Remaining Useful Life Prediction of Rolling Bearings Based on Cumulative Feature Representation and Adaptive Degradation Identification

Accurate rolling-bearing remaining useful life (RUL) prediction is important for reliable equipment operation. This study proposes an RUL prediction framework integrating cumulative damage representation, adaptive degradation identification, and intelligent optimization. Time-domain features are cumulatively summed along the lifetime axis, and a multi-scale temporal convolutional autoencoder (MS-TCN-AE) is used to construct a health indicator (HI). The adaptive threshold maximum rate of change (ATMROC) algorithm then estimates the initial degradation time (IDT) and extracts the degradation stage. Finally, BiGRU-MHSA predicts RUL, while Student Psychology-Based Optimization (SPBO) optimizes key hyperparameters using validation error. Experiments on the PHM2012 and XJTU-SY datasets verify the effectiveness of the method. Independent retraining on the NASA IMS dataset further indicates adaptability to different data sources and fault types.