A deep learning-based method for early crack diagnosis in nonstandard spur gear pairs

Abstract

Gears are the key components of modern industry and have been widely employed in the automotive, wind turbine, and aviation fields. From the engineering point of view, an intelligent method that can automatically extract fault features from the vibration signals would be precious since failing in early diagnosis of root cracks may result in a tooth broken rapidly. In this regard, deep learning (DL) is increasingly popular in achieving early fault diagnostics tasks in geared systems with the wide availability of sensors and ever-increasing computation power. With this in mind, the asymmetric tooth concept offers higher load-carrying capacity, long fatigue propagation life, and the ability to lessen vibration and noise than the standard (symmetric) involute profile spur gears. From this standpoint, this study aims to determine the tooth root crack and its degree for both symmetric (20°/20°) and asymmetric (20°/30°) involute spur gears with a DL-based approach using vibration data. To this end, the single tooth stiffness values of the designed gears were obtained with ANSYS software for healthy and cracked gears (50%-100%), and then the time-varying mesh stiffness was calculated. Besides, a six-degree-of-freedom dynamic model was developed by deriving the equations of motion of a one-stage spur gear transmission. The vibration responses were collected for the healthy state, 50%, and 100% crack degrees for symmetric and asymmetric tooth profiles. Three different signal-to-noise ratios were considered to complicate the early crack diagnosis task and evaluate its influence on the DL algorithm’s classification performance. The obtained findings were then evaluated and interpreted in time and frequency domains. To this end, the Fourier transform was applied to the simulated time-sequence acceleration data in the time domain. As a supplementary finding, the present research also benefited from three statistical indicators, namely, (1) root mean square, (2) kurtosis, and (3) crest factor, to investigate whether the configuration of tooth profiles would provide an advantage in detecting tooth-root cracks. The present study also evaluated the influence of residual signals on the proposed DL-based method’s classification accuracy and further expanded the scope of research work. The findings indicated that the overall classification accuracy could be improved by 5.1% using asymmetric (20°/30°) gearing.