A Health Condition Assessment and Safety Early Warning Framework for Hydropower Equipment Using Multi-Source Data Fusion and Decision Support Systems

Abstract

Hydropower plants heavily depend on the performance of turbine generators, where unexpected equipment failure can cause downtime, loss, and even safety issues in the process. The traditional condition monitoring methods mainly focus on vibration signals detected by a single sensor, which is not integrated with multiple sources and lacks the ability to perceive the risk at the plant level. In addition, the majority of the studies do not consider the equipment level health condition and structural risk factors to develop unified decision support for safety issues. To address the issues, the paper proposes a framework for the integrated health assessment and safety early warning framework, which is developed by integrating the equipment level fault diagnosis model based on the CNN-LSTM network and the structural risk assessment model developed by the GloHydroRes dataset. The novelty of the proposed multi-level data fusion framework integrates equipment-level and plant-level information for safety assessment that incorporates multi-channel vibration and torque signals along with structural attributes such as dam height, reservoir volume, and installed capacity to derive a unified safety risk index. Experimental validation using the proposed approach on the SEU multi-sensor dataset achieved high classification performance with Accuracy of 0.9634, Precision of 0.8642, Recall of 0.9613, F1-score of 0.9065, and MCC of 0.8662. Moreover, multi-class ROC analysis indicated high performance with AUC values of 0.9936 (Low Risk), 0.9930 (Medium Risk), and 0.9997 (High Risk), outperforming individual CNN, LSTM, and Random Forest approaches. The unified safety index revealed that 46.6% of samples were classified as Low Risk, 34.1% of samples were classified as Medium Risk, and 19.3% of samples were classified as High Risk, thus validating the efficacy of the proposed decision support mechanism. Compared with traditional fault detection methods, the proposed approach improves the reliability of fault prediction results, reduces false alarm rates, and enables proactive maintenance decisions based on risk considerations.