Uma nova arquitetura de rede neural artificial para abordagens end-to-end de classificação de sinais

Abstract

The increasing computer power, the introduction of specialized processors designed to accelerate graphics rendering, and the high availability of new data have supported the adoption of artificial neural networks (ANN) to solve real-world problems in different areas, such as computer vision natural language processing. The application of such networks to deal with data characterized by temporal dependencies, as signals (time series), produces better results when a preprocessing stage is considered to extract implicit information. In this work, by taking into account the different preprocessing methods, we have noticed that wavelet transforms have been widely used due to, in general, the remarkable results. However, such transformations have two main drawbacks by modeling signals. Firstly, it is necessary to define a set of parameters to extract better information from different levels of scale and resolution. Secondly, the application of such transformations requires an initial step to convert, in batch, all analyzed signals before starting the network training. In our scenario, such limitations were overcome by our new ANN architecture, which analyzes signals and yields as output a set of characteristics similar to those produced by wavelet transforms. Therefore, our ANN can be combined with others to model signals without requiring an execution a priori of any wavelet transformation. The proposed ANN was assessed against a pre-trained network to classify signals from two real-world applications: local field potential (LFP) from monitoring monkey brains and seismic signals from the Llaima volcano. The final accuracy obtained to model both applications using traditional preprocessing, and our approach were, respectively: (i) LFP -- 0.655 and 0.649, and (ii) Llaima -- 0.976 and 0.974.