Intelligent siltation diagnosis for drainage pipelines using weak-form analysis and theory-guided neural networks in geo-infrastructure

Abstract

Siltation diagnosis of drainage pipelines is crucial for preventing urban flooding. However, the existing intelligent siltation diagnosis algorithms often exhibits limitations in handling multivariate data sequences and extracting multifaceted features, leading to partial distortion in outputs. To address these shortcomings, a neural network architecture consisting of inception network (BCI), residual network, multichannel long short-term memory network (MLSTM), and deep neural network (DNN) is constructed. It employs multichannel technique and bidirectional causal dilation convolution kernels with varying dilation factor steps to extract multiscale features. Weak-form analysis and theory-guided loss function error correction method are introduced to further enhance the accuracy of diagnosis. Then, a knowledge-algorithm collaborative driven model for pipeline siltation diagnosis is proposed. Its accuracy and robustness are verified by testing against typical prediction models with differing types of noise. Results underscore the method's potential for accurately detecting diverse municipal infrastructure defects, implying applicability in geo-infrastructure scenarios.