Fatigue detection in EMG signals - ULisboa · Neste estudo foram analisadas 3 provas de remo...

Fatigue detection in EMG signals

Ines Flores Mendes de Freitas

Dissertation for the Graduation in the Master Degree in

Electrical and Computer Engineering

Jury

President: Prof. Jose Bioucas Dias

Advisor: Prof. Ana Luisa Nobre Fred

Co-advisor: Prof. Antonio Veloso

Vowel: Prof. Agostinho Rosa

November, 2008

Acknowledgements

At first, I would like to thank Professor Ana Fred, for being my advisor, for giving me the opportunity of

studying this theme and for all the knowledge that I achieved. I would also want to thank Professor Antonio

Veloso for being my co-advisor.

I would also like to thank my friends for all the support and motivation, specially Luıs Gomes, Joao

Nobre, Joao Roque, Luıs Roque, Filipe Leonardo and Bruno Lopes, that through the last five years were

friends and co-workers, helping me to overcome the difficulties.

I want to thank my friends from Tomar, that were with me not only in the last five years but through all

my school years, giving support and friendship.

Finally, I want to thank my family, in particular my father and his wife, Lurdes, for the unconditional

support and constant dedication, not only in this five years but always, and for providing the means to do

this thesis.

1

Resumo

A fadiga muscular em atletas e uma das principais causas de lesoes, que normalmente so e detectada

depois de o musculo ja estar lesionado. E importante conseguir detectar a fadiga muscular antes de

ser visıvel exteriormente, nao so para prevenir futuras lesoes como tambem para se poder melhorar o

desempenho dos atletas.

O principal objectivo desta tese e detectar e caracterizar a fadiga muscular em atletas a fazer remo.

Os sinais em estudo sao os impulsos electricos provenientes do musculo (electromiografia). A analise

destes sinais serve de base para avaliar situacoes de fadiga. Propoe-se nesta tese a analise da dinamica

da evolucao da potencia do sinal ao longo de cada remada.

Neste estudo foram analisadas 3 provas de remo diferentes. Usando tecnicas de reconhecimento de

padroes classificaram-se as remadas de acordo com a prova a que pertenciam. Com base na metodologia

e algoritmos propostos foi possıvel determinar que as provas tem um padrao temporal tıpico e que esse

padrao e alterado pela fadiga muscular.

Usando tecnicas de reconhecimento de padroes, uma outra abordagem foi feita com vista a avaliar se

os padroes de movimento de remada podem ser usados como modalidade biometrica para autenticacao

ou identificacao humana. Tendo em conta a possibilidade de haver um padrao de remada tıpico de cada

atleta, foram classificadas as remadas segundo o atleta que as realizou.

Os resultados obtidos permitiram-nos verificar a presenca de fadiga muscular nos atletas depois de

realizado esforco, tendo sido identificadas as caracterısticas do sinal afectadas pela fadiga. Foi ainda

possıvel classificar as remadas tanto de acordo com as provas realizadas como de acordo com a identi-

dade do atleta que a realizou.

Palavras-chave: electromiografia, fadiga muscular, reconhecimento de padroes, processamento de sinal

3

Abstract

Muscular fatigue in athletes is one of the major causes of injuries, which is normally detected after the

muscle is already injured. It is important to detect muscular fatigue before it is visible, not only to prevent

future injuries but also to improve athletes’ performance.

The main objective of this thesis is to detect and characterize muscular fatigue in athletes perform-

ing rowing. The signals under study are electrical impulses produced by the muscle (electromyography).

Analyzing these signals allows us to evaluate if fatigue is present in athletes. In this work we propose a

temporal analysis of the evolution of the mean power across each cycle of the activity.

In this study we addressed 3 different tasks. Using pattern recognition techniques each active zone

(cycle of the activity) was classified according to its task. With our proposed methodologies and algorithms

it was possible to determine that the tasks have a typical temporal pattern which is altered by fatigue.

Also using pattern recognition techniques, we addressed the problem of human identification based on

these signals. Being aware that there could exist a pattern of the active zone for each athlete, the active

zones were classified according to the individual that performed it.

The results achieved allowed us to identify features that lead to the detection of fatigue in athletes.

Furthermore, according to the proposed methodologies applied to the analysis of active zones, we were

able to classify them according to the tasks, and also according to the identity of the performing athlete.

Keywords: electromyography, muscular fatigue, pattern recognition, signal processing

5

Contents

1 Introduction 13

1.1 Activity and muscles involved in the study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.2 EMG analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.3 Organization of the thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 Overview of EMG analysis 17

2.1 EMG . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

2.2 Survey of technologies: frequency-based analysis . . . . . . . . . . . . . . . . . . . . . . . 18

2.3 EMG and fatigue . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

3 Proposed methodology 21

3.1 EMG and Force signals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.2 Pre-processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.3 Signal segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.4 Signal Processing & Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.4.1 Global Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.4.2 Dynamic temporal evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

4 Results and Discussion 37

4.1 Experimental Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

7

4.2 Data selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.3 Pre-processing results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

4.4 Fatigue Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

4.4.1 Global features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

4.4.2 Dynamic temporal evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4.5 Identifying Individuals and Tasks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46

4.5.1 Tasks classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.5.2 Individuals classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

5 Conclusions 55

Appendices 59

A Mean Active Zones 59

B Tasks identification 63

C Individuals Classification - Confusion Matrix 75

8

List of Tables

4.1 Codification for the muscles. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.2 Number of active zones identified by channel, for all individuals and tasks. . . . . . . . . . . 39

4.3 Scenario A1 with 1 feature (AA), global matrix. . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.4 Scenario A1 with 2 features (AA and tch), global matrix. . . . . . . . . . . . . . . . . . . . . 48

4.5 Scenario A1 with 3 features (AA, tch and dact), global matrix. . . . . . . . . . . . . . . . . . 48

4.6 Scenario A2 with 1 feature (AA), global matrix. . . . . . . . . . . . . . . . . . . . . . . . . . 49

4.7 Scenario A2 with 2 features (AA and tch), global matrix. . . . . . . . . . . . . . . . . . . . . 49

4.8 Scenario A2 with 3 features (AA, tch and dact), global matrix. . . . . . . . . . . . . . . . . . 49

4.9 Scenario B with 1 feature (AB), global matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . 50

4.10 Scenario B with 2 features (AB and tch), global matrix. . . . . . . . . . . . . . . . . . . . . . 50

4.11 Scenario C with 1 feature (AC), global matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4.12 Scenario D with 1 feature (AD), global matrix. . . . . . . . . . . . . . . . . . . . . . . . . . . 51

B.1 Scenario A1 with 1 feature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

B.2 Scenario A1 with 2 features. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65


B.4 Scenario A2 with 1 feature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67



9

B.7 Scenario B with 1 feature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

B.8 Scenario B with 2 features. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71

B.9 Scenario C with 1 feature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

B.10 Scenario D with 1 feature. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73

C.1 Feature: Active zone duration. Task: PRE2000 . . . . . . . . . . . . . . . . . . . . . . . . . 76

C.2 Feature: Active zone duration. Task: POS2000 . . . . . . . . . . . . . . . . . . . . . . . . . 76

C.3 Feature: Active zone duration. Task: 2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76

