Trackside Acoustic Fault Diagnosis of Bearing Based on Doppler Knowledge Embedded in Domain Adaptation Network

Trackside acoustic fault diagnosis is widely used for early fault detection in train bearings. However, signal distortion due to the Doppler effect poses a challenge to effective fault diagnosis. Currently, signal correction, the mainstream solution to the Doppler effect, is limited by known kinematic parameters and high noise interference. In this article, a novel learning model based on pseudo-Doppler time domain demodulation (PDDD) and domain adaptation network (DAN) is proposed. It attempts to construct the relationship between Doppler effect distortion signals and train bearing fault diagnosis under unsupervised learning (UL). The method first approximates the Doppler effect-free signals using PDDD. The PDDD use weighted fusion mechanism based on segmented energy ratio (SER) to make fused signal close to original signal, which reduces the distributional differences between samples. Then DAN is used at the feature extraction layer to reduce the approximation error of Doppler time-domain demodulation and the feature bias caused by Doppler frequency shift. The proposed method extracts domain-invariant features of the source-domain signal and the target-domain signal containing Doppler effect using an unsupervised domain adaptation model. On the one hand, the problem of known kinematic parameters and noise interference of existing methods is solved. On the other hand, it successfully transfers the existing knowledge of constant state bearing fault diagnosis to bearing fault diagnosis under Doppler effect.