Domain generalization using pseudo triplet network learning for vibration signal-based fault diagnosis

Abstract

Domain generalization (DG) based intelligent fault diagnosis has developed rapidly in recent years owing to the need for applying trained neural networks to unseen domains. However, models trained using DG often suffer from performance degradation when in presence of nonstationary working conditions. To address this challenge, this work proposes a DG based intelligent fault diagnosis approach based on a vibration response mechanism guided pseudo triplet network, which extracts suitable features that correlate well with the health conditions. Firstly, the proposed approach estimates the cyclic spectral correlation maps of vibration signals to provide vibration response mechanism of different health conditions. Then, a pseudo triplet neural network is designed to calculate the distance between the representations of the prior input, the negative input from the representation of the main input. The prior input is the specific part of the cyclic spectral correlation map with the selected carrier band and it guides the network focus on the fault-related features. Finally, the proposed approach is evaluated through experiments conducted on data collected from nonstationary working conditions.