A Transformer-Based Variational Autoencoder for Training Data Generation in Spindle Motor Vibration-Based Anomaly...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• High-speed spindle motors (above 10,000 rpm) require vibration analysis for mechanical anomaly detection.
• Deep learning models for anomaly detection suffer from limited data, especially for rare fault conditions.
• Sample scarcity (limited real labeled vibration sequences) hinders model performance in spindle motor health monitoring.

Purpose of the Study:

• To propose a Transformer-based Variational Autoencoder (T-VAE) for generating realistic triaxial acceleration sequences.
• To overcome data scarcity and imbalance issues in spindle motor vibration datasets.
• To enhance the accuracy of anomaly detection models for high-speed spindle motors.

Main Methods:

• Developed a Transformer-based Variational Autoencoder (T-VAE) integrating positional encoding and multi-head self-attention.
• Employed a KL annealing strategy to stabilize T-VAE training for multivariate time-series data.
• Generated 100,000 synthetic samples per class (normal and faulty) to augment a CNN-LSTM classifier.

Main Results:

• Augmentation with T-VAE-generated data improved classifier performance from 95.73% (normal) / 81.40% (faulty) to 98.07% (normal) / 97.99% (faulty).
• The T-VAE effectively addressed data scarcity, demonstrating significant accuracy improvements on cross-spindle validation.
• The proposed method showed robust performance on an independent dataset of 50,000 sequences from eleven different spindle motors.

Conclusions:

• The T-VAE is an effective solution for alleviating data scarcity in high-speed spindle motor vibration analysis.
• The generated synthetic data significantly enhances the accuracy of downstream anomaly detection models.
• This approach is directly applicable to real-world predictive maintenance systems in manufacturing.