Efficient Model Updating Method for System Identification Using a Convolutional Neural Network

Model updating processes are important for improving a model's accuracy by considering experimental data. Structural system identification was achieved here by applying the degree-of-freedom-based reduction method and the inverse perturbation method. Experimental data were obtained using the specific sensor location selection method. Experimental vibration data were restored to a full finite element model using the reduction method to compare and update the numerical model. Applied iteratively, the improved reduced system method boosts model accuracy during full model restoration; however, iterative processes are time consuming. The computation efficiency was improved using the system equivalent reduction-expansion process in concert with proper orthogonal decomposition. A convolutional neural network was trained and applied to the updating process. We propose the use of an efficient model updating method using a convolutional neural network to reduce computation time. Experimental and numerical examples were adopted to examine the efficiency and accuracy of the model updating method using a convolutional neural network.