The Application Research Of Convolutional Neural Network Based Method Of Motor Fault Diagnosis

In today’s society,motor,as one kind of energy conversion device,is being widely used in people’s producing process and daily life.Once it fails,economic loss and production efficiency will be caused directly.Heavily,it even may threaten operators’ life safety.Therefore,the failure or unstable operating state of the motor should be eliminated preliminary.This is of great significance for ensuring the normal operation of the production and living system.However,traditional methods of fault diagnosis only work well as the data sets were not that big.In today’s new mission of which there are always huge data sets and in which the faults are often happen in the mean time.Traditional methods won’t have an excellent accurate and they always need a great number of time.In this context,this thesis focused on multiple fault diagnosis task of large vibration data of traction motor.To improve the accuracy and timeliness,both feature engineering and feature learning method were used in this thesis.To solve the precision and timeliness problems under large data set environment of traditional feature engineering methods,a Cross Validation optimized Stochastic Gradient Decent based motor fault diagnosis method were proposed in this thesis.This method solved the problem of local minimum and saddle points which caused by SGD’s uncertainty direction of falling steps.Then,this thesis has conducted the thorough research to deep learning theories.To design a suitable network for multiple fault diagnosis,Le Net was studied and improved into a new convolutional neural network.And this network has 2convolution layers in series.LRN layers were added in both convolution layers.In the second convolution layer,LRN and pooling layer were transpositioned in the second convolutional layer.Finally,traditional feature engineering,CV+SGDBPNN and improved CNN were compared under Tensor Flow.In large dataset environment,CV+SGDBPNN’s performance was going up as the amount of data’s growing.It finally got an accuracy of 88.24%.Had a reconfiguration signal as result,de-noising auto encoder got an accuracy of 93.61%.And the improved convolutional neural network,which whith 2tandemed convolutional layers achieved the best proformance,accuracy score 97.71%on average.There wre even sometimes scored more than 98%.Also,of otherindicatiors,such as precision,recall and F1-score,they were all superior to traditional feature engineering methods of motor fault diagnosis.