A degradation representation learning-based LSTM framework for gear remaining useful life prediction

Abstract

Predicting the useful lifespan of gears is essential for enhancing predictive maintenance reliability and operational efficiency. A novel MDRL-BiLSTM framework is proposed to improve the modeling of gear degradation processes by jointly learning nonlinear degradation characteristics and long-term temporal dependencies from vibration signals. The framework first constructs multimodal degradation features from raw vibration signals, then employs a dual parallel one-dimensional convolutional branch for multiscale feature extraction and fusion, followed by temporal modeling through a BiLSTM network to generate RUL predictions. Experimental results show that the coefficient of determination R2 of MDRL-BiLSTM is 0.9088, significantly outperforming the baseline model, validating its effectiveness in industrial gear RUL prediction.