Unlocking Corrosion Secrets: A Comprehensive Analysis of Acoustic Emission Data Using Continuous Wavelet Transform and Deep Learning for Steel Pipeline Monitoring

doi:10.21203/rs.3.rs-3571590/v1

Download PDF

Research Article

Unlocking Corrosion Secrets: A Comprehensive Analysis of Acoustic Emission Data Using Continuous Wavelet Transform and Deep Learning for Steel Pipeline Monitoring

https://doi.org/10.21203/rs.3.rs-3571590/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

The analysis of acoustic emission data in the time and frequency domains can provide useful insights about the state of steel pipelines, although research in this field is limited. The research shortage has been triggered due to specific problems provided by elements such as irregular structure geometries, material dependencies, and insufficient training data. To address the lack of training data, we created a new dataset of scalogram images for deep learning-based classification approaches. Monitoring a 500-meter-long steel oil and gas pipeline provided experimental data. The acoustic emission waveforms were preprocessed, tagged, and classified based on three critical characteristics: AE-mean, kurtosis, and amplitude. The continuous wavelet transform was used to translate these waveform instances into the time-frequency domain. To. The abstract should be an objective representation of the article and it must not contain results that are not presented and substantiated in the main text and should not exaggerate the main conclusions. State-of-the-art deep convolutional neural networks were used as benchmarks to assess the effectiveness of the proposed dataset. Surprisingly, the unique dataset achieved a classification accuracy of 91.0%. Furthermore, the impact of initial learning rate and L2 regularization hyperparameters were evaluated for EfficientNet-b0. The availability of this dataset opens the door to a plethora of corrosion detection applications, providing substantial prospects in this field.

Deep learning

Feature extraction

Machine learning

Corrosion classification

Acoustic emission

Automatic damage identification and diagnostic is an effective tool to make sure that the industrial processes such as machines, chemical operations, pipeline transportation, and software systems are running smoothly. The characteristics of the damage can be found by figuring out how the damage and the operation parameters are linked. There is a dire need to assess the initiation and expansion of such damages in these engineering applications under strict safety [1]. These applications are being monitored regularly by Non-destructive testing (NDT) methods. Acoustic emission (AE) is one of the most common NDT methods because it has a high degree of sensitivity and is easy to configure without having to be involved in the system. AE sensors record the AE signals that follow damage initiation or the existing damage propagation events within the materials and transform them into electric waveforms. The study of these AE waveforms can help in understanding the damage characteristics. The AE technique has the capacity for global monitoring without any external excitation [2][3]. As a non-destructive testing and tool condition monitoring method, AE lacks a one-size-fits-all formula or set of models that can be applied to every application scenario, especially about different materials and failure mechanisms. Various types of damage sources (pencil lead-break, impact, fatigue) have been localized as artificial AE-source by the researchers [4]. There is a need to develop advanced Structural Health Monitoring (SHM) methods for automatic characterization of corrosion damage in steel pipelines.

In the present era, researchers are focused on developing automated monitoring techniques using experimentally measured structural responses [5][6]. Among those techniques, image-based SHM has shown potential for rapid inspection [7][8]. Efficiency of classification algorithms is heavily dependent on getting access to high-quality datasets with a reasonable set of classes [9]. This holds true, in particular, for significant achievements in deep convolutional neural networks (CNNs) [10], where ample data is needed to train the network. The necessity for substantial computing resources, particularly during training, is a challenge in the implementation of CNN. The growing complexity of CNN structures (in terms of layers) adds to this need.

Transfer learning approaches are used to retrain CNNs that have been learnt in one domain to be able to function on different datasets. The main advantage is that the model may be reused to some extent, with various situations either to retrain the network from scratch or to retrain merely a few selected components of the network. Even though, deep-learning networks achieved outstanding results in other research areas there are currently no large-scale AE datasets that can be used for monitoring of steel pipelines, which restricts the application of deep-learning technology and the wider development of intelligence in this area. These situations highlight the importance of utilizing diverse collection devices for deeper and larger networks in order to provide the best results possible.

The objective of this paper is to provide support to open research in the field of SHM to identify corrosion damages:

To propose and develop a publicly available dataset for corrosion detection and present the estimations of classification accuracy of this dataset.

Evaluate the results of the classification from different networks in terms of consistency.

The structure of the document is as follows. The classification of AE scalogram domain is briefly examined in Section 2, and the CNN evolution is investigated in Section 2.2. Section 2.3 describes transfer learning, including the hyperparameters involved. The methods used are discussed in Section 3. In Section 3.1, the workflow is presented. Section 3.2 includes the AE dataset preprocessing steps towards intermediate formats that can be fed into the CNNs for training. The hyperparameters for retraining and their values are presented in Section 3.3. The classification accuracy and time of the preprocessing and training processes are evaluated in Section 4. The observations based on the results are discussed in Section 5, while the conclusions and recommendations are discussed in Section 6.