C.4 Feature: Active zone duration. Task: all . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

C.5 Feature: Time of fall. Task: PRE2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

C.6 Feature: Time of fall. Task: POS2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

C.7 Feature: Time of fall. Task: 2000 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

C.8 Feature: Time of fall. Task: all . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78

C.9 Feature: Active zone duration and time of fall. Task: PRE2000 . . . . . . . . . . . . . . . . 78

C.10 Feature: Active zone duration and time of fall. Task: POS2000 . . . . . . . . . . . . . . . . 79

C.11 Feature: Active zone duration and time of fall. Task: 2000 . . . . . . . . . . . . . . . . . . . 79

C.12 Feature: Active zone duration and time of fall. Task: all . . . . . . . . . . . . . . . . . . . . . 79

10

List of Figures

1.1 Location of the muscles where the signals were achieved. . . . . . . . . . . . . . . . . . . . 14

1.2 Example of the frequency spectrum of an EMG signal. . . . . . . . . . . . . . . . . . . . . . 15

3.1 Schematic description of the steps of this work. . . . . . . . . . . . . . . . . . . . . . . . . . 22

3.2 Example of an EMG and force signals, as well as the active zone and cycle respectively,

associated with a rowing activity. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.3 Example of two signals with anomalous peaks. . . . . . . . . . . . . . . . . . . . . . . . . . 23

3.4 Fluxogram of the two algorithms used to eliminate the noise. . . . . . . . . . . . . . . . . . 24

3.5 Histogram of the first derivative of the EMG signal. . . . . . . . . . . . . . . . . . . . . . . . 24

3.6 Signals obtained using the two threshold. Blue - First Method. Red - Second Method. . . . 25

3.7 Example of two signals where anomalous peaks were eliminated, by applying the First

Method to the data in figure 3.2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

3.8 Force signal, first and second derivative with the beginning a cycle. . . . . . . . . . . . . . . 27

3.9 Result of the application of the filter. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

3.10 Example of two cases where the algorithm to separate active zones doesn’t work. . . . . . 29

3.11 Spectrogram of an active zone. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

3.12 Representation of the power spectrum at an instant of time for an active zone. . . . . . . . . 31

3.13 Evolution of the mean power along the time, Pmean(ti), in three active zones, for individual

1, channel 1. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.14 Example of the area AA for task POS2000 (green). . . . . . . . . . . . . . . . . . . . . . . . 34

11

3.15 Example of the area AB for the task POS2000 (green). . . . . . . . . . . . . . . . . . . . . 34

3.16 Example of the area AC for the task POS2000 (green). . . . . . . . . . . . . . . . . . . . . 35

3.17 Example of the alignment of the active zones for one individual performing one task. . . . . 35

4.1 Example of some signals that can not be used. . . . . . . . . . . . . . . . . . . . . . . . . . 38

4.2 Evolution of the mean power across active zones for task PRE2000. . . . . . . . . . . . . . 40

4.3 Evolution of the mean power across active zones for individual 1 and task PRE2000. . . . . 41

4.4 Mean power for each individual for tasks PRE2000 and POS2000. . . . . . . . . . . . . . . 42

4.5 Evolution of the dominant frequency across active zones for task PRE2000. . . . . . . . . . 43

4.6 Dominant frequency by instant of time in three active zones. . . . . . . . . . . . . . . . . . . 43

4.7 Mean power by instant of time in three active zones. . . . . . . . . . . . . . . . . . . . . . . 44

4.8 Mean active zones for individual 1, channel 1. Blue - PRE2000. Red - 2000. Green -

POS2000. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45

4.9 The difference between the areas PRE2000 and POS2000 versus the difference between

the change instants PRE2000 and POS2000 . . . . . . . . . . . . . . . . . . . . . . . . . . 46

A.1 Mean active zones of all individuals and tasks for channel 1. . . . . . . . . . . . . . . . . . . 60

A.1 Mean active zones of all individuals and tasks for channel 1. . . . . . . . . . . . . . . . . . . 61

12

Chapter 1

Introduction

This chapter presents a brief overview of the thesis. The objectives and motivations of this work are

presented as well as the structure of this document.

Fatigue in athletes is, sometimes the cause of injuries. The detection of fatigue is done when the

athlete is already injured. In this thesis, the goal is to detect fatigue in athletes based on the analysis of

EMG (Electromyography) signals. The EMG signals, resulting from impulse-like electrical muscle activity,

are acquired through surface electrodes.

1.1 Activity and muscles involved in the study

In this work we focus on the rowing activity. It was chosen because it is a very complete activity and the

results obtained with it can be extended to other activities. The signals used in this work were provided by

the ”Faculdade de Motricidade Humana”, where the acquisition of the signals took place.

Rowing activity is a competition sport in which athletes have to do 2000 meters in the least time possible.

Since this sport is very complete and demanding, it requires a high level of muscular force and resistance,

all the big muscles being involved in this activity. Rowing requires the use of a repetitive force in which

each muscular group is used to do the action.

In this case each athlete performed 3 tasks, using the Concept2, an indoor rower that simulates the

resistance of the water. The principal task is a 2000 meters proof that normally takes approximately 6-8

minutes [14]. In this task athletes do their best at a maximum effort. Before and after the main task the

rowers perform 10 cycles at maximum power. The three tasks performed by the individuals are mentioned

13

as PRE2000 (10 cycles before the 2000 meters task), 2000 (the 2000 meters task) and POS2000 (10

cycles after the 2000 meters task).

Tasks PRE2000 and POS2000 are useful to compare how the athletes are before and after the main

task. Since these two tasks are very similar in terms of technique, they can be compared to see the

changes in the EMG signal due to fatigue. The more relevant analysis is done with the signals achieved

with these two tasks, because they are similar to each other and they are comparable. During the 2000

meters task athletes have a different behavior from the other two ones because it is a longer and more

demanding task.

To acquire the signal, surface electrodes were used in the main muscles involved in the action of rowing.

The muscles selected to this study are Posterior Deltoid (PD), Vastus Lateralis (VL), Biceps Femoris (BF)

and Biceps Brachii (BB). In figure 1.1 it is shown the location of these muscles in an athlete performing

rowing.

Figure 1.1: Location of the muscles where the signals were achieved.

The Deltoid is a muscle forming the round contour of the shoulder, the posterior deltoid is the primary

shoulder hyperextensor. The Vastus Lateralis is situated on the thigh and it extends and stabilizes the

knee. The Biceps Femoris (Hamstrings) is a muscle of the posterior thigh and it allows knee flexion. At

last the Biceps Brachii is a muscle located on the upper arm, and the principal function of this muscle is to

allow the flexion of the elbow and to rotate the forearm.

1.2 EMG analysis

Typically the analysis of fatigue using EMG signals is performed with the frequency spectrum computed

with Fourier Transforms. The EMG signal can have frequencies near 500Hz, so typically the sampling

14

frequency used is 1000Hz. An example of the frequency spectrum achieved with this method is presented

in figure 1.2.

(a) EMG signal. (b) Frequency spectrum.

Figure 1.2: Example of the frequency spectrum of an EMG signal.

