From Black-Box to Physics-Aware Cognitive Agents: A Robust XAI Framework for Predictive Maintenance

Predictive Maintenance (PdM) in the aerospace industry increasingly employs Deep Learning (DL) architectures to estimate the Remaining Useful Life (RUL) of critical components, such as turbofan engines. However, the opaque nature of models such as Long Short-Term Memory (LSTM) networks limits operational trustworthiness and may compromise human situational awareness in safety-critical contexts. This study introduces a novel, explainable, and resilient autonomous-agent framework that advances PdM from purely statistical predictions to physics-consistent diagnostics. The proposed SHAP-Informed Dynamic Loss (SIDL) framework integrates Explainable AI (XAI) directly into the neural optimization process. By validating degradation drivers through a dual-method XAI consensus (Gradient SHAP and Bootstrap Kernel SHAP), the SIDL framework compels the network to prioritize verified thermodynamic parameters, such as high-pressure turbine temperatures and pressures (specifically sensors 𝑠4 and 𝑠13), over stochastic noise. When applied to the NASA C-MAPSS FD001 dataset, this approach achieves a fundamental "reliability shift": while maintaining a competitive Mean Absolute Error (MAE) of 12.15 cycles on unseen data, the model’s physical alignment increases from 15.00% (baseline) to 99.75%. Furthermore, a Large Language Model (LLM) is incorporated as a cognitive agent to dynamically translate high-dimensional XAI outputs into actionable, human-centered maintenance reports. This integration enhances decision reliability and ethical alignment, demonstrating that AI can reduce human cognitive load without compromising methodological transparency.