Gearbox Composite Fault Diagnosis Based On Multi-task Deep Learning

Mechanical fault diagnosis has entered an era of big data.Besides,with the powerful capabilities in adaptive feature extraction and classification,deep learning also has made great achievement in mechanical big data processing.However,these researches are all applied in single-label system to diagnose faults of single target.In the context of big data,single-label system not only split the connection among various faults in machinery,but also be difficult to fully describe various health status information including the location,type and degree of faults.This paper aims at this problem to present a diagnostic method based on multi-task deep learning model,which can simultaneously diagnose faults in two targets of gear case,bearings and gears.At first,due to insufficient experimental data,this paper takes a reference to the bearing data acquisition methods raised by Case Western Reserve University.Taking the integrated bench for power transmission fault diagnosis as research object,this paper designs and collects data of composite fault in 30 types,which provides data guarantee for the following research.At the same time,the model pre-study shows that the one-dimensional convolution method has the best feature extraction ability on the data set.Based on the data characteristics of composite faults in gearbox,this paper firstly puts forward a mutli-task deep learning model base on one-dimensional convolutional neural network,Multi task Convolutional Neural Network(MT-CNN).In this model,shared layers extracts data's shared features at first,then,task layers extract the fault features of gears and bearings respectively,and diagnosed data's vibration signal in frequency domain.As the result,the combined accuracy of recognition reaches 94.6%.Targeted on the problem of slow convergence rate and long training time of the original model,this paper uses batch normalization(BN)to optimize the network.In addition,it increases the kernel size in both shared layers and task layers,so as to increase the receptive field.Therefore,model in this paper advances to MT-CNN based on BN.The result of 4-fold cross validation shows that,after batch normalization,the convergence rate of MT-CNN gets significantly improved,and the average combined accuracy of recognition reaches 96%,which is a huge increase when compared with the original network model without batch normalization.Considering the difficulty to understand deep learning network,this paper also applies model visualization technology to visualize the kernels in task layers and verify its feature extraction ability.Therefore,it is able to present model's classification process by visualization technology.In order to analyze the performance of model better,extended experiments are carried out as well.By means of training with missing data,this paper studies the generalization ability of network model when training in the absence of load or speed data.The result shows that combined accuracy still stay 92% in the absence of load data;while without speed data,the network turns out severe over-fitting,and its accuracy is only 76%.These shows that model's adaptability to rotation speed is poor and there is still room for further improvement.