An improvement of visualized images from vibration for plastic gear early failure detection using convolutional neural network

Recently, data-driven machine health monitoring has become more popular due to the wide-spread deployment of lowcost sensors and deep learning algorithms’ achievements. The detection of failures of machines can be determined based on failure classification results using deep learning architectures. On this tendency, we constructed a plastic gear failure detection structure using a convolutional neural network. In this study, raw vibration data was converted to frequencydomain data. Amplitudes of frequencies in the monitored frequency band were transferred into images, which then were labeled as crack or non-crack by a high-speed camera. Although deep learning architectures have great potential to automatically learn from complex features of input data, the high-amplitude frequencies reflecting the main vibration causes such as gear meshing frequency and its harmonics or shaft frequency affect the accuracy of learning. Besides, the low-amplitude frequencies in a low-frequency band, which are sensitive to gear failures, show efficiency in early failure signs of the plastic gear. Thus, this paper proposed an image visualization and labeling method by focusing on lowamplitude frequency features in the low-frequency band and lessening high-amplitude frequency features. The results show that the proposed system learning from new visualized images can detect plastic gear’s early failure situation before the initial crack happened.