A knowledge-data integration framework for rolling element bearing RUL prediction across its life cycle.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Prognostics and Health Management (PHM)

Background:

• Accurate Remaining Useful Life (RUL) prediction for Rolling Element Bearings (REB) is crucial for industrial maintenance and operational efficiency.
• Existing methods face challenges in multi-sensor data utilization, model interpretability, and full life-cycle prediction accuracy.
• Bottlenecks in current RUL prediction limit its effectiveness in real-world applications.

Purpose of the Study:

• To develop a knowledge-based, explainable framework for accurate, life-cycle RUL prediction of REBs.
• To address the limitations of current models in handling multi-sensor data and predicting across the entire operational lifespan.
• To improve the reliability and interpretability of RUL prediction models.

Main Methods:

• An Improved Graph Convolutional Autoencoder incorporating Pearson correlation and feature transformation for fast-changing signal fusion.
• A Cascaded Multi-head Self-attention Autoencoder with Characteristic Guidance for multi-source signal integration and health indicator construction.
• Life-cycle stage division using Continuous Gradient Recognition with Outlier Detection, coupled with Measurement-based Correction Life Formula and Bidirectional Recursive Gated Dual Attention Unit.

Main Results:

• The proposed framework demonstrated significant improvements in RUL prediction accuracy compared to existing models.
• For a self-designed test rig, a 1.66% enhancement in Corrected Cumulative Relative Accuracy (CCRA) and a 49.00% improvement in Coefficient of Determination (R²) were achieved.
• On PHM 2012 datasets, the framework resulted in a 4.04% increase in CCRA and a 120.72% boost in R².

Conclusions:

• The developed knowledge-based explainable framework effectively overcomes limitations in multi-sensor data utilization and life-cycle RUL prediction.
• The proposed methods for feature fusion, health indicator construction, and life-cycle segmentation lead to superior prediction performance.
• This framework offers a promising solution for accurate and interpretable RUL prediction in Rolling Element Bearings.