long-short-term memory (LSTM).
CNN is the most widely used method of deep learning. Its overall
architecture consists of the input layer, the convolutional layer (Recti­
fied Linear Units layer, pooling layer, and fully connected layer), and the
output layer, as shown in Fig. 8. The accuracy and robustness of the
model depend on the type of network layers, the depth of the network,
the arrangement of various types of layers in the network, the function
chosen for each layer, and the training data. Bastian et al. (2019) pro­
Fig. 7. The artificial neural network architecture. posed a computer vision-based method for the corrosion classification

78
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

Table1 attention in the field of oil and gas pipeline corrosion, due to the large
Summary of Artificial neural networks models used in relevant studies. amount of data required in the modeling process of deep learning, this
Author Method Input variables Output method is still in the exploratory stage in corrosion growth modeling.
variable

(Ren et al., BP neural Natural gas pipeline The 5. The concept of modeling with the hybrid approach
2012) network mileage, elevation maximum
difference, pipe inclination, average On the one hand, deterministic and probabilistic models in industrial
pressure, Reynolds number corrosion
applications are constantly improved. On the other hand, as big data
rate
(Chamkalani ANN pH, velocity, temperature, Corrosion
technologies and artificial intelligence algorithms advance, new models
et al., and partial pressure of the rate based on data-driven methods continue to arise (Yuan et al., 2020).
2013) CO2 Modeling exclusively using empirical models cannot accurately account
(Mazzella ANN Soil parameters, AC/DC Corrosion for the complexity and randomness of the corrosion process, and
et al., interference, cathodic rate
modeling exclusively using data-driven methods lacks good physical
2019) protection, geographic and
climatic factors significance explanations. It is difficult to combine the strengths of both
(Sosa et al., Grey relational Pipeline age, soil Corrosion and mitigate the weaknesses of each in single methods, so this paper
2020) analysis and ANN resistivity, soil pH, cathodic depth presents the concept of modeling with the hybrid approach incorpo­
protection potential, and
rating empirical models and data driven models.
reduction–oxidation
potential
The hybrid approach has achieved some applications in corrosion
(Ossai, 2020) Feed-Forward Temperature, operating Corrosion growth prediction, but no studies have systematically classified and
neural network pressure, partial pressure of depth summarized hybrid approaches used in corrosion growth modeling.
and PSO CO2, basic sediment and Therefore, considering the degradation process of corrosion and the
algorithm water, oil production rate,
application of the hybrid approach, this paper proposes three frame­
water production rate, gas
production rate and gas works of the hybrid approach, under which different algorithms can be
specific gravity logically combined. The hybrid approaches are divided into the hybrid
(Li et al., Kernel principal pH, temperature, moisture Corrosion approach incorporating multiple empirical models, the hybrid approach
2022a) component content, CO2 content, H2S rate incorporating multiple data-driven models, and the hybrid approach
analysis and content, Cl- content, Ca2+
Bayesian and Mg2+ concentration,
incorporating multiple hybrid models.
regularized ANN dissolved oxygen content,
methods pressure, and flow rate 5.1. The hybrid approach incorporating multiple empirical models

and localization of oil and gas pipeline. The study built a dataset of more The corrosion condition of pipelines varies greatly at different phases
than 140,000 images of pipelines with varying degrees of corrosion and of the corrosion process. In order to better describe the corrosion dam­
used CNN to classify pipeline images according to corrosion intensity. age process, different models should be used to represent different
CNN has the ability to train large-scale neural networks and solve big corrosion stages. In pitting corrosion, there are two main processes: pit
data problems in modeling (Matthaiou et al., 2021; Zhu et al., 2019). initiation and pit growth. In order to integrate them, Hong (1999)
Several common multi-layer CNN methods are AlexNet (Krizhevsky proposed a model that used the Poisson process to describe pit initiation
et al., 2017), GoogleNet (Szegedy et al., 2015), and VGG (Liu et al., and the Markov process to describe pit growth. This model combined the
2017a). Ajmi et al. (2020) used an approach based on the AlexNet ar­ homogeneous Poisson process and the homogeneous Markov process to
chitecture of a pretrained network to classify defects in welds. obtain the function of probability distribution of corrosion pit depth,
RNN is widely used in time series prediction modeling and has strong and used the so-called time transformation–condensation method to
performance advantages in computing and storing large-scale data. In non-homogenize this model. The shortcoming of this model is that the
the Prognostics Health Management (PHM) 08 Data Challenge, two- proposed probability distribution of the maximum pit depth does not
thirds of the winners used RNN (Saxena et al., 2008). As an extension conform to any of the extreme distribution functions.
of RNN, LSTM has recently come into view. In the PHM 18 Data Chal­ Therefore, combining the knowledge of extreme value statistical
lenge, two of the top three winners used LSTM (Vishnu et al., 2018). theory, Valor et al. (2007) developed a new method to link the pit
Ahuja et al. (2021) proposed a novel approach to resolve the problems of initiation and growth stages when considering multiple pits. Pit initia­
pitting corrosion diagnosing systems and used the Bidirectional tion was modeled as a non-homogeneous Poisson process in which the
Conv-LSTM method to analyze various images to provide better classi­ distribution of pit nucleation times was imitated using the exponential
fication results. Although deep learning has attracted extensive and the Weibull distributions, while pit growth was modeled as a
non-homogeneous Markov process (Valor et al., 2013a). The significant

Fig. 8. The convolutional neural network architecture.

