Multi-Modal Transformer Fault Diagnosis And Identification

With the rapid development of domestic electric power industry,the voltage level of transmission and transformation lines has been increasing,and ultra-high voltage and large capacity power transformers have become an important cornerstone of electric power transmission.And in the context of big data,multi-modal transformer fault diagnosis based on multi-modal is more advantageous than traditional unimodal diagnosis such as oil chromatography and vibration.Meanwhile,compared with the single data-driven method,the mechanism-based fault diagnosis method can analyze the fault mechanisms of different transformer components more deeply and establish corresponding physical models,which fundamentally improves the diagnosis accuracy.To address the shortcomings of traditional transformer fault diagnosis,this paper investigates a multi-modal,data and mechanism fusion-based transformer fault diagnosis and identification method.Firstly,multiple data of transformers are collected through multi-modal sensors,and the data are cleaned and pre-processed.Subsequently,DS evidence theory is used as the base model,Person correlation coefficient and fuzzy set theory are introduced,and the reliability between the gas production rate and each evidence body is quantified and analyzed by using Person coefficient to improve DS evidence theory.Then the fault threshold interval is fuzzified and the fuzzy theory approach is used to generate the affiliation function,which then acts as the confidence function of the transformer fault.Finally,the multi-modal data are fused with the improved DS evidence theory to diagnose the presence of faults in power transformers.Based on the initial diagnosis,if a fault exists,the type of fault is further analyzed.At this point a multi-modal based transformer fault identification model is constructed,which is divided into two layers of intelligent algorithm training.The first layer learns the classification of each attribute based on the feature values to obtain the attribute values.The second layer performs classification training directly on the upper layer of learners to obtain specific faults.On this basis,the concept of zero-sample learning is introduced to combine the supervised attributes with the transformer knowledge base to construct a fault attribute matrix,and to achieve the diagnosis of invisible class faults with the help of zero-sample migration learning method,which effectively improves the flexibility of the system.To address the problem of low accuracy in the diagnosis process due to sample imbalance,this paper introduces a sample imbalance optimization method combining SMOTE and Easy Ensemble.The method employs the SMOTE algorithm to perform oversampling at the data level,and then uses the Easy Ensemble algorithm to divide the majority class samples of the dataset,and trains the divided samples with the minority class samples to merge into a weak classifier for fault identification by weighted averaging.This method effectively mitigates the impact of sample imbalance on transformer fault diagnosis.