A magnetic flux leakage inspection defect identification method for hydrogen-blended pipelines based on a hybrid 1D-CNN-LSTM neural network

Hydrogen-induced degradation phenomena, including hydrogen embrittlement and hydrogen-induced cracking, pose a significant threat to the structural integrity of carbon steel pipelines in high-pressure hydrogen environments. Conventional non-destructive testing (NDT) methods may exhibit limited reliability due to complex signal interference caused by the coupling of localized stress concentration and hydrogen damage. This study adopts a deep learning-based solution employing a hybrid 1D Convolutional Neural Network and Long Short-Term Memory (1D-CNN-LSTM) model to enhance defect classification accuracy in hydrogen transmission pipelines. To monitor stress-related magnetic field signals in hydrogen-blended steel pipelines, a magnetic flux leakage (MFL) inspection system equipped with Hall sensors was developed. Physically, the synergistic effect of hydrogen and stress in hydrogen transmission pipelines alters the magnetic properties of the material, thereby generating detectable magnetic field perturbations. The system utilizes a one-dimensional convolutional neural network (1D-CNN) to adaptively extract local spatial features from MFL signals, while long short-term memory (LSTM) layers capture the inherent long-term temporal dependencies within scanning sequences. Comparative experimental results show that the adopted 1D-CNN-LSTM model achieves a classification accuracy of 98.88% for four types of defects—pipeline corrosion, weld seams, holes, and cracks—significantly outperforming standalone 1D-CNN (94.68%) and LSTM (91.68%) architectures. On an independent validation dataset, despite occasional misclassification between holes and cracks due to their similar magnetic signatures, the model achieved a robust prediction accuracy of 92%. The analysis results indicate that this technology provides a reliable and efficient solution for the safety monitoring of hydrogen-blended pipelines.