Online Black-Box Modeling for the IoT Digital Twins Through Machine Learning

Abstract

Many applications involving physical systems, such as system control or fault detection, call for a behavioral, black-box, or digital twin of the real system. By observing input-output pairs, a nonlinear system's black-box twinning model can be built, thus enabling real-time accurate estimation of the system's health and status. We propose a modeling approach that can be implemented with little hardware resources and predicts system output with acceptable accuracy for a wide range of applications in the IoT and Industry 4.0 application domains, such as cloud and distributed predictive control, maintenance, fault detection, and model drift avoidance. This approach consists of building a compact numerical model, based on the concept of sum-decomposability, with reduced computational complexity and memory requirements, well suited for microcontroller-based IoT applications. The black-box modeling theory, the sizing process, and the learning method are reported. The outputs of two examples of non-linear systems are replicated in real-time using a pioneer experimental setup built around a microcontroller. According to experimental results, online learning and prediction are performed at 1 kS/s with a prediction error comparable to the resolution of the digitalized input-output data. The reduced size of the obtained model calls for real-time sharing and update with cloud and edge-based simulation ecosystems enabling a near real-time digital twinning of field systems.