Multimodal Fusion Method Migration Application From Sentiment Analysis To Transformer Fault Diagnosis

As one of the core equipment of substation,transformer plays an important role in connecting power grid,regulating reactive power flow and stabilizing load center voltage.Its safety and reliability are directly related to the safe and stable operation of power grid.As the scale of the power grid continues to expand,the importance of transformers continues to increase,and the accuracy requirements for transformer fault diagnosis are gradually increasing.Traditional transformer fault diagnosis is based on a mode of gas data,but when the transformer fails,it is not only manifested in a mode of gas.A single mode carries limited fault information,which limits the improvement of accuracy.Therefore,this paper proposes to construct a multi-modal fault data set and perform transformer fault diagnosis through multi-modal fusion.The research idea of this paper is to learn and innovate the multi-modal fusion model in the field of multi-modal sentiment analysis,and apply it to the transformer fault diagnosis scenario based on multi-modal fusion.The innovative research results of this paper are summarized as follows :Multimodal fusion model : Firstly,this paper draws on the Transformer model,the Multimodal Transformer(Mul T)model and the low-rank multi-modal fusion model LMF,which is inspired by the LMF-Mul T model.After connecting the LMF model to the Mul T model,an innovative model Mul T-LMF is proposed,which transforms the multi-modal fusion network from multi-modal feature splicing to tensor fusion.The model can well solve the problem of heterogeneity and redundancy of multi-modal data encountered in multi-modal fusion,but when applied to multiple modal scenarios of transformer fault diagnosis,the model will bring the problem of parameter surge.Secondly,in order to solve the problem of model parameters and the heterogeneity and redundancy of multi-modal data,this paper draws on the idea of feature separation in the MISA model based on modal common and private features,maps the features of different modes to the modal common and private feature space respectively,and separates the modal common and private features.The cross-modal Transformer model is used to enhance the cross-modal features of common and private modal features,and an innovative model MISA-CT is proposed to solve the problem of modal parameters and improve the efficiency of multi-modal fusion.Multi-modal transformer fault diagnosis : when the transformer fails,it will not only be manifested in one mode of gas solubility,but also in other modal data.The amount of information carried by a mode cannot contain all the fault information,and the accuracy of fault diagnosis is difficult to improve.Multi-modal data sources contain more fault information,which can improve the accuracy of fault diagnosis.In other multi-modal fields such as multi-modal sentiment analysis,multi-modal ideas have been well proved.In this paper,the collected multi-sensor data of 110 KV transformer is used as the multi-modal transformer fault data set.Five modal data of gas,UHF,vibration,infrared image and surveillance video are collected.Six states of normal,overheating fault,bushing rupture,oil leakage,partial discharge and overload are selected as experimental labels.The innovative model MISA-CT of multi-modal sentiment analysis is applied to the multi-modal transformer fault diagnosis scenario.Compared with the latest IEC three-ratio method,support vector machine SVM and CWGAN-GP based on adversarial neural network,the fault diagnosis accuracy of the innovative model MISA-CT has been significantly improved.