AE sensors record the AE signals that follow damage initiation or the existing damage propagation events within the materials and transform them into electric waveforms. The study of these AE waveforms can help in understanding the damage characteristics.

Traditional machine learning approaches are used in AE research for the classification of damages [11][12][13], which requires the extraction of AE features that are given as input to the algorithms. Features are based on time domain properties of AE waveform such as Amplitude, Energy and Duration, while frequency features [14] are also extensively employed.

Recently, deep learning techniques have been employed for the damage identification using AE technique [15][16]. Chen et al. [17] presented a transfer learning approach for evaluating structural conditions of rail in a progressive manner, by using acoustic emission monitoring data and knowledge transferred from an acoustic-related database. The proposed CNN model was used to lower-layer knowledge from a pre-trained AudioSet model, to help in extracting the acoustic-specific features of the time–frequency spectrograms. The proposed approach outperformed the existing techniques by a high macro-F1 score of 97.5% and relatively short computation time. Xin et al. [18] proposed a novel binary AE signal classification framework to identify the fracture within stay-cables, using a TL CNNs based on GoogLeNet. The proposed model effectively identified noise & fracture AE signals in static tests of a stay-cable with an overall accuracy (99.05%). An autoencoder for anomaly detection and a convolutional neural network for classification were examined in a novel two-staged strategy by König et al. [19]. They used GoogleNet to detect three different wear failure scenarios in sliding bearings: running-in, insufficient lubrication, and particle-contaminated fluid. The proposed method's overall accuracy was 82.5%. Hesser et al. [20] used the pretrained VGG16 architecture is used with minor modifications such as the parameters of the last convolutional layer, the two fully connected layers and the output layer. Furthermore, the architecture has also been modified to produce and detect the three kinds of AE sources. During the training with the AE images, all the remaining settings in the previous layers are blocked and can't be adjusted. Conv2D, which uses deep transfer learning to efficiently train the architecture, achieves the best outcome of 100%. Ai et al. [21] observed the Alkali-silica reaction (ASR) in a concrete specimen reinforced with steel rebars. The recorded AE signals were filtered and split into two categories. Two deep learning algorithms convolutional neural network (CNN) and stacked autoencoder and two shallow machine learning networks SVM and KNN were employed to classify the AE signals into the corresponding category. The basic GoogLeNet was modified by extending the number of layers up to 22. CNN using the data from a single sensor outperformed other networks in terms the classification accuracy with the highest accuracy (85.2%). Li et al. [22] compared the classification accuracies of three CNN models, including one AlexNet model input with the SWT plots of 0–800 kHz, one AlexNet model input with the SWT plots of 200–700 kHz, and the proposed MBCNN to classify different types of AE waves in rail crack monitoring. The testing accuracy of 99.52% was achieved by the proposed MBCNN.

The difference between traditional and deep learning approaches is that the former is based on the extraction of a set of values from the time, frequency, and perceptual domains to be fed into the ML algorithms, whereas, in deep learning, the CNNs transform their representation more accurately. In this study, we used a different configuration of transfer learning on a set of CNNs and compared the performance in terms of classification accuracy and training time.

In this section, the physical properties of the gas pipeline, experimental settings for AE data collection, specification of the data acquisition systems and methodology for classification of the AE signals are described in detail. The detailed diagram of the whole analysis has been described in Fig. 1.

3.1 Experimental Setup for Data Acquisition

AE data collected from an experimental setup that has been mentioned in detail [23].

3.2 Data Fusion

AE signals are collected from the pipeline using two separate systems synchronized.

3.3 Wavelet Denoising

Wavelet denoising based on Hybrid Particle Swarm Optimization (HPSO) algorithm for the selection of noise threshold was also performed and used the wden function from the Matlab Wavelet Toolbox. Specific parameters have been set: HPSO threshold selection [23] is used to select the wavelet coefficients in combination with a soft thresholding being rescaled using level dependent estimation of level noise. Decomposition is made at level 6 with db5 as the mother wavelet.

3.4 Feature Selection

The SelectKBest (SKB) algorithm is used for selecting the most significant features in the dataset [24]. The corrosion AE signatures obtained from previous work[25]. The ${Chi}^{2}$ function is used by the SelectKBest method to estimate the importance of features in classification methods, that is given by Eq. 1:

$${\text{Ӽ}}_{\varvec{c}}^{2}= \sum \frac{{({\varvec{O}}_{\varvec{j}}-{\varvec{ϵ}}_{\varvec{j}})}^{2}}{{\varvec{ϵ}}_{\varvec{j}}}$$

Where c is the number of classes, ${\varvec{O}}_{\varvec{j}}$ is the observed value and ${\varvec{ϵ}}_{\varvec{j}}$ is the estimated value. The top-most three features AE-mean, Kurtosis, and Amplitude were selected as the significant features for labelling the experimental data. AE-mean secured the highest score with 4.422. Kurtosis, and Amplitude stood at 2nd and 3rd position by securing scores of 3.903 and 2.147 respectively. The results of the SelectKBest (SKB) algorithm have been given in Table 1.

