Abstract: To address the problems that existing leakage fault diagnosis methods for hydraulic cylinders mostly rely on manually designed features, have difficulty in fully characterizing sample differences among multi-specification hydraulic cylinders of hydraulic supports in fully mechanized top-coal caving faces, and struggle to model dynamic coupling between rod-chamber and rodless-chamber pressures, this study proposed a leakage fault diagnosis method for hydraulic cylinders of hydraulic supports in fully mechanized top-coal caving faces based on a Physics-Prior Dual Cross-Attention Network (PPDCAN). Using a simulation experiment system for hydraulic-cylinder leakage faults, dual-channel pressure signals were collected from hydraulic cylinders of six specifications under five operating conditions: normal operation, internal leakage, external leakage from the rodless chamber, external leakage from the rod chamber, and compound leakage. To address the large differences in original signal sample length, the difficulty of long-sequence modeling, and obvious distribution shifts among different hydraulic cylinders, Z-Score standardization, segmentation, linear interpolation, and the Short-Time Fourier Transform were used to construct a unified two-dimensional time-frequency feature representation. According to the characteristics of the two-dimensional time-frequency representation, a PPDCAN was designed. The Dual Cross-Attention (DCA) module explicitly established bidirectional interactions between the time-frequency features of rod-chamber and rodless-chamber pressures, enhancing the network’s ability to represent cross-channel coupling anomalies caused by leakage faults. Meanwhile, a Transformer encoder was used to extract deep time-frequency features. Cross-entropy loss and Physics Prior Guided Leakage Consistency Loss (PPLCLoss) were introduced so that the network could account for both fault discrimination capability and leakage-mechanism constraints. The experimental results showed that the optimal diagnostic accuracy was obtained when the DCA module and PPLCLoss acted jointly. The accuracy of the PPDCAN reached 91.76%, and the accuracy after merging external-leakage fault classes reached 96.73%, outperforming the other baseline networks. The network confined misclassifications to external-leakage fault classes with similar physical features, reduced confusion among normal, internal-leakage, and external-leakage states, and demonstrated stronger leakage fault identification capability. The extracted features also showed better intra-class compactness and inter-class separability.