Vibration-sensor-free diagnosis for aviation piston pumps: A CWT-ViT approach with high-dimensional tensor fusion

Diagnosing faults in axial piston pumps within space-constrained Electro-Hydrostatic Actuators (EHA) is challenging without dedicated vibration sensors. Traditional methods often rely on such sensors or struggle to model complex faults using single-source data. To overcome these limitations, this paper introduces a novel, vibration-sensor-free fault diagnosis method based on multi-source data fusion and a modified Vision Transformer (ViT). The proposed method transforms inherent sensor signals (flow, pressure, and leakage) into time–frequency spectrograms using Continuous Wavelet Transform. Crucially, these are fused into a high-dimensional nine-channel tensor, a strategy that — unlike conventional grayscale conversion — preserves the intricate dynamic coupling between different physical quantities, thereby preventing information loss. To effectively analyze this rich tensor representation, we introduce a modified ViT. Its architecture is adapted with a multi-channel input layer to process the tensor directly, while its Multi-Head Self-Attention mechanism excels at capturing the global, cross-channel dependencies often missed by conventional CNNs with their limited receptive fields. Experimental results demonstrate an outstanding classification accuracy of 98.59% across eight fault states, significantly outperforming competing methods. By successfully integrating a high-dimensional fusion strategy with a global feature-learning architecture, this study establishes a powerful and practical paradigm for intelligent fault diagnosis in sensor-limited aviation systems without requiring additional hardware.