Process Safety and Environmental Protection 174 (2023) 320–327

Contents lists available at ScienceDirect

Process Safety and Environmental Protection

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

Predictive deep learning for pitting corrosion modeling in buried

transmission pipelines
Behnam Akhlaghi a, Hassan Mesghali a, Majid Ehteshami a, Javad Mohammadpour b,
Fatemeh Salehi b, Rouzbeh Abbassi b, *
a
Department of Civil Engineering, K. N. Toosi University of Technology, Tehran, Islamic Republic of Iran
b
School of Engineering, Macquarie University, Sydney, NSW 2109, Australia

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

Keywords: Despite significant efforts and investments in the renewable energy sector, fossil fuels continue to provide the
Transmission pipelines majority of the world’s energy supply. Transmission pipelines, which are extensively used in the oil and gas
Pitting corrosion industry, are vulnerable to various failure mechanisms, such as corrosion. Among these, pitting corrosion in
Deep learning
offshore pipelines is the most prevalent type of external corrosion. This study explores the potential of deep
Generalization model
learning models (Generalization and Generalization-Memorization models) to predict the maximum depth of
Generalization-Memorization model
pitting corrosion in oil and gas pipelines. The models are trained considering various characteristics of the soil
where the pipe is buried and different types of the protective coating of the pipes. The application of deep neural
networks resulted in a mean squared error of prediction of 0.0055 in training data and 0.0037 in test data. These
results demonstrate that deep learning models outperform all empirical and hybrid models applied in previous
studies on the same dataset. The proposed model in this study has the potential to predict failure rates of the
pipelines due to external corrosion and enhance the safety and reliability of these facilities.

1. Introduction pipelines, due to chemical and electrochemical reactions in their sur­

rounding environment (Bhandari et al., 2015, 2017). Corrosion causes a
Despite consistent efforts and investments in the renewable energy reduction in material resistance due to the gradual loss of thickness.
sector, it is still not meeting the requirements of the power grid, as fossil Without timely inspection, corrosion has the potential to ultimately
fuels continue to be the main source of energy production worldwide result in structural failure (Arzaghi et al., 2020). Corrosion is a
(Chandrasekaran and Sricharan, 2021; Chandrasekaran et al., 2022). destructive phenomenon that can result in significant structural failures
The transportation sector is critical to the oil and gas industry, with and economic losses, and it can occur either slowly or rapidly depending
pipelines carrying over 60 % of products, such as crude oil, natural gas, on the environmental conditions. Corrosion of pipelines is a significant
refined petroleum products, and slurry (El et al., 2020; Bagheri et al., threat that can lead to catastrophic environmental damages in the
2021). ocean, which most of these pollution forms remain untracebale and
Transmission pipelines are prone to various failures resulting from highly toxic, such as pollution of soil and sea, toxic releases, and fire and
various factors, including environmental parameters. These failures explosions (Wang et al., 2019; Chandrasekaran and Jain, 2017). The
have detrimental effects on both the environment and the economy. integrity and safety of oil and gas pipelines can greatly affect a country’s
Previous research has shown that corrosion is the most critical failure social and economic progress (Shaik et al., 2022). Furthermore, corro­
mechanism in transmission pipelines, which may occur either externally sion may lead to tragic accidents that threaten both human safety and
or internally (el Amine Ben Seghier et al., 2018; Keshtegar et al., 2019). the environment. For example, a corrosion leakage in 2013 caused the
According to the European Gas Pipeline Incident Data Group (EGIG), death of 62 people and incurred a loss of 118 million USD in China (Ma
nearly 27 % of gas pipeline incidents were caused by corrosion (Ma et al., 2023).
et al., 2023). Corrosion is a natural process leading to the degradation of Pipelines are vulnerable to various forms of corrosion, including
metallic structures, particularly those associated with oil and gas internal and external corrosion. Internal corrosion refers to corrosion

* Corresponding author.
E-mail address: [email protected] (R. Abbassi).

https://doi.org/10.1016/j.psep.2023.04.010
Received 20 February 2023; Received in revised form 19 March 2023; Accepted 4 April 2023
Available online 5 April 2023
0957-5820/© 2023 The Author(s). Published by Elsevier Ltd on behalf of Institution of Chemical Engineers. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
B. Akhlaghi et al. Process Safety and Environmental Protection 174 (2023) 320–327