79
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

variables in this model are the pitting initial time and the corrosion Table 2
depth as a function of exposure time to the corrosion environment. Summaries of ensemble learning utilized in relevant papers.
Compared with the models of AZIZ (1956) and Provan and Rodriguez Ensemble Base model Outcome Ref.
(1989), the model obtained by this method was more in line with the learning
actual evolution of maximum pitting corrosion depth with time. Based Voting Artificial neural The ensemble model (Masi et al., 2015)
on Valor’s method, Shekari et al. (2017) and Yazdi et al. (2022a) further networks combined with
improved the initiate time and pit density models, and demonstrated the artificial neural
practical application of the proposed methods. networks was used to
predict metal loss and
corrosion rates along
5.2. The hybrid approach incorporating multiple data-driven models pipelines. The results
show that the method
Ensemble learning is a popular strategy for incorporating multiple improves the prediction
performance and can
data-driven models, in which many base models are created and their
specify the mean and
outputs are meta-heuristically merged (Polikar, 2006). Common standard deviation of
ensemble learning includes voting, bagging, stacking, and boosting. the predicted values.
Voting is the simplest kind of ensemble learning. In prediction Gaussian process The mix GPR model can (Liu et al., 2017b)
problems, the outputs of each base model are combined and the class regression (GPR) handle non-linear
relationships between
with the most votes is chosen as the final decision. variables and
Stacking (Zeng et al., 2020) is a hierarchical construction of a uncertainties due to
multi-level training base in which the first level consists of several base unexplained factors,
models that are then output to a second level of stacked base models that and achieves the best
performance in
determines the final decision.
predicting corrosion
Bagging constructs the original training set by random sampling into rates compared to
N training subsets for different base models, and votes on the prediction Linear regression and
results of all base models to produce the final prediction results. This RBF.
method is suitable for solving the situation where the available training Linear The ensemble learning (Cai et al., 2020b)
regression; RF; model that combine RF,
data set is relatively small. The Random Forest (RF) algorithm is a SVM; ANN; GPR; ANN and SVM meta-
typical decision tree ensemble method using the bagging mechanism heuristically yields
(Polikar, 2012). higher prediction
In boosting (Ahmadianfar et al., 2022), weights are assigned to each accuracy and time-
efficiency than all
training tuple and updated after the base model is learned, so that the
standalone models.
subsequent base model pays more attention to the misclassified training Stacking LR; ANN; RF; Considering soil (Shi et al., 2017,
tuple. Popular algorithms are Gradient Boosting Decision Tree (GBDT), SVM; stacking properties, four single 2020)
Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting ensemble models and a stacking
(LGBoost). model were established
to predict the
The applications of the above methods are presented in Table 2. remaining wall
Ensemble learning can combine multiple base models, so that the overall thickness of the
result is better than the base models used alone. Chou et al. (2017) used pipeline. Compared to
single models and ensemble learning models composed of four machine the predictive
performance of single
learning algorithms (ANN, SVM, classification and regression tree, and
models, the superiority
linear regression) to predict the corrosion rate of carbon steel in the of stacking ensemble is
marine environment, as shown in Fig. 9. The results showed that demonstrated.
ensemble learning has better performance. Zounemat-Kermani et al. Bagging Decision tree The simplest decision (Camacho-Navarro
(2020) studied the ability of network-based ensemble methods and tree algorithm in the et al., 2017)
bagging scheme was
tree-based ensemble methods to predict concrete corrosion and found used to improve the
that tree-based ensemble models perform better than network-based accuracy of damage
ensemble models. detection;
In addition, ensemble learning can also be used as a simple data Boosting GBDT; RF The accuracy of GBDT (Gu et al., 2019)
in predicting the
fusion solution by integrating features from different information
service life of oil and
sources. For multiple datasets from different sources, due to discrep­ gas pipelines was
ancies of different features, it is impossible to use a single algorithm to slightly better than that
learn what is in all the data. In this case, ensemble learning can be used of RF.
to enhance the relationship between data and make better use of existing XGBoost; RF; XGBoost is the best (Mazumder et al.,
AdaBoost; algorithm for 2021)
data resources. Wang et al. (2021) developed a new artificial neural
LGBoost; predicting oil and gas
network-based data fusion model to provide tailored corrosion pre­ decision tree; pipeline failures by
dictions for the decks of bulk carriers. The study used ship measured Naive Bayes comparing other
thickness data and corrosion experimental data. The model framework machine learning
methods.
is shown in Fig. 10.

5.3. The hybrid approach incorporating multiple hybrid models make predictions to generate a prediction distribution for the uncer­
tainty of each feature in the input data. Xie et al. (2019) combined
Corrosion is the result of a combination of physical, chemical, and data-driven models with statistical models to realize the prediction of oil
biological processes (Yazdi et al., 2022c). One of the strategies for the and gas equipment failure rates. Oumouni and Schoefs (2021) proposed
hybrid approach incorporating multiple hybrid models is to use a new statistical, data-driven state dependence and perturbed model to
data-driven models to establish nonlinear or linear relationships be­ simulate hidden level-degradation both in time increments and
tween input variables and responses and use probabilistic models to

80
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

Fig. 9. Ensemble learning framework (Chou et al., 2017).

Fig. 10. The ANN-based data fusion architecture, with the measured thickness trained network (model 1) in yellow, the experimental data trained network (model 2)
in blue, and a fused layer in green to combine the predictions(Wang et al., 2021).

81
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

measurement uncertainty during pitting. Table 3

Model form uncertainty can also occur when using advanced tech­ Summaries of empirical models and data driven models introduced in this
niques such as machine learning to model corrosion processes, which is review.
caused by the biased understanding of the corrosion process. Hybrid Corrosion growth models Strengths Limitations
approach models can enhance the physical understanding of the corro­ Deterministic Single-value It is easy to use and The corrosion rate
sion process to reduce model form uncertainty. Cai et al. (2020a) pro­ models corrosion rate widely used in can be overly
posed a hybrid method for estimating the remaining useful life of subsea model industry. conservative.
pipelines based on physical models and data-driven models. The method Linear The model is simple The corrosion depth
corrosion rate and shows data and its
uses a DBN to simulate the degradation process of subsea pipelines
model relationship with distribution are
affected by multiple causes. When the training data for DBN is insuffi­ time. highly dependent on
cient, the structural model and parametric model for DBN are con­ the accuracy of the
structed using empirical or theoretical physical models. Heidary and detection tool.
Groth (2021) proposed a hierarchical Bayesian model based on a Power function This model is easy to The model
corrosion rate be applied in parameters need to
non-uniform gamma process to incorporate the uncertain ILI data and model practice. be determined
physical properties of failure in pitting corrosion. according to the
actual working
6. Discussions and suggestions environment.
Probabilistic Generalized GEVD is the best It depends on a large
models extreme value distribution to number of corrosion
As shown in the above sections, the four questions posed in the distribution characterize depth data.
introduction of this paper are answered sequentially in each section. The model corrosion rate
data sources used for corrosion growth modeling can be field data, uncertainty with the
experimental data and simulation data. Aleatory uncertainty and lowest Kolmogorov-
Smirnov statistic (
epistemic uncertainty generally exist in the modeling process. The Wang et al., 2015a).
strengths and limitations of empirical models and data-driven models Gamma Analytical The increment of
are summarized as Table 3. The deterministic model is simple and easy process/ tractability and corrosion depth is
to use, but critical issue is that the predicted value can be very conser­ Inverse physical independent of time.
Gaussian interpretability.
vative, which will can lead to unnecessary economic losses and risks.
process
Probabilistic models can account for temporal and spatial variability and Markov Not only the Model accuracy is
can provide probability distributions of corrosion depth or rates. How­ process variation of pit depth sensitive to the
ever, each probabilistic model has certain limitations due to its own with time but also the number of discrete
theoretical characteristics of the method, as shown in Table 3. The data- distribution of pit states.
depth and pit density
driven model can efficiently solve complex nonlinear problems with the
at each time can be
help of advanced intelligent algorithms and superior computing per­ obtained.
formance, and is suitable for old pipelines that cannot use in-line in­ Brownian It can ensure the The variance of the
spection technologies. But the learning process of artificial intelligence motion process temporal and spatial logarithm of the
continuity of the corrosion rate will
algorithms has the characteristics of “black box”, so the models have no
corrosion process. increase
physical meaning and the prediction accuracy greatly depends on the unconstrained,
quantity and quality of the dataset. If there is plenty of data available, which does not
machine learning or deep learning methods are the best options. match the actual.
When the available data is insufficient, different models will show Brownian Preserving the The knowledge of
bridge process advantages of the advanced stochastic
different effects, and in this situation the hybrid approach will be a best
Brownian motion process theory is
choice. Section 5 presented the concept of modeling with the hybrid process while required.
approach and three frameworks of the hybrid approach. Compared with considering the
a single model, the hybrid approach can not only incorporate the het­ passive states that
metal structures may
erogeneous data from different sources to make better use of data re­
experience.
sources, but also increase the understanding of the physical process of Machine Bayesian Insensitivity to Validation of
corrosion to reduce model uncertainty. Therefore, the hybrid approach learning networks missing data; it can Bayesian networks
models should be given more attention in the future development of incorporating prior are limited in
corrosion growth prediction. knowledge into industry practice (
modeling; Users can George and Renjith,
monitor and update 2021).
6.1. Challenge and deficiencies to the whole system
over time.
Some challenges and deficiencies in the related research are identi­ Artificial The difficulty of The learning process
neural modeling is reduced; has the
fied through this review work, and they are:
networks it can show characteristics of a
The data issues: Due to the employment of different types of sen­ sensitivity of each black box;
sors, the field data collected is multi-source, heterogeneous, and variable.
imbalanced. For example, the types of data include discrete samples Deep learning RNN, CNN, It can achieve precise The model lacks
such as temperature, pressure, and soil pH, continuous waveforms ob­ LSTM prediction results. adequate theoretical
foundation; it
tained by in-line inspection technology, images obtained by infrared involves complex
nondestructive testing technology, and text logs in management or hyper-parameters
operation. In addition, the data also suffers from the problems of data and network design
scarcity and missing values. There are far too few publicly available
datasets, and the majority of studies are exclusive to individual cases. As
a single data source cannot provide complete information, the estab­
lished model has certain limitations and conditions. The performance of
the model cannot be guaranteed when applied to pipelines with other

