Transformer Winding Fault Detection Based On Multi-feature Fusion Algorithm And Unbalanced Data Set

Transformers provide important functions in the power system such as conversion of volt-age level and distribution of electricity.In the process of transportation and operation,deforma-tion may occur due to collisions,short-circuit impacts,etc.in transformer windings.Discovery of potential transformer winding deformation faults and maintenance in time is conducive to the stable operation of the power system.At present,there are many methods for non-intrusive detection of transformer winding state,such as frequency response analysis method,short circuit impedance method,low voltage impulse method,vibration analysis method and so on.Among them,the frequency response analysis method is widely used because of its advantages such as simple operation,sensitivity to minor winding deformation,high safety performance,and resistance to environmental inter-ference.The methods of analyzing the frequency response curve mainly include methods using mathematical statistics,methods based on simulation and modeling,and methods applying ar-tificial intelligence.In order to combine the advantages of different statistical indicators in the feature extrac-tion of transformer frequency response curves,an adaptive multi-feature fusion algorithm for transformer winding fault identification is proposed.When the data set is relatively balanced,the model training results show that the recognition accuracy of the single-feature SVM model is about 80%,and the accuracy of the proposed algorithm reaches 89.4%,which proves the ef-fectiveness of the algorithm.When a single feature SVM model is used for fault identification,the SVM model is easy to tolerate misclassification and cause misjudgment.Using the adap-tive multi-feature fusion algorithm can make the advantages of different feature SVM models complementary,and ultimately get a higher transformer fault recognition rate.The algorithm is simple to calculate and can effectively improve the efficiency of transformer fault diagnosis,but it has the disadvantage of a decline in the prediction accuracy when the data set is unbalanced.To solve the problems of low prediction accuracy uder unbalanced data set,long calcula-tion time,and large requirement of sample when using some machine learning methods for fault diagnosis,a fault detection technique for transformer windings using Siamese network under unbalanced data set is proposed in this study.And a technique to extract the characteristics of transformer frequency response curve using equivalent second-derivative-area paramaters is also proposed.By analyzing the fault sample data obtained in this paper,compared with the ex-isting judgment methods of FRA curve offset,the equivalent second-derivative-area paramaters proposed in this paper has higher accuracy and feasibility in frequency response curve feature extraction.By building parameter-optimized models of CNN,SVM,and correlation coefficient method,and calculating the absolute sum of Absolute Sum Logarithmic Error and the Min-Max Index as input features,the fault recognition effect is compared.and the equivalent second-order derivative area parameter is obtained.It is concluded that the model training effect is better when using the equivalent second-derivative-area paramaters as input,and the Siamese network with this feature as the input feature has the best effect in the transformer winding fault identification under the unbalanced data set in this study.The fault identification accuracy rate reaches 89.2%.The adaptive multi-feature fusion algorithm proposed in this study can improve the accu-racy of transformer windings fault detection effectively when the data set is relatively balanced.The equivalent second-derivative-area parameter can effectively extract the features of the fre-quency response curve.The fault detection technique for transformer windings using Siamese network can identify the fault more accurately under unbalanced data set.The research in this study is of great significance to improve the accuracy of transformer winding deformation fault diagnosis,which is conducive to early detection and repair of transformer faults,and to ensure the safe and stable operation of the power system.