2.3 Etapa 1: Aquisição e Segmentação de dados
2.3.2 Segmentação de Dados EMG
De acordo com (Oskoei e Hu, 2008a), existem dois métodos de segmentação EMG: segmentação separada e segmentação sobreposta. Na segmentação separada, segmentos
disjuntos com um comprimento predefinido são utilizados para a extração de caracte- rísticas. Por outro lado, na segmentação sobreposta, o novo segmento desliza-se sobre o segmento atual, com um incremento (intervalo de tempo entre dois segmentos conse- cutivos) de tempo menor do que o comprimento do segmento, e maior do que o tempo de processamento. Assim, a segmentação separada está associada com o comprimento do segmento, enquanto que a segmentação sobreposta está associada com o tempo e o
incremento, como mostrado na Figura 2.11.1958 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 55, NO. 8, AUGUST 2008
Fig. 1. Disjoint (left) and overlapped (right) segmentation.
III. M
ETHODOLOGYThis section introduces methods applied for data collection,
experiments, and analysis. It is comprised of different sections
for data acquisition, data segmentation, feature selection, clas-
sification, postprocessing, and evaluation.
A. Data Acquisition
A four-channel MES was collected from four locations on a
forearm (i.e., biarticulate wrist flexor and triarticulate and biar-
ticulate wrist extensor muscles), using bipolar active electrodes
(Biometrics, Ltd., SX230). An active electrode has a preampli-
fier with gain 1000, which can differentiate between a small
signal of interest and much larger interference signals that are
present on the skin. It also has a very high input impedance
to cope with mismatches in skin contact resistance. Signals are
passed through a high-pass filter with a cutoff frequency of
20 Hz to remove dc offsets due to membrane potentials, and to
minimize interference due to electrode movement. A low-pass
filter is used to remove unwanted frequencies above 450 Hz,
and a notch filter used to remove unwanted line-frequencies
(50/60 Hz). An electrode was also placed on the wrist, provid-
ing a common ground reference. Signals were sampled at 1000
Hz using a 12-b A/D converter.
Data were collected from 11 healthy subjects. Each subject
performed five limb motions, and rest to provide six distinct
states (i.e., classes). The motions were isotonic and comprised of
flexion, extension, abduction, adduction, and keeping the hand
straight. Two sequences of six motions in which each motion
was held fixed for the five seconds are called a block. Four
blocks of data were gathered from subjects in each session. Two
sessions were conducted for each subject, and in each session,
the accuracy of classification was computed using a fourfold
cross-validation method.
B. Data Segmentation
A segment is a sequence of data limited in a time slot,
which is used to estimate signal features. A short length of
segment leads to bias and variance in feature estimation; while
a long one imposes high computational load and a likely fail-
ure to perform real-time operation. Real-time constraints en-
force a delay time of less than 300 ms between the onset of
muscle contraction made by a subject, and a corresponding
motion in a device [3]. It should be noted that the minimum
interval between two distinct contractions is approximately
200 ms [29], [30]. This means that a segment of MES data
with a length of 200 ms (or more) contains enough information
to estimate a motion state of the hand. A segment length equal
to or less than 200 ms leaves enough time (at least 100 ms)
for the computation of features, classification, and generation
of control commands, plus a device response time to maintain a
real-time smooth motion control scheme.
However, a segment larger than 200 ms necessitates over-
lapped segmentation [1], [2] in order to avoid failure in real-time
operation. The application of overlapped segmentation facili-
tates the employment of large segments (greater than 200 ms)
for real-time control; however, computational load is still a mat-
ter of fact. Two methods of segmentation, namely, disjoint and
overlapped segmentation, are illustrated in Fig. 1. Disjoint seg-
mentation is associated with segment length, while overlapped
segmentation is associated with length and increment. The in-
crement is the time interval between two consecutive segments.
It should be less than the segment length, and more than the
processing time.
The first experiment in which the accuracy of classifications
with a segment length of 50, 100, 150, 200, 300, and 500 ms
were examined, investigates the influence of the segment length
on classification (for different features). For segment lengths of
50, 100, 150, and 200 ms, disjoint segmentation was applied,
while for lengths of 300 and 500 ms, overlapped segmentation
with an increment of 200 ms was applied. Since an MES has an
undetermined state between two levels of contraction, most clas-
sification error belongs to the transition period between classes.
Hence, to avoid contradictory data during transition between
motions, one segment of the transition period was eliminated
from the training data. This means that classification relies on a
steady state of muscle contraction.