82
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

types of data sources. contrast, the probabilistic models take into account the variability
The measurement uncertainty issues: Due to the high cost of ILI over time and space and generate the probability distribution of
and its limitations in engineering applications, it is uncommon to have corrosion depth/rate.
two sets of ILI data in the same pipeline. In addition, the ILI data may be • Each probabilistic model has its own characteristics. Gamma process
measured by different ILI tools. Changes in ILI tools, odometer-related or Inverse Gaussian process based-corrosion growth models involve
errors, and measurement errors related to the size of corrosion defects specific assumptions about the average growth path; Markov
will all have an impact on the final results. process-based corrosion growth models need to discretize damage
The hyper-parameters issues: The choice of hyper-parameters has states and evaluate the possibility of transitions; Brownian motion
a significant impact on the accuracy of the results. However, in some process based-corrosion growth models cannot characterize the
studies, specific data preprocessing process and hyper-parameter opti­ positive and monotonically increasing nature of corrosion growth.
mization process were not provided, although their respective models • Compared with empirical models, data-driven models can be built
were explained and validated (Soomro et al., 2022). faster and easier with the help of artificial intelligence algorithms
through pipeline historical data sets and implemented with the
6.2. Directions and suggestions for future development actual operation of the pipeline. Data driven models can relatively
simply quantify the relationship between factors and corrosion
The properties that an excellent corrosion growth model should damage without the need for specialized knowledge of the corrosion
possess have been described in detail in the introduction section, so this domain and improve the performance of models. However, the
section gives some suggestions for the development of future models. learning process of artificial intelligence algorithms has the charac­
Among the existing models, the hybrid approach model can combine teristics of being a ‘black box’ and the prediction accuracy greatly
the advantages of multiple models, so we suggest that more options can depends on the quantity and quality of the dataset for a pipeline
continue to be explored in this area. For example, to make full use of system.
multiple forms of data, optimal data fusion mechanisms for heteroge­ • This paper introduces the concept of modeling with the hybrid
neous information should be devised, and the redundancy and approach and proposes three frameworks for the hybrid approach.
complementarity of multi-source information should be analyzed and All three frameworks have their own unique advantages. Compared
fused. Using the pipeline defect data collected by electromagnetic ul­ with a single model, they can not only incorporate the heterogeneous
trasonic guided wave, magnetic flux leakage and eddy current testing data from different sources to make better use of data resources, but
equipment, Liang et al. (2022) proposed a data fusion method also increase the understanding of the physical process of corrosion,
combining improved bird swarm algorithm and weighted regularized which can lead to reduce model uncertainty. The applications under
extreme learning machine, which effectively improved the data fusion these three frameworks are also listed separately. We think that
accuracy of pipeline defect; For dealing with data scarcity and missing hybrid approach models have greater potential than empirical
values, data reconfiguration and transfer learning between similar cat­ models and data driven models and thus it should be given more
egories can be used to increase the required sample size, while the attention in the future development of corrosion growth prediction.
hybrid approach that makes reasonable use of available data and in­
formation should be designed. Canonaco et al. (2022) design an algo­
rithmic solution using transfer learning for corrosion prediction when Declaration of Competing Interest
the target pipeline does not have sufficient supervisory information; To
improve the interpretability of the model, interaction mechanisms be­ The authors declare that they have no known competing financial
tween expert-experience and existing knowledge with the data-driven interests or personal relationships that could have appeared to influence
model can be established to make the reasoning and decision-making the work reported in this paper.
process of the learning model more transparent, which is explained in
detail in Section 5.3. Acknowledgements
While improving control theory performance, it is also vital to pro­
mote the advancement of various sensors and technologies. The quality This work was supported by the Key-Area Research and Develop­
and quantity of corrosion data and source of uncertainty are directly ment Program of Guangdong Province (Number: 2019B111102001), the
linked with them. In-line inspection technology must continue to evolve Beijing Municipal Science and Technology Commission (Number
to reduce measurement uncertainties and to gather accurate defect in­ Z201100004520011), the Fundamental Research Funds for the Central
formation. Simultaneously, cloud computing technology should be used Universities of China (Grant No. FRF-GF-20–24B, FRF-MP-19–014), and
to strengthen the management and storage of field data and operational Innovation Group Project of Southern Marine Science and Engineering
parameter data, and high-quality and diverse datasets and testbeds are Guangdong Laboratory (Zhuhai) (No. 311021013).
needed to build for validating algorithms or models.
References
7. Conclusion
(EGIG), E.G.P.I.D.G., 2020. 11th Report of the European Gas Pipeline Incident Data
This paper reviews corrosion growth models and draws the following Group (Period 1970–2019). European Gas Pipeline Incident Data Group (EGIG),
https://www.egig.eu/reports.
conclusions: ASTM G46–94. (2005). Standard guide for examination and evaluation of pitting
corrosion. ASTM International.
• The data, which is the foundation for modeling, should also be given Abbas, M.H., Norman, R., Charles, A., 2018. Neural network modelling of high pressure
CO2 corrosion in pipeline steels. Process Saf. Environ. Prot. 119, 36–45.
more consideration. In-Line Inspection technology is an excellent Agarwal, A., Ayello, F., Raghunathan, V., Demay, V., Sridhar, N., 2021. Quantitative
tool to determine the location and the size of corrosion defects in assessment of failure probability of underground natural gas storage wells using an
pipeline. Aleatory uncertainty cannot be reduced, but epistemic integrated bow-tie bayesian network approach. J. Fail. Anal. Prev. 21, 107–123.
Ahammed, M., Melchers, R.E., 1997. Probabilistic analysis of underground pipelines
uncertainty should be minimized as much as possible to improve the
subject to combined stresses and corrosion. Eng. Struct. 19, 988–994.
accuracy of results. Ahmadianfar, I., Jamei, M., Karbasi, M., Sharafati, A., Gharabaghi, B., 2022. A novel
• The deterministic models are simple and easy to use, but they do not boosting ensemble committee-based model for local scour depth around non-
consider the influence of uncertain factors, including measurement uniformly spaced pile groups. Eng. Comput. 38, 3439–3461.
Ahuja, S.K., Shukla, M.K., Ravulakollu, K.K., 2021. Optimized deep learning framework
errors and randomness related to corrosion growth. At the same time, for detecting pitting corrosion based on image segmentation. Int. J. Perform. Eng.
these models only give a point estimate at a specified time. In 17, 627–637.

