Simple diagnosis for layered structure using convolutional neural networks

In this study, we propose a structural health monitoring and diagnostic method for layered (multi-story) structures using a convolutional neural network (CNN). The proposed method is a primary diagnostic one, and its purpose is to allow quick identification of the location of an abnormality after detecting it. An abnormality is defined as a decrease in the stiffness characteristics (spring constant) of the outer wall of a multi-story structure when it deteriorates or is damaged. The proposed method has the following features. A modal circle is generated by multiplying the frequency response functions (FRFs) simulated by a mathematical model and the FRFs from the actual structure, in frequency space, and then a CNN learns the features of the abnormality from the modal circle and diagnoses it in the actual multi-story structure. We first verified the validity of the proposed method by considering a three-story structure as a numerical example. When the method was applied to three types of abnormal conditions, it was shown that the abnormal diagnosis could be performed correctly. Next, we constructed an experimental model of a three-story structure, and realized three types of abnormal conditions similar to those in the numerical model. We verified the applicability of the proposed method and showed that correct diagnosis of an abnormality was possible. Both the validity and applicability of the proposed method were thus confirmed.