Fault Data Preprocessing and Diagnostic Technology for Hydropower Units Based on Deep Learning Algorithms

Abstract

Fault diagnosis of hydropower generating units faces the dual challenges of strong noise interference in monitoring data and scarcity of fault samples. These issues are also encountered in intelligent sensing systems operating under complex electromagnetic environments, where electromagnetic interference and signal degradation may obscure weak fault characteristics and reduce monitoring reliability. To address the lack of theoretical guidance for parameter selection in improved wavelet threshold functions and the poor training stability of generative adversarial networks, this paper proposes a comprehensive fault diagnosis framework integrating adaptive wavelet denoising with stable generative adversarial networks. A continuous and differentiable threshold function with an adjustment factor is first constructed, and a quantitative relationship between the adjustment factor and noise level is established, enabling an output SNR above 5.8 dB even when the input SNR decreases to -2 dB. Subsequently, a gradient-penalized generative adversarial network based on Wasserstein distance is developed, with convergence analysis demonstrating that stable critic loss reflects reliable estimation of the Wasserstein distance between data distributions. Finally, a CNN-BiLSTM-Attention hybrid network is introduced to combine spatial feature extraction, temporal dependency modeling, and adaptive attention to emphasize critical fault information. A physical constraint loss is further incorporated into GAN training to improve the realism of generated samples. Validation using operational data from nine hydropower generating units demonstrates that the proposed denoising strategy improves the output SNR by 7.32 dB at an input SNR of 6 dB and maintains a 5.1 dB gain at -2 dB, while the complete framework achieves a fault diagnosis accuracy of 97.8%, outperforming models without preprocessing by 12.1 percentage points. The proposed approach provides an effective solution for intelligent fault diagnosis in noisy sensing environments and offers valuable support for reliable signal processing in electromagnetic monitoring and industrial condition assessment.