83
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

Ajmi, C., Zapata, J., Elferchichi, S., Zaafouri, A., Laabidi, K., 2020. Deep learning El-Abbasy, M.S., Senouci, A., Zayed, T., Parvizsedghy, L., Mirahadi, F., 2016. Unpiggable
technology for weld defects classification based on transfer learning and activation oil and gas pipeline condition forecasting models. J. Perform. Constr. Facil. 30,
features. Adv. Mater. Sci. Eng. 2020, 1574350. 04014202.
Al-Amin, M., Zhou, W., Zhang, S., Kariyawasam, S., Wang, H., 2014. Hierarchical ELDREDGE, G.G., 1957. Analysis of corrosion pitting by extreme-value statistics and its
Bayesian corrosion growth model based on in-line inspection data. J. Press. Vessel application to oil well tubing caliper surveys. Corrosion 13, 67–76.
Technol. 136, 041401. Elsayed, E., Liao, H., 2004. A geometric Brownian motion model for field degradation
Amaya-Gómez, R., Sánchez-Silva, M., Muñoz, F., 2019. Integrity assessment of corroded data. Int. J. Mater. Prod. Technol. 20, 51–72.
pipelines using dynamic segmentation and clustering. Process Saf. Environ. Prot. Enning, D., Garrelfs, J., 2014. Corrosion of iron by sulfate-reducing bacteria: new views
128, 284–294. of an old problem. Appl. Environ. Microbiol. 80, 1226–1236.
Arzaghi, E., Abbassi, R., Garaniya, V., Binns, J., Chin, C., Khakzad, N., Reniers, G., 2018. Fan, Z.-Y., Ju, H., Sun, F.-B., 2015. Improved gamma process for degradation analysis
Developing a dynamic model for pitting and corrosion-fatigue damage of subsea under nonlinear condition. Int. J. Reliab., Qual. Saf. Eng. 22, 1550030.
pipelines. Ocean Eng. 150, 391–396. Foorginezhad, S., Mohseni-Dargah, M., Firoozirad, K., Aryai, V., Razmjou, A.,
Arzaghi, E., Chia, B.H., Abaei, M.M., Abbassi, R., Garaniya, V., 2020. Pitting corrosion Abbassi, R., Garaniya, V., Beheshti, A., Asadnia, M., 2021. Recent advances in
modelling of X80 steel utilized in offshore petroleum pipelines. Process Saf. Environ. sensing and assessment of corrosion in sewage pipelines. Process Saf. Environ. Prot.
Prot. 141, 135–139. 147, 192–213.
ASME B31.8S. (2010). Managing System Integrity of Gas Pipelines. ASME Code for Fougeres, A.L., Holm, S., Rootzen, H., 2006. Pitting corrosion: comparison of treatments
Pressure Piping. with extreme-value-distributed responses. Technometrics 48, 262–272.
AZIZ, P.M., 1956. Application of the statistical theory of extreme values to the analysis of George, P.G., Renjith, V.R., 2021. Evolution of Safety and Security Risk Assessment
maximum pit depth data for aluminum. Corrosion 12, 35–46. methodologies towards the use of Bayesian Networks in Process Industries. Process
Balekelayi, N., Tesfamariam, S., 2020. External corrosion pitting depth prediction using Saf. Environ. Prot. 149, 758–775.
Bayesian spectral analysis on bare oil and gas pipelines. Int. J. Press. Vessels Pip. Gu, J., Zhang, H., Chen, L., Lian, S., 2019. The application of the big data algorithm for
188, 104224. pipeline lifetime. Anal., 2019 Chin. Autom. Congr. (CAC) 824–829.
Barbalat, M., Lanarde, L., Caron, D., Meyer, M., Vittonato, J., Castillon, F., Fontaine, S., Guan, S., Ayello, F., Sánchez, A.N., Khare, V., Sridhar, N., Yang, Y., Hu, J., Xu, C., Zou, B.,
Refait, P., 2012. Electrochemical study of the corrosion rate of carbon steel in soil: 2016. Internal corrosion direct assessment using Bayesian networks modeling with
evolution with time and determination of residual corrosion rates under cathodic limited data: a case study. CORROSION 2016.
protection. Corros. Sci. 55, 246–253. Guida, M., Pulcini, G., 2013. The inverse Gamma process: a family of continuous
Bastian, B.T., N, J., Ranjith, S.K., Jiji, C.V., 2019. Visual inspection and characterization stochastic models for describing state-dependent deterioration phenomena. Reliab.
of external corrosion in pipelines using deep neural network. NDT E Int. 107, Eng. Syst. Saf. 120, 72–79.
102134. Guo, J., Wang, C., Cabrera, J., Elsayed, E.A., 2018. Improved inverse Gaussian process
Ben Seghier, M.E.A., Keshtegar, B., Taleb-Berrouane, M., Abbassi, R., Trung, N.-T., 2021. and bootstrap: Degradation and reliability metrics. Reliab. Eng. Syst. Saf. 178,
Advanced intelligence frameworks for predicting maximum pitting corrosion depth 269–277.
in oil and gas pipelines. Process Saf. Environ. Prot. 147, 818–833. Guo, J., Li, Z., Li, M., 2020. A review on prognostics methods for engineering systems.
Bhandari, J., Khan, F., Abbassi, R., Garaniya, V., Ojeda, R., 2015. Modelling of pitting IEEE Trans. Reliab. 69, 1110–1129.
corrosion in marine and offshore steel structures – a technical review. J. Loss Prev. Heidary, R., Groth, K.M., 2021. A hybrid population-based degradation model for
Process Ind. 37, 39–62. pipeline pitting corrosion. Reliab. Eng. Syst. Saf. 214, 107740.
Bhandari, J., Lau, S., Abbassi, R., Garaniya, V., Ojeda, R., Lisson, D., Khan, F., 2017. Helton, J.C., Johnson, J.D., Oberkampf, W.L., Sallaberry, C.J., 2010. Representation of
Accelerated pitting corrosion test of 304 stainless steel using ASTM G48; analysis results involving aleatory and epistemic uncertainty. Int. J. Gen. Syst. 39,
Experimental investigation and concomitant challenges. J. Loss Prev. Process Ind. 605–646.
47, 10–21. Hong, H.-P., 1999. Application of the stochastic process to pitting corrosion. Corrosion
Cai, B., Shao, X., Liu, Y., Kong, X., Wang, H., Xu, H., Ge, W., 2020a. Remaining useful life 55, 10–16.
estimation of structure systems under the influence of multiple causes: subsea Jain, S., Sánchez, A.N., Guan, S., Wu, S., Ayello, F., Sridhar, N., Feng, Q., Yan, B., 2015.
pipelines as a case study. IEEE Trans. Ind. Electron. 67, 5737–5747. Probabilistic assessment of external corrosion rates in buried oil and gas pipelines.
Cai, R., Han, T., Liao, W., Huang, J., Li, D., Kumar, A., Ma, H., 2020b. Prediction of Corros. 2015. March 15 2015, Dallas, Tex.
