Multimodal Twin Intelligence for Resilient Power Fleets

Abstract

Digital Twin Intelligence stands to benefit substantially from the integration of multimodal data sources in industrial processes. Ensemble models can reduce the false-negative rate of failure-detection systems, thereby improving overall resilience. This work presents a visual-fault-diagnosis pipeline that develops visual-inspection AI from multimodal data, using a model-ensemble strategy to support digital twin applications. In asset-prognostics settings, failure signals are typically rare, data are often distributed across separate administrative domains, and detection modalities tend to be imbalanced. By orchestrating multimodal data within an ensemble framework, this approach improves the training and validation of neural networks for final-event-time detection across diverse assets, ultimately strengthening the resilience of the digital twin.

A digital twin can, in principle, model system behavior within a safe, simulated environment. In practice, however, this requires extensive sensor deployment across the physical asset to capture potential failure modes in sufficient detail. In electrical power generation and distribution, most production plants—particularly renewable energy facilities—lack dedicated visual-fault-diagnosis models, despite being connected to centralized intelligence systems responsible for control, detection, diagnosis, and failure prediction. Without such capabilities, continuous power generation and distribution cannot be guaranteed. Fleets at scale stand to benefit from inspection systems that enable easier maintenance, reduced downtime, and uninterrupted service through standardized operations. While prior research has introduced inspection models for various fault types in electrical components, these models typically address only a narrow set of safety-critical components.