In this thesis we propose an analysis of the dynamic of the pattern of the power evolution across

time/active zone, corresponding to a cycle of the EMG signal. This work has several levels, which are

presented next:

• first the data was structured and the signals were kept in a data structure to facilitate the access to

the signals,

• as the signals had noise, it was made a pre-processing of data to eliminate the noise in signals,

• then the signals were segmented and the segments were analyzed,

• the analysis conducted to the extraction of features from the signal and,

• those features were used to classify the segments of signal.

The previous approach is used, with variants, to analyze and characterize fatigue, as well as to identify

the tasks and to recognize the identity.

1.3 Organization of the thesis

This thesis is composed of 6 chapters, including the present one.

15

In Chapter 2, one presents an overview of all the topics addressed in this project. First it is explained

what is the EMG signal, how the acquisition is usually made and the fields where this kind of signal is

used. Next it is mentioned the technologies involved in the processing of the EMG signal, and the methods

already used. In the last section it is described the relation between the EMG signal and fatigue. The work

already performed in this field and how the EMG signal changes with fatigue.

Chapter 3 describes the proposed methodology and algorithms used in this work. First it is explained

what kind of signals are involved, and the pre-processing done to de-noise those signals. Next it is ex-

plained how the signals were divided into cycles for further analysis. After the segmentation of the signals

it is explained the method used to process the cycles obtained. First the parameters used to study the

signal are defined, and then the scenarios used to identify the tasks. Finally the method used to identify

the individuals on the EMG signal is presented.

Chapter 4 describes the experimental conditions and results. It is described how the acquisition of

the signal was made and the population under study. It is also referred which are the muscles involved. In

another section it is shown the results obtained after the pre-processing of the signal and the data selection

for this experiment. The last section shows the results obtained with the division of the signal into cycles.

In Chapter 5 it is made a discussion of the results achieved with the methodologies proposed. First

are shown the results obtained with the fatigue analysis and the conclusions achieved with the methods

proposed. Next are presented the results achieved with the identification of tasks and of individuals, and it

is discussed which is the best scenario to identify the tasks and the individuals.

Chapter 6 presents the conclusions and future work.

16

Chapter 2

Overview of EMG analysis

This chapter gives an overview of signal processing issues involved in the EMG analysis. It provides

a description of the EMG signal as well as the fields where the EMG is used. Next it is explained the

technologies normally used to do the frequency analysis of EMG signals. At the end of the chapter it is

explained the relation between the EMG signal and the fatigue, ie, previous work in this area and the results

achieved so far.

2.1 EMG

The electromyographic signal (EMG) is a electrophysiological signal that measures electrical currents gen-

erated in muscles during its contraction [18]. Invasive and non-invasive methods have been used to

acquire the electrical impulses generated by muscles. In the invasive method a needle is inserted directly

into the muscle through the skin. The non-invasive way is recorded with electrodes on the skin surface.

Needle EMG recording provides a more exact representation and finer resolution of the electrical activity

of the muscle fibers than that possible with surface EMG. However, surface EMG signals can still be used

to extract a useable representation of muscle status. These two methods complement each other and

sometimes they are not alternative. They provide different views and different information of the muscle

status.

The surface EMG signal is achieving more importance in several fields, being used for the diagnosis

of muscle or nerve disorders, and for the analysis of the neuromuscular system [19]. Studying EMG

signals can help controlling prosthesis because each signal, generated by muscles performing different

tasks, contains information about the direction of movements, and speed of action. As the amplitude of

17

EMG signals is directly correlated with the force generated by the muscle, it is possible to know how to

control the prosthesis [1]. The ones that are commercially available, according to [5], don’t have the same

flexibility as the organ that they replace, so the study of electromyography can improve the development of

prosthesis. EMG can also be used to detect muscle fatigue, but this issue is going to be further addressed

in section 2.3.

2.2 Survey of technologies: frequency-based analysis

The EMG signal is a complex one which is affected by the anatomical and physiological properties of the

muscle under study. There is a random variation in the effectiveness of the body tissues in conducting the

signal from the muscle to the surface electrodes, therefore the EMG signal has a nondeterministic nature.

It is important to study both slow and fast changes on the signal because they contain different information.

The slow variations provide information related to the body movements and tissue properties. The fast

variations of the signal are useful for understanding the muscle activity [10].

The most common methods used to determine the frequency spectrum of EMG signals are the fast

and short-time Fourier transforms (FFT and STFT). The problem with these transformations is that they

assume that the signal is stationary [9]. Other tools, used to provide basic knowledge of these signals, are

rectification of the signal, ”integration” of the signal, and zero crossing count. Rectification of the signal is

a convenient tool to measure the ” relative strength” of the signal and ”relative strength of contraction” of

the muscle. Using the short-time Fourier transform with a relatively short time window we can obtain the

spectral variation with time but does not adopt an optimal time or frequency resolution for the non-stationary

signal. Parametric identification methods can also be used (AR or ARMA models) [16].

None of the previously mentioned techniques offer analysis that will take both time and frequency

variation into account in the optimal sense. The wavelet transform (WT) is an alternative to other time

frequency representations with the advantage of being linear, capable of a multi-resolution representation

and giving a good local analysis of non-stationary signals [11]. One of the main properties of the wavelet

transform is that it can be implemented by means of a discrete time filter bank.

The WT decomposes the signal into several multi-resolution components according to a basis function

called the ”wavelet function”. This method provides a tool to detect and characterize a short time compo-

nent within a non-stationary signal. It is a technique that provides information related to the time-frequency

variation of the signal and it is an extremely flexible approach to signal decomposition [10]. The WT is

classified into discrete wavelet transform (DWT) and continuous wavelet transform (CWT). CWT operates

over every possible scale and translation and it is mostly used for data compression while DWT uses a

18

specific subset of scale and translation values and it is used for signal analysis.

2.3 EMG and fatigue

The EMG signal can also be used to detect muscle fatigue. According to [12] the issue of fatigue is

complex due to the various physiological and psychological phenomena which contribute to it, and which

demonstrate it. The ways of measuring fatigue are subjective because they rely on the cooperation of the

individual. Some authors refer fatigue as the point at which an individual can no more perform a specific

task. A more exact definition of fatigue is a condition when the ability of the muscle to contract and produce

force is reduced.

It is very important to detect this condition because early detection can avoid irreversible injuries to

the subject [10]. The muscle fatigue is considered the incapacity to maintain the desirable level of force

performing a specific task. Commonly, the force output of a muscle has been used as a indicator of muscle

fatigue, ie, the point at which the contraction can no longer be maintained has been used as the point at

which the muscle is fatigued. The problem with this approach, it is that fatigue is detected when it had

already occurred and, in general, we want to detect the failure before it occurs in order to prevent it to

happen [13].

Using the EMG signal, to detect muscle fatigue, the parameters normally used are the amplitude and

the frequency of the signal. The studies in the literature show that the amplitude of EMG signals increases

progressively as a function of time when the fatigue increases [16]. According to [16] muscle fatigue can

be monitored by changes in the EMG frequency properties such as mean and median frequency.

In this thesis, we propose to analyze the EMG signal by using the signal divided into cycles. The

analysis proposed is based on the evolution of the mean power across each cycle, which is used to detect

fatigue and to identify individuals. Fatigue is detected by the classification of each cycle according to the

task involved and the individuals are identified by classifying each cycle according to the individuals.