3.5 Random Under Sampling (RUS)

Class imbalance is a method in which models tend to be biased towards the dominant class when the training data is unbalanced. It is also known as the long-tailed data distribution problem [11]. Acevedo [9] demonstrated that the model's performance is dependent on the balance of the training set of blood cells. Therefore, achieving an effective classification with imbalanced data is crucial [12].

3.6 Data Transformation

Wavelets Transform (WT) is a time–frequency analysis method for capturing the time–frequency features of non-stationary signals like AE signals [26]. Continuous Wavelet

Transform (CWT) has demonstrated excellent performance in the field of signal processing, both in terms of time and frequency [27]. The scalogram image is the visual representation of continuous wavelet coefficients. Two-dimensional scalogram images serve as input for CNN models.

Morlet wavelet is one of commonly used mother wavelets, and it allows for the measurement of the localization frequency for signals with quicker and slower oscillations, resulting in a flexible window that narrows at high frequencies and widens at low frequencies, respectively. In this study, a CWT with source wavelet Morlet (parameter = 8) was applied to the pipeline AE signals and AE images are obtained, which can be expressed mathematically by Eq. 2.

$$\varPsi \left(\tau \right)={e}^{i{\omega }_{0}\tau }{e}^{\frac{{-\tau }^{2}}{\left(2{\vartheta }^{2}\right)}}$$

In frequency domain, the dilated version of Morlet wavelet is given by Eq. 3.

$$\varPsi \left(\text{Ꙍ}\right)= {\vartheta \sqrt{2\pi e}}^{\frac{{-(\mathfrak{a}\omega -\omega 0)}^{2}{\vartheta }^{2}}{2}}$$

Where $\vartheta$ and $\omega 0$ are constants; The wavelet center frequency is denoted by $\omega 0$.

Only the object detection data sets have been tested using the current methodologies. On a texture-based data collection, it's unclear how they function. The proposed dataset AE_CORBI has been assessed by several sample damage classification algorithms in this section.

4.1 AlexNet

AlexNet was the convolutional network that won the ImageNet Challenge in 2012, a year after LeCun's handwritten digit recognition network [28] set a new benchmark for the development of fast CNNs. AlexNet not only includes more layers, but it also employs Rectified Linear Units (ReLUs) as activation functions instead of the sigmoid or tanh, resulting in speedier training. The size of the input image is 227 × 227 × 3. AlexNet has eight layers, five of which are convolutional and three of which are fully connected. There are 60 million parameters in AlexNet [10].

4.2 VGG Group

The VGG group of neural networks was introduced in 2014 as an advancement of AlexNet, in which the ReLU activation function have been employed, but with smaller receptive fields (3 × 3 instead of 11 × 11 and 5 × 5 in AlexNet) with a fixed stride of 1 resulting in improved performance [29]. The most popular are VGG-16 and VGG-19, which have three completely connected layers and 13 and 16 convolutional layers, respectively. Both have the same input size of 224 × 224 × 3, with VGG16 having 138 million parameters and VGG-19 having 144 million.

4.3 GoogleNet

GoogleNet (Inception-v1) received the best performance in ImageNet Challenge in the same year. It includes nine inception modules with 22-layers and 7 million parameters. The originality of this (a) (b) Fig. 3. Scalogram of the (a) Corroded (b) non-corroded class approach is the parallel application of various-sized filters, with the outputs integrated into an inception module. A 1 × 1 convolutional layer is also used by the Inception module, resulting in lowering the cost of calculation. Finally, the number of parameters is considerably reduced by performing average pooling before the classification layer [30]. Its image input size is 224 × 224 × 3. The design of the inception was improved in 2015 with convolution factorization, resulting in Inception-v2 with 42 layers depth, and auxiliary classifiers with batch normalization, leading to Inception-v3 [31]. The input image size for Inception-v3 is 299 × 299 × 3. It includes a total of 48 layers and 23.9 million parameters.

4.4 ResNets

We selected ResNet18 and ResNet50 because they are lightweight networks with 18, 50 and 101 layers, respectively. By using ResNet-based deep learning methods, high success rates have been achieved. ResNet18, ResNet50, and ResNet101 have 11, 25.6, and 44.5 million parameters, respectively. By bypassing nonlinear levels and fitting identity mappings with residual blocks, ResNets avoid the growing training error as the number of layers grows [32]. ResNets input image size is 224 × 224 × 3.

4.5 NASNet-Mobile

NASNet stands for neural architecture search network. It was developed by Google's brain team, the NASNet offers two primary functions such as normal cell and reduction cell. Initial operations on the tiny dataset are performed by NASNet before being transferred to the larger dataset, which has a greater mean Average Precision. Scheduled drop-path for effective regularization is used in NASNet to increase efficiency [33]. NASNet-Mobile which uses 224 × 224 x 3 as input image size, contains 913 layers, while and NASNet-Large model uses 331 × 331 x 3 as the input image size.

