BASIC STUDY ON A DETECTION METHOD FOR BREAKING SOUNDS OF STRUCTURAL MEMBERS BY USING MACHINE LEARNING

Abstract

In case of large earthquake, it is important to grasp the damage distribution over a city because such information is very useful for taking emergency activities. Therefore, many researches have been carried out to develop a system for estimating the damage of structure. In many researches, acceleration responses or other motions are observed with sensors and the measured time series data are analyzed to find out the change of vibration characteristics due to the damage or to estimate the deformation of structure. On the other hand, we focus on the breaking sounds of structural members emitted in a large earthquake to assess the damage of structure. In this research, frist, breaking sounds of timber and other sounds were observed in a silent laboratory, and the recorded data were used for machine learning of NN (neural network). In the analysis, about 80% of samples were correctly judged by using the learned NN model. The echoes of breaking sounds were also investigated whether they could be categorized into the breaking sounds or not. The k-Shape, that is one of the clustering methods, were also applied to the recorded sounds to study the ability for classification. Comparison between the analyzed results of them show that the NN model gives better success rates of detecting breaking sounds of timber in this research.