Data-Driven Digital Twins for Real-Time Machine Monitoring: A Case Study on a Rotating Machine

Abstract In this work, we present a framework for data-driven digital twins for real-time machine monitoring. Data-driven digital twins are gaining prominence in a variety of industrial applications owing to their ability to capture complex relationships between sensor data and system behavior. The computational efficiency gained using such twins is critical for real-time machine monitoring and diagnostics with timely and interactive human intervention. One of the fundamental challenges in the current data-driven digital twins is a lack of understanding of how different data synthesis strategies of the same sensor data affect the predictive power of the twin models typically obtained through statistical learning. As a result, the interactive support for enabling human intervention and machine health monitoring is not generalized for different machine configurations and fault conditions. Using turbomachinery as a concrete demonstrative context, we investigate two fundamentally different data synthesis strategies, namely, integrated and combinatorial, as digital twins for a rotating machine. Specifically, we consider a rotor kit as a machine component, develop a synthetic dataset using simulations, and conduct systematic studies on the predictive performance of reduced-order models trained using the different data synthesis strategies. Our experiments show that the combinatorial dataset offers higher prediction accuracy in comparison to randomized data generation. Moreover, we created a cloud-based augmented reality (AR) mobile tool to show the feasibility of our methodology in developing potential machine monitoring applications with human-in-the-loop components.