XMDCA-TL: An Explainable Multi-Domain Channel Attention Transfer Learning Framework for Fault Diagnosis in Industrial Gas Turbines

Transfer learning bridges the gap between lab and real-world conditions: The framework successfully transferred knowledge from controlled laboratory data to actual industrial equipment (a 42.2 MW gas turbine), addressing a critical challenge in industrial AI where labeled real-world data is scarce and expensive to obtain.

Explainability is essential for industrial adoption: The paper demonstrates that AI systems for critical infrastructure can provide "physical interpretation of the model's decision-making process" through XAI analysis, enabling engineers to understand why the system diagnosed a particular fault rather than treating it as a "black box"—crucial for safety-critical applications.

Multi-perspective data analysis improves reliability: By converting vibration signals into three complementary representations (time, frequency, and time-frequency domains) and modeling their interactions, the system achieved "robustness to noise" and better fault detection, suggesting that responsible AI systems should leverage multiple data perspectives rather than single-view approaches.

Self-supervised learning reduces dependence on labeled data: The framework used hybrid masking and reconstruction techniques to improve model robustness without requiring extensive manual labeling, addressing a practical barrier to deploying AI in industrial settings where expert annotation is costly and time-consuming.

Interpretability and performance are compatible goals: The results show that XMDCA-TL achieved "satisfactory fault diagnosis performance" while maintaining explainability, countering the common assumption that interpretable AI requires sacrificing accuracy—important for building trustworthy systems in high-stakes domains.