Research On Bearing Fault Diagnosis Method Based On Raw Vibration Data And Deep Learning

As one of the core components of mechanical equipment,rolling bearings are widely used in the field of industrial manufacturing and people’s daily life.It is of great significance to carry out fault diagnosis research on rolling bearings.Traditional fault diagnosis methods require manual extraction of features,rely on expert experience,and are not intelligent enough.With the development of science and technology,the era of mechanical big data has arrived,and industrial machinery systems are becoming more and more complex and intelligent.Research on data-driven intelligent fault diagnosis methods has become an urgent need to achieve fast,stable,and high-quality bearing fault diagnosis.In response to the above problems,this thesis takes rolling bearings as the research object,focuses on the input of raw bearing vibration data,and designs and implements three rolling bearing fault diagnosis methods based on deep learning theory.Firstly,the autoencoder is applied to bearing fault diagnosis.To solve the problem of insufficient feature expression ability of the three-layer autoencoder,this thesis uses the stack autoencoder.Considering that too many neurons in the middle layer will produce redundancy,and too few neurons in the middle layer will not be able to effectively rebuild the input sample,this thesis introduces the Dropout regularization technology in the middle layer.An autoencoder was used to extract bearing fault features,and support vector machine was used as classifier to realize bearing fault state recognition.Existing deep learning fault diagnosis models are generally based on time-frequency diagram input.However,the step of transforming time-frequency diagram requires the professional background knowledge of researchers and is prone to cause the loss of useful information.This thesis improves the existing fault diagnosis model and designs a twodimensional convolutional neural network fault diagnosis model based on single-channel grayscale image input.This method avoids the method of signal processing to obtain the input data in frequency domain,and directly splices the one-dimensional time-domain data into a two-dimensional matrix with a single channel.A neural network model based on single-channel two-dimensional matrix input was built.Adam gradient descent algorithm was used to update the model parameters,and Dropout layer was added to avoid overfitting of the model.High precision bearing fault state identification was realized.In order to further improve the accuracy and efficiency of fault diagnosis,a bearing fault state identification method based on one-dimensional convolutional neural network is proposed in this thesis.The reduction of the number of model parameters significantly speeds up the training time.In view of the periodicity of input signals,the convolution kernel of the first convolution layer of the model is set as a wide convolution with a large length,so that the receptive field of the output neuron to the input signal is greater than one period.The data argumentation technology greatly expands the training data,which makes the model fit faster and has better robustness.The batch normalization algorithm and average pooling algorithm are introduced,which makes the model easy to converge and the accuracy is improved.By comparing and analyzing the existing published research results,the fault diagnosis accuracy and model training time of the proposed method are improved and increased.