Make Gating Fairer: Fault Attribute-Driven Bias Calibration for Generalized Zero-Shot Industrial Fault Diagnosis

Abstract

Generalized zero-shot diagnosis (GZSD) for industrial processes has gained increasing attention as it aims to diagnose both seen and unseen faults based on fault attributes. Nevertheless, the lack of unseen fault data for training poses a critical domain shift problem (DSP), where a biased model prediction toward seen faults is caused by overfitting available seen fault data so that unseen faults tend to be falsely identified as seen faults. Therefore, we propose a Fault Attribute-driven bIas calibRation (FAIR) model to address the model bias toward seen faults, which consists of a bias calibration module and a generation module based on attribute feature recombination. The proposed bias calibration module is the first gating mechanism designed in the attribute space to differentiate between seen and unseen faults. Different from existing methods that construct the boundary for both faults in the feature space, we first propose to employ attributes of unseen faults as substitutes for unavailable training samples to construct the classification boundary in the attribute space, thereby mitigating the model bias caused by the lack of unseen fault features. To further alleviate the DSP, we propose a generation module to generate samples for unseen faults. In particular, to improve the quality of generated samples, the generation module leverages fault attribute features extracted from seen faults and recombines them under the guidance of fault attributes. We conduct GZSD experiments on a real hydraulic system and the Tennessee-Eastman process (TEP) benchmark dataset. In comparison with the SOTA methods, the proposed method achieves an average improvement of 11.84% in terms of the harmonic mean of the accuracy of seen and unseen faults on the TEP, and 6.22% on the real hydraulic system.