In the gear fault diagnostics, conventional methods have largely depended on the complicated signal processing and feature extraction skills, which are often cumbersome for engineers to implement easily. Recently, the convolutional neural networks, which is a kind of deep learning techniques, have found increased success in this field by taking advantage of minimal engagement of signal processing and automated features extraction for the fault diagnosis. In the previous studies, however, the accuracy of the method was often assessed by means of cross-validation from the data sets at the same faulted tooth, which may not be the case in the real gear, since the fault may appear at different teeth from the training. This article proposes a convolutional neural network method based on the signal segmentation to solve this problem, which is to divide the original signal into those at each tooth of the gear. The effectiveness of the method is validated by the data made from the gearbox test rig, in which the vibration and transmission errors are measured, respectively. The training and test data sets are prepared at different fault locations. The performances of the convolutional neural network with signal segmentation are compared and discussed with those by the ordinary convolutional neural network without segmentation. As a whole, the results with the signal segmentation suggest that the fault can be successfully identified even when the fault location in the test is different from that of the training, which provides great feasibility toward the real applications.
A Signal Segmentation Approach to Identify Incident/Reflected Traveling-Waves for Fault Location in Half-Bridge MMC-HVDC Grids
