Failure Mode Classification for Rolling Element Bearings Using Time-Domain Transformer-Based Encoder.

Area of Science:

• Machine Learning
• Signal Processing
• Mechanical Engineering

Background:

• Transformer models often require data transformation or significant computational resources.
• Effective analysis of vibration signals is crucial for machinery health monitoring.
• Existing methods may lack interpretability and trustworthiness.

Purpose of the Study:

• To develop a Transformer-based encoder for learning sparse representations of vibration signals.
• To integrate unsupervised denoising for direct time-domain analysis.
• To enhance failure mode classification with an interpretable and trustworthy model.

Main Methods:

• A novel Transformer encoder architecture with integrated unsupervised denoising.
• Direct time-domain processing of vibration signals, avoiding data transformation.
• Training and validation on IMS and CWRU benchmark datasets.

Main Results:

• The proposed model learns meaningful and sparse representations of vibration signals.
• Competitive performance achieved in failure mode classification, particularly on unbalanced datasets.
• Demonstrated effectiveness with a lightweight architecture.

Conclusions:

• The Transformer-based encoder with unsupervised denoising offers an efficient and effective approach for vibration signal analysis.
• The model provides interpretable and trustworthy results for failure mode classification.
• This method reduces the need for data pre-processing and extensive computational power.