Intelligent fault diagnosis and localization of hydropower station equipment using multi-source data fusion and deep residual networks

Abstract

Reliable fault diagnosis is important for rotating equipment used in energy infrastructure, including hydropower turbine–generator systems. A hydropower-oriented framework is presented that combines a physics-constrained simulation-assisted augmentation module, modality-aware feature fusion, a physics-informed deep residual network, and adversarial domain alignment. The physical formulation incorporates hydraulic-power, rotor-dynamic, and generator-electrical residuals; however, the available evaluation data comprise public rotating-machinery benchmarks rather than synchronized measurements from an operating hydropower plant. The experimental evidence therefore evaluates diagnostic performance and cross-benchmark transfer on CWRU, Paderborn, MFPT, and compound rotating-machinery data, not direct hydropower deployment. The reported aggregate results are 98.6% accuracy and 98.4% macro-F1, while the reported cross-benchmark average is 97.4% accuracy. At an SNR of 5 dB, the reported accuracy is 94.2%. Per-sample latency is 8.2 ms on an NVIDIA A100 GPU; this measurement is treated as a server-class reference and does not establish edge-device or plant-level real-time suitability. The results indicate the potential value of integrating physical residuals, simulation-assisted augmentation, and modality-aware representation learning for rotating-equipment diagnosis, while plant-specific digital-twin calibration, multi-sensor hydropower validation, quantitative domain-discrepancy analysis, and edge deployment remain necessary before operational use.