surface chloride concentration of marine concrete using ensemble machine learning. Joshi, A.V., 2011. Statistical analysis of in-line inspection performance with gamma
Cem. Concr. Res. 136, 106164. distribution. Corros. 2011. March 13 2011, Houst., Tex.
Caleyo, F., Velázquez, J.C., Valor, A., Hallen, J.M., 2009a. Markov chain modelling of Kamil, M.Z., Taleb-Berrouane, M., Khan, F., Amyotte, P., 2021. Data-driven operational
pitting corrosion in underground pipelines. Corros. Sci. 51, 2197–2207. failure likelihood model for microbiologically influenced corrosion. Process Saf.
Caleyo, F., Velázquez, J.C., Valor, A., Hallen, J.M., 2009b. Probability distribution of Environ. Prot. 153, 472–485.
pitting corrosion depth and rate in underground pipelines: A Monte Carlo study. Keshtegar, B., Seghier, Ben, Zhu, M.E.A., Abbassi, S.-P., Trung, N.-T, R., 2019. Reliability
Corros. Sci. 51, 1925–1934. analysis of corroded pipelines: novel adaptive conjugate first order reliability
Camacho-Navarro, J., Ruiz, M., Villamizar, R., Mujica, L., Moreno-Beltrán, G., 2017. method. J. Loss Prev. Process Ind. 62, 103986.
Ensemble learning as approach for pipeline condition assessment. Journal of Physics: Khan, F., Yarveisy, R., Abbassi, R., 2021. Cross-country pipeline inspection data analysis
Conference Series 842, 012019. and testing of probabilistic degradation models. J. Pipeline Sci. Eng. 1, 308–320.
Canonaco, G., Roveri, M., Alippi, C., Podenzani, F., Bennardo, A., Conti, M., Mancini, N., Kim, C., Chen, L., Wang, H., Castaneda, H., 2021. Global and local parameters for
2022. A transfer-learning approach for corrosion prediction in pipeline characterizing and modeling external corrosion in underground coated steel
infrastructures. Appl. Intell. 52, 7622–7637. pipelines: a review of critical factors. J. Pipeline Sci. Eng. 1, 17–35.
Chamkalani, A., Nareh’ei, M.A., Chamkalani, R., Zargari, M.H., Dehestani-Ardakani, M. Kiureghian, A.D., Ditlevsen, O., 2009. Aleatory or epistemic? Does it matter? Struct. Saf.
R., Farzam, M., 2013. soft computing method for prediction of co2 corrosion in flow 31, 105–112.
lines based on neural network approach. Chem. Eng. Commun. 200, 731–747. Koch, G., Ayello, F., Khare, V., Sridhar, N., Moosavi, A., 2015. Corrosion threat
Chen, S.H., Pollino, C.A., 2012. Good practice in Bayesian network modelling. Environ. assessment of crude oil flow lines using Bayesian network model. Corros. Eng., Sci.
Model. Softw. 37, 134–145. Technol. 50, 236–247.
Chou, J.-S., Ngo, N.-T., Chong, W.K., 2017. The use of artificial intelligence combiners Krizhevsky, A., Sutskever, I., Hinton, G.E., 2017. Imagenet classification with deep
for modeling steel pitting risk and corrosion rate. Eng. Appl. Artif. Intell. 65, convolutional neural networks. Commun. ACM 60, 84–90.
471–483. Lawless, J., Crowder, M., 2004. Covariates and random effects in a gamma process model
Cockburn, G., Tesfamariam, S., 2012. Earthquake disaster risk index for Canadian cities with application to degradation and failure. Lifetime Data Anal. 10, 213–227.
using Bayesian belief networks. Georisk: Assess. Manag. Risk Eng. Syst. Geohazards Layouni, M., Hamdi, M.S., Tahar, S., 2017. Detection and sizing of metal-loss defects in
6, 128–140. oil and gas pipelines using pattern-adapted wavelets and machine learning. Appl.
Cunha Lins, Vd.F., Magalhães Ferreira, M.L., Saliba, P.A., 2012. Corrosion resistance of Soft Comput. 52, 247–261.
API X52 carbon steel in soil environment. J. Mater. Res. Technol. 1, 161–166. Lee, J., Cameron, I., Hassall, M., 2019. Improving process safety: what roles for
Dawson, J., Ganim, L., 2016. Applying Corrosion Growth Rates Derived From Repeat ILI digitalization and industry 4.0? Process Saf. Environ. Prot. 132, 325–339.
Runs to Predict Future Severity, International Pipeline Conference. American Society Leis, B.N., Bubenik, T.A., 2001. Periodic re-verification intervals for high-consequence
of Mechanical Engineers, p. V001T003A004. areas. Battelle Rep. GTI.
Dawuda, A.-W., Taleb-berrouane, M., Khan, F., 2021. A probabilistic model to estimate Li, C.Q., Mahmoodian, M., 2013. Risk based service life prediction of underground cast
microbiologically influenced corrosion rate. Process Saf. Environ. Prot. 148, iron pipes subjected to corrosion. Reliab. Eng. Syst. Saf. 119, 102–108.
908–926. Li, X., Castaneda, H., 2014. Influence of soil parameters on coating damage evolution of
Demissie, G., Tesfamariam, S., Sadiq, R., 2016. Considering soil parameters in prediction X52 pipeline steel under cathodic protection conditions. Corros. 2014. March 09
of remaining service life of metallic pipes: bayesian belief network model. J. Pipeline 2014, San. Antonio, Tex., USA.
Syst. Eng. Pract. 7, 04015028. Li, X., Han, Z., Zhang, R., Abbassi, R., Chang, D., 2020. An integrated methodology to
Desjardins, G., 2001. Corrosion rate and severity results from in-line inspection data. manage risk factors of aging urban oil and gas pipelines. J. Loss Prev. Process Ind.
Corros. 2001, Houst., Tex. 66, 104154.
El Amine Ben Seghier, M., Keshtegar, B., Tee, K.F., Zayed, T., Abbassi, R., Trung, N.T., Li, X., Zhang, L., Khan, F., Han, Z., 2021a. A data-driven corrosion prediction model to
2020. Prediction of maximum pitting corrosion depth in oil and gas pipelines. Eng. support digitization of subsea operations. Process Saf. Environ. Prot. 153, 413–421.
Fail. Anal. 112, 104505. Li, X., Zhang, Y., Abbassi, R., Khan, F., Chen, G., 2021b. Probabilistic fatigue failure
El-Abbasy, M.S., Senouci, A., Zayed, T., Mirahadi, F., Parvizsedghy, L., 2014. Artificial assessment of free spanning subsea pipeline using dynamic Bayesian network. Ocean
neural network models for predicting condition of offshore oil and gas pipelines. Eng. 234, 109323.
Autom. Constr. 45, 50–65.

