Digital twin structural health monitoring driven by multi-fidelity time-series surrogate models

Abstract

In structural health monitoring (SHM), real-time observation of a system's structural performance is conventionally undertaken. This involves the utilization of indirect measurement techniques to monitor targets that are not directly quantifiable through sensor readings. The digital twin (DT) concept offers a novel comprehensive structural performance monitoring approach by establishing a virtual-to-physical mapping. Nevertheless, challenges persist in predicting autocorrelated time-series data in dynamic systems and achieving a balance between the costs and modeling accuracy. This paper proposes a Digital Twin framework based on a multi-fidelity time-series surrogate mode. Taking the real-time monitoring of the suspension structure stress of an autonomous mobile robot (AMR) as an example, the application process of this framework is illustrated. First, a load identification model was constructed based on measured data to recognize the load applied to the suspension structure, providing input parameters for the subsequent stress prediction model. The second stage involves the construction of an autoregressive least squares multi-fidelity Gaussian process regression model. Incorporating an autoregressive term within this model enables the integration of the autocorrelation characteristics of the data during prediction whilst simultaneously automating the optimization of hyperparameters based on heuristic rules. This process mitigates the influence of initial hyperparameter settings on the model's training. Finally, a DT of the AMR suspension structure was constructed, and its effectiveness in real-time stress monitoring of the suspension structure was verified under five different working conditions. The work improves the prediction accuracy of time-series monitoring targets in dynamic systems, offering a new solution for applying DT in SHM processes.