Machinery Prognostics and High-Dimensional Data Feature Extraction Based on a Transformer Self-Attention Transfer...

Area of Science:

• Engineering
• Artificial Intelligence
• Data Science

Background:

• Machinery degradation assessment is crucial for prognosis and health management.
• Current AI models face challenges in manual feature extraction, long sequence prediction, and cross-operational RUL prediction with high-dimensional data.

Purpose of the Study:

• To propose a novel health indicator construction methodology using a Transformer Self-Attention Transfer Network (TSTN).
• To enable direct processing of high-dimensional raw datasets for accurate Remaining Useful Life (RUL) prediction.

Main Methods:

• Designed an encoder with long-term and short-term self-attention mechanisms to capture time-varying information.
• Developed an estimator to map encoder outputs to degradation trends.
• Implemented a domain discriminator for extracting invariant features across operating conditions.
• Utilized the FEMTO-ST bearing dataset and Monte Carlo methods for RUL prediction.

Main Results:

• The TSTN method directly processes high-dimensional data, preserving all information.
• Achieved superior RUL prediction accuracy compared to RNN, LSTM, and traditional methods.
• Demonstrated a notable SCORE of 0.4017 in RUL prediction accuracy.

Conclusions:

• The proposed TSTN methodology offers significant advantages for machinery health prognostics.
• TSTN shows strong potential for state-of-the-art Remaining Useful Life prediction in complex environments.