第五部門：応用数理問題（C)

[S05C-01] Comparison between the classification methods using NN and CNN for the detection of wood destruction sounds.

*Tsukasa Takemori¹, Masayuki Saeki¹ (1. Tokyo University of Science)

Keywords:Machine Learning, Convolutional Neural Network, Disaster prevention engineering

This paper shows the results of precision comparison between the classification methods using NN and CNN which are developed for detecting a wood destruction sound in case of a large earthquake. In this research, first, MFCC of wood breaking sounds are calculated as sound characteristics and are input to NN as training and evaluation data. Next, the spectrogram or the mel-spectrogram of the same wood breaking sounds are calculated and input to CNN. Then, the precisions of the classification methods are evaluated. The result shows that the method using CNN gives better precision than that using NN. Furthermore, the other sounds are synthesized by mixing the wood breaking sounds with other environmental sounds and are classified using the trained model. Even in this case, the method using CNN gives better precision in our analysis.