Remaining Useful Life Prediction of Rolling Bearings Based on SSA-Optimized TCN-CBAM-BiLSTM Model

Abstract

In the field of remaining useful life (RUL) prediction for rolling bearings, existing deep learning models suffer from issues such as insufficient capability to capture long-term temporal dependencies, poor focus on key degradation features, reliance on manual experience for hyperparameter tuning, and weak generalization ability under complex operating conditions. To address these challenges, this paper proposes a RUL prediction method for rolling bearings based on a Sparrow Search Algorithm (SSA)-optimized Temporal Convolutional Network-Convolutional Block Attention Module-Bidirectional Long Short-Term Memory Network (TCN-CBAM-BiLSTM). Firstly, a basic TCN-CBAM-BiLSTM prediction model is constructed. The dilated convolutions and residual connections in TCN are employed to capture long-span degradation correlations throughout the bearing's lifecycle. The CBAM's dual attention mechanism dynamically suppresses noise and enhances key degradation features, while BiLSTM's bidirectional temporal modeling capability fully extracts contextual information from sequences. Building on this foundation, SSA is utilized to adaptively optimize core hyperparameters such as the model's learning rate and the number of TCN convolutional kernels, resulting in the development of a high-precision SSA-TCN-CBAM-BiLSTM prediction model that resolves the subjectivity and inefficiency associated with manual parameter tuning. Multiple comparative experiments are conducted using the XJTU-SY rolling bearing full-lifecycle degradation dataset. The results demonstrate that the proposed model significantly reduces both the root mean square error and mean absolute error, two key metrics. Moreover, its predicted curves closely align with the actual RUL curves under multiple operating conditions, showcasing superior prediction accuracy and cross-condition generalization capability.