Wavelet-based techniques for adaptive feature extraction and pattern recognition.

AUTOR(ES)
DATA DE PUBLICAÇÃO

1999

RESUMO

In this work, wavelet-based techniques are studxied for adaptive feature extraction in the time-frequency plane. Emphasis is placed on pattern recognition problems, in particular fault detection in control systems and classification / clustering electrocardiographic signals. In the context of fault detection, a technique for residue generation using wavelet filter banks is introduced. Numerical simulations show that the proposed method exhibits good noise rejection characteristics and robustness to transient inputs, either from commands or unmeasured exogenous disturbances. Results are compared to those obtained by a standard observer-based technique. Classification of electrocardiographic patterns is performed with a wavelet neural network, which employs an adaptive wavelet layer as a pre-processing stage to a perceptron classifier. Basic concepts involved, as well as aspects of training and initialization are discussed. Two modifications to the basic supervised training algorithm are proposed, namely the introduction of a bias component in the wavelets and the adoption of a weight decay policy. Results are interpreted with basis on the concept of superposition wavelets. A competitive wavelet network, inspired in Kohonen lavers, is proposed for means of pattern clustering. It was verified that this paradigm has some advantages over the conventional neural layers when patterns to be analyzed have a low signal to noise ratio. A hierarchical, multiresolutional procedure for performing clustering is also presented. Classification / clustering tests are carried out on signals taken from the MIT-BIH Arrhythmia Database.

ASSUNTO(S)

sistemas dinÃmicos processamento de sinais reconhecimento de padrÃes inteligÃncia artificial anÃlise de ondas localizadas eletrocardiografia redes neurais anÃlise de falhas

ACESSO AO ARTIGO

Documentos Relacionados