inside a pipeline by corrosive fluids or gases such as CO2 and/or H2S model the corrosion phenomenon (Valor et al., 2013).
passing through. However, external corrosion is caused by a range of In recent years, researchers have developed machine learning algo­
environmental factors, including the presence of an acidic atmosphere rithms and the learning-based Bayesian network (LBN) models to
for above-ground pipelines, corrosive soils in buried pipelines, and address various research challenges in this field of research (Seghier
marine life and seawater temperature in submerged pipelines (Bachega et al., 2018; Jung et al., 2018; Kamil et al., 2021; Yazdi et al., 2022). It
Cruz et al., 2022; Taleb-Berrouane et al., 2021). Despite the use of High was demonstrated that the LBN model by utilizing microbiologically
Strength Steels (HSS), such as X80 steel, in the manufacturing of pipe­ influenced corrosion (MIC) data could accurately estimate likelihood
lines, external corrosion remains a concern (Arzaghi et al., 2020). despite the complex nature of microorganisms (Kamil et al., 2021; Yazdi
Studies have shown nearly 73 % of corrosion incidents are due to et al., 2022). In addition, the successful use of MLAs has drawn attention
external Corrosion (Lam and Zhou, 2016). Among external corrosion to their potential application in the development of chemical safety and
incidents, pitting corrosion is the most damaging form and can lead to health-related models (Seghier et al., 2018; Wang and Cai). These al­
other forms of external corrosion (Chen et al., 2013; Scully et al., 2008). gorithms have been utilized for a range of purposes, including predicting
Pitting corrosion is a form of low-temperature corrosion that occurs at properties and toxicity, predicting consequences, and detecting faults in
temperatures lower than 260 ℃ in the presence of water (Mollapour and various fields. Artificial neural networks (ANNs), a subset of MLAs, have
Poursaeidi, 2021). The corrosion process consists of three essential been widely used in academia and industry (Jung et al., 2018). Recent
components: (i) the anode, which represents the metal under corrosion, studies have focused on using hybrid and metaheuristic algorithms.
(ii) the electrolyte, which is often a corrosive medium that transfers Researchers (El et al., 2020) introduced three hybrid algorithms that
electrons from the anode to the cathode, and (iii) the cathode, which is combine SVR with particle swarm, genetic, and firefly algorithms to
an electrical conductor. Crude oil and natural gas can carry various predict the maximum pitting corrosion depth. They indicated that the
corrosive impurities under different circumstances. In the oil and gas SVR-firefly hybrid algorithm outperformed the other algorithms. Ma­
industry, these impurities may include free water, hydrogen sulfide chine learning and deep learning techniques have been widely applied
(H2S), and carbon dioxide (CO2) (Frankel and Sridhar, 2008). The in various contexts related to transmission pipes. In a recent article
decomposition of metals, resulting in the deterioration of mechanical (Eastvedt et al., 2022), the efficacy of machine learning algorithms in
properties, can lead to material depletion and possible failures in the oil detecting faults by analyzing the relationship between pressure change,
and gas industry (Sharland, 1987). Environmental conditions and soil velocity change, and crude oil temperature was investigated. The study
corrosivity are among the factors that influence the corrosion process. highlighted that machine learning offers a safe, accurate, and
Although various corrosion control strategies aim to prolong the lifespan cost-effective means of monitoring subsea pipelines, without requiring
of pipelines, they cannot completely prevent external localized corro­ the introduction of specialized equipment to the pipeline network. This
sion (Chandrasekaran et al., 2020; Chandrasekaran, 2020). The useful approach can hold significant promise for enhancing process safety and
life of pipelines is impacted when the mitigation strategies fail (e.g. safeguarding ocean environments. Eventually, the paper demonstrated
coating disbandment) and expose the pipelines to direct contact with the the potential contributions of machine learning in terms of safety, cost,
environment (Velázquez et al., 2009). Predictive corrosion models for and environmental protection to the pipeline industry. Moreover, ma­
oil and gas pipelines have gained importance in implementing preven­ chine learning and deep learning techniques have also been utilized in
tative technologies (passive and active protections) (Melchers, 2004). estimating the condition of multi-product pipelines, given the cata­
The initiation of new localized corrosion pitting is difficult to replicate strophic consequences of pipeline abnormalities, including economic
using mathematical formulations due to its stochastic nature. In prac­ losses, personal injury, and environmental degradation. Therefore,
tice, more focus is given to studying the growth of existing localized timely detection of such abnormalities is paramount (Seghier et al.,
corrosion rather than the initiation of new ones. 2018; Jung et al., 2018). In addition to transmission pipes, neural net­
Accurate prediction of the maximum depth of pitting corrosion in works and deep learning have been applied to predict the conditions of
transmission pipelines is critical for reducing the damages caused by this related structures, such as offshore platforms, using vibration response
mechanism, lowering maintenance costs, and minimizing the environ­ data. A recent study highlighted the application of deep learning
mental impacts of pipeline leaks due to failures (Keshtegar et al., 2019; methods in detecting damage to offshore platforms, even in noisy con­
Scully et al., 2008). The prediction of pitting corrosion depth can be ditions, using measured vibration responses from structures subjected to
utilized as a management tool to schedule maintenance activities. random excitations. The study demonstrated the promising potential of
Over the years, various studies have been conducted on predicting deep learning techniques in industrial process safety and operational
the maximum corrosion depth of pipelines using different methods. risk management (Bao et al., 2021).
Rossum (Sim et al., 2014) proposed an empirical formulation for the The present study adopts deep neural networks (DNNs) and deep
maximum depth of pitting corrosion (d) based on the Power-law model, learning algorithms to propose an accurate model for predicting the
expressed as follows: maximum pitting corrosion depth in oil & gas transmission pipelines.
The following sections comprehensively discuss these methods and their
d(T) = kT n (1)
performance improvement techniques. This study aims to improve and
where k and n are constant coefficients, and T is the pipeline construc­ preserve environmental quality and avoid heavy maintenance costs
tion duration in years. incurred on transmission pipelines in the oil and gas industry by using
The power law model has serious limitations in its ability to account different deep-learning models to design an accurate model for
for the impact of pipe coating and pipe-to-soil potential, as well as the maximum pit depth estimation.
potential effects of soil and pipe properties on both the initiation time
and exponent of pitting (Velázquez et al., 2009). The above equation 2. Developed methodology
was developed by adopting a new parameter (T0 ), referring to the pit
initiation time (Velázquez et al., 2009). Also, Alamilla et al (Alamilla The ANN is a data processing system based on the structure of the
and Sosa, 2008). proposed a new model to predict the pitting growth human brain, compromising a large number of interconnected pro­
rate in buried steel pipes using the Taylor series expansion. They re­ cessing elements (neurons) that work together to solve a problem
ported that the initial conditions and temporal corrosion rate could (Oostwal et al., 2021; Mohammadpour et al., 2022a). The ANN structure
affect the development and enhancement of corrosion. Since pitting generally comprises three parts: inputs, hidden, and output layers, as
corrosion is a stochastic process with the Markov property and inde­ illustrated in Fig. 1. In the neural network, the weights are critical and
pendent of the past states, researchers have adopted the Markov chain to are defined based on the significance of features in prediction by
transforming the input data into hidden layers. Activation functions,

