Deep Learning-Based Crack Detection in Spiral-Welded Pipelines: A Novel Adaptive Convolutional Feature Extraction Method

This article proposes a new time-series signal processing model for crack detection and localization in spiral-welded pipelines. Using the time-series signal of pipeline cracks as the original dataset, local features of the original dataset are extracted through 1-D convolutional neural network (1D-CNN), and the global dependency relationship of the data is established by multihead attention (MHA) mechanism. The processed dataset is defined as a convolutional kernel, and a dynamic adjustment strategy p is adopted to make the size of the convolutional kernel suitable for the processed dataset. The model dynamically adjusts the adaptive convolution kernel size based on extracted data features, calculates the mean and standard deviation, and captures complex sequence patterns. The output high-dimensional feature data are trained in the residual network (ResNet34) to achieve crack detection in classification tasks and localization in regression tasks. Experimental results demonstrate that the model performs well under three datasets, especially under normal data conditions, the model achieves 99.3% detection accuracy (ACC) and $0.9993~{R}^{{2}}$ in regression, reducing localization mean absolute error (MAE) to 0.0893, outperforming benchmarks (e.g., Transformer) by 12.2% in ACC and 20% in inference speed. The lightweight design (35% fewer parameters) ensures computational efficiency, making it suitable for real-time monitoring of aging heating pipelines in Beijing. This work provides a robust solution for industrial nondestructive testing, advancing the reliability analysis of spiral-welded pipelines.