Explainable AI-Driven Quality and Condition Monitoring in Smart Manufacturing

Area of Science:

• Industrial AI
• Explainable Artificial Intelligence (XAI)
• Manufacturing Technology

Background:

• Artificial intelligence (AI) adoption in manufacturing faces trust and regulatory barriers due to opaque "black-box" models.
• Explainable AI (XAI) is crucial for understanding AI decision-making in industrial settings.

Purpose of the Study:

• To investigate XAI techniques for interpreting AI systems in manufacturing, focusing on explainability rather than performance optimization.
• To propose and apply a unified, explainability-centered framework across diverse manufacturing use cases.

Main Methods:

• Applied established XAI techniques (Grad-CAM, Integrated Gradients, SHAP, etc.) to heterogeneous data modalities (vision, acoustics).
• Evaluated XAI consistency across different AI models, tasks, and supervision strategies.
• Utilized three manufacturing use cases: casting defect classification, metal surface defect localization, and acoustic anomaly detection.

Main Results:

• XAI techniques successfully exposed model attention, feature relevance, and decision drivers.
• Demonstrated alignment between AI model behavior and physically interpretable defect/fault mechanisms.
• XAI facilitated transparent, auditable, and human-interpretable AI decision-making.

Conclusions:

• XAI is essential for trustworthy AI in manufacturing, enabling human-in-the-loop oversight and accountability.
• The proposed cross-modal framework supports Industry 5.0 principles for transparent industrial AI.
• Positioning explainability as a core requirement enhances AI acceptance and regulatory compliance.