19

Chapter 3

Proposed methodology

This chapter presents the proposed methodology to analyze the EMG signal. It also refers all the proce-

dures done with the signal to do the pre-processing, the processing and the analysis of the EMG signal.

The diagram in figure 3.1 gives a schematic description on the steps of this work. First it is done the

denoising of the signal in order to eliminate anomalous spikes. Then, both EMG and force signals are

divided into segments; for the force signal those segments are called cycles while for the EMG signal are

called active zones. The cycles of the force signal are used to help dividing the EMG signal into active

zones. Then it is computed the spectrogram of each active zone, and it is achieved the mean power

evolution of each active zone. Based on the mean power evolution it is performed the task identification

(to analyze fatigue in individuals) and the human identification (to identify the individuals according to the

mean power evolution of each active zone). The mean power is used because it has a typical temporal

pattern for all the individuals.

21

Figure 3.1: Schematic description of the steps of this work.

3.1 EMG and Force signals

Two types of signals were recorded and used simultaneously: the EMG signal and the force signal. As the

typology of these two signals is different, the procedures to process them are also different. In figure 3.2

are shown examples of the two signals.

Due to the repetitive nature of the activity, both the EMG and the force signals can be divided into

pieces or segments. In the case of the force signal those pieces are the cycles of the activity. As can be

seen in figure 3.2(b), this is a periodic signal which is divided into cycles. A cycle can be defined as a set

of movements which are repeated during the activity. The EMG signal can also be divided into parts which

are called active zones. Active zones correspond to the period in which the muscle is active. Each active

zone from the EMG signal correspond to a part of the corresponding cycle in the force signal.

22

(a) EMG signal. (b) Force signal.

(c) EMG active zone. (d) Force cycle.

Figure 3.2: Example of an EMG and force signals, as well as the active zone and cycle respectively,associated with a rowing activity.

3.2 Pre-processing

Some of the EMG signals achieved had spikes, considered as noise. The noise of those signals was the

result of problems in the acquisition. An example of these signals with artifacts is presented in figure 3.3.

(a) (b)

Figure 3.3: Example of two signals with anomalous peaks.

23

As can be seen, these spikes do not form part of the signal and may hamper the correct processing

of the signal. To solve this problem two algorithms were created. The goal of them is to choose the best

threshold to eliminate noisy data, ie, if the derivative of the signal, at each point, is greater than the chosen

threshold, the value of the signal in that point is put zero. Thus the anomalous spikes are eliminated and

one has a clearer signal which is better to analyze. In figure 3.4 can be seen a fluxogram of the two

algorithms.

Figure 3.4: Fluxogram of the two algorithms used to eliminate the noise.

The two algorithms are based on the histogram of the first derivative of the signal. The first one uses

the cumulative sum of the histogram to calculate the threshold. The threshold corresponds to the value at

which the cumulative sum of the histogram reaches 99% of the total samples of the signal. An example of

the result of the histogram is shown in figure 3.5.

Figure 3.5: Histogram of the first derivative of the EMG signal.

As can be seen in figure 3.5 the major concentration of points is located in lower values. As the

derivative of the signal is the difference between two consecutive points just when we have an anomalous

peak the derivative takes higher values. So the threshold was chosen based on the value associated with

99% of the samples.

In the second method is used the derivative of the histogram (mentioned before). As can be seen in

24

figure 3.5 the histogram has a period where the change is more abrupt, so this was the point used as

threshold. To achieve that point it was calculated the maximum of the derivative of the histogram.

The thresholds achieved with those two algorithms are used to eliminate the spikes. When the derivative

of the signal is greater than the threshold it is put zero for the value of the signal at that point. This way we

have two different results for the two thresholds.

As two algorithms were created one has two different results with the application of these two methods.

In order to choose the best method, and as is not obvious to see which one is the best one, some plots

were made to see which threshold is going to be used. The figure 3.6 shows, as an example, that the best

method is the first one.

Figure 3.6: Signals obtained using the two threshold. Blue - First Method. Red - Second Method.

In figure 3.6 the blue signal represents the one which was processed by the First Method and the red

one by the Second Method. As can be seen the Second Method that is shown in red eliminates relevant

data besides the noisy data. So the method chosen to be used was the Frist Method because it eliminates

the noise but it does not eliminate important data. An example of the application of this method is shown

next.

The two signals shown in figure 3.7 are the same ones as in the figure 3.3 but without the noise. As

can be seen the algorithm works well for this kind of signals. It clears the signals maintaining its nature.

This way it was used in all the signals for all the individuals. All the work done after was using the signals

achieved after the denoising.

25

(a) (b)

Figure 3.7: Example of two signals where anomalous peaks were eliminated, by applying the First Methodto the data in figure 3.2

3.3 Signal segmentation

The previous work allowed the signals to be prepared to be processed. The EMG and the force signals

were divided into active zones and cycles, respectively . In this work it is expected to analyze each active

zone separately and then compare the results achieved.

The force signal has a more constant and clear behavior. All the force signals are clear and well defined.

For these reasons it should be the base of all the study. From it can properly be obtained the beginning,

end, and duration of each cycle. And it can also be obtained the amplitude and power of the force realized

by each athlete.

The force signal was divided into cycles based on signal’s derivatives:

• In figure 3.8(a) it is shown the force signal, where it can be seen an example of the beginning of a

cycle;

• compute the first derivative of the force signal shown in figure 3.8(b);

• compute the second derivative of the force signal shown in figure 3.8(c);

• the beginning of each cycle is obtained by the intersection of a maximum of the second derivative,

shown in figure 3.8(c), with a positive value of the first derivative, shown in figure 3.8(b);

• the end of each cycle is considered the value of the next beginning;

• the duration of each cycle is considered the difference between two consecutive beginnings, shown

in figure 3.8(d);

26

• the amplitude corresponds to the maximum value of each cycle, shown in figure 3.8(d);

• the power equals the area under the curve of each cycle, shown in figure 3.8(d).

Figure 3.8: Force signal, first and second derivative with the beginning a cycle.

A similar work was done with the EMG signal, but since this is a more complex one, the methods were

different. Although the anomalous spikes were eliminated, the EMG signal is not so clean and so well

behaved as the force one, so the previous procedure failed to divided it into active zones. The methods

used are explained next.

To divide the EMG signal into active zones it was necessary to detect the envelope of the signal. For

that reason the signals were rectified and filtered. The filter used was a band-pass filter, the value for the

pass-band ripple used was 1.5dB and the value for the stopband attenuation was 2dB. The values for the

cut-off frequency were calculated from the equation 3.1 and the values for the stopband corner frequencies

from the equation 3.2. Those parameters were tuned based on the observation of the results.

Wp = [0.005 4]/(N/2) ∗ π (3.1)

Ws = [0.001 5]/(N/2) ∗ π (3.2)

27

In figure 3.9 is presented an example of the result of the rectification and filtering of the signal. After

filtering the signal is easier to process because the higher frequencies are eliminated.

Figure 3.9: Result of the application of the filter.

With this method one obtains a signal with values near zero when the muscle is not making effort, and

with higher values in the other case (when the muscle is active). With the filtered signal it is possible to

detected the beginning and the end of each active zone. The method used to do this is explained in the

next pseudo-code.