321
B. Akhlaghi et al. Process Safety and Environmental Protection 174 (2023) 320–327

each batch to the system separately to accelerate calculations. The size

of data in each batch should be between 10 and 32. One epoch is defined
as the entire dataset being passed once through the neural network,
while iterations mean the number of batches required to complete one
epoch. At each stage of training, the prediction accuracy of the system is
tested after training. The training process stops when the system no
longer improves its prediction compared to the previous stage.

2.2. Activation function

Fig. 1. The schematic of the model of a neural network. Activation functions are selected based on the function of the
network. They can be altered in the process of designing and optimizing
such as Rectified Linear Unit (ReLU), Scaled Exponential Linear Unit the network by trial and error to achieve the best result (Oostwal et al.,
(SELU), and sigmoid, are used to activate or deactivate a neuron 2021).
(Oostwal et al., 2021). If the input data is greater than the thresholds, the Important activation functions are expressed by Eqs. (3)–(7):
neuron’s status is considered active, and data are sent to the next layer.
1
These hidden layers eventually calculate the output layer based on the Sigmoid : f (x) = (3)
1 + ex
weights. This process aims to determine a logical relationship between
inputs and outputs to reduce network errors. Modifying weights and {
0:x≤0
applying different methods, such as backpropagation, can help the Rectified Linear Unit(ReLU) : f (x) = (4)
x:x>0
network converge and reduce errors.
In this model, there are n binary inputs (usually given as a vector) {
α∗x :x≤0
and the same number of weights w1,…, wn, which are multiplied Leaky Rectified Linear Unit(LeakyReLU) : f (x) = (5)
x:x>0
together and summed up. It is noted that z in Eq. (2) is called the pre-
activation. Another term, called bias (b), is a constant factor (Ma and {
α(ex − 1) : x ≤ 0
Mei, 2021; Mohammadpour et al., 2022b). Leaky Rectified Linear Unit(ELU) : f (x) = (6)
x:x>0
∑n
z= x i wi + b (2) {
i=1 α(ex − 1) : x ≤ 0
Scaled Exponential Linear Unit(SELU) : f (x) = λ (7)
Neural networks propagate information through the use of non- x:x>0
linear activation functions and parameters {w, b}. Within each layer,
the neurons calculate a weighted sum of the input from the preceding where x is the function input, and f(x) is the activation function. The
layer by performing the operation wx + b. The resulting output is then hyperparameter of α is the negative slope coefficient, which is mainly set
passed through an activation function (Ma and Mei, 2021). between 0.1 and 0.3.
Neural networks are built through three main steps: training, vali­
dating, and testing. The design of neural networks involves determining 2.3. Vanishing or exploding gradient
the number of hidden layers and the number of neurons on each hidden
layer and selecting the activation function. In the training process, the The gradient decreases as it progresses from the end layers to the
output layer generates the ultimate results after the design of neural beginning layers, resulting in stopping weight updates in these layers
network layers is complete. The predicted results are then compared to and the non-optimal performance of the network training. The following
actual data in an iterative process to minimize the deviations. Deep methods are applied to solve such problems (Abuqaddom et al., 2021).
learning is a promising alternative to traditional machine learning
methods, operating based on multi-layer and deep neural networks 2.3.1. Drop-out
(Mohammed and Kora, 2021). Considering a neural network with a number of neurons, drop-out
means that during the training of these neurons, some of them are
2.1. Neural network architecture randomly ignored in the forward or return path with the aim of reducing
overfitting. Drop-out is a simple and practical technique for avoiding
In artificial neural networks, the weights are initially chosen over-fitting in deep neural networks. During each stage of training, each
randomly, and the search continues to find the optimal weights. The neuron has a probability (p) of being temporarily removed and not
most important parameters in the neural network are explained below. considered in the network, and then re-introduced. This way, at each
In most problems, the model is well-trained by including one hidden stage of training, a unique neural network is trained, allowing all the
layer. However, for problems requiring multiple hidden layers, a neurons to adapt to the data and have the desired contribution in
shallow model is initially considered to determine the number of hidden modeling the problem. After the training and during the execution of the
layers. If the training process is unsuccessful, a new hidden layer is model, the drop-outs of the model are removed. However, in the final
added, and the model is re-evaluated. This iterative process continues model, all the neurons are present during testing and execution.
until acceptable accuracy in prediction is reached, at least on the
training data. As the number of layers increases, the neural network 2.3.2. Adam’s optimization algorithm
becomes deeper. The decreasing gradient may remain in the local extremum or have a
The number of neurons in the input layer, hidden layers, and output low optimization speed. To address these challenges, the Adam opti­
layer depends on the type of problem and is determined according to the mization algorithm is an adaptive moment estimation (Fei et al., 2020),
number of features and labels. Another parameter is the learning rate of defined using Eqs. (8)–(12):
neural network architecture, which is usually taken as a large number. If
m = β1 m − (1 − β1 )∇θ J(θ) (8)
the system diverges, the learning rate is divided by three in each itera­
tion, and the model is re-evaluated to meet convergence criteria with the
s = β2 s + (1 − β2 )∇θ J(θ) ⊗ ∇θ J(θ) (9)
optimal learning value (Li et al., 2021).
It is recommended to divide the data into smaller batches and feed

