Bearing Anomaly Detection Method Based on Multimodal Fusion and Self-Adversarial Learning

Area of Science:

• Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Bearing anomaly detection faces challenges from imbalanced data and complex conditions, leading to model overfitting and high false positive rates.
• Existing data-driven deep learning models struggle with the noisy, real-world operational environments of intelligent railways.

Purpose of the Study:

• To develop a robust deep learning strategy for bearing anomaly detection in high-noise railway environments.
• To address data scarcity and class imbalance issues in bearing anomaly detection.

Main Methods:

• Converted 1D vibration time-series data into Gramian Angular Difference Field (GADF) images for multimodal feature fusion with original time-series data.
• Implemented a composite data augmentation strategy (time-domain and image-domain) to expand anomaly samples.
• Introduced Self-Adversarial Training (SAT) within the fused feature space to generate adversarial samples, promoting generalized and robust feature learning.

Main Results:

• The proposed method significantly outperformed traditional baseline models in accuracy, precision, recall, and F1-score for bearing anomaly detection.
• Demonstrated exceptional robustness against rail-specific interferences and noise.
• Effectively mitigated issues of data scarcity and class imbalance through advanced augmentation and training techniques.

Conclusions:

• The multimodal fusion and SAT strategy provides a specialized and effective solution for bearing anomaly detection in challenging intelligent railway maintenance scenarios.
• The approach enhances model generalization and robustness, crucial for reliable performance in realistic, noisy operational conditions.