Algorithm 1 Dividing the EMG signal into active zones1: Rectified signal=abs(signal)/max(signal)2: Signal filtered=BPass(Rectified Signal)3: threshold=mean value of the filtered signal4: active initialized 05: if signal filtered at each point ≥ threshold then6: active at that point = 17: else8: active = 09: end if

10: if active(i-1)==0 and active(i)==1 then11: beginning of active zone=i12: end if13: if active(i-1)==1 and active(i)==0 then14: end of active zone=i15: end if

In a few cases it is not possible to automatically separate the active zones because the filter did not

eliminate all the high frequencies An example of those signals is presented in figure 3.10. In red is shown

the pieces of the signal that were incorrectly considered as an active zone.

28

(a) (b)

Figure 3.10: Example of two cases where the algorithm to separate active zones doesn’t work.

In those examples, when the division in active zones was made, it was obtained more than 10 active

zones. To solve this problem the cycles of the force signal were used. The duration of the active zones

was compared with the duration of the corresponding cycle of the force signal. When the percentage of

the duration of the active zone by the duration of the cycle of the force signal was smaller than 0.3, two

consecutive active zones were joined. With this method it was possible to separate correctly the active

zones.

3.4 Signal Processing & Classification

The basis of all the analysis was done with the spectrogram of each active zone. An example of the

spectrogram of an active zone is presented in figure 3.11.

As can be seen in figure 3.11, the concentration of the power spectrum density is in lower frequencies,

and it decreases across time. In this graphic red means higher power values while blue means lower power

values. Let P (ti, fj) be the power spectrum density at instant ti and frequency fj for a given active zone.

Based on the spectrogram, two kinds of features were calculated: global features characterizing the global

behavior of each active zone; and features associated with the temporal evolution along each active zone.

The techniques used to calculated those features are explained next.

The analysis made in this work is based on the active zones of each muscle. The observation of the

active zones allows us to see that the first active zone sometimes is incomplete as well as the last one. So

the work done excludes these two ones to avoid compromising the results.

29

(a) (b)

Figure 3.11: Spectrogram of an active zone.

3.4.1 Global Features

Global features were calculated for each active zone. These features are important to see the evolution of

fatigue along each task. The most important parameters to analyze are the power and the frequency. It

was calculated the maximum and the mean power of each active zone and the dominant frequency of the

active zone. The dominant frequency, Fdom, is the frequency associated with the maximum power value,

as shown in equation 3.3.

Fdom = arg(maxtimaxfj

P (ti, fj)) (3.3)

The mean power of each active zone, Pmeanglobal, is calculated according to equation 3.4.

Pmeanglobal =1n2·

n∑i=1

n∑j=1

P (ti, fj) (3.4)

In the equation 3.4, n is the number of samples considered in the calculation.

3.4.2 Dynamic temporal evolution

In order to see the temporal evolution along each active zone it is important to analyze each instant of

time of each active zone. So we calculated features associated with each instant of time ti. The parame-

ters calculated were the dominant frequency (frequency associated with the maximum power) Fdomi, the

maximum power Pmaxi, the mean power Pmeani, the mode power (value that occurs more frequently),

30

the median power (number which separates the higher half of a sample) and the standard deviation of the

power.

Fdom(ti) = arg(maxfjP (ti, fj)) (3.5)

Pmax(ti) = maxfjP (ti, fj) (3.6)

Pmean(ti) =1n·

n∑j=1

P (ti, fj) (3.7)

In figure 3.12 can be seen, for an active zone and a particular instance of time, the maximum power,

mean power and dominant frequency identified.

Figure 3.12: Representation of the power spectrum at an instant of time for an active zone.

The evolution of the mean power along each active zone presents a typical pattern, as illustrated in

figure 3.13. It consists of an initiated period with high mean power values followed by a period where the

mean power decreases abruptly, and an ending period where the mean power stays in lower values. For all

the individuals it can be seen the same pattern, what changes from individual to individual and task to task

is the instant when the mean power decreases or the duration of the active zone. For those reasons, this

was the pattern chosen to continue the analysis. This result is more relevant in the first channel, where the

graphics of the evolution of this feature have a more consistent pattern. According to this fact the proposed

work has focused on the muscle named Posterior Deltoid.

Based on the mean power signal, the mean active zone of each task, for each individual, was calculated.

The mean active zone, Pmean(ti) was calculated for each individual and the next explanation is based on

this fact. For each instant of time was calculated the value of the mean power, using the different mean

31

Figure 3.13: Evolution of the mean power along the time, Pmean(ti), in three active zones, for individual1, channel 1.

powers associated with each active zone, according to equation.

Pmean(ti) =activezones∑

l=1

Pmeanl(ti) (3.8)

The reason to calculate the mean active zone was to see if it was possible to distinguish between

individuals based on their active zones.

Based on all the information collected and the results obtained one can say that each individual could

have a different pattern of active zone and also each task could also have some different characteristics

from the others tasks. According to this purpose, analyzing each active zone, one tried to identify the

individual and the task of each active zone. The algorithm used was the k-Nearest Neighbor (7-NN), in

this specific case 7 nearest neighbors. This algorithm, classifies signals based on the closest training

examples. To classify them are used features extracted from the signals. A signal is classified by the

majority vote of its neighbors, with the signal being assigned to the class most common amongst its k

nearest neighbors (7 in this case), using the Euclidean-distance between feature vectors. In this work was

used for training signals all of them except the one being tested (leave-one-out method) [6].

This algorithm was used to classify the signals in two ways. First to see to which task correspond the

signal and then to see to which individual correspond the signal. The methodologies used for the two

classifications are presented next.

The features used in the task and individual classification were the instant of change (tch), the duration

of the active zone (dact), and the area under the curve. The value of tch was considered as the minimum

of the second derivative of the active zone in region of the maximum of the signal, and was calculated for

all the active zones concerned to the tasks PRE2000 and POS2000, for channel 1. It was also calculated

the mean value of the change instant, ¯tch=0.392, as well as the standard deviation of the same instant,

σtch=0.0967, using both PRE2000 and POS2000 signals of all individuals. The area under the curve is

32

calculated until different values as shown next by the different scenarios.

Task classification - Fatigue related analysis

As it was mentioned before, tasks PRE2000 and POS2000 are similar to each other. They both consist

in performing 10 rowing cycles at the maximum effort. Each athlete has his own way of doing those tasks

but it is expected that they perform the two tasks with the same technique. However, fatigue could change

the EMG signal and, consequently, the active zones from task PRE2000 could be different from those

achieved in task POS2000. It was to prove this fact that one tried to identify the tasks of the active zones.

The classification algorithm used was the k-NN as explained before.

Visual inspection of the active zones, show that the most significant changes, between PRE2000 and

POS2000 EMG signals segments occur in the first part, when the mean power takes higher values. So

it was the first part of the active zone which was analyzed. The algorithm k-NN was used in 4 different

scenarios, named as A, B, C and D. Those scenarios and corresponding features used are explained next.

The difference between the scenarios is just the way the features are calculated because the methods

used are the same.

Scenario A In the first scenario were used three features: the area under the curve (AA), the change

instant (tch) and the duration of the active zone (dact). The AA was calculated until the end of the first part

of the active zone. To define this instant two experiments were made, called A1 and A2.

