摘要详情

ID / 提交时间

72 / 2025-03-30 21:52:45

标题

Fault Diagnosis of Load Rejection Conditions in Pump-Turbines via Integration of Slow Feature Analysis and Physical Constraints

关键字

HardPINN, Fault diagnosis, SFA, Mechanism-data collaborative modeling, Pump-turbine load rejection conditions

主题及专题

默认会议主题

状态

全文待审

作者

雨辰何 / 中国计量大学

云王 / 浙江同济科技职业技术学院

家杰胡 / 中国计量大学

杰陈 / 中国计量大学

晓松王 / 浙江华电乌溪江混合抽水蓄能发电有限公司

摘要

This paper proposes a Hard Constrained Physics-informed Neural Network (HardPINN) integrated with Slow Feature Analysis (SFA) to address the challenging problem of fault diagnosis in pump-turbines under load rejection transient conditions. During load rejection, the strong multi-physics field coupling characteristics of the water-machine-electric system result in inaccurate transient flow patterns modeling using traditional physical models. Meanwhile, existing data-driven methods suffer from poor generalization performance and insufficient extraction of weak fault features due to neglecting physical mechanism constraints. To address these issues, HardPINN embeds core physical laws, such as the continuity and momentum equations, into the neural network loss function as strong constraints, enabling dynamic modeling of multi-physics field coupling. Additionally, SFA is introduced to suppress transient high-frequency noise interference and enhance sensitivity to slow-varying fault features. Experiments on pump-turbine datasets of load rejection condition validate the effectiveness of HardPINN, which achieves an average accuracy of 90.72% in diagnosing three typical faults (penstock rupture, guide vane asynchrony, and runner imbalance) and outperforms SFA (85.87%), CNN (80.69%), and GMM (38.66%). By implementing mechanism-data collaborative modeling, HardPINN resolves the gradient conflict between transient responses and long-term degradation features, providing a robust solution for the intelligent operation and maintenance of hydropower systems.