322
B. Akhlaghi et al. Process Safety and Environmental Protection 174 (2023) 320–327

m
̂ =
m
(10) 2.5. Generalization and Generalization-Memorization models
1 − β1 t
Fig. 2 depicts two popular models utilized in deep learning: 1-
s
̂s = t (11) Generalization model, in which all neurons of one layer are connected to
1 − β2
all neurons of the next layer, and 2- Generalization-Memorization
√̅̅̅̅̅̅̅̅̅̅̅ model, in which all or some outputs of the last hidden layer are fed
θ = θ + ηm
̂ ∅ ̂s + ε (12) back to the input layer to facilitate the memorization process within the
network.
where β1 and β2 are parts of the architecture of neural networks. The
The processes considered in a deep neural network are shown in
value of e is 10− 7 , and s correctly detects the direction of movement.
Fig. 3.

2.4. Optimization
3. Comparative criteria
Since datasets include parameters with different orders, normaliza­
tion (xscaled ) and standardization (z) methods are applied to uniform the Three comparative criteria, including root mean square error
variation range of variables expressed and minimize the prediction (RMSE), mean absolute error (MAE), and mean squared error (MSE), are
error, respectively, as follows. adopted to determine the accuracy and performance of two models
(Generalization and Generalization-Memorization), as defined in Scully
x − xmin et al. (2008).
xscaled = (13)
xmax − xmin √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
1∑ n

where x is the initial data, xmin is the minimum value of data, and xmax is RMSE = (di − ̂ d i )2 (15)
n i=1
the maximum value in the dataset.
x− μ 1∑ n
z= (14) MAE = |di − ̂
di| (16)
σ n i=1
where μ is the average of data, and σ is the standard deviation of data. In
this study, both methods are adopted to obtain better results. 1∑ n
MSE = d i )2
(di − ̂ (17)
n i=1

where di and ̂ d i represent the actual and predicted values of pitting

corrosion depth, respectively. Also, n is the given number of samples.
It is worth noting that the minimum values of RMSE, MAE, and MSE
present more accurate performance.

4. Implementation of developed models

4.1. Data collection

Using relatively extensive and accurate databases is essential for

leveraging the full potential of ANNs to create high-accuracy models.
Table 1 presents a large database of information on pitting corrosion in
oil and gas pipelines. This database includes 259 maximum pit depths
that were gathered from coated steel pipes buried in various types of soil
over a period of approximately 23 years (Velázquez et al., 2009, 2010).
The types of soil in which the pipes were buried included clay, sandy
clay loam, clay loam, silty clay loam, silty clay, and silt loam. In addi­
tion, the database includes data on the time that each pipe was exposed

Fig. 3. Flowchart of deep learning models (Generalization and Generalization-

Fig. 2. Generalization and Generalization-Memorization models. Memorization model).

