Multi-Objective Optimization Method to Search for the Optimal Convolutional Neural Network Architecture for Long-Term Structural Health Monitoring

To ensure the safety of structures, structural health monitoring (SHM) techniques that use cutting-edge sensing technologies have been developed. However, in the process of long-term structural health monitoring, sensor defects and data loss commonly occur, which pose limitations in the current SHM technique. To recover lost data and predict structural responses, convolutional neural networks (CNNs) have been used in SHM, but no obvious technique or rule for configuring CNN architecture with optimal performance has been presented yet. This study proposes a method for searching for the optimal CNN architecture capable of predicting the structural response of structures to evaluate their long-term safety. In this method, multi-objective optimization, considering both prediction performance and CNN training efficiency, is presented as a strategy. The optimization method using the two objective functions is applied to the structural response estimation, and the characteristics of the derived solutions are examined. Furthermore, the solutions derived using the two objective functions are classified into two solution groups that are biased to each objective function, and a strategy for minimizing the discrepancy between the two solution groups is additionally presented based on their trade-off relationship. The architecture characteristics, estimation performance, and training efficiency of the solutions derived by setting the discrepancy as the third objective function are investigated. The CNN derived by the proposed method with the third objective function reduced 40.35% of computational cost compared with that derived with two objective functions while they showed similar accuracies for the response estimation.