Artigo

APRENDIZADO DE MÁQUINA APLICADO À CLASSIFICAÇÃO DE SINAIS OBTIDOS POR EQUIPAMENTO ELETROENCEFALOGRÁFICO NÃO-INVASIVO

BECCHI, Gabriel Chaves1; SANTOS, Alisson Ravaglio3; COELHO, Leandro Dos Santos2;

Resumo

Introdução:There has been intensive research from academics and practitioners regarding models for neurosciences.

Objetivo:This project aims to compare different classifiers of the machine learning field involving single approach and ensemble approaches to classify electroencephalography (EEG) data obtained using an Emotiv neuroheadset. Furthermore, an optimization tool using particle swarm intelligence, a swarm intelligence approach, is adopted for tuning the machine learning classifiers.

Metodologia:Several tests were realized with several people using Emotiv neuroheadset. The output signals are divided in 5 classes based in thoughts, being 0 for neutral, 1 for up, 2 for down, 3 for left and 4 for right.

Resultados:The final result is a comparative study between the general classification algorithms, such as k-nearest neighbors, deicsion tree, and support vector machine, for the RRG data. Accuracy of over 98% and a classification average time of 0.7 milliseconds with the tuned custom ensemble was obtained. Modeling the data as wave has presented the best results in building a patterned model. The noisy signal, after the moving average and Savitzky-Golay filter, has presented good final result in amplitude and frequency information pattern. As for the classification algorithms, ensemble of weak learners has the best final accuracy. Giving this weak learner a random sample of the train data and setting the cross-validating accuracy as weight in the voting system does a generalization necessary for multi-individual real-time application.

Conclusões:There are still many other methods that need investigation for a final global optimal algorithm for this EEG data movement classification. The most import considerations are always having a good data set, pre-process the data, for removal of noisy signals, and prepare a generalize classification machine, since individuals have different brain activities, over fitting can happen easily.

Palavras-chave:Machine learning. Classifiers. Emotiv neuroheadset.

Legendas

    1. Estudante
    2. Orientador
    3. Colaborador