Classificação de sinais eletromiográficos do músculo masseter de bovinos baseada em dicionários para reconhecimento de padrões ingestivos

Abstract

This work presents a proposal for a new pattern recognition method of surface electromyography signals of cows masseter muscle for classification of ingestion and rumination patterns, as well as the definition of a segmentation methodology for this signal. The method, called Fisher Discriminant Dictionary Learning (FDDL), is based on the training of class-specific dictionaries, which is a matrix composed of signal prototypes. Dictionaries define a sparse vector that encodes the signal for the purpose of signal reconstruction. The reconstruction error information from each dictionary is used as a metric for classification, thus eliminating the feature extraction step. Results in cows with 2000 examples of chewing and cross-validation showed a significantly higher performance rating (p < 0.05) in relation to the current methods found in the literature, with an average accuracy of 90 %. The method proved to be robust in the presence of noise, with a performance gain of 2.45% with addition of severe noise (0 dB) in the training and test and 14.75% with addition in the test only, being superior to all methods in the range of 0-20 dB. Further works should evaluate the method’s ability to be implemented for real-time opperation and expand the classification to other ingestive patterns, such as identification of pasture height.