Motor Bearing Fault Diagnosis Based On Convolutional Neural Network

The important position of the motor in the industrial production is self-evident,greatly promoting the productivity,the harvest of huge economic benefits,but in the bad working conditions for a long time the motor often failure,resulting in serious safety accidents.With the development of equipment monitoring technology,a large number of motor running state data have been accumulated.With the help of deep neural network model,fault features in the running state data can be extracted automatically.Therefore,it is of practical value to study the application of deep neural network model in motor fault diagnosis.In order to ensure the safe operation of the motor and get rid of the shortcomings of the traditional diagnosis technology,this paper applied the deep diagnosis model based on convolutional neural network to the vibration signal recognition of the motor bearing,and successfully realized the end-to-end fast diagnosis.The main research work and success of this paper are as follows:Aiming at one-dimensional time-domain vibration signals,the convolutional neural network model is studied and modified,and a one-dimensional three-layer convolutional neural network model is built,and its structure and parameter setting are explained.Sliding window resampling and other methods are used to preprocess CWRU open source data.By inputting the one-dimensional time-domain vibration signal into the model,the recognition rate of 99% is better than that of the same model,which realizes the end-to-end convenient diagnosis and reduces the complexity of motor fault diagnosis.On the basis of the three-layer convolutional neural network,the number of layers is deepened to six layers.According to the design idea of the convolutional kernel proposed in this paper,the first layer convolutional kernel is widened to expand the receptive field.In order to make the model learn the characteristics independent of mechanical displacement,the basis of setting some superparameters of the model is put forward.After the introduction of BN layer,the recognition rate of CWRU data by the new model is improved to 100%,it is verified that the reasonable deepening of convolutional neural network can indeed improve the ability of model recognition.The relationship between the amount of training data and the effectiveness of the model is reflected through comparative experiments,which provides a basis for the sampling and production of data sets in engineering practice.In order to broaden the available application fields of convolutional neural network in this paper,the VGG model of convolutional neural network is modified to build a two-dimensional convolutional neural network model.The open source bearing data of the University of Turin were preprocessed,including classification and labeling,and a new data set was produced.The data of three channels in the new dataset are extracted to generate two-dimensional visual images to reflect their features.The new data set was input into the model,and the recognition rate reached 99.993%.The T-SNE visualization tool was used to show the classification.The two-dimensional model is optimized and transformed.The characteristics of transfer learning are utilized to solve the remaining problems in three aspects of time consumption,data volume and computer resources,so as to obtain a more efficient and universal two-dimensional model,which is verified by comparative experiments.