C. Feature Selection
Because of the significance of features during classification,
feature selection is an essential stage in myoelectric control de-
sign. A significant amount of literatures investigates or compares
various features of TD, frequency-domain (FD), and time-scale,
for myoelectric control [1], [2], [25]. Features should be capa-
ble of presenting the characteristics or properties of a signal
for different limb motions. Computational load should also be
considered in the real-time applications. In this work, the rel-
ative performance of various single features and feature sets
(multifeatures) is determined in the context of an SVM-based
classifier.
(a)
1958 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 55, NO. 8, AUGUST 2008
Fig. 1.
Disjoint (left) and overlapped (right) segmentation.
III. METHODOLOGY
This section introduces methods applied for data collection,
experiments, and analysis. It is comprised of different sections
for data acquisition, data segmentation, feature selection, clas-
sification, postprocessing, and evaluation.
A. Data Acquisition
A four-channel MES was collected from four locations on a
forearm (i.e., biarticulate wrist flexor and triarticulate and biar-
ticulate wrist extensor muscles), using bipolar active electrodes
(Biometrics, Ltd., SX230). An active electrode has a preampli-
fier with gain 1000, which can differentiate between a small
signal of interest and much larger interference signals that are
present on the skin. It also has a very high input impedance
to cope with mismatches in skin contact resistance. Signals are
passed through a high-pass filter with a cutoff frequency of
20 Hz to remove dc offsets due to membrane potentials, and to
minimize interference due to electrode movement. A low-pass
filter is used to remove unwanted frequencies above 450 Hz,
and a notch filter used to remove unwanted line-frequencies
(50/60 Hz). An electrode was also placed on the wrist, provid-
ing a common ground reference. Signals were sampled at 1000
Hz using a 12-b A/D converter.
Data were collected from 11 healthy subjects. Each subject
performed five limb motions, and rest to provide six distinct
states (i.e., classes). The motions were isotonic and comprised of
flexion, extension, abduction, adduction, and keeping the hand
straight. Two sequences of six motions in which each motion
was held fixed for the five seconds are called a block. Four
blocks of data were gathered from subjects in each session. Two
sessions were conducted for each subject, and in each session,
the accuracy of classification was computed using a fourfold
cross-validation method.
B. Data Segmentation
A segment is a sequence of data limited in a time slot,
which is used to estimate signal features. A short length of
segment leads to bias and variance in feature estimation; while
a long one imposes high computational load and a likely fail-
ure to perform real-time operation. Real-time constraints en-
force a delay time of less than 300 ms between the onset of
muscle contraction made by a subject, and a corresponding
motion in a device [3]. It should be noted that the minimum
interval between two distinct contractions is approximately
200 ms [29], [30]. This means that a segment of MES data
with a length of 200 ms (or more) contains enough information
to estimate a motion state of the hand. A segment length equal
to or less than 200 ms leaves enough time (at least 100 ms)
for the computation of features, classification, and generation
of control commands, plus a device response time to maintain a
real-time smooth motion control scheme.
However, a segment larger than 200 ms necessitates over-
lapped segmentation [1], [2] in order to avoid failure in real-time
operation. The application of overlapped segmentation facili-
tates the employment of large segments (greater than 200 ms)
for real-time control; however, computational load is still a mat-
ter of fact. Two methods of segmentation, namely, disjoint and
overlapped segmentation, are illustrated in Fig. 1. Disjoint seg-
mentation is associated with segment length, while overlapped
segmentation is associated with length and increment. The in-
crement is the time interval between two consecutive segments.
It should be less than the segment length, and more than the
processing time.
The first experiment in which the accuracy of classifications
with a segment length of 50, 100, 150, 200, 300, and 500 ms
were examined, investigates the influence of the segment length
on classification (for different features). For segment lengths of
50, 100, 150, and 200 ms, disjoint segmentation was applied,
while for lengths of 300 and 500 ms, overlapped segmentation
with an increment of 200 ms was applied. Since an MES has an
undetermined state between two levels of contraction, most clas-
sification error belongs to the transition period between classes.
Hence, to avoid contradictory data during transition between
motions, one segment of the transition period was eliminated
from the training data. This means that classification relies on a
steady state of muscle contraction.
C. Feature Selection
Because of the significance of features during classification,
feature selection is an essential stage in myoelectric control de-
sign. A significant amount of literatures investigates or compares
various features of TD, frequency-domain (FD), and time-scale,
for myoelectric control [1], [2], [25]. Features should be capa-
ble of presenting the characteristics or properties of a signal
for different limb motions. Computational load should also be
considered in the real-time applications. In this work, the rel-
ative performance of various single features and feature sets
(multifeatures) is determined in the context of an SVM-based
classifier.