84
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

Li, X., Zhang, Y., Abbassi, R., Yang, M., Zhang, R., Chen, G., 2021c. Dynamic probability Quej-Ake, L., Contreras, A., Liu, H., Alamilla, J., Sosa, E., 2018. Assessment on external
assessment of urban natural gas pipeline accidents considering integrated external corrosion rates for API pipeline steels exposed to acidic sand-clay soil. Anti-Corros.
activities. J. Loss Prev. Process Ind. 69, 104388. Methods Mater. 65, 281–291.
Li, X., Jia, R., Zhang, R., Yang, S., Chen, G., 2022a. A KPCA-BRANN based data-driven Rachman, A., Zhang, T., Ratnayake, R.C., 2021. Applications of machine learning in
approach to model corrosion degradation of subsea oil pipelines. Reliab. Eng. Syst. pipeline integrity management: a state-of-the-art review. Int. J. Press. Vessels Pip.
Saf. 219, 108231. 193, 104471.
Li, X., Wang, J., Abbassi, R., Chen, G., 2022b. A risk assessment framework considering Ren, C.-y, Qiao, W., Tian, X., 2012. Natural Gas Pipeline Corrosion Rate Prediction Model
uncertainty for corrosion-induced natural gas pipeline accidents. J. Loss Prev. Based on BP Neural Network. In: Cao, B.-Y., Xie, X.-J. (Eds.), Fuzzy Engineering and
Process Ind. 75, 104718. Operations Research. Springer, Berlin Heidelberg, Berlin, Heidelberg, pp. 449–455.
Li, X., Zhu, Y., Abbassi, R., Chen, G., 2022c. A probabilistic framework for risk Romanoff, M., 1957. Underground Corrosion. US Government Printing Office.
management and emergency decision-making of marine oil spill accidents. Process Sambo, C., Yin, Y., Djuraev, U., Ghosh, D., 2018. Application of adaptive neuro-fuzzy
Saf. Environ. Prot. 162, 932–943. inference system and optimization algorithms for predicting methane gas viscosity at
Li, Y.-T., He, X.-N., Shuai, J., 2022d. Risk analysis and maintenance decision making of high pressures and high temperatures conditions. Arab. J. Sci. Eng. 43, 6627–6638.
natural gas pipelines with external corrosion based on Bayesian network. Pet. Sci. Saxena, A., Goebel, K., Simon, D., Eklund, N., 2008. Damage propagation modeling for
19, 1250–1261. aircraft engine run-to-failure simulation, 2008 international conference on
Liang, H., Cheng, G., Zhang, Z., Yang, H., Luo, S., 2022. Research on data fusion method prognostics and health management. IEEE, pp. 1–9.
for multi -sensor detection of pipeline defects. Laser & Optoelectronics Progress Shabarchin, O., Tesfamariam, S., 2016. Internal corrosion hazard assessment of oil & gas
https://kns.cnki.net/kcms/detail/31.1690.tn.20220713.1856.521.htm〈l〉. pipelines using Bayesian belief network model. J. loss Prev. Process Ind. 40,
Liao, K.X., Yao, Q.K., Wu, X., Jia, W.L., 2012. A numerical corrosion rate prediction 479–495.
method for direct assessment of wet gas gathering pipelines internal corrosion. Shaik, N.B., Pedapati, S.R., Taqvi, S.A.A., Othman, A.R., Dzubir, F.A.A., 2020. A feed-
Energies 5, 3892–3907. forward back propagation neural network approach to predict the life condition of
Liu, B., Zhang, X., Gao, Z., Chen, L., 2017a. Weld defect images classification with vgg16- crude oil pipeline. Processes 8, 661.
based neural network. International Forum on Digital TV and Wireless Multimedia Shekari, E., Khan, F., Ahmed, S., 2016. A predictive approach to fitness-for-service
Communications. Springer,, pp. 215–223. assessment of pitting corrosion. Int. J. Press. Vessels Pip. 137, 13–21.
Liu, Y., Song, Y., Keller, J., Bond, P., Jiang, G., 2017b. Prediction of concrete corrosion in Shekari, E., Khan, F., Ahmed, S., 2017. Probabilistic modeling of pitting corrosion in
sewers with hybrid Gaussian processes regression model. Rsc Adv. 7, 30894–30903. insulated components operating in offshore facilities. ASME. ASME J. Risk
Ma, Y., Zheng, J., Liang, Y., Klemeš, J.J., Du, J., Liao, Q., Lu, H., Wang, B., 2022. Uncertain. Part B 3, 011003.
Deeppipe: theory-guided neural network method for predicting burst pressure of Shi, F., Liu, Z., Li, E., 2017. Prediction of pipe performance with ensemble machine
corroded pipelines. Process Saf. Environ. Prot. 162, 595–609. learning based approaches. 2017 Int. Conf. Sens., Diagn., Progn., Control (SDPC)
Marsh, G., Taylor, K., 1988. An assessment of carbon steel containers for radioactive 408–414.
waste disposal. Corros. Sci. 28, 289–320. Shi, F., Peng, X., Liu, Z., Li, E., Hu, Y., 2020. A data-driven approach for pipe deformation
Masi, G.D., Gentile, M., Vichi, R., Bruschi, R., Gabetta, G., 2015. Machine learning prediction based on soil properties and weather conditions. Sustain. Cities Soc. 55,
approach to corrosion assessment in subsea pipelines. OCEANS 2015 - Genova 1–6. 102012.
Matthaiou, A., Papalambrou, G., Samuelides, M., 2021. Corrosion Detection with Soomro, A.A., Mokhtar, A.A., Kurnia, J.C., Lashari, N., Lu, H., Sambo, C., 2022. Integrity
Computer Vision and Deep Learning. Developments in the Analysis and Design of assessment of corroded oil and gas pipelines using machine learning: a systematic
Marine Structures. CRC Press, pp. 289–296. review. Eng. Fail. Anal. 131, 105810.
Mazumder, R.K., Salman, A.M., Li, Y., 2021. Failure risk analysis of pipelines using data- Sosa, E., Martinez, A.V., Alamilla, J.L., Contreras, A., Quej, L.M., Liu, H., 2020.
driven machine learning algorithms. Struct. Saf. 89, 102047. A numerical external pitting damage prediction method of buried pipelines. Corros.
Mazzella, J., Hayden, T., Krissa, L., Tsaprailis, H., 2019. Estimating corrosion growth rate Rev. 38, 433–444.
for underground pipeline: a machine learning based approach. Corros. 2019. March Spark, A., Wang, K., Cole, I., Law, D., Ward, L., 2020. Microbiologically influenced
24 2019, Nashv., Tenn., USA. corrosion: a review of the studies conducted on buried pipelines. Corros. Rev. 38,
