A CNN-LSTM-BASED APPROACH FOR FAULT DIAGNOSIS IN ASYMMETRIC GEARS UNDER RANDOM SPEED VARIATION

Gearboxes are oftentimes utilized in diverse strategic fields, including wind energy, military services, rail-road vehicles, and aero engines. A gearbox condition monitoring system is expected to detect localized faults by utilizing vibrational signals collected at divergent speed profiles. Still, only a few studies have focused on random speed variations in the field, and most are limited by the assumption that the pinion speed does not alter significantly within a rotational cycle. Besides, these works all focus on gear pairs with symmetric teeth. The present study, on the other hand, combined a 1-D convolutional neural network (CNN) with a long short-term memory (LSTM) network to classify tooth crack faults under random speed variations for spur gears with symmetrical (20 degrees/20 degrees) and asymmetric (20 degrees/30 degrees) teeth. The randomness was achieved by segmenting randomly selected portions from a time-varying speed profile while creating the testing dataset. Further, the hyperparameters of the 1-D CNN-LSTM were optimized with the help of grid search. The performance of the proposed model was tested by comparing its results with those of a standalone 1D CNN. The findings obtained showed that the average accuracy could be enhanced by up to 4.400% by using spur gear pairs with asymmetric teeth instead of one with symmetric teeth.