Classification of acoustical alarm signals with CNN using wavelet transformation

This paper presents a wavelet transformation (WT) based technique for reducing the size of cellular neural network (CNN) used for an acoustic alarm signals classification system proposed by Osuna et al. The system consists of three processing units: i) transformation of a 1-dimensional (1-D) signal into a sequence of 2-dimensional (2-D) signals, so called images obtained by a low pass filter cascade incorporated with a grid like correlation process ii) concentrating an image sequence into a single image by a linear threshold template CNN, iii) classification of the resulting image by discrete-valued perceptrons. In this paper, a discrete WT incorporating a grid like correlation process has been used for transforming a 1-D acoustic signal into an image sequence. All other operations needed for the classification has been performed for the sake of comparison. The WT based technique proposed in this paper gives the possibility of acoustic alarm signal classification by using CNNs of small size, e.g., 13/spl times/13. By using the WT based technique, CNN of size 13/spl times/13 becomes sufficient.