Experiment A1 was calculating the area until ¯tch, this was the first feature used. Next the same algorithm

was used but with two features, the area mentioned before and tch of each active zone. At last another

feature was added which was dact and the algorithm was used with the three features. In figure 3.14 can be

seen an example of these features, are identified the tch for each active zone and the line in black shows

the value of ¯tch.

Experiment A2 was using the area until the mean value of the change instant minus the standard

deviation (σtch), this was the first feature used. The features used next were the same as in experiment

A1, and can be seen an example in figure 3.14, the only difference is that the line in black takes a lower

value, but it is not visually significant.

Scenario B In scenario B was used two features: the area under the curve (AB) and tch. The

difference from the scenario A is that the area is calculated differently. In this case to choose until where

the area was calculated it was compared the two correspondent active zones of each task, for example,

the second active zone of task PRE2000 was compared with the second active zone of task POS2000.

33

Figure 3.14: Example of the area AA for task POS2000 (green).

Comparing tch of the two active zones it was chosen the maximum of the two. The area was calculated

until this value minus the σtch, for the two active zones correspondent to the two tasks. In order to calculate

this area both time and power were normalized by the maximum, as can be seen in figure 3.15.

Figure 3.15: Example of the area AB for the task POS2000 (green).

In figure 3.15 we see that the curve in blue has a higher ¯tch so was that one that was used. First the

algorithm was used with just one feature which was AB . Then it was used the two features together, AB

and tch.

Scenario C In scenario C just one feature was used because the others could not be used. The

feature used was the area under the curve (AC). In this case the area was calculated until the tch of

each active zone minus σtch. In order to calculate this area both time and power were normalized by the

maximum as shown in figure 3.16.

34

Figure 3.16: Example of the area AC for the task POS2000 (green).

Scenario D In scenario D the method used was sightly different. It was calculated the correlation

index of all the signals, of the same task, for each individual, taking as reference the second active zone.

Based on the correlation indexes was calculated the shift, of each active zone, when compared to the

second active zone. With the deviation mentioned before the active zones were shifted according to the

value of the deviation. After this procedure the active zones were aligned as can be seen an example in

figure 3.17.

(a) Before alignment. (b) After alignment

Figure 3.17: Example of the alignment of the active zones for one individual performing one task.

After applying this method all the active zones were aligned, then it was calculated the mean active

zone and then tch of it. The goal of this method was to have the instant to calculate the area under each

curve. The area (AD) was calculated , for each active zone (without the alignment), until the tch mentioned

before. For this process both time and power were normalized. In scenario D was used just one feature:

AD.

35

Individual Identification

As it was mentioned before, each individual has his own way of performing the tasks, so it could be possible

to identify an individual by his active zones. The active zones acquired in the process were classified using

the k-NN algorithm to identify the individuals. Two features were used, which are the duration of the active

zone and the change instant (tch).

The algorithm was used with each feature separately and with both simultaneously. This means that 3

scenarios were used, one with dact, other with the fall instant and other with these two features together.

For the 3 scenarios was applied the algorithm using just the signals from the 3 tasks isolated and then the

signals from the 3 tasks simultaneously.

36

Chapter 4

Results and Discussion

In this chapter will be shown the results achieved with the methodology presented in chapter 3. It is also

made a discussion of the results. The work done was divided into two types of analysis: fatigue evaluation

and identification of individuals and tasks. These are presented in the sequel after a summary of the

experimental data and pre-processing results achieved.

4.1 Experimental Setup

The acquisition of the signal was made at the ”Faculdade de Motricidade Humana”, using several athletes

to perform a specific activity, in this case rowing. The population consists of 11 subjects that have been

doing rowing for several years. Six of those rowers are athletes of high competition, recognized by the

”Federacao Portuguesa de Remo”; the others were taken from two national clubs. All the athletes have

more than 5 years of experience. The fact that the athletes have several years of experience is very

important to assure that their technique is stable and the results are reliable [14]. As this study must be

anonymous, the individuals are mentioned in this work by numbers from 1 to 11.

To acquire the EMG signal were used surface electrodes in the main muscles involved in the task of

rowing: Posterior Deltoid, Vastus Lateralis, Biceps Femoris and Biceps Brachii. So it was obtained 4 EMG

signals and one force signal for each task and each individual.

Each signal generated by each muscle was recorded in a different channel, the first 4 channels are

related to each muscle as it is shown in table 4.1. The last channel is the force signal.

To identify the muscle that is being in focus it is mentioned the channel where it was recorded.

37

Table 4.1: Codification for the muscles.Channel Muscles

1 Posterior Deltoid2 Vastus Lateralis3 Biceps Femoris4 Biceps Brachii

4.2 Data selection

In an experimental environment the acquisition is not perfect, because there are several facts that are

involved in the acquisition. So it is normal that some signals had noise and others that could not be used.

As it was mentioned before in chapter 3, it was done the de-noising of the signals. After the de-noising

almost all the signals were prepared to be used but some of them were just noise.

In order to see if all the signals could be used in this work all of them were visualized. Some could not

be used because, as can be seen in figure 4.1, of acquisition problems.

(a) (b)

Figure 4.1: Example of some signals that can not be used.

In these signals the active zones are not visible probably due to a bad acquisition of the signal. If

they were used they would change the results because it is not possible to have a good process to those

signals. To avoid the distortion of the results the following signals were discarded:

• 2000, individual 1, channel 3

• PRE2000, individual 4, channel 2

• PRE2000, individual 8, channel 2

38

• POS2000, individual 11, channel 3

Besides those signals another one was corrupted from a particular point on, which was the signal

achieved in task 2000 from the individual 1, channel 1. This one was considered in all work but was

truncated to eliminate the corrupted part.

4.3 Pre-processing results

In this section it is presented an overview of the results achieved in the pre-processing. First it was used

a method to eliminate the noise of the signals, that algorithm was explained in section 3.2 as well as the

results achieved. As shown, the method cleared the signals and it was possible to correctly divide them

into active zones.

The division into active zones was explained in section 3.3. In table 4.2 is presented the number of

active zones obtained for the different individuals and tasks.

Table 4.2: Number of active zones identified by channel, for all individuals and tasks.

PRE2000 2000 POS20001 10 230 102 10 324 113 10 336 104 10 205 105 10 198 106 11 199 107 11 195 108 10 208 109 10 146 8

10 9 218 911 10 158 10

After the division into active zones it was calculated the spectrogram of all the active zones. In the next

sections are presented the results achieved with the analysis of the active zones.

39

4.4 Fatigue Analysis

4.4.1 Global features

As mentioned in chapter 2 the literature say that the amplitude of EMG signals increases progressively as

a function of time when the fatigue increases and the mean power frequency decreases as a function of

time [16]. In this study is proposed another method to identify fatigue as explained in chapter 3.

The global features, detailed in section 3.4.1, are calculated for each active zone and then compared

to see the evolution of the feature along each task. These provide a global overview of the evolution of

the variation of the power and frequency during each task. The results obtained show that there is not a

constant behavior with these features. Some results are presented in the next figures.

