Optimal architecture of a convolutional neural network to estimate structural responses for safety evaluation of the structures

A convolutional neural network (CNN) is a deep learning algorithm, which can be utilized in various engineering fields due to its superior prediction and classification performance. In recent years, CNN that is known to be outstanding to handle large volumes of data, it is has been in the spotlight to solve the problems of sensor defects and data loss, which may have resulted from the limitations of the current structural health monitoring (SHM) techniques. However, although the architecture of CNN should be constructed differently depending on the characteristics of each problem, there is no rational nor reasonable method for the construction of the architecture. In this regard, this study seeks to propose a technique for constructing an optimal architecture for the effective utilization of CNN in recovery and estimation of measured data dealt in the field of SHM. In the proposed technique, the number of kernels, the kernel size, and the subsampling size are set as the decision variables, among the variables that determine the CNN architecture. To prevent CNN from being trained with bias toward specific datasets, root mean square errors are calculated for each of the training datasets and validation datasets, and set as objective functions, respectively. Then these two objective functions are minimized at the same time. In this case, non-dominated sorting genetic algorithm-II, a multi-objective optimization algorithm, is introduced to minimize these two objective functions. The proposed technique is verified by a numerical study on beam-like structures and an experimental study on reinforced concrete structures. These two studies explore the optimal CNN architecture, which estimates the dynamic strain of the structure, and evaluates the performance of the explored architecture.