Research On Bearing Fault Diagnosis Methods Based On Convolution Neural Network And Model Fusion

Rolling bearing is one of the most important parts of rotating machinery.With the rolling bearing working a long time under the harsh conditions of high speed and heavy load,it will inevitably produce performance degradation and cause faults.In order to ensure that there are no serious problems,it is necessary to collect data regularly to analyze its status.With the rapid development of rail transit industry,relying on the traditional artificial fault diagnosis technology cannot meet the needs at this era.Techniques of artificial intelligence have good performance in feature extraction and classification,so they are widely used in many fields.However,it is still an emerging industry in the field of fault diagnosis in complex and heavy industries,therefore the research on artificial intelligence deep learning and machine learning technology will be of great significance in the field of bearing fault diagnosis.This thesis will study the fault diagnosis methods of rolling bearings in rail transit.Firstly,the thesis analyzes the development of this field,the principle of bearing and the cause of failure.Subsequently,a convolutional neural network with three convolutional layers is proposed,which can realize end-to-end diagnosis,that is,directly use time-domain vibration signals as the input of this network model for diagnosis,and use data enhancement technology to effectively expand and batch the data set.Standardized technology.In this thesis,a standard data set of rolling bearings from Case Western Reserve University is used as the test data,and the network model can achieve an accuracy of more than 99% for its fault diagnosis identification.Through further analysis of rolling bearing vibration signals,a single-wide convolution kernel convolutional neural network SWKCNN model is proposed according to its characteristics to diagnose rolling bearing faults.This network model only uses one wide convolution kernel in the first convolutional layer,and uses small convolution kernels in all subsequent convolutional layers,and the batch standardization technology is also used.The author finds that the network model has a faster training speed and can achieve better fault diagnosis.The recognition rate of SWKCNN on the data set can reach 99.53%.Finally,the XGBoost(extreme gradient boosting algorithm)in Boosting algorithm,which performs a second-order Taylor expansion approximation of the loss function,can invoke CPU multi-threading to achieve parallel computation was used,thus improving the training speed.Through the fusion of the convolutional neural network model and XGBoost,the SWKCNN+XGBoost fusion model is proposed.The single-width convolutional neural network is used to extract features automatically,and then the XGBoost algorithm is used for classification.The accuracy of the fusion model can reach99.74 %,and the time required for diagnosis can be reduced while the accuracy is higher,and better bearing fault diagnosis performance can be obtained.