Multimodal Sensor Data Fusion for Hydraulic Piston Pump Fault Diagnosis Using a Spectrographic Residual Attention Network

Accurate fault diagnosis of hydraulic piston pumps operating under variable speeds and fluctuating loads remains a complex measurement challenge due to the highly coupled dynamics of fluidic and mechanical degradation. In aviation hydraulic systems, undetected pump degradation can lead to catastrophic flight control failures, making robust real-time fault diagnosis a fundamental safety-critical engineering requirement. Traditional single-domain analysis frequently fails to capture the complete fault signature. To overcome this, this paper proposes a novel multimodal sensor fusion framework, Res-MSCNN-SE-BiLSTM, for the robust diagnosis of four pump health states: nominal operation, regulation valve wear, servo piston wear, and swashplate bearing bush wear. Addressing the limitations of 1D time-series analysis, synchronized pressure and vibration measurements are transformed into 2D Short-Time Fourier Transform (STFT) RGB spectrograms, yielding a unified time-frequency representation that effectively handles non-stationary speed profiles. Within the architecture, a Squeeze-and-Excitation (SE) attention block adaptively recalibrates the sensor channels, dynamically weighting the most relevant physical domain for each fault type. Parallel multi-scale residual convolutions extract heterogeneous spatial fault signatures, while a Bidirectional LSTM encoder explicitly models the chronological sequence of mechanical-hydraulic transient events within the diagnostic window. Evaluated on a comprehensive multimodal pump dataset using a strict sequence-based train-test data split to prevent data leakage, the proposed framework achieves a classification accuracy of 99.60%, a Macro G-mean of 99.73%, and an AUC of 0.9999. The empirical results confirm that the proposed multi-sensor fusion strategy successfully disentangles complex degradation patterns, offering a highly reliable methodology for the intelligent condition monitoring of critical hydraulic machinery.