基于MHSA时-空域特征融合的金属氧化物避雷器故障诊断方法研究OA

To address the limitations of insufficient feature representation in single-modal approaches and the relatively high computational complexity of dual-stream fusion models in fault diagnosis of metal oxide arresters(MOAs),a temporal-spatial feature fusion diagnosis method based on the multi-head self-attention(MHSA)mechanism is proposed.First,partial discharge(PD)signals under four operating states are collected,and temporal samples are constructed using Hamming-window segmentation.Then,the signals are mapped into two-dimensional images through symmetrized dot pattern(SDP)transformation.In the proposed method,a bidirectional gated recurrent unit(Bi-GRU)is used to extract temporal features from PD signals,while EfficientNet-B0 is employed to extract spatial features from SDP images.MHSA is then introduced to achieve dynamically weighted fusion,and fault identification is finally performed using a Softmax classifier.Experimental results show that the proposed method achieves a diagnostic accuracy of 98.79%on the test set,with an average loss of 0.0126 and floating-point operations(FLOPs)of 623 M.Compared with single temporal-feature,single spatial-feature,and simple feature-concatenation fusion models,the accuracy is improved by approximately 7.22%,6.08%,and 2.36%,respectively.Feature contribution weight analysis further shows that the contribution weights of the temporal and spatial features are 38.12%and 61.87%,respectively,indicating their complementary roles in MOA fault diagnosis.