The rapid growth of urbanization has led to an increase in the size of sewage pipelines. However, due to factors such as environmental conditions, misuse, and age, many pipelines are defective to varying degrees and thus impact the ecosystem. Modern vision-based defect detection is limited at high sewage levels and often struggles with low accuracy in classifying defects. In order to solve these problems, acoustics-based defect detection coupled with defect classification using a one-dimensional convolutional neural network (1D-CNN) for municipal sewage pipelines is proposed. Different pipeline defects exhibit varying occurrence rates, leading to an imbalance in the data set of classification categories, which in turn affects classification performance. The introduction of synthetic minority oversampling technology (SMOTE) into 1D-CNN effectively balances the data set. Field experimental results show that the proposed approach successfully classifies rupture, misalignment, and deformation, and identifies pipeline ports. The classification model is robust, with evaluation metrics of accuracy, macro average, micro average, and weighted average exceeding 90%. It requires simple equipment setup with low implementation costs. The computational efficiency of 1D-CNN enables real-time defect classification, with a single-sample classification time of approximately 2×10−5 s, making it a practical solution for municipal sewage pipeline maintenance.
Wang et al. (Fri,) studied this question.