方位(导航)
计算机科学
信号处理
生物医学工程
工程类
控制理论(社会学)
电子工程
人工神经网络
模式识别(心理学)
可靠性工程
作者
Guoxu Wang,Siyu Liu,Weidong Jiao,Zhilin Dong,Wanxiu Xu,Jianfeng Sun,Xiaohao Chen,Yonghua Jiang
标识
DOI:10.1109/tim.2026.3676093
摘要
The prediction of the remaining useful life (RUL) of rolling bearings plays a crucial role in preventing major failures and enabling predictive maintenance. To address bearing fault prediction, a novel hybrid method is proposed. The developed methodology integrates a two-stage prognostic framework. Initially, complete ensemble empirical mode decomposition with adaptive noise adaptively decomposes vibration signals to extract multi-domain features from IMFs, which are transformed into a robust health indicator using a kernel principal component analysis-based approach for nonlinear dimensionality reduction, whereby principal components accounting for 85% of the cumulative variance are retained, with the first principal component selected for precise degradation onset detection using the 3δ rule. Subsequently, the preprocessed sequences are fed into a dual-branch network where BiLSTM captures bidirectional temporal dependencies while Transformer encoder models global contextual patterns through self-attention for RUL prediction. The proposed model is ultimately validated using two publicly available rolling bearing datasets. The analysis results show that our method achieves relatively high prediction accuracy. Compared with other methods, it reduces MAE and RMSE by approximately 10%-30% . Under consistent testing conditions, it outperforms comparative methods, confirming its robust generalization capability across different operational environments and high precision in RUL prediction.
科研通智能强力驱动
Strongly Powered by AbleSci AI