CNN-LSTM Hybrid Model to Promote Signal Processing of Ultrasonic Guided Lamb Waves for Damage Detection in Metallic Pipelines

Abstract

The ultrasonic guided lamb wave approach is an effective non-destructive testing (NDT) method used for detecting localized mechanical damage, corrosion, and welding defects in metallic pipelines. The signal processing of guided waves is often challenging due to the complexity of the operational conditions and environment in the pipelines. Machine learning approaches in recent years, including convolutional neural networks (CNN) and long short-term memory (LSTM), have exhibited their advantages to overcome these challenges for the signal processing and data classification of complex systems, thus showing great potential for damage detection in critical oil/gas pipeline structures. In this study, a CNN-LSTM hybrid model was utilized for decoding ultrasonic guided waves for damage detection in metallic pipelines, and twenty-nine features were extracted as input to classify different types of defects in metallic pipes. The prediction capacity of the CNN-LSTM model was assessed by comparing it to those of CNN and LSTM. The results demonstrated that the CNN-LSTM hybrid model exhibited much higher accuracy, reaching 94.8%, as compared to CNN and LSTM. Interestingly, the results also revealed that predetermined features, including the time, frequency, and time–frequency domains, could significantly improve the robustness of deep learning approaches, even though deep learning approaches are often believed to include automated feature extraction, without hand-crafted steps as in shallow learning. Furthermore, the CNN-LSTM model displayed higher performance when the noise level was relatively low (e.g., SNR = 9 or higher), as compared to the other two models, but its prediction dropped gradually with the increase of the noise.

Signal Process of Ultrasonic Guided Wave for Damage Detection of Localized Defects in Plates: From Shallow Learning to Deep Learning

“…A higher AUC is associated with superior model performance. The AUC reaches 1 when a model achieves 100% accuracy (Shang et al, 2023a).…”

“…Time and frequency domain feature parameters were chosen to better signify the information for the signals, as shown in Table 1. And y 1 stands for the frequency vibration energy, y 2 À y 4 , y 6 and y 10 À y 13 stand for spectrum dispersion or concentration; and y 5 and y 7 À y 9 represent the position difference of the main frequency (Chen 2014); they have been selected based on Chen (2014), Shang et al (2023a), and Shang et al (2023b).…”

“…The ROC curves were used as the performance evaluation tool in machine learning and DL (Shang et al, 2023a). ROC curves are generated by contrasting the true positive rate with the false positive rate at various threshold levels (Zhang et al, 2020).…”

“…ROC curves are generated by contrasting the true positive rate with the false positive rate at various threshold levels (Zhang et al, 2020). The degree or measure of separability was described by the AUC (Shang et al, 2023a;Zhang et al, 2020). A higher AUC is associated with superior model performance.…”

“…In many signal classification systems, waveform-based deep neural networks are now required (Abdel-Hamid et al, 2012;Mohamed et al, 2011;Nagrani et al, 2017;Variani et al, 2014;Zhang & Koishida, 2017). These advanced neural networks directly utilize unprocessed signals as input in various tasks (Shang et al, 2023a), such as identifying stress levels of structures, assessing the condition of infrastructure, identifying structural damage, and detecting structural damage (Zhang et al, 2020), diagnosing structural damage (Lin et al, 2018;Pan et al, 2018), and monitoring structural health (Gui et al, 2017;Zhang et al, 2022).…”