Towards learning digital twin: case study on an anisotropic non-ideal rotor system

TL;DR

The paper addresses maintaining accurate digital twins under lifelong degradation by introducing a learning digital twin (LDT) that continuously updates the digital model through memory, a detector, and an adaptor. It formalizes concept drift for DTs, distinguishing real drift via $p(y|X)$, and presents two modeling options: a linear state-space model and a hybrid model that combines a physical model with a Gaussian process data model. Active detectors (threshold-based and LDDM) trigger adaptation, which re‑identifies the model or retrains the data component on warning data, and the framework is validated on a simulated anisotropic non‑ideal rotor showing improved prediction accuracy and faster adaptation, especially with the LDDM detector. The results demonstrate the practicality of lifelong learning for DTs in degraded systems and highlight robustness advantages of LDDM over window-based approaches. The work has implications for real‑time, self‑managing DTs in manufacturing contexts, enabling more reliable predictions amid continual wear and degradation.

Abstract

In the manufacturing industry, the digital twin (DT) is becoming a central topic. It has the potential to enhance the efficiency of manufacturing machines and reduce the frequency of errors. In order to fulfill its purpose, a DT must be an exact enough replica of its corresponding physical object. Nevertheless, the physical object endures a lifelong process of degradation. As a result, the digital twin must be modified accordingly in order to satisfy the accuracy requirement. This article introduces the novel concept of "learning digital twin (LDT)," which concentrates on the temporal behavior of the physical object and highlights the digital twin's capacity for lifelong learning. The structure of a LDT is first described. Then, in-depth descriptions of various algorithms for implementing each component of a LDT are provided. The proposed LDT is validated on the simulated degradation process of an anisotropic non-ideal rotor system.

Figures (11)

• Figure 1: The structure of a digital twin b7
• Figure 2: The structure of a LDT and its connection with the physical twin. A LDT consists of four parts: memory, a digital model, a detector, and an adaptor. The backwards influence from LDT to physical twin is not implemented in this paper and is hence denoted by a dashed arrow.
• Figure 3: Hybrid model structure. It consists of two parts, physical model and a data model. In this work, a Gaussian process model is implemented as the data model.
• Figure 4: An eccentric rotor driven by a DC motor, which is placed on an anisotropic nonrigid foundation b22.
• Figure 5: Input and output in an experiment. The supply voltage $V_s$ is chosen as chosen as system input $u$ and the motor velocity $w$ is chosen as output $y$, respectively.