Does Your Model Think Like an Engineer? Explainable AI for Bearing Fault Detection with Deep Learning

XAI can validate whether AI models reason like human experts: The paper demonstrates that explainable AI techniques can be used to check if deep learning models make decisions based on the same logic that domain experts (engineers) use, rather than relying on spurious patterns. This alignment with expert reasoning is crucial for trustworthiness in real-world industrial applications.

Standard XAI methods alone are insufficient for domain-specific problems: Simply applying off-the-shelf feature attribution techniques often fails to provide meaningful insights to specialists. The authors show that combining XAI with domain-specific knowledge (in this case, signal processing tools) is necessary to generate truly informative explanations that experts can understand and validate.

Model explainability can predict generalization performance: By analyzing whether a model's decision-making logic matches engineering principles for bearing fault detection, the researchers were able to anticipate which models would generalize well to new data—suggesting that trustworthiness and performance are connected.

Explainability is essential for deploying AI in safety-critical industries: The paper highlights that the "opaque nature" of high-performing deep learning models is a major barrier to real-world deployment in industrial settings, where understanding why a system makes a decision (especially for fault detection) is as important as accuracy.

Domain expertise should guide XAI framework design: The proposed methodology acts as a template showing that effective explainability requires tailoring techniques to specific problems by incorporating the knowledge and reasoning patterns of domain experts, rather than applying generic solutions.