323
B. Akhlaghi et al. Process Safety and Environmental Protection 174 (2023) 320–327

Table 1 Table 2
Statistical attributes in the dataset (Velázquez et al., 2009, 2010). Defaults for model parameters.
Variable (unit) Xmin Xmax Algorithm Parameters Setting values

Max pit depth (mm) 0.41 13.44 GA (El et al., 2020) Population size 30
Exposure time (years) 5 50 Crossover’s probability 90 %
Resistivity (X-m) 1.9 399.5 Mutation’s probability 70 %
Water content (%) 8.8 66 Type of replacement Elitism
Sulfate (ppm) 0.99 1370.2 Type of selection Linear
Bicarbonate (ppm) 0.99 195.2 ranking
Chloride (ppm) 0.99 672.7 Max number of generation 30
pH 4.14 9.88 PSO (El et al., 2020) Number of particles 30
Pipe/soil potential, (V) -1.97 -0.42 Maximum number of 30
Bulk density (g/ml) 1.1 1.56 iterations
Redox potential (mV) 2.1 348 c1,c2 2.05
Coating type 0.3 1 ωMax 1.2
ωMin 0.1
FFA (El et al., 2020) Number of fireflies 30
to soil, the type of coating used on each pipe, and the pipe to soil po­ Maximum number of 30
tential. The database also provides data on eight important soil prop­ iterations
Alpha 0.5
erties, which are redox potential, pH, resistivity, water content, bulk Beta 4
density, chloride, bicarbonate, and sulfate content. The 259 datasets can Gamma 1
be categorized based on the type of soil, with 110 datasets from clay soil, Generalization model Number of hidden layers 4
79 datasets from clay loam soil, 61 datasets from sandy clay loam soil, Number of neurons 15
Epochs 1000
and the remaining datasets from silty clay and silty clay loam. It is worth
Drop-out rate 0.2
noting that Xmin and Xmax represent the minimum and maximum values Batch-size 20
of each variable. Activation function ELU
Optimization type MinMax
4.2. Empirical models scaler
Optimizer Adam
Generalization-Memorization Number of hidden layers 3
In recent studies, different methods and techniques have been used model
to estimate the maximum depth of corrosion depth in pipes. Mollapour Number of neurons 15
and Poursaeidi (2021), Frankel and Sridhar (2008). used a modified Epochs 1000
Drop-out rate 0.2
equation based on the Rossum model to estimate the maximum pit
Batch-size 20
depth. Arzaghi et al. (2020), Sharland (1987). attempted to estimate the Activation function RELU
long-term corrosion rate in the tubes using the calculation of the pitting Optimization type MinMax
corrosion rate and the pipe environmental conditions. In addition, El scaler
et al. (2020) used hybrid intelligent algorithms to estimate the Optimizer Adam

maximum depth of pitting corrosion in pipes. They combined the SVR

algorithm with PSO, FFA, and GA algorithms. The results of the present higher than that of the empirical models (Velázquez et al., 2009; Ala­
study will be compared with previous studies’ findings using the same milla and Sosa, 2008) and the hybrid models (SVR-GA, SVR-PSO, and
dataset in order to better demonstrate the accuracy of the developed SVR-FFA). The lowest MSE values predicted by the empirical models
model. (Velázquez et al., 2009) are 2.7445 and 1.1311 for training and vali­
dation data, respectively. The hybrid models perform better, with min­
4.3. Implementation of the developed models imum MSE values of 0.009 and 0.5588 for training and validation
datasets. MSE for the best case of the deep learning model is 0.0055 and
In this study, deep learning, as a branch of machine learning, is 0.0037 for training and validation data, respectively. Table 3 compares
utilized through Generalization and Generalization-Memorization Generalization and Generalization-Memorization models, which use
models. The adopted dataset is randomly split into 60 % data for different activation functions with/without dropout techniques.
training, 20 % for testing, and 20 % for validation. It is worth noting that According to Table 3, all three hybrid models of El et al. (2020)
the machine learning process is developed using Python, which is a accurately predict the maximum pitting corrosion depth. However,
widely used programming language for this purpose. Table 2 lists the there is an overfitting issue in the comparison of their results in the
control parameters of the developed hybrid models, including SVR-GA, training and test datasets. This research uses the drop-out and Adam
SVR-PSO, SVR-FFA, and deep learning models (Generalization and techniques to solve this problem. The process of training and reducing
Generalization-Memorization models). the error can be seen in Fig. 4. In each phase of the training, the accuracy
of the prediction system is tested after training. The training is stopped
5. Comparative predicted results of models when there is no improvement in forecasting compared to the previous
stage. The figure demonstrates accurate trends without overfitting, as
Generalization and Generalization-Memorization models are adop­ the behavior is similar for both models.
ted to model deep learning. By considering network architectures de­ Fig. 5 displays the residuals for the training and testing datasets.
signs with varying numbers of hidden layers and activation functions, as Since residuals are defined as the difference between actual values and
well as the use of Dropout and Adam techniques, the prediction accuracy predicted results (Residual = yactual − ypredicted), the positive residuals
of the Generalization model in terms of MSE is determined as 0.0056 and show that the predicted data are slightly lower than real values. In
0.0610 for training and validation data, respectively. The best predic­ contrast, the negative residuals are related to overestimation. It is
tion accuracy for Generalization-Memorization modeling, measured in evident that highly accurate prediction results in zero difference. Fig. 5
terms of MSE, is found to be 0.0055 for both training and validation data presents an even distribution of residuals for both deep learning models
after multiple steps of deep neural network design. considered in this study, with a tendency to cluster towards the middle
As demonstrated in Table 3, the accuracy of Deep Learning models of the scatter plot. It affirms the accuracy of the prediction without
and DNNs for predicting the maximum pitting corrosion depth is much