(a) Individual 1, channel 1 (b) Individual 2, channel 1

(c) Individual 3, channel 1 (d) Individual 4, channel 1

Figure 4.2: Evolution of the mean power across active zones for task PRE2000.

In figure 4.2 is presented the behavior of the Pmeanglobal across active zones for four individuals,

referred to channel 1 and task PRE2000. For the examples 4.2(a) and 4.2(b), which correspond to individ-

uals 1 and 2 respectively, in the firsts 3 active zones the mean power assumes higher values and then one

sees that the last active zones have lower values. This fact could have led us to conclude that probably

the mean power decreased with the evolution of the task. But then, observing the other two individuals,

40

figure 4.2(c) and 4.2(d), one see that the behavior explained before is not visible in these cases. Overall

the available data, there is no consistent pattern of mean power evolution along the tasks for the various

individuals.

It is also important to see which is the behavior of the mean power, for the same individual in the different

channels (muscles). Those results are presented in figure 4.3. In figure 4.3 it is presented the mean power

of signals from individual 1 for the four channels. These examples show that the different channels don’t

have a similar behavior in terms of the feature in study. By the observation of these figures one sees that

channel 1 has a similar behavior to channel 4. But this conclusions could not be generalized because this

is a single example and the analysis of all the individuals and channels do not present consistent relevant

results.

(a) Channel 1 (b) Channel 2

(c) Channel 3 (d) Channel 4

Figure 4.3: Evolution of the mean power across active zones for individual 1 and task PRE2000.

In figure 4.4 can be seen the mean power of each task for each individual. The blue line corresponds

to the values of task PRE2000 for each individual, while the red line corresponds to task POS2000. As it

can be seen task PRE2000, performed before, has higher values for the mean power. As the two tasks are

similar the only difference between them is the effort made between them, consequently the fatigue. If it

was used, as feature, the difference between Pmean of task PRE2000 and Pmean of task POS2000, we

would obtain an error probability of 13.6%.

41

(a) Channel 1 (b) Channel 2

(c) Channel 3 (d) Channel 4

Figure 4.4: Mean power for each individual for tasks PRE2000 and POS2000.

As was mentioned before it was also calculated the dominant frequency, Fdom, of each active zone.

An example of the results achieved with this feature is shown in figure 4.5. In this figure one shows the

evolution of the dominant frequency across active zones for four individuals for channel 1. In this case

it is even more difficult to see a significant pattern or behavior with this feature. Although this is just an

example, the observation of all the graphics obtained with this feature lead us to the same conclusions

presented before.

The more relevant and consistent results were achieved with the time related features which are going

to be presented and discussed in the next section.

42

(a) Individual 1, channel 1 (b) Individual 2, channel 1

(c) Individual 3, channel 1 (d) Individual 4, channel 1

Figure 4.5: Evolution of the dominant frequency across active zones for task PRE2000.

4.4.2 Dynamic temporal evolution

Time related features are considered the features related to each instant of time inside each active zone.

These ones are useful to see how the parameters evolve across the active zone. The more relevant results

were achieved with this analysis. Some examples, of the results achieved, can be seen in the next figures.

Figure 4.6: Dominant frequency by instant of time in three active zones.

The analysis of the graphics produced based on the dominant frequency, which can be seen in figure

4.6, are not conclusive because they don’t show a clear behavior through the active zones. It can not be

seen a clear pattern in this feature, and therefore the relation between fatigue and the dominant frequency

43

is not obvious as a dynamic pattern.

Otherwise the mean power is very useful because it can be seen a typical temporal pattern in this

feature. In figure 4.7 are shown only the mean power results because the mode and median power have

the same behavior. The work done after, was all based on the mean power.

Figure 4.7: Mean power by instant of time in three active zones.

The graphics shown in figure 4.7 are just an example, taken from 3 active zones from individual 1.

But all the graphics achieved for all the individuals exhibit the same pattern. It consists of a period of

the active zone where the mean power has higher values, followed by a period where the mean power

decreases, and at the end a period where the mean power stays in lower values. The pattern is the same

but from individual to individual and task to task some parameters change. For example a parameter that

changes from individual to individual is the duration of the active zone. Some athletes have longer active

zones than others. Another parameter that differs is the fall instant, considered the instant when the mean

power decreases, which tends to be similar for the same individual but different when comparing two or

more individuals. These two instants are very important to the work performed and they were used in the

identification of the individuals and tasks (see section 3.4.1).

As mentioned in chapter 3, the mean active zone Pmean, based on the evolution of the mean power, of

each task for each individual was calculated to compare the type of active zones that each individual have.

An example of the results obtained is presented in figure 4.8, which corresponds to the mean active

zones of tasks PRE2000, 2000 and POS2000, for individual 1, channel 1. The results for the other indi-

viduals are presented in annex A. Only the results according to channel 1 are presented in this document

because this channel has a clearer pattern so it was the focus of the study.

From the observation of figures in annex A we can draw several conclusions.

First of all one observes that, as it was referred before, the mean power has approximately the same

behavior for all the individuals and tasks. It starts with higher values, which stay approximately constant for

a while, and then the values decrease abruptly.

44

Figure 4.8: Mean active zones for individual 1, channel 1. Blue - PRE2000. Red - 2000. Green - POS2000.

The graphics show that task 2000 has a smoother curve, in all individuals, than the tasks PRE2000 and

POS2000. This fact is explained because the mean active zone is achieved with more samples in this task.

In PRE2000 and POS2000 one just has 8 samples, which are the active zones considered. Task 2000 has

more active zones, typically around 200, so consequently more samples to be considered to calculate the

mean active zone.

Another fact observed is that, almost in all cases, the mean active zone of tasks PRE2000 and POS2000

are more similar to each other than to task 2000. Tasks PRE2000 and POS2000 are performed in the same

conditions because the nature of the tasks is the same. Task 2000 doesn’t have the same nature, it is a

longer one in which athletes could have different methods to perform this task. This leads to the differences

observed between tasks PRE2000 and POS2000 and the task 2000.

Comparing the mean active zones in tasks PRE2000 and POS2000, almost all the individuals have the

curve of task PRE2000 above POS2000, a possible explanation for this is being fatigue. As the difference

between task PRE2000 and POS2000 is the effort made in task 2000, it is reasonable to say that the

differences between the signals from the two tasks are due to fatigue. Task POS2000 is the last one to be

performed so the degradation of the mean power could be explained by the athletes being fatigued. Ahead

in this thesis the results of the study to try to classify the active zones according to its task are presented.

Another fact that we tried to observe was the behavior of the areas, mentioned in section 3.4.2, and the

change instants.

To acquire the graphics shown in figure 4.9(a) it was calculated the area AA, as explained before, until

¯tch for both PRE2000 and POS2000 tasks. In the graphic is used the subtraction between AA of PRE2000

45

(a) Scenario A1 (b) Scenario A2

Figure 4.9: The difference between the areas PRE2000 and POS2000 versus the difference between thechange instants PRE2000 and POS2000

and POS2000 and is used the subtraction between tch of PRE2000 and tch of POS2000, for each active

zone. As can be seen the area of active zones from task PRE2000 are higher than from task POS2000.

The graphic 4.9(b) is acquire as the graphic 4.9(a) but the area is calculated until ¯tch minus σtch.

