| Induction motors are widely used in modern transportation systems,industrial production,and daily life of residents because of their huge advantages of simple structure,reliable operation,and low price.The bearing is the most important part of the induction motor;therefore,in order to ensure the safety of the mechanical equipment,a timely,accurate and effective method for motor bearing fault diagnosis is urgently needed.This paper firstly adopts the method of fault feature extraction based on Teager-Kaiser energy operator to identify the weak feature signal under strong noise when the motor bearing fails.It is theoretically analyzed that the application of the Teager-Kaiser energy operator to the stray magnetic field can not only demodulate the characteristic frequency of the fault,but also strengthen the amplitude of the weaker characteristic frequency of the fault,thereby improving the ability of fault detection.On this basis,This article has used two convolutional neural network models.The first one is the classic one-dimensional convolutional neural network model.The model uses the time domain and frequency domain characteristics of magnetic field signals to identify faults.In order to enhance the robustness and generalization.ability of the model,the data set strengthening technology is used to strengthen the collected magnetic field data.The time and frequency domain features extracted by the Teager-Kaiser energy operator are input into the model for training and fault identification.The results prove that the fault identification rate of the applied frequency domain is about 2%higher than the time domain,reaching 85%.Through further analysis of the characteristics of the.stray magnetic field signal,after many adjustments and experiments,a second convolutional neural network model is proposed:a one dimensional wide convolution kernel deep convolution neural network model.The characteristic of this model is that the first layer of large convolution kernels is connected with multiple layers of small convolution kernels.Accumulating cores.can speed up the training process better.Deepening the number of network layers not only reduces the model training parameters,but also expands the receptive field of each layer of the network.Use batch normalization algorithm to improve network model training efficiency and enhance model generalization ability.Compared with the classic CNN,the one-dimensional wide convolution kernel deep convolutional neural network model uses the frequency domain features extracted by the Teager-Kaiser energy operator,and the fault classification accuracy is improved to 97%.The excellent adaptive feature learning features of convolutional neural networks are used to realize end-to-end fault detection of frequency domain signals.Finally,the experiment proves that the one-dimensional wide convolutional deep neural network can still complete the advantage of bearing fault diagnosis under the background of strong noise.Through comparison experiments with various methods,it is verified that the neural network model proposed in this paper shows superior performance in terms of training efficiency and recognition accuracy. |
