摘要详情

ID / 提交时间

92 / 2025-05-14 22:19:02

标题

Bayesian PPO-Enhanced Feature Optimization for Transformer-Based Bearing Compound Fault Diagnosis Under Variable Speed Conditions

关键字

Bearing fault diagnosis, variable speed condition, Vold-Kalman filter, Bayesian Proximal Policy Optimization, Transformer

主题及专题

4、System Monitoring, Fault Diagnosis and Prediction

状态

全文待审

作者

Liren Pan / Zhenjiang Technician College Jiangsu Province

Jialin Li / Zhenjiang Technician College Jiangsu Province

Leng Xue / Zhenjiang Technician College Jiangsu Province

Tong Xu / Zhenjiang Technician College Jiangsu Province

Yongwei Su / Zhenjiang Technician College Jiangsu Province

Ying Zhong / Zhenjiang Technician College Jiangsu Province

摘要

Bearing fault diagnosis under variable speed condition remains a persistent challenge due to the non-stationary, complex nature of vibration signals. Bearing fault diagnosis under variable speed conditions poses a formidable challenge due to the non-stationary and intricate nature of vibration signals. This study introduces a sophisticated feature engineering pipeline designed specifically for Transformer-based diagnostic models. The proposed methodology leverages Vold-Kalman Filtering to enable adaptive signal decomposition and accurate speed profile inference. Subsequently, an enhanced Fast Independent Component Analysis, integrated with wavelet denoising and stability-driven feature selection, effectively isolates independent signal sources. To detect subtle anomalies across multiple temporal scales, Multi-Scale Exponentially Weighted Moving Average control charts with adaptive smoothing and dynamic thresholding are employed. A Bayesian Proximal Policy Optimization algorithm facilitates intelligent feature selection, adeptly managing uncertainty and ensuring feature robustness. Furthermore, a dynamic feature combination strategy generates higher-order features, enhancing diagnostic precision. The resulting normalized feature matrix, comprising optimized base features, interaction terms, and positional encodings, is meticulously tailored for Transformer input. Validated across 11 distinct bearing health states, this framework achieves a classification accuracy of 97.3%, markedly outperforming conventional approaches and providing a robust solution for fault diagnosis in non-stationary environments.