Causal Representation Decoupling for Fault Diagnosis Using Superpixel-Based Graph Neural Networks

Existing Graph Neural Network (GNN)-based methods enjoy remarkable capability in capturing the fault features in mechanical diagnosing systems. However, these methods suffer from the following two drawbacks: difficulty in constructing reliable graphs from time-frequency images due to non-uniform feature distribution, and limited generalization and robustness corresponding to the probability drift amongst data collected with respect to different working conditions, such as load and speed changes. To address these, this paper proposes a novel framework integrating superpixel processing and causal feature decoupling for GNN-based diagnosis. Firstly, it leverages Simple Linear Iterative Clustering (SLIC) to segment time-frequency images, constructing graphs via feature similarities between superpixel nodes. Secondly, a causal extraction module then uses soft masking, backdoor adjustment, and independence constraints to disentangle causal features, enhancing the model’s ability to generalize across varying conditions. Finally, these features are input to a Vision GNN (ViG) to generate causal subgraphs, improving fault classification accuracy. Experimental results on a dataset generated from a laboratory test bench and a publicly available dataset indicate that our fault diagnosis framework yields excellent performance comparing with existing methods.