324
B. Akhlaghi et al. Process Safety and Environmental Protection 174 (2023) 320–327

Table 3
The comparative statistical indicators for different models.
Models MSE RMSE MAE

Train Test Train Test Train Test

Alamilla and Sosa (2008) 34.1810 37.2137 5.6998 5.8525 4.3805 4.8449
Velázquez et al. (2009) 2.7445 1.1311 1.1957 2.3153 0.7011 1.2165
El et al. (2020) (SVR) 0.0122 1.3533 0.1107 1.1633 0.0972 0.8925
El et al. (2020) (SVR-GA) 0.0097 0.7038 0.0988 0.8389 0.0955 0.5848
El et al. (2020) (SVR-PSO) 0.0111 0.6403 0.1057 0.8002 0.0986 0.5719
El et al. (2020) (SVR-FFA) 0.009 0.5588 0.0949 0.2909 0.0935 0.2359
Present models Generalization 0.0056 0.0037 0.0748 0.0610 0.0559 0.0591
Generalization-Memorization 0.0055 0.0055 0.0741 0.0743 0.0695 0.0645

Fig. 5. Partial residual plot in Generalization and Generalization-

Memorization model.
Fig. 4. The process of training and reducing the error in models.
various parameters of buried transmission pipelines and the surrounding
outliers. soil. The following are conducting remarks of this study:
Fig. 6 compares the actual maximum pit depth values against the
predicted results by the Generalization model and Generalization- • The results indicated that deep learning has high accuracy and effi­
Memorization model. An excellent agreement in the figure confirms ciency in predicting the maximum pitting corrosion depth. This
the high accuracy of the present model shown in Table 3. In addition, the model not only improved the prediction accuracy for the training
figure shows no overfitting problem in predicting the maximum pit dataset but also resolved the overfitting problem of the SVR-FFA
depth due to the different techniques adopted in this research. algorithm, resulting in increased accuracy for the test dataset.
• The results show that deep learning models far outperformed other
6. Conclusion models used in previous studies to predict the maximum pitting
corrosion depth of oil & gas transmission pipelines. The best case of
The ability to predict the maximum pitting corrosion depth in oil and the deep learning model has MSE values of 0.0055 and 0.0037 for
gas transmission pipelines can provide a valuable tool for cost preven­ training and validation data, respectively.
tion and environmental protection in the oil and gas industry. In this
study, deep learning models are used to accurately predict the maximum
pit depth. A comprehensive and real-world dataset is utilized, including

325
B. Akhlaghi et al. Process Safety and Environmental Protection 174 (2023) 320–327

Fig. 6. Comparison of actual data and predicted results.