4.6 Xception

The Extreme Inception convolutional neural network, is an improved version of Inception-v3 by replacing the inception modules with depth-wise separable convolution layers, resulting in improved accuracy [34]. Xception takes in 299 × 299 RGB images, is 71 layers deep, and has 22.9 million parameters.

4.7 ShuffleNet

ShuffleNet contains 50 layers, and 1.4 million parameters with an image size of 224 × 224 × 3. Channel shuffling and pointwise convolution are employed in this architecture, resulting in improved accuracy and reduced computing cost in small networks [35].

4.8 MobileNet

Google introduced MobileNet for mobile and embedded applications in 2017 to limit the demand on resources [36]. By decreasing the amount of memory required while retaining the same degree of accuracy, MobileNet-v2 was developed in 2018 and quickly became the standard. Shortcut connections between linear bottlenecks and an inverted residual structure were used to make this happen [37]. It comprises 53 layers, 3.5 million parameters, and accepts images with dimensions of 224 × 224 × 3.

4.9 EfficientNet-b0

EfficientNet, as demonstrated by Google AI research, is a collection of CNN models that, with slight adjustments, outperforms its predecessors [38]. EfficientNet is a family of models derived from a baseline model, which was developed through a neural architecture search [39]. The neural architecture search is a framework for finding the optimal architecture in a given search space while keeping a set of constraints in mind. Sandler et al. [37] [40] proposed the mobile inverted bottleneck (MBConv) layer as the network's basic building element. EfficientNet has 8 variations, ranging from B0 to B7. Each following model number refers to versions with more parameters and improved precision. It comprises 290 layers, 4 million parameters, and accepts images with dimensions of 224 × 224 ×3.

We used the AE scalograms dataset to compare the performance of the existing state-of-the-art deep learning algorithms. Classification is used to highlight the results.

Resizing is the process of changing the dimensions of an image to the network. Before training the network, the input image size diversified, we resize these images according to the input dimensions required by the network. To make a comprehensive evaluation, 80–20% training–testing ratios are considered: the data set was randomly divided into 80% for training and 20% for testing (26,343 training samples and 6,586 testing samples per class). Out of training data 20% was reserved for validation set. The experiments with ten deep-learning classification architectures including VGG16, VGG19, DarkNet19, DarkNet DarkNet53, ShuffleNet, GoogleNet, NASNet-mobile, SqueezeNet, InceptionV3, MobileNet-v2, Densenet201, Resnet18, Resnet50, Resnet101—were carried out. All the models were trained and tested on a laptop having the following configuration: Intel Core i7 2.80 GHz processor with NVIDIA GeForce GTX 1070 GPUs and 16 GB RAM. MATLAB version R2020b was used in this experiment.

Thereafter, the evaluation is carried out on the validation set to minimize overfitting. The unknown testing set is used to evaluate the performance once the training procedure and parameter selection have been completed. Instead of randomly selecting photos from the classes, uniform sampling was used to support training batches of 128, 64, and 16. With a learning rate of 0.0001, decaying exponentially by 0.1% every 50 iterations, and Stochastic Gradient Descent optimization with a validation patience 5.

5.1 Performance Comparison

Image classification provides three well accepted standard evaluation metrics: overall accuracy, average accuracy, and a confusion matrix. The performance evaluation metrices derived from the confusion metrics are given by Equation 4 to 6. The sensitivity of a classification model can be defined as the percentage of true cases (TP), which are correctly classified cases as defined by Equation 4 The classification model's specificity relates to the percentage of true non-cases, or AE signals from corrosion sources that have been classified as non-corrosion sources. The specificity can be expressed by Equation 4. F1 score as represented by Equation 6, was used for the comparison of two classification models. A weighted harmonic average of precision and recall values could be used to calculate F1 score [18]. The precision was defined as the likelihood that a randomly chosen AE signal processed as a corrosion case was the corrosion case. The recall, on the other hand, was defined as the chance that a randomly picked corrosion instance was accurately identified as such. Because there was no historical data on precision or recall, the beta value was set to 1, and the F-score was termed the F1-score

Table 2. Accuracy measures of the deep learning classification models

Model	Precision	Recall	F1-score	Accuracy (%)
NASNet-Mobile	0.83	0.94	0.88	87.2
ResNet-50	0.84	0.90	0.87	86.4
Inception-V3	0.89	0.90	0.89	89.2
ResNet-18	0.82	0.94	0.88	86.6
MobileNet-v2	0.89	0.92	0.91	90.3
ShuffleNet	0.87	0.91	0.89	88.8
Xception	0.85	0.91	0.88	87.6
EfficientNet-b0	0.88	0.95	0.91	91.0

$Precision = \frac{TP}{TP+FP}$

(4)

$Recall = \frac{TN}{FN+TN}$

(5)

