基于物理引导多模态映射融合网络的抽水蓄能机组暂态故障诊断策略OA

Aiming at the problems of insufficient fusion of multi-source data and inadequate utilization of physical mechanisms in transient fault diagnosis of pumped storage units,this paper proposes a transient fault diagnosis strategy for pumped storage units based on a Physics-guided Multi-Modal Graph Fusion Network(PG-MGFN).This approach deeply integrates physical model constraints,graph representation learning,and adaptive fusion mechanisms to overcome the limitations of traditional Multi-Modal Deep Fusion Networks(MDFN)in complex systems.First,an a priori constraint model is constructed based on the electromagnetic transient characteristics of the unit's operation to preprocess multi-modal data such as vibration,current,and pressure.Raw signals are calibrated using physical indicators to enhance the physical interpretability of features.Second,data from each modality is mapped into dynamic graph structures that include node attributes and cross-modal correlations.Graph Neural Networks(GNNs)are employed to capture hierarchical features within modalities and non-linear coupling relationships between modalities.Finally,a physics-guided attention mechanism is designed to adaptively adjust the fusion weights of multi-modal graph features,achieving directional enhancement of fault features.Experimental results show that compared with the traditional MDFN,PG-MGFN has achieved remarkable improvements in fault diagnosis accuracy,model generalization ability,and robustness.In typical fault diagnosis tasks such as bearing and blade faults,PG-MGFN achieves an average accuracy of 93.2％,which is 18.7％higher than that of the traditional MDFN.It maintains a recognition accuracy of 99.5％in a 20 dB noise environment,verifying its excellent anti-interference capability.This provides new technical support for the intelligent operation and maintenance of pumped storage units.