CIGCN: a domain generalisation fault diagnosis method for train bearing based on causal inference graph convolutional network

The bogie bearings of metro trains are key components to guarantee the stability and safety of train operation. Once it fails, it will accelerate the degradation of equipment performance and cause serious operational risks. The graph neural networks show a broad application prospect in train bearing state recognition. However, in the face of changing conditions such as metro train speed, track conditions, and environmental interference, there is still a problem of insufficient generalisation performance. Therefore, a new causal inference graph convolutional network (CIGCN) is presented to enhance the robustness and adaptability of train bearing fault diagnosis under unseen conditions. Firstly, a new structural causal model is proposed to analyse the causal relationship between fault-related variables. Secondly, a multi-source information fusion graph construction method is proposed, which converts multi-modal signals such as vibration, acoustic, and acoustic emission into graph structure data. Finally, a decoupling transformation module and a sample weighting mechanism are proposed to achieve effective decoupling of fault causal factors and non-causal factors at the representation level and suppress structural bias caused by non-causal interference. At the same time, the backdoor adjustment regularisation strategy based on random combination is proposed to decrease the negative influence of non-causal factors, promoting the learning of cross-domain invariant features.