$F1-score= \frac{2\left(Precision\text{*}Recall\right)}{(Precision+Recall)}$

(6)

$Accuracy = \frac{TP+TN}{TP+TN+FN+FP}$

(7)

Table 3. Complexity Comparison of deep-learning detection networks

Model	No. of Layers	No. of Parameters (million)	Processing Time (Seconds)
NASNet-Mobile	913	4.2	7.70
ResNet-50	177	23.5	4.90
Inception-V3	315	21.8	8.77
ResNet-18	71	11.1	2.40
MobileNet-v2	154	2.2	5.50
ShuffleNet	172	.862	2.30
Xception	170	20.8	14.32
EfficientNet-b0	290	4	7.709

Table 2. summarizes the accuracy performance of the models. ‘Average accuracy’ indicates the average number of training images per class for each class. We observe a large difference in performance across the different models. Additionally, precision, recall and F1-score of all the deep learning models have been presented in Table 2. Figure 4. Displays a line graph which shows that EfficientNet-b0 secured the highest accuracy.

Besides the pretrained CNNs listed in Table 4, VGG16, VGG19, AlexNet, and GoogleNet were being used. Although their training accuracy was around 50% but the training loss was unknown. NASNet-Large is a costly network in terms of computational resources and training time due to its large number of layers. Even with a minimum batch size of 4, the system used in this investigation did not support it. EfficientNet-b0 achieved the highest accuracy of 91.0%, followed by MobileNet-V2 yielded 90.3% accuracy. Furthermore ShuffleNet, and inception-V3 also got a reasonable accuracy with 88.8% and 89.2% respectively. The highest recall and F1 score have been achieved by Efficient-b0, but the best precision value has been achieved by A good balance between precision and recall value is found in the case of inception-V3.

5.2 Computational Time Evaluation

Our next step was to investigate how fast a given dataset affects detection performance. Because the initial goal of dataset is to build an intelligent platform that can be used by a variety of autonomous robots and equipment to improve the timeliness and efficiency of precision structural health management. This platform must be run in real time and with high accuracy. Since each detection network was tested on a single machine, the evaluation metric is the processing time (PT) required to process 1000 images. Figure 4 shows the results of the detection time assessment.

The results in Table 3. demonstrate that Xception takes the most time to compute with a CT of 14.32 seconds followed by Inception-V3, EfficientNet-b0 with CT value 8.77, and 7.7 seconds respectively. ResNet18 and ShuffleNet are the fastest classification models with computation speed 2.4 and 2.3 seconds respectively. However, EfficientNet-b0 with the best performance of classification accuracy with a comparatively low computational time of 7.7 seconds. The number of learnable parameters. Furthermore, the processing time has been displayed diagrammatically in Fig. 4.

5.3 Effect of learning rate on EfficientNet-b0

When training deep neural networks, the initial learning rate plays a key role in determining the step size at which the optimizer modifies the model's weights. For the model to function at its best and achieve effective convergence, choosing the right initial learning rate is crucial.

The findings in Fig. 6 show the necessity of balancing convergence speed and stability during training by showing that a learning rate of 0.01 leads in the maximum testing accuracy. A greater testing accuracy of 90.510% is obtained with a learning rate of 0.00001, however this number is still below that attained with a learning rate of 0.01. This can be as a result of the fact that a learning rate that is too low can result in slower convergence, which might cause the model to underfit the data. It is important to keep in mind, nevertheless, that selecting the ideal learning rate involves making a choice between preventing underfitting and overfitting. The testing accuracy significantly increases to 89.200% when the learning rate is set too low, such at 0.001. This might be as a result of a slower convergence rate, which enables the model to sensitively adjust its weights. On the other hand, a learning rate of 0.0001 results in testing accuracy that is little lower, 88.760%, suggesting that the model may converge.

The model's performance could be further improved by adjusting the learning rate hyperparameter using strategies like learning rate schedules or adaptive learning rate optimizers.

5.4 Effect of L2 Regularization on EfficientNet-b0

We examined the outcomes of evaluating the EfficientNet-b0 model's accuracy using various L2 regularization hyperparameter settings. By inserting a penalty term to the loss function based on the magnitudes of the model's weights, L2 regularization, also known as weight decay, is a popular method for preventing overfitting in deep learning models is displayed in Fig. 6.

The testing accuracy results for EfficientNet-b0 for various L2 regularization values, spanning from 0.00001 to 0.05, are shown in Fig. 6. Let's explore the ramifications of these findings and discover how L2 regularization affects the model's functionality.

