Reliability analysis of corroded pipes using MFL signals and Residual Neural Networks

• Published in: Process safety and environmental protection 2024-04, Vol.184, p.1131-1142
• Main Authors: Chen, Yinuo, Tian, Zhigang, Wei, Haotian, Dong, Shaohua
• Format: Article
• Language: English
• Subjects: MFL inspection; Pipeline corrosion; ResNet; Risk management
• ISSN: 0957-5820
• DOI: 10.1016/j.psep.2024.02.052

A magnetic flux leakage (MFL) tool may identify most pipeline corrosion defects. Therefore, analyzing the MFL signals to obtain helpful information for pipeline safety analysis is significant. However, existing methods need many steps to get pipeline reliability results via MFL detection, resulting in a substantial time investment. This study proposes a reliability prediction method based on Residual Neural Networks (ResNet) that can directly map the MFL inspection data to the pipeline’s reliability. Pipeline defect effective area model, rather than those based on just depth, is effectively integrated with deep learning models. Due to the limited practical data sources, the finite element (FE) method is used to simulate a large amount of data for ResNet training. It is found that the ResNet family can improve both the model’s performance and training efficiency. Compared to traditional methods, the proposed model’s accuracy is more than 20% higher, and the computational efficiency has been increased by 200 times. Case studies of FE simulations and industrial applications illustrate that the suggested approach is capable of assessing the reliability of corroded pipes in a more timely and accurate manner than traditional methods. The proposed method is also helpful for pipeline operators to understand the pipeline risk condition and obtain suggestions for optimizing costs and re-assessment intervals, providing a foundation for pipeline digital integrity management.