McNealy, R., McCann, R., Van Hook, M., Stiff, A., Kania, R., 2010. In-line inspection 231–262.
performance III: effect of in-ditch errors in determining ILI performance, 2010 8th Strutt, J.E., Nicholls, J.R., Barbier, B., 1985. The prediction of corrosion by statistical
Int. Pipeline Conf. 469–473. analysis of corrosion profiles. Corros. Sci. 25, 305–315.
Melo, C., Dann, M.R., Hugo, R.J., Janeta, A., 2019. A framework for risk-based integrity Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov, D., Erhan, D.,
assessment of unpiggable pipelines subject to internal corrosion. J. Press. Vessel Vanhoucke, V., Rabinovich, A., 2015. Going deeper with convolutions. Proc. IEEE
Technol. 141, 021702. Conf. Comput. Vis. Pattern Recognit. 1–9.
Mihajlovic, V., Petkovic, M., 2001. Dynamic bayesian networks: a state of the art. Univ. Tesfamariam, S., Woldesellasse, H., Xu, M., Asselin, E., 2021. General corrosion
Twente Doc. Repos. vulnerability assessment using a Bayesian belief network model incorporating
Mullins, J., Ling, Y., Mahadevan, S., Sun, L., Strachan, A., 2016. Separation of aleatory experimental corrosion data for X60 pipe steel. J. Pipeline Sci. Eng. 1, 329–338.
and epistemic uncertainty in probabilistic model validation. Reliab. Eng. Syst. Saf. Timashev, S., Anna, B., 2016. Diagnostics and reliability of pipeline systems. Springer,,
147, 49–59. New York.
NACE Standard SP0502. (2010). Pipeline External Corrosion Direct Assessment Timashev, S.A., Bushinskaya, A.V., 2015. Markov approach to early diagnostics,
Methodology. reliability assessment, residual life and optimal maintenance of pipeline systems.
Nešić, S., 2007. Key issues related to modelling of internal corrosion of oil and gas Struct. Saf. 56, 68–79.
pipelines–a review. Corros. Sci. 49, 4308–4338. Valor, A., Caleyo, F., Alfonso, L., Rivas, D., Hallen, J.M., 2007. Stochastic modeling of
Ossai, C.I., 2020. Corrosion defect modelling of aged pipelines with a feed-forward multi- pitting corrosion: a new model for initiation and growth of multiple corrosion pits.
layer neural network for leak and burst failure estimation. Eng. Fail. Anal. 110, Corros. Sci. 49, 559–579.
104397. Valor, A., Caleyo, F., Alfonso, L., Velázquez, J.C., Hallen, J.M., 2013a. Markov Chain
Ossai, C.I., Boswell, B., Davies, I.J., 2015. Predictive modelling of internal pitting Models for the Stochastic Modeling of Pitting Corrosion. Math. Probl. Eng. 2013,
corrosion of aged non-piggable pipelines. J. Electrochem. Soc. 162, C251–C259. 108386.
Ossai, C.I., Boswell, B., Davies, I., 2016. Markov chain modelling for time evolution of Valor, A., Caleyo, F., Hallen, J.M., Velázquez, J.C., 2013b. Reliability assessment of
internal pitting corrosion distribution of oil and gas pipelines. Eng. Fail. Anal. 60, buried pipelines based on different corrosion rate models. Corros. Sci. 66, 78–87.
209–228. van Noortwijk, J.M., van der Weide, J.A.M., Kallen, M.J., Pandey, M.D., 2007. Gamma
Oumouni, M., Schoefs, F., 2021. A Perturbed Markovian process with state-dependent processes and peaks-over-threshold distributions for time-dependent reliability.
increments and measurement uncertainty in degradation modeling. Comput. Civ. Reliab. Eng. Syst. Saf. 92, 1651–1658.
Infrastruct. Eng. 36, 978–995. Vanaei, H.R., Eslami, A., Egbewande, A., 2017. A review on pipeline corrosion, in-line
Pandey, M., Yuan, X., Van Noortwijk, J., 2005. Gamma process model for reliability inspection (ILI), and corrosion growth rate models. Int. J. Press. Vessels Pip. 149,
analysis and replacement of aging structural components, Proceedings ICOSSAR. 43–54.
Citeseer. Velázquez, J.C., Cruz-Ramirez, J.C., Valor, A., Venegas, V., Caleyo, F., Hallen, J.M.,
Papavinasam, S., Revie, R.W., Friesen, W.I., Doiron, A., Panneerselvan, T., 2006. Review 2017. Modeling localized corrosion of pipeline steels in oilfield produced water
of models to predict internal pitting corrosion of oil and gas pipelines. Corros. Rev. environments. Eng. Fail. Anal. 79, 216–231.
24, 173–230. Vishnu, T., Gupta, P., Malhotra, P., Vig, L., Shroff, G., 2018. Recurrent neural networks
Park, C., Padgett, W., 2005. Accelerated degradation models for failure based on for online remaining useful life estimation in ion mill etching system, Proceedings of
geometric Brownian motion and gamma processes. Lifetime Data Anal. 11, 511–527. the Annual Conference of the PHM Society, Philadelphia, PA, USA.
Peng, X., Anyaoha, U., Liu, Z., Tsukada, K., 2020. Analysis of magnetic-flux leakage Wang, H., Yajima, A., Liang, R.Y., Castaneda, H., 2015a. Bayesian modeling of external
(MFL) data for pipeline corrosion assessment. IEEE Trans. Magn. 56, 1–15. corrosion in underground pipelines based on the integration of markov chain Monte
Polikar, R., 2006. Ensemble based systems in decision making. IEEE Circuits Syst. Mag. 6, Carlo techniques and clustered inspection data. Comput. -Aided Civ. Infrastruct. Eng.
21–45. 30, 300–316.
Polikar, R., 2012. Ensemble learning. Ensemble machine learning. Springer, pp. 1–34. Wang, H., Yajima, A., Liang, R.Y., Castaneda, H., 2015b. Reliability-based temporal and
Provan, J.W., Rodriguez, E.S., 1989. 1. Development of a Markov Description of Pitting. spatial maintenance strategy for integrity management of corroded underground
Corros Corros 45, 178–192. pipelines. Struct. Infrastruct. Eng. 12, 1281–1294.
Qin, M., Liao, K., He, G., Huang, Y., Wang, M., Zhang, S., 2022. Main control factors and Wang, H., Yajima, A., Liang, Y., Castaneda, H, R., 2015c. A Bayesian model framework
prediction model of flow-accelerated CO2/H2S synergistic corrosion for X65 steel. for calibrating ultrasonic in-line inspection data and estimating actual external
Process Saf. Environ. Prot. 160, 749–762. corrosion depth in buried pipeline utilizing a clustering technique. Struct. Saf. 54,
19–31.

