Remaining Useful Life Prediction of Rolling Bearings Using mLSTM (Matrix LSTM)

Accurately modeling long-term dependencies and facilitating cross-timestep feature interaction are two major challenges in predicting the Remaining Useful Life (RUL) of rolling bearings from vibration data. To address these shortcomings, we propose a novel RUL prediction method based on the mLSTM architecture. The method initially extracts features from raw signals at multiple scales, encompassing time-domain, frequency-domain, and complexity metrics. This process constructs a degradation-sensitive feature set. A Health Indicator (HI) is established to quantify the bearing's performance degradation trajectory throughout its lifecycle. The proposed method introduces an mLSTM network architecture. In mLSTM, the hidden state vector of traditional LSTM is extended to a matrix form. Cross-timestep feature interaction is then achieved using multi-head attention. To verify the effectiveness of our approach, experiments were performed using the PHM2012 bearing dataset. Compared to the best baseline model, GRU, the proposed method reduced MSE, RMSE, and MAE by 35.997%, 19.158%, and 25.506%, respectively. This outcome confirmed the superiority of the proposed method.