Application of Acoustic Signal Hybrid Filtering with 1D-CNN for Fault Diagnosis of Natural Gas Pipeline Leakage

Abstract

In acoustic signal–based natural gas pipeline leak detection, it is a technical difficulty to effectively remove the strong background noise hidden in the acoustic signal. Unfortunately, existing techniques usually study denoising in isolation from the model, and do not consider the synergistic effect of integrating denoising with the model to update filtering parameters adaptively, which limits the performance development of the diagnostic model. To solve this problem, this paper proposes a fault diagnosis method that incorporates acoustic signal hybrid filtering and a one-dimensional convolutional neural network (1D-CNN) to jointly update the denoising parameters. Firstly, the acoustic signal is sampled at a fixed frequency by the sensor unit. Secondly, a hybrid filtering method with low-frequency characteristics is designed, which integrates low-pass filtering and median filtering to eliminate background noise. Finally, an adaptive fault diagnosis network that fuses acoustic signal denoising with 1D-CNN feature extraction is developed. The network updates the filtering parameters during training based on the model recognition accuracy, so that the denoising parameters and model weights change dynamically. The experimental results show that the method is capable of performing adaptive parameter search with 97.07% identification accuracy of fault type.