A Class-Specific Prototype and Multivariate Coupling-Aware Method for EHA Fault Time-Series Diagnosis

In the multivariate time-series fault diagnosis task for aviation electro-hydrostatic actuators (EHA), the overall signal morphologies of different fault categories are relatively similar, while the key discriminative differences are hidden in local segments and variations in variable coupling. Therefore, existing Transformer-based methods usually have difficulty characterizing local specificity. To address this issue, this paper proposes a Local Prototype-Global Generic Dual-branch Transformer (LPG-Former). First, to obtain local information capable of characterizing class differences, a class-specific discriminative prototype (CDP) is constructed. The CDP selects discriminative time points from the time-series samples of each class to capture key local morphological variations, and constructs local prototypes carrying class-related local differential features. To further improve the ability of the CDP to capture multivariate fault coupling relationships, a multivariate coupling-aware prototype matching strategy (MCPM) is designed. The MCPM extends univariate prototypes into multivariate local prototype blocks and jointly measures local dissimilarity, variable correlation, and trend consistency, thereby enabling prototype learning with awareness of multivariate coupling relationships. Finally, to fuse local discriminative information and global temporal information, a dual-branch Transformer is constructed. LPG-Former encodes the differential features between the CDP and the best-fit subsequence (BFS) of the input sample through a Local Prototype Transformer, and complements global generic information through a Global Generic Transformer, thereby achieving collaborative dual-branch representation. Experimental results on an eight-class EHA operating-state dataset show that LPG-Former achieves an accuracy of 98.74% and an F1-score of 98.78%, significantly outperforming classical methods such as InceptionTime and TapNet.