Impact load identification by training a convolutional neural network (CNN) with physics-based simulator under uncertainty

Abstract

Identifying an impact load applied to structural systems and taking a prompt maintenance action is critical in structural health management. While various approaches have been investigated, this impact load identification is still a challenging task, considering associated high-level measurement requirements and sophisticated dynamic characteristics of uncertain structural systems. In addressing this challenge, we propose a novel impact load identification method based on a convolutional neural network (CNN) that incorporates the system model parameter uncertainties. This method aims to estimate the location, magnitude, and direction of an impact load using just one multi-axis accelerometer. While CNNs are powerful DL classifier models capable of mapping a structure’s response to input excitation, their effectiveness is frequently hindered by the scarcity of training data. In this study, we proposed to utilize a Physics-Based simulator to generate the necessary training dataset. Physics-based simulators show great performance in robotics, building energy management, and cooling systems and it can provide valuable training data. Consequently, a physics-based simulator for uncertain systems was developed to bridge the gap between simulated and actual structures and evaluated using uncertainty analysis. To validate the proposed method, both numerical and laboratory case studies have been conducted. The Physics-Based simulator demonstrated its ability to effectively train the CNN model, addressing uncertainty issues in model parameters. Moreover, the applicability of this method was tested by utilizing a single measurement point in different locations, showcasing its versatility and potential for real-world engineering applications.