Research On Multi-classification Method Of Smart Meters Fault Based On Deep Learning

As one of the important terminal devices of the power grid system,smart meters play an important role in supporting the realization of informatization and intelligence of the power grid.With the widespread application of smart meters,the increasing complexity of their functions has brought diversified types of operating faults.The accurate and rapid classification of fault types of smart meters is of great significance to improve the reliable operation of the electricity collection system and reduce operating costs.At present,smart meter fault data has the characteristics of diverse types and obvious differences in the number of categories.At the same time,as an embedded system,its on-chip computing and storage resources are limited.It is difficult to accurately and quickly classify the types of power meter faults.Based on this,this thesis studies a multi-classification method of smart meters fault based on deep learning.The main work of the thesis is as follows:Firstly,data preprocessing and feature selection technology are studied on of smart meters fault.The data set is cleaned with duplicate value removal,missing value deletion or filling to make the data regular;feature attributes are selected through correlation analysis,and the importance of feature variables is sorted by random forest algorithm to remove redundancy features and irrelevant features to filter out the highcorrelation meter fault feature attribute set;in order to further improve the data quality,design an autoencoder model to learn the distribution of various data samples.It uses the mean square error as the restoration error and sets different threshold for each class to realize the detection and deletion of abnormal values,making the fault data set more conducive to the construction of subsequent classification models.Secondly,the multi-classification method of smart meter fault type based on deep learning under imbalanced data sets is studied.Aiming at the characteristics of diverse fault types and imbalance in the number of samples between categories,an imbalanced multi-classification method based on the CVAE-CNN model is proposed.The model takes category labels as constraints and builds a CVAE network composed of fully connected layers to generate minority class samples.Modeling the hidden variables that obey the multi-dimensional and independent Gaussian distribution in each dimension according to the lower bound of variation.The network learns various distribution characteristics and the global characteristics of the data set to improve the quality of the generated data.The balanced data is classified using the CNN network,a one-dimensional convolutional layer is designed to extract the complex features hidden in the data,the maximum pooling method is constructed to improve the fault tolerance rate of the model,and the classification is processed according to various distribution characteristics to improve the recognition rate of the minority class.The advancement and effectiveness of the proposed method are verified on the public imbalance data set and the actual electricity meter fault data set.Finally,the compression method of the smart meter fault multiclassification model based on knowledge distillation is studied.Aiming at the characteristics of smart meters with limited resources such as on-chip computing and storage,a multi-classification method for knowledge distillation of smart meters based on feature learning is proposed.This method is designed to transfer the output response and the hidden layer features of the teacher network CNN to the student network MLP,and establish the relationship between the extension MLP hidden layer feature dimension and the CNN hidden layer feature output,and construct the mean square deviation between the two as a "feature target" added to the student network loss function.Controlling the weight of the "hard target","soft target" and feature learning loss in the overall loss of MLP to dynamically adjust their importance to obtain more effective information,so that the student network can not only learn the knowledge in the output results of the teacher network,but also fit the learning according to the hidden layer features of the teacher network,which improves the classification accuracy and reduces the computational complexity.The public data set and the smart meter fault data set are used to verify the superiority of the proposed method,which provides a reference for the subsequent migration and application of the classification model to the smart meter platform.