Trans-D3: A Novel Hybrid Transformer-Based Actor-Critic Approach for Remaining Useful Life Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Predictive Maintenance

Background:

• Accurate Remaining Useful Life (RUL) prediction is crucial for industrial system reliability and safety.
• Existing methods face challenges in handling complex, variable operational data.
• Industry 4.0 necessitates advanced prognostics for optimized maintenance strategies.

Purpose of the Study:

• Introduce TRANS-D3, a novel hybrid model for enhanced RUL prediction.
• Improve prediction accuracy and reliability in diverse operational contexts.
• Establish a robust optimization paradigm for industrial prognostics.

Main Methods:

• Hybrid approach combining Twin Delayed Deep Deterministic Policy Gradient (TD3) and Transformer architecture.
• Optimized reward function using Linear Quadratic Regulator (LQR) for dynamic error correction.
• Validation on the CMAPSS dataset across various operational scenarios (FD001, FD003, FD004).

Main Results:

• Significant Root Mean Square Error (RMSE) reductions: 84-90% (FD001) and 23-45% (FD003/FD004).
• High statistical reliability confirmed by R2 values > 0.93 (max 0.9984) and unbiased estimation.
• Penalty reductions of 80-95% compared to DAST and STAR, ensuring stable predictions.

Conclusions:

• TRANS-D3 offers a robust and accurate RUL prediction framework.
• The method demonstrates superior performance in both baseline and highly variable conditions.
• This advancement supports safer and more reliable operations in Industry 4.0 environments.