Intelligent Defect Recognition for Oil and Gas Pipeline Safety Based on GA-TQWT and SAM-Inception

Abstract The safety precautions for on-site pipeline defects vary depending on the type of defect. Hence, accurate recognition of the specific defect is a prerequisite for pipeline safety and reliability. In this paper, the Tunable Q-factor Wavelet Transform denoising method optimized by Genetic Algorithm (GA-TQWT) and an Inception convolutional neural network with a Spatial Attention Module (SAM-Inception) are proposed to improve the defect diagnosis effect. Firstly, we utilize our own magnetic eddy current inspection equipment to collect 4 types of defective signals of pipelines in the field, including pit, scratch, weld defect and normal signal data, 340 samples in total. The inspection equipment can perform non-destructive and effective inspection of pipelines without removing the corrosion protection layer and insulation layer. Then, the noisy field signals are adaptively decomposed and reconstructed by TQWT, during which the genetic algorithm seeks optimization for the two hyperparameters of Q-factor and decomposition level. Furtherly, the denoised signal is converted into a two-dimensional image by Gramian Angular Field (GAF) to emphasize the defect details and accommodate the input of the self-designed SAM-Inception network. Finally, the wide-receptive-field inception extracts the abstract features of the defective images, and the spatial attention module assigns different weights to the extracted spatial features to retain important features and weaken irrelevant information. The result shows that the optimized GA-TQWT denoising algorithm demonstrates superior performance in both noise reduction and signal fidelity, achieving an SNR of 11.3367 dB and an RMSE of 0.0650, outperforming other denoising methods such as TQWT, MODWT (Maximal Overlap Discrete Wavelet Transform), EMD (Empirical Mode Decomposition) and VMD (Variational Mode Decomposition). Additionally, the SAM-Inception network has high sensitivity to the involved 4 type defects, achieving 100% training accuracy and 98.53% testing accuracy. Therefore, the combination of GA-TQWT and SAM-Inception can realize the intelligent diagnosis of different pipeline defects, and lay the foundation for the reasonable maintenance recommendations and further quantitative study of the same defect.