Hücresel yapay sinir ağları ile alarm seslerinin sınıflandırılması

Abstract

Bu tezde, Hücresel Yapay Sinir Ağlan (HYSA)'nın alarm seslerinin sınıflandırılmasında kullanılabilirliği incelenmiş ve bu amaçla dalgacık ("wavelet") dönüşümü (DD)'ne dayalı bir teknik geliştirilmiştir. Böylece, iki boyutlu ızgara biçimindeki mimarisi ile doğal uygulama alanım görüntü işlemede bulan HYSA'nın ses işareti gibi bir boyutlu işaretlerin işlenmesinde kullanımına yönelik ilk adımlardan biri atılmıştır. Tez çalışmasında önerilen yaklaşım, araba korna sesleri, elektronik ve elektromekanik telefon zili sesleri gibi göreceli olarak basit seslerin sınıflandırılması problemine uygulanmış ve sonuçlar literatürdeki benzer çalışmalar ile karşılaştırılmıştır. Sınıflandırma probleminin çözümünde yapılan işlemler üç ana bölüme ayrılarak incelenebilir. Ön işlem birimi olarak adlandırılan birinci bölümde ses işaretleri görüntüye dönüştürülür. Bu birimin çıkışında, dizinin her bir elemanı bir ses çerçevesine ("frame")'e karşı düşmek üzere bir görüntü dizisi elde edilir. İkinci birim ise zaman içerisinde gelen görüntü dizisinden kolay sınıflandırılabilir tek bir işaret elde etmek amacıyla kullanılır. Bu birim içerisinde bir HYSA ve basit bir mantık işlemcisi bulunmaktadır. Üçüncü ve son birim ise sınıflandırma problemlerinde sıkça kullanılan çok çıkışlı, tek-katmanlı doğrusal bir algılayıcıdan oluşmaktadır. Elde edilen sonuçlar aynı uygulama üzerine yapılan diğer çalışmalar ile karşılaştırıldığında onların verdiklerine yakın mertebede iyi hata oranlan gözlenmiştir. Bu çalışmadaki yaklaşımın bir çok noktada iyileştirmeye olanak tanıması bir üstünlüğü olarak belirtilebilir.

This thesis presents Cellular Neural Networks (CNNs)[l],[2] based classification system for recognizing acoustic alarm signals. The proposed classification system uses wavelet transformation (WT) for transforming a one dimensional (1-D) acoustic signal into a sequence of two dimensional (2-D) signals, so called input images, which will be fed to the CNN. Test samples from three sets of acoustical alarm signals are recognized with an error rate of less than 16%, which shows that CNN can be used as a tool for acoustic signal recognition. The extensive works done in the literature show that CNNs with their 2-D grid architecture are well suited for image processing. Up to now, a very limited work is devoted to the applications of CNNs to process 1-D signals such as acoustic signals. This work extends by introducing a wavelet based technique, the field of CNN applications on the classification of acoustic alarm signals which was initiated by Osuna et. al [3]. The proposed solution in this thesis for the classification of acoustical alarm signals is divided into three parts. The first, a transformation of 1-D into a sequence of images is done by WT together with a grid like correlation process. The produced image sequences serve as the inputs to the CNN. CNN is responsible for the extraction of the relevant information carried by the image sequences, and it also performs a time-to-space mapping via concentrating the information distributed over an image sequence into a single image. Thirdly, a simple classification device based on linear perceptron is used for separating the VIdifferent classes of input signals by sorting the output images of the CNN into the correct classes. (See Figure 1.) The first unit in Figure 1, performs a transformation of the 1-D acoustical signal into image sequences. Every input signal is then characterized by its corresponding image sequence. Image sequences produced by the first unit are applied to the second unit, CNN. This block essentially consists of two parts: i) Analysis and ii) Correlation. Signal 1-D to 2-D Transform CNN Classifier Class Sequence of Images Two Images Figure 1. Processing units for the classification of acoustical signals. To perform analysis, WT is used, since there are many similarities between the WT and hearing. In the modelling of cochlea and auditory processing, WT is used by several researchers[6],[7]. Although continuous wavelet transformation can work in this method, discrete wavelet transform (DWT) which is an application of WT for sampled signals, has been used. The reason is that there are no fully programmable CNN chips available yet, and CNN must be simulated on a digital computer. vnWT, simply, decomposes an arbitrary function into a two-parameter family of elementary wavelets that are obtained by shifts in time variable but also dilations (or compressions) that act both on time and the frequency variables. WT can be defined as follows; WTx(x,a) = 4TJx(t)-hf^dt (1) VH ^ a ' where h(t) is a basic wavelet prototype and called as mother wavelet or analysing wavelet, h ft-x is scaled and translated version of basic wavelet V a j prototype, a is scaling parameter, and only positive dilations a>0 is taken into account, x shifts the wavelet in time. h(t) can be any band-pass function. In particular one can dispense with complex-valued transforms and deal only with real-valued ones. The DWT is defined as DWT{x[n];2j,k2j} = cik = 2>(n)h*(n-k2j) (2) where hj(n-k2J)'s are analysis discrete wavelets. In this equation scale, a, is 2İ and time lo cation, x, is k2J. Consider two filter impulse responses g(n) and h(n). Here h stands for high-pass or discrete wavelet and g stands for low-pass. The wavelets and scaling sequences are obtained iteratively as g1(n) = g(n), h1(n) = h(n), (3a) Vlllgj+1(n) = £gj(k)-g(n-2k), hj+1(n) = 2hj(k)-g(n-2k)3 (3b) (3c) (3b) and (3 c) means that one goes from one octave j to the next (j+1) by applying the interpolation operator fj+1(n) = Zfi(k)'g(n-2k)> (4) which should be thought of as the discrete equivalent to the dilation; f(t)->2"1/2f(t/2) The DWT corresponds to the analysis filter bank. The filters present in the filter bank are precisely g(n), h(n). There are several algorithms to implement DWT [12]-[14]. Direct implementation will be employed in this thesis. (See Figure 2.) LPF -® Decimate! by2 HPF © LPF -©- HPF LPF -d>-> k+1 HPF Figure 2. DWT processing block (dotted rectangle) and DWT structure.