Vision-Fused Twins for Resilient Power Assets

Abstract

Industrial processes stand to benefit substantially from Digital Twin Intelligence, particularly through the integration of multimodal datasets. Ensemble models can reduce a failure-detection model's false-negative rate, thereby improving overall system resilience. The visual-fault-diagnosis pipeline presented here introduces methods for developing visual-inspection AI from multimodal data sources, with an underlying model-ensemble strategy that supports digital twin architectures. In asset-prognostics pipelines, failure signals are typically rare, data are often collected across distinct administrative domains, and detection modalities are frequently imbalanced. By orchestrating multimodal data, ensemble models improve the training and validation of neural networks for final-event-time detection across diverse assets — enhancing the resilience of the digital twin as a whole.

In principle, a digital twin can serve as a model of system behavior within a safe, simulated environment. In practice, however, this requires extensive sensor deployment across the physical counterpart to capture potential failure events in sufficient detail. In electrical power generation and distribution, most production plants — particularly renewable ones — lack visual-fault-diagnosis models, despite being connected to centralized intelligence systems responsible for control, detection, diagnosis, and failure prediction. Without such capabilities, power generation and distribution cannot proceed reliably or without interruption. Fleets adopting these inspection systems stand to gain simplified maintenance, reduced downtime, and continuous, standardized service. While prior research has introduced inspection models for various fault types in electrical components, these models have generally been limited to a narrow set of safety-critical components.