The L2 regularization hyperparameter has a changing trend as does the testing accuracy. From a high of 91.0% at L2 = 0.0001 to a low of 80.8% at L2 = 0.05, the accuracy varies. This shows that although some levels of L2 regularization aid in accuracy improvement, excessive regularization can result in considerable accuracy losses. With L2 = 0.0001, the model obtains an accuracy of 91.0%, which is the highest testing accuracy. This shows that a moderate dose of L2 regularization is advantageous for avoiding overfitting and enhancing the model's generalization to new data. In comparison to bigger values (e.g., 0.01, 0.05), accuracy is improved with small values (e.g., 0.00001) and moderate values (e.g., 0.0001). The significance of hyperparameter tuning to determine the appropriate regularization intensity for a certain dataset and model architecture is highlighted by this sensitivity. It is interesting to note that accuracy decreases when L2 is set either too high (e.g., 0.05) or too low (e.g., 0.005, 0.01). This exemplifies the trade-off between under-regularization, which results in overfitting and poor generalization, and over-regularization, which inhibits the model's capacity to learn complicated patterns.

In this research work, a domain-specific vision dataset for corrosion diagnosis in the field of structural health monitoring is presented. In this dataset, the innovation lies in the fact that it is intended to serve as a data benchmark for the development of deep learning-based classification models that are based on genuine properties of acoustic emission signals caused by corrosion in steel pipelines. In comparison to many other existing vision datasets, the advantages of AE-based datasets include the following characteristics:

With the advent of the dataset, we are now able to investigate both online and offline SHM applications in a genuine real-world environment. We conducted trials with a variety of deep-learning-based frameworks to confirm that this dataset was effective for classification. Although our baseline classification findings are encouraging, there is still potential for improvement when using state-of-the- art deep-learning models to corrosion identification and testing. Finally, after considering all aspects of speed, accuracy, and resilience, we conclude that the inception-v3 network has plenty of opportunity to improve on end-to-end fine-grained classification tasks in terms of both speed and accuracy.

However, there is still a need for a comprehensive evaluation of the literature on data sets and methods for AE-based classification. Furthermore, practically all available data sets have several constraints, such as the lack of large-scale datasets, material and geometry dependency, and the defect nature. The development of data-driven algorithms, particularly deep-learning-based approaches, is significantly hampered by these restrictions.

We intend to include more images and comments of numerous corrosion damages in the future, particularly for fine-gained categories that were difficult to describe in the past. Also, on the agenda is the investigation of ways to improve the performance of deep-learning classification and detection frameworks for use in additional SHM applications. And finally, we anticipate that this dataset will be beneficial in researching how to teach fine-grained visual categories to people.

Author Contributions

Conceptualization, F.H.; methodology, F.H.; software, F.H., and L.A.R.; validation, F.H., L.A.R. and M.I.K.; formal analysis, L.A.R. and F.H.; investigation, L.A.R.; resources, L.A.R.; data curation, F.H., M.I.K..; writing—original draft preparation, F.H., M.H.M.A., S.M.A., F.Q., writing—review and editing, L.A.R., M.I.K., M.H.M.A., S.M.A., F.Q.; visualization, F.H.; supervision, L.A.R. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Data Availability Statement

Data is not publicly available. Public access to the data used for this study is limited by Universiti Teknologi Petronas, Malaysia, however, can be applied for by researchers after contacting Dr Lukman AB Rahim ([email protected]).

Acknowledgments

The authors would like to thank Universiti Teknologi PETRONAS, Perak, Malaysia for providing the labs and research facilities in the form graduate assistance (GA) scheme.

Conflicts of Interest

The authors declare no conflict of interest.

