Non-invasive rotor fault measurement in pumped storage systems via enhanced energy entropy analysis and hybrid deep learning with metrological validation

Abstract

To advance measurement methodologies for rotor health monitoring in pumped storage systems, this study proposes a novel non-invasive measurement framework integrating Motor Current Signature Analysis (MCSA) with hybrid deep learning. This method quantifies energy entropy shifts by employing Cyclic Autocorrelation Function (CAF) and Empirical Mode Decomposition (EMD), thereby addressing key metrological challenges including signal noise suppression and uncertainty propagation. Principal Component Analysis (PCA) is rigorously applied to reduce feature dimensionality while preserving measurement integrity. A novel SCSO-CNN-BiLSTM-Attention model is proposed, combining CNN’s spatial feature extraction, BiLSTM’s temporal dependency modeling, and attention-based critical weighting, optimized via Sand Cat Swarm Optimization (SCSO) for hyperparameter tuning with traceable uncertainty budgets. Validated on seven fault conditions, the model achieves 95.24% testing accuracy, outperforming traditional methods in both metrological robustness and computational efficiency. Key innovations include IMF1’s high-frequency sensitivity and IMF4’s low-frequency discriminability. T-SNE visualization confirms enhanced clustering through integrated RMS and IMF entropy features. Despite residual errors in cavitation-coupled faults, the framework demonstrates robust generalizability and real-time potential. This work provides a physics-informed AI solution for centrifugal pump health monitoring, balancing diagnostic precision with computational efficiency for industrial deployment.