The class imbalance of crack defects in oil and gas pipelines presents a significant challenge to the diagnosis of pipeline integrity, and the scarcity of high-risk cracks can easily lead to the recognition models tending to predict them as the low-risk cracks. Therefore, this paper proposes an end-to-end triple attention mechanism and dual-modal deep convolutional network (TAMDDCN). First, the co-optimized visual modal branch and sequential modal branch respectively dig deep into the essential features of crack defects. Then, the features are dynamically assigned with weights according to their importance using the Convolutional Block Attention Module (CBAM), and the feature interactive relationships are established by self-attention (SA). Finally, the cross-modal and weighted defect information is integrated organically for crack defect diagnosis. The results demonstrate that (1) the TAMDDCN model achieves 100% training accuracy, 99% validation accuracy, and 100% test accuracy when the ratio of different depth cracks is 4∶3∶2∶1; (2) the TAMDDCN, with the minimal computational cost and highest classification accuracy, showcases its superior scalability over state-of-the-art methods; (3) the TAMDDCN outperforms the ablation models by approximately 4%–12% and the classical models by approximately 18%–26% in terms of test accuracy; and (4) the TAMDDCN still keeps high test accuracy of 94%, 92%, and 88%, respectively, given the addition of Gaussian noise with signal-to-noise ratios of 15, 10, and 5 dB.
Liao et al. (Mon,) studied this question.