A multimodal explainable AI framework for industrial turbine vibration health monitoring and regulatory decision support in finance
Abstract
Industrial rotating machinery plays a pivotal role in global energy infrastructure, yet conventional vibration monitoring systems often operate as black boxes, providing limited interpretability and failing to leverage the rich multi-sensor data available in modern plants. This paper introduces a novel framework that integrates multimodal sensor fusion—combining accelerometer, acoustic, thermal, and operational data—with explainable artificial intelligence (XAI) and multi-criteria decision analysis. The core engine is a domain-collaborative multimodal transformer that jointly processes heterogeneous time-series and image-based streams, producing fault classifications alongside SHapley Additive exPlanations (SHAP)-based feature attributions and natural-language diagnostic narratives. The framework is validated on a 250 MW combined-cycle gas turbine power plant with 24 months of operational data. Experimental results demonstrate a fault detection accuracy of 94.2%, a 14.5% improvement over vibration-only baselines, while achieving the highest interpretability score (5/5) among compared methods. Decision Making Trial and Evaluation Laboratory (DEMATEL) causal analysis identifies diagnostic transparency and system reliability as primary drivers of regulatory compliance. The primary contribution is an open-source, scalable blueprint for trustworthy AI in industrial vibration monitoring, addressing the urgent need for transparent, auditable, and human-centered decision support in critical energy assets. The proposed framework achieves a balanced integration of three critical dimensions: diagnostic accuracy and interpretability, technical performance and regulatory compliance, and automated inference and human oversight. Based on these findings, we recommend that industrial operators for finance risk optimization: (1) deploy multimodal sensor arrays combining vibration, acoustic, thermal, and operational sensors; (2) implement explainable AI protocols utilizing SHAP-based feature attribution; and (3) adopt DEMATEL-derived priorities for risk-informed maintenance scheduling.