E. Y. Dris, R. Drai, M. Bentahar, and D. Berkani, “Adaptive Algorithm for Estimating and Tracking the Location of Multiple Impacts on a Plate-Like Structure,” Res. Nondestruct. Eval., vol. 31, no. 1, pp. 1–23, 2020, doi: 10.1080/09349847.2019.1617913.
M. Rahimi, A. Alghassi, M. Ahsan, and J. Haider, “Deep Learning Model for Industrial Leakage Detection Using Acoustic Emission Signal,” Informatics, vol. 7, no. 4, pp. 1–18, 2020, doi: 10.3390/informatics7040049.
G. N. Morscher and Z. Han, “Damage determination in ceramic composites subject to tensile fatigue using acoustic emission,” Materials (Basel)., vol. 11, no. 12, pp. 1–18, 2018, doi: 10.3390/ma11122477.
F. Hassan et al., “State-of-the-Art Review on the Acoustic Emission Source Localization Techniques,” IEEE Access, vol. 9, pp. 101246–101266, 2021, doi: 10.1109/ACCESS.2021.3096930.
H. Liu and Y. Zhang, “Image-driven structural steel damage condition assessment method using deep learning algorithm,” Meas. J. Int. Meas. Confed., vol. 133, pp. 168–181, 2019, doi: 10.1016/j.measurement.2018.09.081.
F. Liu, S. Gao, Z. Tian, and D. Liu, “A new time-frequency analysis method based on single mode function decomposition for offshore wind turbines,” Mar. Struct., vol. 72, no. April, p. 102782, 2020, doi: 10.1016/j.marstruc.2020.102782.
S. Sikdar and J. Pal, “Bag of visual words based machine learning framework for disbond characterisation in composite sandwich structures using guided waves,” Smart Mater. Struct., vol. 30, no. 7, p. 75016, 2021, doi: 10.1088/1361-665x/ac01a8.
M. T. Pham, J. M. Kim, and C. H. Kim, “Intelligent fault diagnosis method using acoustic emission signals for bearings under complex working conditions,” Appl. Sci., vol. 10, no. 20, pp. 1–14, 2020, doi: 10.3390/app10207068.
O. Russakovsky et al., “ImageNet Large Scale Visual Recognition Challenge,” Int. J. Comput. Vis., vol. 115, no. 3, pp. 211–252, 2015, doi: 10.1007/s11263-015-0816-y.
B. A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet Classification with Deep Convolutional Neural Networks,” Commun. ACM, vol. 60, no. 6, pp. 84–90, 2012.
M. Sohaib, M. Islam, J. Kim, D. C. Jeon, and J. M. Kim, “Leakage detection of a spherical water storage tank in a chemical industry using acoustic emissions,” Appl. Sci., vol. 9, no. 1, 2019, doi: 10.3390/app9010196.
N. Morizet, N. Godin, J. Tang, E. Maillet, M. Fregonese, and B. Normand, “Classification of acoustic emission signals using wavelets and Random Forests: Application to localized corrosion,” Mech. Syst. Signal Process., vol. 70–71, pp. 1026–1037, 2016, doi: 10.1016/j.ymssp.2015.09.025.
M. Elforjani and S. Shanbr, “Prognosis of Bearing Acoustic Emission Signals Using Supervised Machine Learning,” IEEE Trans. Ind. Electron., vol. 65, no. 7, pp. 5864–5871, 2018, doi: 10.1109/TIE.2017.2767551.
A. Meserkhani, S. M. Jafari, and A. Rahi, “Experimental comparison of acoustic emission sensors in the detection of outer race defect of angular contact ball bearings by artificial neural network,” Meas. J. Int. Meas. Confed., vol. 168, p. 108198, 2021, doi: 10.1016/j.measurement.2020.108198.
L. Ai, V. Soltangharaei, M. Bayat, B. Greer, and P. Ziehl, “Source localization on large-scale canisters for used nuclear fuel storage using optimal number of acoustic emission sensors,” Nucl. Eng. Des., vol. 375, no. January, p. 111097, 2021, doi: 10.1016/j.nucengdes.2021.111097.
D. J. Atha and M. R. Jahanshahi, “Evaluation of deep learning approaches based on convolutional neural networks for corrosion detection,” Struct. Heal. Monit., pp. 1–19, 2017, doi: 10.1177/1475921717737051.
S. X. Chen, L. Zhou, Y. Q. Ni, and X. Z. Liu, “An acoustic-homologous transfer learning approach for acoustic emission–based rail condition evaluation,” Struct. Heal. Monit., vol. 20, no. 4, pp. 2161–2181, 2021, doi: 10.1177/1475921720976941.
H. Xin, L. Cheng, R. Diender, and M. Veljkovic, “Fracture acoustic emission signals identification of stay cables in bridge engineering application using deep transfer learning and wavelet analysis,” Adv. Bridg. Eng., vol. 1, no. 1, pp. 1–16, 2020, doi: 10.1186/s43251-020-00006-7.
G. J. F. Konig, C. Sous, A. Ouald1. G. J. F. Konig, C. Sous, A. Ouald Chaib, “Machine learning based anomaly detection and classification of acoustic emission events for wear monitoring in sliding bearing systems,” Tribol. Int., vol. 155, pp. 285–287, 2020, doi, “Machine learning based anomaly detection and classification of acoustic emission events for wear monitoring in sliding bearing systems,” Tribol. Int., vol. 155, pp. 285–287, 2020, doi: 10.1016/B978-0-323-60984-5.00062-7.
D. F. Hesser, S. Mostafavi, G. K. Kocur, and B. Markert, “Neurocomputing Identification of acoustic emission sources for structural health monitoring applications based on convolutional neural networks and deep transfer learning,” Neurocomputing, vol. 453, pp. 1–12, 2021, doi: 10.1016/j.neucom.2021.04.108.
L. Ai, V. Soltangharaei, and P. Ziehl, “Evaluation of ASR in concrete using acoustic emission and deep learning,” Nucl. Eng. Des., vol. 380, no. March, 2021, doi: 10.1016/j.nucengdes.2021.111328.
