Fault Detection Based on Causal Discovery and Graph Convolutional Network for Complex Mechatronic Systems

Detecting precisely any incipient fault is of critical importance for a mechatronic system, which may improve its safety, availability, and maintainability. However, high dimensionality and data imbalance are big obstacles for establishing an effective fault detection model. In this article, an interpretable fault detection model is proposed by designing a causal-based graph convolution network (GCN). First, the biased data distribution of imbalanced monitoring data is adjusted according to prior knowledge, to obtain a more trustable causal network. Second, average causal effects (ACEs) are calculated according to the obtained causal network to quantitatively describe the causal relations. Finally, a directed weighted GCN (DWGCN) is designed for the causal network with ACEs to build the fault detection model. Spectral clustering and graph coarsening methods for directed graphs are integrated in DWGCN. And, to improve the modeling efficiency, the model structure parameters are carefully designed. Based on both simulation data and real data, the proposed priori-constrained causal discovery has been proven to be able to mine causal networks that meet more than 50% of prior-constraints. And, based on real data, the average F-
measure
and
G-mean
of the DWGCN reach 84.7% and 85.3%, which is superior to the conventional GCNs and machine learning methods.