Declaration of Competing Interest Fei, Z., Wu, Z., Xiao, Y., Ma, J., He, W., 2020. A new short-arc fitting method with high
precision using Adam optimization algorithm. Optik 212, 164788. https://doi.org/
10.1016/j.ijleo.2020.164788.
The authors declare that they have no known competing financial Frankel, G.S., Sridhar, N., 2008. Understanding localized corrosion. Mater. Today 11
interests or personal relationships that could have appeared to influence (10), 38–44. https://doi.org/10.1016/S1369-7021(08)70206-2.
the work reported in this paper. Jung, H., Jo, H., Kim, S., Lee, K., Choe, J., 2018. Geological model sampling using PCA-
assisted support vector machine for reliable channel reservoir characterization.
J. Pet. Sci. Eng. 167 (April), 396–405. https://doi.org/10.1016/j.
References petrol.2018.04.017.
Kamil, M.Z., Taleb-Berrouane, M., Khan, F., Amyotte, P., 2021. Data-driven operational
Abuqaddom, I., Mahafzah, B.A., Faris, H., 2021. Oriented stochastic loss descent failure likelihood model for microbiologically influenced corrosion. Process Saf.
algorithm to train very deep multi-layer neural networks without vanishing Environ. Prot. 153, 472–485. https://doi.org/10.1016/j.psep.2021.07.040.
gradients. Knowl. Based Syst. 230, 107391 https://doi.org/10.1016/j. Keshtegar, B., ben Seghier, M.E.A., Zhu, S.P., Abbassi, R., Trung, N.T., 2019. Reliability
knosys.2021.107391. analysis of corroded pipelines: novel adaptive conjugate first order reliability
Alamilla, J.L., Sosa, E., 2008. Stochastic modelling of corrosion damage propagation in method. J. Loss Prev. Process Ind. 62 (May), 103986 https://doi.org/10.1016/j.
active sites from field inspection data. Corros. Sci. 50 (7), 1811–1819. https://doi. jlp.2019.103986.
org/10.1016/j.corsci.2008.03.005. Lam, C., Zhou, W., 2016. Statistical analyses of incidents on onshore gas transmission
el Amine Ben Seghier, M., Keshtegar, B., Elahmoune, B., 2018. Reliability analysis of low, pipelines based on PHMSA database. Int. J. Press. Vessels Pip. 145, 29–40. https://
mid and high-grade strength corroded pipes based on plastic flow theory using doi.org/10.1016/j.ijpvp.2016.06.003.
adaptive nonlinear conjugate map. Eng. Fail. Anal. 90 (February), 245–261. https:// Li, Y., Zhang, Q., Yoon, S.W., 2021. Gaussian process regression-based learning rate
doi.org/10.1016/j.engfailanal.2018.03.029. optimization in convolutional neural networks for medical images classification.
Arzaghi, E., Chia, B.H., Abaei, M.M., Abbassi, R., Garaniya, V., 2020. Pitting corrosion Expert Syst. Appl. 184, 115357 https://doi.org/10.1016/j.eswa.2021.115357.
modelling of X80 steel utilized in offshore petroleum pipelines. Process Saf. Environ. Ma, H., Zhang, W., Wang, Y., Ai, Y., Zheng, W., 2023. Advances in corrosion growth
Prot. 141, 135–139. https://doi.org/10.1016/j.psep.2020.05.024. modeling for oil and gas pipelines: a review. Process Saf. Environ. Prot. 171, 71–86.
Bachega Cruz, J.P., et al., 2022. Uniform corrosion assessment in oil and gas pipelines https://doi.org/10.1016/j.psep.2022.12.054.
using corrosion prediction models – Part 1: models performance and limitations for Ma, Z., Mei, G., 2021. Deep learning for geological hazards analysis: data, models,
operational field cases. Process Saf. Environ. Prot. 167, 500–515. https://doi.org/ applications, and opportunities. Earth Sci. Rev. 223, 103858 https://doi.org/
10.1016/j.psep.2022.09.034. 10.1016/j.earscirev.2021.103858.
Bagheri, M., Zhu, S.P., ben Seghier, M.E.A., Keshtegar, B., Trung, N.T., 2021. Hybrid Melchers, R.E., 2004. Pitting Corrosion of Mild Steel in Marine Immersion
intelligent method for fuzzy reliability analysis of corroded X100 steel pipelines. Environment—Part 2: Variability of Maximum Pit Depth. October 2004.
Eng. Comput. 37 (4), 2559–2573. https://doi.org/10.1007/s00366-020-00969-1. Mohammadpour, J., Ghouchani, Z., Salehi, F., Lee, A., 2022a. ANN and CFD-DPM
Bao, X., Fan, T., Shi, C., Yang, G., 2021. Deep learning methods for damage detection of modeling of alumina-water nanofluid heat transfer in a double synthetic jet
jacket-type offshore platforms. Process Saf. Environ. Prot. 154, 249–261. https://doi. microchannel. In: Ferreira Martins, M., Ramos, R., Belich, H. (Eds.), Multiphase Flow
org/10.1016/j.psep.2021.08.031. Dynamics: A Perspective from the Brazilian Academy and Industry. Springer
Bhandari, J., et al., 2017. Accelerated pitting corrosion test of 304 stainless steel using International Publishing, Cham, pp. 231–241. https://doi.org/10.1007/978-3-030-
ASTM G48; experimental investigation and concomitant challenges. J. Loss Prev. 93456-9_21.
Process Ind. 47, 10–21. https://doi.org/10.1016/j.jlp.2017.02.025. Mohammadpour, J., Husain, S., Salehi, F., Lee, A., 2022b. Machine learning regression-
Bhandari, J., Khan, F., Abbassi, R., Garaniya, V., Ojeda, R., 2015. Modelling of pitting CFD models for the nanofluid heat transfer of a microchannel heat sink with double
corrosion in marine and offshore steel structures – a technical review. J. Loss Prev. synthetic jets. Int. Commun. Heat Mass Transf. 130, 105808 https://doi.org/
Process Ind. 37, 39–62. https://doi.org/10.1016/j.jlp.2015.06.008. 10.1016/j.icheatmasstransfer.2021.105808.
Chandrasekaran, S., 2020. Design of Marine Risers with Functionally Graded Materials. Mohammed, A., Kora, R., 2021. An effective ensemble deep learning framework for text
Woodhead Publishing. classification. J. King Saud. Univ. - Comput. Inf. Sci. https://doi.org/10.1016/j.
Chandrasekaran, S., Jain, A.K., 2017. Ocean Structures: Construction, Materials, and jksuci.2021.11.001.
Operations. Crc Press. Mollapour, Y., Poursaeidi, E., 2021. Experimental and numerical analysis of pitting
Chandrasekaran, S., Sricharan, V.V.S.%J.O.E., 2021. Numerical study of bean-float wave corrosion in CUSTOM 450 stainless steel. Eng. Fail. Anal. 128 (March), 105589
energy converter with float number parametrization using WEC-Sim in regular https://doi.org/10.1016/j.engfailanal.2021.105589.
waves with the Levelized Cost of Electricity assessment for Indian sea states. vol. Oostwal, E., Straat, M., Biehl, M., 2021. Hidden unit specialization in layered neural
237, p. 109591. networks: ReLU vs. sigmoidal activation. Phys. A: Stat. Mech. Appl. 564, 125517
Chandrasekaran, S., Hari, S., Amirthalingam, M., J.M.S. Engineering, A., 2020. Wire arc https://doi.org/10.1016/j.physa.2020.125517.
additive manufacturing of functionally graded material for marine risers. vol. 792, p. Scully, J.R., Budiansky, N.D., Tiwary, Y., Mikhailov, A.S., Hudson, J.L., 2008. An
139530. alternate explanation for the abrupt current increase at the pitting potential. Corros.
Chandrasekaran, S., Khan, F., Abbassi, R., 2022. Wave Energy Devices: Design, Sci. 50 (2), 316–324. https://doi.org/10.1016/j.corsci.2007.08.002.
Development, and Experimental Studies. CRC Press. Seghier, M., Keshtegar, B., Correia, J., de Jesus, A., Lesiuk, G., 2018. Structural reliability
Chen, A., et al., 2013. Study of pitting corrosion on mild steel during wet-dry cycles by analysis of corroded pipeline made in X60 steel based on M5 model tree algorithm
electrochemical noise analysis based on chaos theory. Corros. Sci. 66, 183–195. and Monte Carlo simulation. Procedia Struct. Integr. 13, 1670–1675. https://doi.
https://doi.org/10.1016/j.corsci.2012.09.017. org/10.1016/j.prostr.2018.12.349.
Eastvedt, D., Naterer, G., Duan, X., 2022. Detection of faults in subsea pipelines by flow Shaik, N.B., et al., 2022. Recurrent neural network-based model for estimating the life
monitoring with regression supervised machine learning. Process Saf. Environ. Prot. condition of a dry gas pipeline. Process Saf. Environ. Prot. 164, 639–650. https://
161, 409–420. https://doi.org/10.1016/j.psep.2022.03.049. doi.org/10.1016/j.psep.2022.06.047.
El, M., et al., 2020. Prediction of maximum pitting corrosion depth in oil and gas Sharland, S.M., 1987. A review of the theoretical modelling of crevice and pitting
pipelines. Eng. Fail. Anal. 112 (March), 104505 https://doi.org/10.1016/j. corrosion. Corros. Sci. 27 (3), 289–323. https://doi.org/10.1016/0010-938X(87)
engfailanal.2020.104505. 90024-2.

