The aim of this paper is to classify two kind of signals recorded by seismic station: artificial explosions and seismic activity. The problem is approached from both the preprocessing and the classification point of view. For the preprocessing stage, instead of the conventional Fourier Transform, we use a Linear Prediction Coding (LPC) algorithm, which allows to compress the data and extract robust features for the signal representation, For the classification stage, we have compared the performance of several neural models. An unsupervised method, based on the Principal Component Analysis (PCA) and the Mixture of Gaussian (MoG) clustering algorithm, gives a 70% percentage of correct classification. The Elman Recurrent Neural Nets (RNN) is able to reach 91% of correct classification on the test set. However this performance is strongly and critically dependent on the order of presentation of the events. Instead a MLP with a single hidden layer gives the 86% of correct classification on the test set, independently of the order of presentation of the patterns.
Signal classification using neural networks
FALANGA, Mariarosaria;MARINARO, Maria;SCARPETTA, Silvia
2002
Abstract
The aim of this paper is to classify two kind of signals recorded by seismic station: artificial explosions and seismic activity. The problem is approached from both the preprocessing and the classification point of view. For the preprocessing stage, instead of the conventional Fourier Transform, we use a Linear Prediction Coding (LPC) algorithm, which allows to compress the data and extract robust features for the signal representation, For the classification stage, we have compared the performance of several neural models. An unsupervised method, based on the Principal Component Analysis (PCA) and the Mixture of Gaussian (MoG) clustering algorithm, gives a 70% percentage of correct classification. The Elman Recurrent Neural Nets (RNN) is able to reach 91% of correct classification on the test set. However this performance is strongly and critically dependent on the order of presentation of the events. Instead a MLP with a single hidden layer gives the 86% of correct classification on the test set, independently of the order of presentation of the patterns.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.