85
H. Ma et al. Process Safety and Environmental Protection 171 (2023) 71–86

Wang, H., Yajima, A., Castaneda, H., 2019. A stochastic defect growth model for Yazdi, M., Khan, F., Abbassi, R., Quddus, N., 2022b. Resilience assessment of a subsea
reliability assessment of corroded underground pipelines. Process Saf. Environ. Prot. pipeline using dynamic Bayesian network. J. Pipeline Sci. Eng. 2, 100053.
123, 179–189. Yazdi, M., Khan, F., Abbassi, R., Quddus, N., Castaneda-Lopez, H., 2022c. A review of
Wang, S., Liu, D., Du, N., Zhao, Q., Xiao, J., 2015d. Analysis of the long-term corrosion risk-based decision-making models for microbiologically influenced corrosion (MIC)
behavior of x80 pipeline steel in acidic red soil using electrical resistance test in offshore pipelines. Reliab. Eng. Syst. Saf. 223, 108474.
technique. Adv. Mater. Sci. Eng. 2015, 931761. Ye, Z.S., Chen, N., 2014. The inverse gaussian process as a degradation model.
Wang, X., 2010. Wiener processes with random effects for degradation data. J. Multivar. Technometrics 56, 302–311.
Anal. 101, 340–351. Yeo, C., Bhandari, J., Abbassi, R., Garaniya, V., Chai, S., Shomali, B., 2016. Dynamic risk
Wang, Y., Su, C., Xie, M., 2020. Remaining useful life prediction of corroded oil pipelines analysis of offloading process in floating liquefied natural gas (FLNG) platform using
based on binary inverse Gaussian process. J. Southeast Univ. (Nat. Sci. Ed. ) 50, Bayesian Network. J. Loss Prev. Process Ind. 41, 259–269.
1038–1044. Yuan, Y., Zhang, Y., Ding, H., 2020. Research on key technology of industrial artificial
Wang, Z., Sobey, A.J., Wang, Y., 2021. Corrosion prediction for bulk carrier via data intelligence and its application in predictive maintenance. Acta Autom. Sin. 46,
fusion of survey and experimental measurements. Mater. Des. 208, 109910. 2013–2030.
Weber, P., Jouffe, L., 2003. Reliability modelling with dynamic bayesian networks. IFAC Zeng, J., Xu, H.Y., Gong, G.L., Xu, C., Tian, C.X., Lu, T., Jiang, R., 2020. Hackem-LIBS: an
Proceedings Volumes 36, 57–62. heterogeneous stacking ensemble model for laser-induced breakdown spectroscopy
Worthingham, R., Morrison, T., Desjardins, G., 2000. Comparison of Estimates From a elemental quantitative analysis. Ieee Access 8, 136141–136150.
Growth Model 5 Years After the Previous Inspection, 2000 3rd International Pipeline Zhang, L., Wen, J., Li, Y., Chen, J., Ye, Y., Fu, Y., Livingood, W., 2021. A review of
Conference, p. V002T006A028. machine learning in building load prediction. Appl. Energy 285, 116452.
Xia, X., Nie, J.F., Davies, C.H.J., Tang, W.N., Xu, S.W., Birbilis, N., 2016. An artificial Zhang, S., 2014. Development of probabilistic corrosion growth models with applications
neural network for predicting corrosion rate and hardness of magnesium alloys. in integrity management of pipelines. The University of Western Ontario.
Mater. Des. 90, 1034–1043. Zhang, S., Zhou, W., 2014. Bayesian dynamic linear model for growth of corrosion
Xiang, W., Zhou, W., 2018. Corrosion Growth Modeling by Learning a Dynamic Bayesian defects on energy pipelines. Reliab. Eng. Syst. Saf. 128, 24–31.
Network From Multiple In-Line Inspection Data, 2018 12th International Pipeline Zhang, S., Zhou, W., 2015. Probabilistic characterisation of metal-loss corrosion growth
Conference, Calgary, Alberta, Canada., p. V001T003A026. on underground pipelines based on geometric Brownian motion process. Struct.
Xiang, W., Zhou, W., 2019. Integrated pipeline corrosion growth modeling and reliability Infrastruct. Eng. 11, 238–252.
analysis using dynamic Bayesian network and parameter learning technique. Struct. Zhang, S., Zhou, W., Qin, H., 2013. Inverse Gaussian process-based corrosion growth
Infrastruct. Eng. 16, 1161–1176. model for energy pipelines considering the sizing error in inspection data. Corros.
Xie, L., Håbrekke, S., Liu, Y., Lundteigen, M.A., 2019. Operational data-driven prediction Sci. 73, 309–320.
for failure rates of equipment in safety instrumented systems: a case study from the Zhang, X., Li, Y., Wang, X., 2017. Maintenance strategy of corroded oil-gas pipeline
oil and gas industry. J. Loss Prev. Process Ind. 60, 96–105. based on inverse Gaussian process. Acta Pet. Sin. 38, 356–362.
Xie, M., Tian, Z., 2018. A review on pipeline integrity management utilizing in-line Zhou, L.N., Pan, S.M., Wang, J.W., Vasilakos, A.V., 2017a. Machine learning on big data:
inspection data. Eng. Fail. Anal. 92, 222–239. opportunities and challenges. Neurocomputing 237, 350–361.
Yang, J., Li, R., Chen, L., Hu, Y., Dou, Z., 2022. Research on equipment corrosion Zhou, W., Hong, H.P., Zhang, S., 2012. Impact of dependent stochastic defect growth on
diagnosis method and prediction model driven by data. Process Saf. Environ. Prot. system reliability of corroding pipelines. Int. J. Press. Vessels Pip. 96–97, 68–77.
158, 418–431. Zhou, W., Xiang, W., Hong, H.P., 2017b. Sensitivity of system reliability of corroding
Yang, Y., Khan, F., Thodi, P., Abbassi, R., 2017. Corrosion induced failure analysis of pipelines to modeling of stochastic growth of corrosion defects. Reliab. Eng. Syst.
subsea pipelines. Reliab. Eng. Syst. Saf. 159, 214–222. Saf. 167, 428–438.
Yarveisy, R., Khan, F., Abbassi, R., 2022. Data-driven predictive corrosion failure model Zhu, H., Ge, W., Liu, Z., 2019. Deep learning-based classification of weld surface defects.
for maintenance planning of process systems. Comput. Chem. Eng. 157, 107612. Appl. Sci. 9, 3312.
Yazdi, M., Khan, F., Abbassi, R., 2021. Microbiologically influenced corrosion (MIC) Zounemat-Kermani, M., Stephan, D., Barjenbruch, M., Hinkelmann, R., 2020. Ensemble
management using Bayesian inference. Ocean Eng. 226, 108852. data mining modeling in corrosion of concrete sewer: a comparative study of
Yazdi, M., Khan, F., Abbassi, R., 2022a. Operational subsea pipeline assessment affected network-based (MLPNN & RBFNN) and tree-based (RF, CHAID, & CART) models.
by multiple defects of microbiologically influenced corrosion. Process Saf. Environ. Adv. Eng. Inform. 43, 101030.
Prot. 158, 159–171.

86