On the Soundness of XAI in Prognostics and Health Management (PHM)

Deep learning models for predicting equipment failure lack transparency: The paper highlights that while Deep Convolutional Neural Networks (DCNN) and Long Short-Term Memory (LSTM) networks are effective at predicting Remaining Useful Life (RUL)—how long equipment will function before failing—these "black box" models are difficult to explain to industrial experts and stakeholders who need to understand why the system made a particular prediction.

Explainability methods for time series predictions need more research: The authors note that explainability (XAI) methods have been "less studied" for time series regression tasks compared to classification tasks, representing a significant gap in responsible AI practices for predictive maintenance applications where understanding predictions is critical for decision-making.

Not all explanation methods are equally reliable: Through comparative testing using multiple quality metrics, the paper found that different XAI methods vary substantially in their robustness and reliability, with GRAD-CAM emerging as the most robust method tested—suggesting that organizations cannot assume all explanation techniques will provide equally trustworthy insights.

Responsible AI requires systematic evaluation of explanation methods: The research demonstrates the importance of rigorously testing and comparing XAI methods using defined metrics rather than deploying them without validation, ensuring that explanations provided to stakeholders actually reflect sound reasoning rather than potentially misleading interpretations.