Dual-space metric fusion network for interpretable few-shot fault diagnosis of industrial processes

While few-shot learning (FSL) addresses sample scarcity in fault diagnosis of industrial processes, existing methods typically rely on temporal features extracted by deep neural networks (DNN), leading to a lack of physical interpretability and making models prone to overfitting. To address this issue, a novel interpretable few-shot fault diagnosis framework, named dual-space metric fusion network (DSMFN), is proposed. First, a pre-trained denoising autoencoder (DAE) is used to learn the multivariate nonlinear correlations under normal operating conditions and extract the fault contribution vectors of physical sensor variables. Second, at the decision layer, a soft fusion between temporal feature distance and physical variable similarity is performed to introduce the explicit physical constraints while retaining the generalization of DNN, effectively alleviating overfitting. Furthermore, a calibrated global variable contribution tracking module is designed to statistically assess the recurrent fault-associated variable deviations without predefined fault semantics. Experiments on the Tennessee Eastman Process (TEP) benchmark dataset demonstrate that DSMFN outperforms existing few-shot baseline models in terms of classification accuracy and robustness. A mechanism-based sanity check further shows that the highly ranked variables are consistent with the affected process variables and closed-loop control responses, demonstrating that the proposed tracking module can provide statistically supported variable-level evidence for subsequent process analysis.