D. Li, Y. Wang, W. J. Yan, and W. X. Ren, “Acoustic emission wave classification for rail crack monitoring based on synchrosqueezed wavelet transform and multi-branch convolutional neural network,” Struct. Heal. Monit., vol. 20, no. 4, pp. 1563–1582, 2021, doi: 10.1177/1475921720922797.
F. Hassan, L. A. Rahim, A. K. Mahmood, and S. A. Abed, “A Hybrid Particle Swarm Optimization-Based Wavelet Threshold Denoising Algorithm for Acoustic Emission Signals,” Symmetry (Basel)., vol. 14, no. 6, p. 1253, 2022.
C. Sun et al., “Machine Learning Allows Calibration Models to Predict Trace Element Concentration in Soils with Generalized LIBS Spectra,” Sci. Rep., vol. 9, no. 1, pp. 1–18, 2019, doi: 10.1038/s41598-019-47751-y.
Z. May, M. K. Alam, N. A. A. Rahman, M. S. Mahmud, and N. A. Nayan, “Denoising of hydrogen evolution acoustic emission signal based on non-decimated stationary wavelet transform,” Processes, vol. 8, no. 11, pp. 1–19, 2020, doi: 10.3390/pr8111460.
L. Gou, H. Li, H. Zheng, H. Li, and X. Pei, “Aeroengine Control System Symmetry Fault Diagnosis Based on CWT and CNN,” Math. Probl. Eng., vol. 2020, 2020, doi: 10.1155/2020/5357146.
J. Pacheco-Chérrez, A. Delgado-Gutiérrez, D. Cárdenas, and O. Probst, “Reliable damage localization in cantilever beams using an image similarity assessment method applied to wavelet-enhanced modal analysis,” Mech. Syst. Signal Process., vol. 149, 2021, doi: 10.1016/j.ymssp.2020.107335.
L. J. Yann LeCun, Bernhard Boser, John Denker, Donnie Henderson, R. Howard, Wayne Hubbard, “Handwritten Digit Recognition with a Back-Propagation Network,” 1989, pp. 396–404.
K. Simonyan and A. Zisserman, “Very deep convolutional networks for large-scale image recognition,” in 3rd International Conference on Learning Representations, ICLR 2015 - Conference Track Proceedings, 2015, pp. 1–14.
S. R. Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet and A. R. Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, “Going Deeper with Convolutions,” in Conference on Computer Vision and Pattern Recognition (CVPR), pp. 1–9, doi: 10.4324/9781410605337-29.
C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the Inception Architecture for Computer Vision,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, vol. 2016-Decem, pp. 2818–2826, doi: 10.1109/CVPR.2016.308.
J. S. Kaiming He, Xiangyu Zhang, Shaoqing Ren, “Deep Residual Learning for Image Recognition,” in Conference on Computer Vision and Pattern Recognition (CVPR), pp. 770–778, doi: 10.1002/chin.200650130.
B. Zoph, V. Vasudevan, J. Shlens, and Q. V. Le, “Learning Transferable Architectures for Scalable Image Recognition,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2018, pp. 8697–8710, doi: 10.1109/CVPR.2018.00907.
F. Chollet, “Xception: Deep learning with depthwise separable convolutions,” in Proceedings - 30th IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, pp. 1251–1258, doi: 10.1109/CVPR.2017.195.
J. S. Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, “ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,” in 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 18326147, doi: 10.4324/9780203491348.
A. G. Howard et al., “MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,” 2017, [Online]. Available: http://arxiv.org/abs/1704.04861.
M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. C. Chen, “MobileNetV2: Inverted Residuals and Linear Bottlenecks,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, p. 18326147, doi: 10.1109/CVPR.2018.00474.
M. Tan and Q. V. Le, “EfficientNet: Rethinking model scaling for convolutional neural networks,” in 36th International Conference on Machine Learning, ICML 2019, 2019, vol. 2019-June, pp. 10691–10700.
B. Zoph and Q. V. Le, “Neural architecture search with reinforcement learning,” in 5th International Conference on Learning Representations, ICLR 2017 - Conference Track Proceedings, 2017, pp. 1–16.
M. Tan et al., “Mnasnet: Platform-aware neural architecture search for mobile,” in Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, 2019, vol. 2019-June, pp. 2815–2823, doi: 10.1109/CVPR.2019.00293.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Unlocking Corrosion Secrets: A Comprehensive Analysis of Acoustic Emission Data Using Continuous Wavelet Transform and Deep Learning for Steel Pipeline Monitoring

Status:

Version 1

Abstract

Figures

1 Introduction

2 Literature Review

3 Materials and Methods

3.1 Experimental Setup for Data Acquisition

3.2 Data Fusion

3.3 Wavelet Denoising

3.4 Feature Selection

3.5 Random Under Sampling (RUS)

3.6 Data Transformation

4 Dataset Benchmarking

4.1 AlexNet

4.2 VGG Group

4.3 GoogleNet

4.4 ResNets

4.5 NASNet-Mobile

4.6 Xception

4.7 ShuffleNet

4.8 MobileNet

4.9 EfficientNet-b0

5 Experiments

5.1 Performance Comparison

5.2 Computational Time Evaluation

5.3 Effect of learning rate on EfficientNet-b0

5.4 Effect of L2 Regularization on EfficientNet-b0

6 Conclusion

Declarations

References

Additional Declarations

Status:

Version 1