(b)
Figura 2.11: (a) Segmentação separada; (b) Segmentação sobreposta. Extraído de (Oskoei e Hu, 2008a).
(Christodoulou e Pattichis, 1999) usaram uma janela de comprimento constante e um algoritmo de segmentação que calcula um limiar dependente do valor máximo e do valor médio absoluto do sinal EMG completo. Picos que excedem o limiar calculado foram considerados como segmentos candidatos. (Gut e Moschytz, 2000) utilizaram uma janela deslizante no tempo para determinar o início e o fim de um segmento. Se o declive médio dentro desta janela ultrapassa um determinado limiar, o início de um segmento é
detectado, enquanto que o fim de um segmento é atingido quando a variação total do sinal EMG, dentro da janela, consegui ficar abaixo de outro limite.
(Oskoei e Hu, 2008a) avaliaram a segmentação disjunta e sobreposta, comparando o desempenho de classificação sobre segmentos disjuntos e sobrepostos (entre 50 ms e 200 ms), e com incrementos de 50 ms. Os resultados mostraram que a segmentação separada com um comprimento de 200 ms fornece um alto desempenho durante classificação de sinais EMG, e um tempo de resposta razoável para aplicações em tempo real, enquanto que a segmentação sobreposta, com um comprimento de 200 ms e um incremento de 50 ms, encurta o tempo de resposta sem degradação perceptível na precisão.
(Kaur et al., 2009) analisaram três técnicas de segmentação EMG: 1) por meio da identificação dos picos das MUAPs; 2) encontrando o Ponto de Inicial de Extração BEP (do inglês, Beginning Extraction Point) e o Ponto Final de Extração EEP (do inglês, En- ding Extraction Point) das MUAPs; e 3) usando Transformada Wavelet Discreta TWD (do inglês, Discrete Fourier Transform). Na primeira técnica, o sinal de EMG foi segmentado utilizando um algoritmo que detecta áreas de baixa atividade e MUAPs candidatas; na segunda técnica identificou-se os BEPs e os EEPs das possíveis MUAPs por meio do deslizamento de uma janela ao longo do sinal; e na terceira técnica, o sinal EMG foi de- composto com a ajuda da Wavelet Daubechies4 (db4) para detectar MUAPs. Em geral, a primeira técnica teve o melhor desempenho, com uma taxa de acerto total de 95, 90%, em comparação com percentagens de 75, 39% e de 66, 64% para a segunda e terceira, respectivamente.
2.4
Etapa 2: Extração de Características
A etapa de extração de características implica a transformação do sinal bruto em uma estrutura de informação relevante, chamado vetor característico, através da eliminação de ruído e destacando os dados importantes do sinal. Além disso, também pode implicar um processo de "redução de dimensionalidade", o qual elimina informação redundante, a partir do vetor característico, com o objetivo de facilitar o processo de classificação.
De acordo com (Zecca et al., 2002), existem três grandes grupos de características em sistemas de controle EMG: a) Domínio do tempo; b) Domínio da frequência; e c) no Domínio tempo-frequência. A Tabela 2.1 mostra as principais técnicas de extração de características encontradas na literatura.
Tabela 2.1: Técnicas de extração de características para sistemas de controle EMG.
Tipos Característica
Integral de EMG (IEMG) Valor Médio Absoluto (MAV)
Valor Médio Absoluto Modificado 1 (MMAV1) Valor Médio Absoluto Modificado 2 (MMAV2) Inclinação do Valor Médio Absoluto (MAVS) Raiz Média Quadrática (RMS)
Domínio do Tempo Variância (VAR)
Comprimento de Forma de Onda (WL) Cruzamentos de Zero (ZC)
Mudanças de Sinal de Inclinação (SSC) Amplitude Willison (WAMP)
Integral Quadrática Simples (SSI) Histograma de EMG (HEMG) Mediana da Frequência (FMD) Frequência Média (FMN)
Mediana da Frequência Modificada (MFMD) Domínio da Frequência Frequência Média Modificada (MFMN)
Relação de frequência (FR)
Densidade espectral de potência (PSD) Potência Média (PMED)
Potência máxima (PMAX) Discriminante Bi-espectral (DBS) Coeficientes Auto-Regressivos (AR) Short Time Fourier Transform (STFT) Domínio Tempo-Frequência Transformada Wavelet (WT)
Transformada Wavelet Packet (WPT) Outros Filtragem Espacial (FS)
Dimensão Fractal (DF)