Research on broken wire identification method based on PoolConv for prestressed concrete cylinder pipe

To enhance the real-time monitoring capability of Distributed Acoustic Sensors (DAS) systems in detecting wire breaks in Prestressed Concrete Cylinder Pipe (PCCP), this study conducted a 1 : 1 scale wire break experiment on an underground PCCP with a diameter of 4000 mm. In response to issues such as the low sliding efficiency of the classical One-Dimensional Convolutional Neural Network (1-D CNN) filtering window and the issue of shared feature point weights leading to the underrepresentation of key features, a novel Pooling Convolution (PoolConv) model for wire break detection was proposed. Through comprehensive comparison with methods such as Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Resnet50, and 1-D CNN, results showed that the performance of PoolConv and 1-D CNN was significantly superior to other approaches.Especially the PoolConv model, the average damage recognition rate can reach 99.59%. Further analysis and comparison with the previous version of the 1-D CNN showed that the PoolConv model not only improved the average validation accuracy and testing accuracy by 0.51 and 2.01%, respectively, but also enhanced the training and testing time efficiency by 20 times and 3 times, respectively. Moreover, in terms of convergence speed, generalization ability, and resistance to noise, the PoolConv model exhibited superior performance and also had a smaller model weight file, effectively circumventing the problem of low efficiency present in the 1-D CNN.