FDBRP: A Data-Model Co-Optimization Framework Towards Higher-Accuracy Bearing RUL Prediction.

Area of Science:

• Engineering
• Machine Learning
• Artificial Intelligence

Background:

• Predicting Remaining Useful Life (RUL) is crucial for effective maintenance of rolling bearings.
• Existing methods often struggle with complex spatiotemporal dependencies and data limitations.

Purpose of the Study:

• To propose an integrated framework, FDBRP, for accurate RUL prediction of rolling bearings.
• To enhance signal representation, spatiotemporal feature consistency, and long-range dependency capture in RUL prediction models.

Main Methods:

• A data augmentation module with sliding-window processing and 2D feature concatenation.
• A feature fusion module combining graph convolutional networks, multi-scale temporal convolution, and attention mechanisms.
• An optimized dilated convolution module with interval sampling and residual connections.

Main Results:

• The proposed FDBRP model achieved a high average score of 0.756564.
• Demonstrated a lower average error of 10.903656 in RUL prediction for test bearings.
• Outperformed state-of-the-art benchmark methods in RUL prediction accuracy.

Conclusions:

• The FDBRP model offers superior RUL prediction capabilities for rolling bearings.
• The integrated approach effectively addresses challenges in spatiotemporal modeling and long-range dependency capture.
• The methodology shows significant potential for industrial applications in predictive maintenance.