Bearing Fault Diagnosis And Prediction Under Multiple Operating Conditions Based On Deep Learning

As an indispensable part of rotating machinery,rolling bearings usually work in the harsh environment of high speed,high temperature and heavy load.Rolling bearing faults and performance degradation often account for the majority of the whole machine failure,and even lead to the entire equipment stop.Operation,resulting in huge economic losses,bearing fault identification and life prediction directly determine the reliability of mechanical equipment operation.Therefore,it is very important to monitor the condition of the bearing as well as fault diagnosis and prediction.The main contents of this paper include the following aspects:The fault characteristics of rolling bearings are usually affected by the working environment.The characteristic changes caused by faults are difficult to separate from the characteristic changes caused by working conditions.For this reason,this paper proposes Supervised Auto-encoder(SAE).The working condition conversion method,the method maps the eigenvalues under different working conditions to the eigenvalue sequence under the reference condition in a supervised manner,thereby eliminating the interference of the fault eigenvalue change caused by the change of the working condition,and the experimental result shows that the method It can better complete the conversion of feature sequences under different working conditions and solve the distortion problem of fault characteristics caused by the change of working conditions.Aiming at the problem that the traditional convolutional neural network has too large dimension of the fully connected layer,resulting in large computational burden,over-fitting and under-utilization of shallow information,this paper proposes a fusion of shallow information to simplify the convolutional neural network(Integrating Shallow Information Simplified)The Convolutional Neural Networks(ISIS-CNN)model extracts the feature vectors obtained by different convolution kernels into a feature value by a full convolution operation and uses it as a new fully connected layer to reduce the data dimension and model parameters.The information extracted from the shallow convolution kernel is incorporated into the new fully connected layer to increase the amount of information input by the classifier.The experimental results show that under the premise of ensuring the accuracy of existing fault diagnosis,ISIS-CNN effectively compresses the training time and improves the computational efficiency of CNN.In the aspect of bearing fault trend prediction,this paper uses the Long Short Term Memory(LSTM)for processing time series in deep learning to predict the root mean square eigenvalues of bearing degradation trends for different time lengths.The purpose of early warning,the experimental results show that for short-and medium-term predictions,the effect is ideal,and the predicted failure time point is not much different from the actual failure time.For long-term prediction,due to the long time step,the error continues to accumulate,and the local prediction effect is not ideal,but the overall trend effect is consistent with the real curve.