Explicit speed-integrated LSTM network for non-stationary gearbox vibration representation and fault detection under varying speed conditions

Abstract

Condition monitoring of the gearbox plays a crucial role in implementing proactive maintenance strategies and minimizing the economic loss of unexpected failures. Gearboxes often operate under variable speed conditions, which makes the collected vibration monitoring signals non-stationary. Existing works did not explore the scientific structures that incorporate speed signals into the long short-term memory (LSTM) networks, and thus leave room for improvement at varying speed conditions. To this end, this paper proposes novel explicit speed-integrated LSTM (SI-LSTM) models to enhance the representation accuracy of non-stationary vibration signals and improve gearbox fault detection capability. The SI-LSTM models with three variants are designed to account for the effects of speed variations on vibration signals. In SI-LSTM model 1, the vibration and speed signals are directly merged and input into the LSTM network. In SI-LSTM model 2, the speed signal is integrated into the network before the final LSTM layer. SI-LSTM model 3 is designed with a dedicated LSTM layer for speed signal, and the state outputs of both speed and vibration LSTMs are then merged and input into a final LSTM layer. Comprehensive experiments are conducted on a helical fixed axis gearbox dataset and a planetary gearbox dataset, and finally SI-LSTM model 3 is the best recommended structure. Spectral analysis is used to demonstrate the effectiveness of SI-LSTM model 3. The performance are also compared with four state-of-the-art methods, and the SI-LSTM model 3 achieves the highest AUCs of 0.9998 and 0.9676 and the best vibration representation accuracy on fixed-axis and planetary gearbox datasets, respectively.