Classification And Recognition Of Transmission Fault Based On Convolutional Neural Network

Time-frequency image of the vibration is a joint distribution information in time domain and frequency domain, directly reflecting the frequency changing over time. Time frequency image contains plenties of information to depict the machine running state, which help us to diagnose the machine state effectively. Time-frequency image of the vibration is a joint distribution in time domain and frequency domain, directly reflecting the frequency changing over time, which also contains plenties of information to depict the machine running state, which help us to diagnose the machine state effectively.As one of the most promising deep learning algorithms, Convolutional neural network(CNN) has good recognition performance on image processing. Therefore, CNN is used for time-frequency image classification in order to diagnose the gearbox faults. The vibration signals of the gearbox under different conditions are collected, and then all kinds of signals are transformed into time-frequency images by using time-frequency transform. Finally, these time-frequency matrices were input to the CNN to classify different types of faults.The choice of time frequency methods directly affects the performance of time-frequency analysis. Therefore, vibration signal models corresponding to the real faults in experiments are established, including the gear impact fault, the bearing impact fault and the hybrid fault. In order to select a suitable time-frequency analysis method, three methods, such as the Continuous Wavelet Transform(CWT), S-transform and the Short Time Fourier transform(STFT) were compared. Experiment results demonstrate that the Morlet wavelet has the best time frequency characteristics, and the S-transform is superior to STFT.The structure parameters of the CNN directly affects its ability of classification. To explore the optimal CNN structure for fault diagnosis, we investigated how the structural parameters, such as convolutional kernel’s size, number of kernels, batch size and number of iterations, influenced the recognition results. At last, the CNN was constructed for gearbox fault diagnosis, which ensured the network had not only the good performance in fault classification but also a high efficiency.When the rotation speed of the gearbox varying in time, the time-frequency image at a certain rotating speed differs greatly from those under other running speeds, which also leads to difficulties in fault classification. The proposed CNN structure is successfully applied in gearbox fault classification under varying speed conditions. The results validated the advantage of the CNN, which can keep consistency in classification in terms of the image shifting, scaling and distortion to some degree, which also reveals that the CNN can be effectively used for vibration signal processing and intelligent fault diagnosis.