Research On Rolling Bearing Fault Diagnosis Algorithm Based On Deep Learning

In recent years,with the development of industrial modernization,mechanical equipment are developing in the direction of large-scale,automation,and intelligence.The data acquisition system of mechanical equipment has become larger and more complex,which has brought new opportunities for the field of fault diagnosis.It also brings major challenges.In order to ensure the reliability,effectiveness and sustainability of the normal operation of mechanical equipment,it is urgent to ensure a real-time,accurate and efficient health monitoring system.Traditional fault diagnosis models based on feature extraction and model classification have been difficult to satisfy the complexity,variability and uncertainty of modern equipment faults.The data-driven deep learning model has achieved amazing results in many fields.It has the incomparable advantages over other data models in handling complex data.Based on this,this thesis takes rolling bearings as the research object and explores the application potential of deep learning in the field of bearing fault diagnosis.Aiming at some problems encountered in the practical application,several bearing fault diagnosis algorithms based on deep neural network are constructed.The main content of the thesis is as follows:First,this thesis verifies the feasibility of deep learning model in the field of bearing fault diagnosis.Three major deep learning algorithms,namely fully connected neural network,recurrent neural network and convolutional neural network,are modeled respectively.Taking the bearing data of Case Western Reserve University as the experimental data,three diagnostic models based on different network structures are presented.These models are trained and tested with raw vibration signal.Experiments show that these three types of deep neural network structures can achieve end-to-end fault diagnosis of raw bearing data.Second,aiming at the situation where bearing data under all loads cannot be obtained in industrial scenarios,a convolutional neural network model based on multiscale convolution kernels and Dropout is designed.This model enhances the ability to extract features by building multi-scale convolution kernels on the first layer.At the same time,the generalization ability of the model is improved through Dropout technology.Experiments under different load verify that the model can overcome the problem that the diagnostic recognition rate will decrease under variable load conditions.Last,aiming at the problem that most of the data-driven algorithms which focus on improving the performance of artificial data sets will degenerate rapidly in practical industrial applications,a fault diagnosis model based on residual learning is presented.The proposed model uses the residual structure to deepen the network,and adopts batch normalization layers and early stopping strategies to help network training.Taking the bearing data of the University of Paderborn as the experimental data,the experiment proves that the model can effectively extract the common features between artificial data and natural data.After training with artificial data,the model can be well adapted to real-time monitoring of real environments.