关键词: 轴承, 健康状态退化趋势, 卷积神经网络, 双向门控循环单元, 注意力机制

Abstract: To better capture the subtle evolution of bearing health degradation, this study introduces a hybrid deeplearning architecture that synergizes convolutional feature extraction, bidirectional gated-recurrent dynamics, and an attention-guided refinement stage. The convolutional neural network ( CNN) is employed to extract local spatiotemporal features from vibration signals, bidirectional gated recurrent unit ( BiGRU ) captures bidirectional long-term temporal dependencies,and attention dynamically highlights critical degradation stages,enabling high-precision trend prediction of the bearing health indicator. Experiments are conducted on the XJTU-SY bearing full-life dataset,using horizontal acceleration
signals together with 13 time-frequency domain features as inputs. The CNN-BiGRU-Attention model achieves a mean absolute error ( EMA) of 0. 015, a root mean square error ( ERMS) of 0. 020, and a coefficient of determination ( R2) of 0. 983,outperforming BiGRU, BiGRU-Attention, and transformer-gated recurrent unit baselines significantly. The proposed approach offers a reliable health assessment tool for predictive maintenance of rotating machinery.