Fault diagnosis for complex systems: A differential sensitivity-aided canonical correlation analysis approach

Abstract

To address the challenges associated with limited fault diagnosis performance in complex systems, attributed to varying feature sensitivities and difficulties in identifying unknown faults, this paper introduces a novel fault diagnosis method based on differential sensitivity-aided canonical correlation analysis. The proposed method consists of several key steps. First, a differential sensitivity processing algorithm is introduced, encompassing two primary stages. In the initial step, the Fisher score algorithm is employed for the differential selection of sensitive features. In the second step, the mixture correlation coefficient is utilized to eliminate redundant features. Next, a well-established fault detection algorithm based on canonical correlation analysis is employed to detect the occurrence of faults. Furthermore, we present an instance confidence evaluator algorithm that integrates a fault model bank to isolate abnormalities and identify potential unknown faults. Experimental results obtained from the application of this method to bearings and a high-speed train braking system demonstrate its superior generalization performance, as well as its exceptional fault detection and fault isolation capabilities.