Enhancing the credibility of an entropy-based structural health monitoring system using convolutional neural networks

Abstract A practical structural health monitoring (SHM) system is proposed based on neural network entropy (NNetEn) and convolutional neural network (CNN) for solving the issue of diagnosis reliability and over-reliance on analysis parameters. As the NNetEn approach is less affected by parameter selection during the analysis process, it is first adopted. In contrast to other entropy analyses, the NNetEn method minimizes the influence of the time series length and amplitude on the analysis outcomes. Furthermore, the CNN model is used for credibility boosting, which enhances the reliability of the overall detection. For verification of the effectiveness of the proposed method, a seven-story numerical model is constructed, and an experiment on a scaled-down steel structure with 16 damage cases is conducted. A confusion matrix is implemented to judge the results, and the effectiveness is evaluated against four metrics: accuracy, precision, recall, and F 1-score. The accuracy obtained from numerical simulation and experiment for the developed SHM system is 95.35% and 93.13%, respectively, and the other three metrics also demonstrate good consistency. Moreover, a comparative analysis with traditional methods and other entropy-based approaches demonstrated the superior performance of the new system. The obtained results have effectively proved the credibility of the proposed SHM system. The developed system has a high potential for practical implementation in structural safety diagnosis, making it a valuable asset.