Abstract: Aiming at the difficulty in accurately predicting the remaining useful life of industrial robots caused by repeat positioning accuracy degradation during long-term operation, traditional time-series prediction models are limited by inadequate mining of nonlinear features. Meanwhile, deep learning prediction models commonly face bottlenecks such as high randomness in manual hyperparameter tuning and susceptibility to local optima. To solve the above problems, this paper proposes a robot accuracy life prediction method combining the Sand Cat Swarm Optimization (SCSO) algorithm and a CNN-BiLSTM-Attention hybrid architecture. First, time-series accuracy data are acquired via multi-level load accelerated degradation tests, and acceleration factor conversion as well as life label construction are implemented based on the inverse power law model. The SCSO algorithm is adopted to adaptively search for the optimal hyperparameters of the network. By integrating the feature extraction ability of 1D-CNN, bidirectional time-series modeling capability of BiLSTM and weighted focusing performance of the Attention mechanism, high-precision tracking of accuracy degradation trends is realized. Ablation experiments and independent repeated experiments verify strong synergistic effects among all submodules, and SCSO can effectively restrain initialization fluctuation and overfitting. Comparative results indicate that compared with the unoptimized baseline model, the SCSO-optimized model achieves reductions of 79.36%, 76.56% and 82.03% in MAE, MAPE and RMSE, respectively; MSE is reduced by 96.77%, and the coefficient of determination R² reaches 0.981. In contrast with PSO, GWO, SABO and random search algorithms, the proposed method possesses prominent advantages in global optimization and convergence stability. It can offer technical support for predictive maintenance and health management of industrial robots.