Concurrent bearing faults diagnosis with CNN-based object detection—An evaluation

With the development of deep learning, convolution neural network (CNN)-based classifiers have already been demonstrated as a powerful method to detect bearing fault in wavelet-transformed time-frequency images. However, the real-world motor operations that suffer from concurrently occurred bearing faults have to define the fault labels that take all possible bearing fault combinations into account. Such a large number of labels significantly increase the labor and time costs in dataset labeling and network training. In this study, the CNN-based object detection is suggested to perform concurrent bearing faults diagnosis instead of conventional classifiers. Since object detection is capable of identifying multiple individual bearing faults independently in a single time-frequency image, it only requires to define the individual bearing fault labels rather than all possible bearing fault combinations. It dramatically reduces the workload in dataset preparation. In addition, to address the issue of relatively high noise level in weak abnormal vibrations, a moving average scheme is applied on the time-domain vibration signal in advance of time-frequency image generation. With the assistance of the moving average, the noise level is effectively suppressed, and x times higher detection accuracy is achieved in concurrent bearing fault diagnosis.