A zero-shot industrial process fault diagnosis method based on domain-shift constraints

Abstract

Background

Fault diagnosis is crucial for industrial maintenance, but existing supervised methods rely on extensive data, which is often difficult to collect. The challenge of gathering comprehensive fault samples limits the performance of traditional fault diagnosis methods.

Method

In this paper, we propose a fault diagnosis method named ZSIDM-OC to address the zero-shot problem in industrial processes, specifically concerning the domain shift issue. This novel framework includes three key modules: the Hierarchical Global-Local Feature Integration Module for capturing both global and local features of the fault data; the Prototype-Based Discriminative Loss Module, which reduces feature redundancy and enhances the model's ability to recognize unknown fault classes; and the Bidirectional Consistency Enforcement Module ensuring consistent data distribution in both low-dimensional and high-dimensional spaces, thereby reducing domain shift.

Significant Findings

Our analysis indicates that the domain shift problem is inevitable in a zero-shot setting and significantly affects the performance of existing methods. Experimental results demonstrate that under zero-shot conditions, ZSIDM-OC offers significant advantages on both the Energy Storage Plant dataset and the Tennessee Eastman dataset. This method effectively mitigates the challenges posed by domain shift and limited fault sample availability, showcasing its potential to improve fault diagnosis in industrial processes.