Research On Fault Diagnosis Of Electrical Equipment Fused With Unbalanced Fault Text And Monitoring Data Sample Set

In order to ensure the safe and stable operation of the power grid,it is of great significance to detect and deal with electrical equipment failures in time.At present,artificial intelligence technology is widely used in transformer fault diagnosis,but the actual implementation has certain requirements for the sample set,and the actual collected electrical equipment fault case reports have problems such as noisy vocabulary,small number of samples,unbalanced distribution of fault types,and noisy labels.,which will greatly limit the practical application effect of the fault diagnosis model,and the existing fault diagnosis model still has the problem of single diagnosis basis.In order to solve the above problems,this paper establishes a BERT(Bidirectional Encoder Representation from Transformers)based electrical equipment fault diagnosis model.Based on this model,the following related research is carried out.First,the impact of textual noise labels on the fault diagnosis performance of the model is clarified.The principle of BERT’s parallel processing of text information,multi-head attention calculation mechanism and downstream label mapping is explored,the text embedding matrix and aggregated semantic vector are obtained,and a fault diagnosis model for electrical equipment is constructed.Collect fault text sample sets that have been preliminarily evaluated and marked by on-site personnel,make fault text sample sets containing 80%,50%and 20%noise labels respectively,and then use the fault text sample sets containing noise labels and label after consultation with experts The fault diagnosis performance of the model is evaluated with the fault text sample set,and the influence degree of different proportions of noise labels on the fault diagnosis performance of the model is compared.The results of the calculation example show that the fault text sample set containing only 20%noise labels has a great negative impact on the fault diagnosis performance of the model,and the higher the noise label content is,the worse the fault diagnosis performance of the model is,and the initial loss value during training is The higher the value,the slower the descent speed and the longer the training time.Then,an information fusion fault diagnosis model based on BERT+1D-CNN is proposed.The classification principles of the BERT model based on text classification and the KNN and 1D-CNN models based on numerical data classification are explored.Analyze the method of KNN model to obtain k value through five-fold cross-validation,study the structure of traditional 1D-CNN,and improve the structure of traditional 1D-CNN according to the characteristics of gas monitoring data,and establish fault diagnosis models for single gas monitoring data.The text fault diagnosis results are used as feature quantities and gas feature quantities for information fusion,and jointly guide the KNN and improved 1D-CNN models to make fault diagnosis decisions.The calculation results prove that the performance of BERT+KNN and BERT+1D-CNN information fusion fault diagnosis model is higher than that of BERT single fault text diagnosis model,KNN,1D-CNN single gas monitoring data diagnosis model.Compared with the average value,BERT+KNN is 0.91%higher in precision than BERT+1D-CNN,and the F1 value is 0.96%higher.Finally,a KNN SMOTE sample balancing method is proposed.The sample balance principle of traditional undersampling and oversampling is explored.Based on the oversampling method and the KNN classifier,focus on boundary class samples to synthesize new samples.The results of the calculation example prove that the KNN SMOTE method has a more positive effect on the classification performance of the model than the traditional sample balancing methods Near Miss and SMOTE.Compared with the average value,the BERT+KNN model processed by KNN SMOTE has a 0.18%increase in precision than the unprocessed P value,and the F1 value has increased by 0.10%.The BERT+1D-CNN model processed by KNN SMOTE is more accurate than the unprocessed Rate P value increased by 5.44%,F1 value increased by 5.36%,the effect is obvious.