Optimizing sample length for fault diagnosis of clutch systems using deep learning and vibration analysis

Abstract

Clutches are prone to failure owing to extended heat exposure and high levels of abrasion during power transfer. Internal damage, downtime, and permanent transmission system lock-up all can result from these faults. To detect and diagnose these faults, this study employs the deep learning approach. Vibration signals were obtained from a test rig that was exposed to various clutch conditions at various loads. The amount of data points (signal length) when collecting vibration signals from a test rig can have a significant effect on the accuracy of results. A shorter sample length can lead to an increased uncertainty in the results, while a longer sample length can lead to more accurate results. A longer sample length also increases the computational complexity of the diagnosis process, which can lead to longer execution times. In this study vibration signals were collected for various sample lengths to find the optimal sample length for systemic clutch fault diagnostics. The collected vibration signals are analyzed and transformed into vibration plots that serve as input to the deep learning pretrained network. VGG-16 model was considered for this study to diagnose the clutch system faults. Based on the outcomes, the optimal sample length for the no load condition was identified as 4000, while for the 5-kg load and 10-kg load conditions 3000 sample length was suggested for fault diagnosis of the clutch system.