Fault diagnosis method for hydraulic cylinders of fully mechanized top-coal caving supports based on an PPDCAN

To address the challenges of strong concealment, complex dual-channel pressure coupling, significant differences in sample lengths across various hydraulic cylinders, and the lack of physical interpretability in traditional data-driven methods, a fault diagnosis method for hydraulic cylinder leakage in fully mechanized top-coal caving supports is proposed, based on physics priors and dual cross-attention networks. The method utilizes a hydraulic cylinder leakage fault simulation experimental system, which collects dual-channel pressure signals from cylinders of six different specifications under five working conditions: normal, internal leakage, external leakage in the rodless chamber, external leakage in the rod chamber, and composite leakage. To address issues such as inconsistent sample lengths, high computational costs of long-sequence modeling, and significant distribution shifts between different cylinders, a unified two-dimensional time-frequency representation is constructed using Z-score normalization, segmentation, linear interpolation, and Short-Time Fourier Transform. A physics-informed dual cross-attention network is then designed, with the dual cross-attention module establishing a symmetric interaction relationship between the dual-channel pressure signals, while a Transformer encoder is employed for deep time-frequency feature extraction. Additionally, a composite objective function is introduced, consisting of a classification loss and a physics-prior-guided leakage consistency loss, allowing the model to learn fault discrimination information while adhering to hydraulic leakage mechanisms. Experimental results show that the proposed method achieves an accuracy of 91.76% on the test set, with a fusion accuracy of 96.73% for external leakage faults, representing improvements of 3.27% and 3.73%, respectively, compared to the best baseline models. Ablation studies further confirm the effectiveness of the dual cross-attention module and the physics-prior loss. This method effectively identifies leakage faults in hydraulic cylinders of fully mechanized top-coal caving supports under sparse fluid field pressure information conditions, demonstrating strong potential for practical engineering applications.