326
B. Akhlaghi et al. Process Safety and Environmental Protection 174 (2023) 320–327

Sim, S., Cole, I.S., Choi, Y.S., Birbilis, N., 2014. A review of the protection strategies Velázquez, J.C., Caleyo, F., Valor, A., Hallen, J.M., 2010. Field study—pitting corrosion
against internal corrosion for the safe transport of supercritical CO2 via steel of underground pipelines related to local soil and pipe characteristics. Corrosion 66
pipelines for CCS purposes. Int. J. Greenh. Gas Control 29, 185–199. https://doi.org/ (1) (pp. 016001–016001-5).
10.1016/j.ijggc.2014.08.010. Wang, Q., Cai, C. Machine Learning in Chemical Safety and Health.
Taleb-Berrouane, M., Khan, F., Hawboldt, K., 2021. Corrosion risk assessment using Wang, Y., Zhang, P., Qin, G., 2019. Non-probabilistic time-dependent reliability analysis
adaptive bow-tie (ABT) analysis. Reliab Eng. Syst. Saf. 214, 107731 https://doi.org/ for suspended pipeline with corrosion defects based on interval model. Process Saf.
10.1016/j.ress.2021.107731. Environ. Prot. 124, 290–298. https://doi.org/10.1016/j.psep.2019.02.028.
Valor, A., Caleyo, F., Alfonso, L., Velázquez, J.C., Hallen, J.M., 2013. Markov chain Yazdi, M., Khan, F., Abbassi, R., 2022. Operational subsea pipeline assessment affected
models for the stochastic modeling of pitting corrosion. Math. Probl. Eng. 2013 by multiple defects of microbiologically influenced corrosion. Process Saf. Environ.
https://doi.org/10.1155/2013/108386. Prot. 158, 159–171. https://doi.org/10.1016/j.psep.2021.11.032.
Velázquez, J.C., Caleyo, F., Valor, A., Hallen, J.M., 2009. Predictive model for pitting
corrosion in buried oil and gas pipelines. Corrosion 65 (5), 332–342. https://doi.
org/10.5006/1.3319138.

327