Despite the evidences mentioned before, which are more or less common to all individuals, they have

different types of active zones.

For example the individual 2 has longer active zones than individual 1. Individual 9 is the one that the

mean power stays less time in high values. Individual 6 seems to have the higher values of mean power

for the first part of the active zone. All of these evidences show the individuality of each athlete. Based on

this fact it was tried to identify individuals by their active zones, the results achieved are presented in the

next section.

4.5 Identifying Individuals and Tasks

In this section will be presented the results obtained with the identification of the individuals and the tasks.

Based on all the information mentioned before one tried to identify the individuals and also the tasks

based on features extracted from the time related patterns of the active zones. All the work done is based

on channel 1 because, as it was explained before, this is the channel that has a more constant behavior.

Classification results achieved with the k-Nearest Neighbor algorithm are listed in matrices called con-

fusion matrices. Each column of the matrix represents the instances in a predicted class, while each row

represents the instances in an actual class. If all the signals are well classified the elements of the diagonal

are 100% while the rest of the elements are zero.

46

4.5.1 Tasks classification

The results achieved in task classification are summarized in confusion matrices. It was calculated confu-

sion matrices for each individual, presented in annex B and a global one which include all the individuals.

The rows correspond to the real class of the signal, in this case the task that the signals correspond to.

The columns refer to the task to which the signal was classified.

Scenario A

In the first scenario were used three features: the area AA, tch and dact, in two experiments.

Experiment A1 Experiment A1, as mentioned before, was calculating the area until ¯tch, this was the

first feature used. Next was used two features, the area mentioned before and tch of each active zone. At

last was added dact.

The results achieved with the first feature, the area, are presented, for each individual table B.1 in annex

B. Next it is presented the table 4.3 with the global results, including all the individuals.

Table 4.3: Scenario A1 with 1 feature (AA), global matrix.

PRE POSPRE 57,1429 42,8571POS 45,4545 54,5455

Using just the area as feature, and analyzing the results, it is observed that the results are different

according to each individual. From the 11 individuals 6 of them have more than 50% of correct identification

of the tasks. Taking in consideration that just one feature is used the results could be considered good.

Commonly a signal is characterized by different parameters and it is the conjunction of several of them

that represents the signal. In this case this feature is enough to characterize individual 4 because it is

achieved 100% of correct identifications. It is expected that when it is joined another features the results

stays the same because the area is always used as feature. The opposite extreme is realized by individual

8 for which none of the active zones are correctly identified. The results presented in table 4.3 composed

averaging of all the individuals. It can be seen that the percentage of a correct identification is about 55%

which means that more than 50% of the tasks are well identified as PRE2000 or POS2000. This means

that the area contains discriminating power about the active zones, but it should be used in conjunction

with other ones in order to obtain better performance.

47

The results achieved for each individual, using as features the area and the tch, are presented in annex

B in table B.2. The global results for all the individuals are presented in table 4.4

Table 4.4: Scenario A1 with 2 features (AA and tch), global matrix.

PRE POSPRE 74,0260 25,9740POS 29,8701 70,1299

In this case the results, in general, are better than using just the area as feature. As it was referred the

individual 4 stays with 100% of correct identification. Focusing the attention on individual 8, one sees that,

adding tch as feature, the results change and the actives zones are all now correctly identified. According

to these results the active zones of individual 8 are well defined by the instant of change. Comparing the

global matrix, in table 4.4, with the one in table 4.3, one concludes that the results are improved. The

percentage of a correct task identification goes from approximately 55% to above 70%. This means that

the two features are important to characterize the active zones, and differentiate them according to the

tasks.

Finally, another feature was added which was dact. In annex B, table B.3 are presented the results

for each individual using the three features and in table 4.5 are presented the average results for all the

individuals.

Table 4.5: Scenario A1 with 3 features (AA, tch and dact), global matrix.

PRE POSPRE 94,8052 5,1948POS 5,1948 94,8052

Analyzing the results achieved, using one, two and three features, one sees that, with the use of the

three features the results are better. In this case the majority of the tasks are well classified, one have

approximately 95% of tasks well classified. The three features together is the best way to have a good

identification of the tasks of the active zones in this experiment.

Experiment A2 Experiment A2, as referred before, was using the area until ¯tch minus σtch, this was

the first feature used. The features used next were the same as in experiment A1.

The results achieved, in this experiment, using the area as feature, for each individual are presented in

table B.4. The global results are shown next in table 4.6.

Comparing experiment A1 and A2, using just one feature, the results are not conclusive in which is the

48

Table 4.6: Scenario A2 with 1 feature (AA), global matrix.

PRE POSPRE 55,8442 44,1558POS 44,1558 55,8442

best choice. With experiment A1 the results are better for some individuals and with experiment A2 are

better for other individuals. For example, individual 4, who had the best results in experiment A1, in exper-

iment A2 the results are approximately 28%, which is a poor result. On the other hand, individual 8, who

had the worst results in experiment A1, in experiment A2 has approximately 57% of correct classification.

Comparing the global matrices, the results are very similar and it is not possible to get a good conclusion,

but as it was mentioned before one feature, usually is not enough to characterize a signal.

The features used next were the same as in experiment A1. So in tables B.5 and B.6 are presented the

results for each individual for 2 features and 3 features, respectively. In tables 4.7 and 4.8 are presented

the global results for the same cases.

Table 4.7: Scenario A2 with 2 features (AA and tch), global matrix.

PRE POSPRE 75,3247 24,6753POS 23,3766 76,6234

Table 4.8: Scenario A2 with 3 features (AA, tch and dact), global matrix.

PRE POSPRE 94,8052 5,1948POS 3,8961 96,1039

As it was expected the results are improved by the use of more than one feature. It is interesting to see

that with this experiment the global results with two and three features are sightly better than in experiment

A2. But it can not be considered a significant difference so it can not be said that on experiment is better

than the other. In some individuals is best with one experiment and for others with the other experiment.

Scenario B

In scenario B were used two features: the area AB and tch.

Using just the area as feature, the results achieved are those presented in table B.7 for each individual.

49

In table 4.9 are presented the overall mean results. Next, the classification algorithm was used with the

two features, the area and the change time instant. The results for each individual are presented in table

B.8 and the global results are presented in table 4.10.

Table 4.9: Scenario B with 1 feature (AB), global matrix.

PRE POSPRE 58,4416 41,5584POS 41,5584 58,4416

Table 4.10: Scenario B with 2 features (AB and tch), global matrix.

PRE POSPRE 72,7273 27,2727POS 27,2727 72,7273

Analyzing table B.7 one see that 6 in 11 individuals have more than 50% of the active zones well

classified. Analyzing specific cases, individual 1 has approximately 85% of correct classification which is

a very good result. Also individuals 4, 5 and 9 have really good recognition ratio. For those individuals the

area as feature, calculated according to this scenario, is almost enough to classify the tasks. The global

matrix (see figure 4.9) shows that all the individuals together have approximately 58% of a correct task

classification.

With the use of two features the results, in general, get better. The percentages of active zones correctly

classified are higher than with just one feature. This fact is also observed in the global matrix, figure 4.10,

where the percentage of a correct classification is approximately 72%, showing that those two feature are

indicated to identify the tasks.

Scenario C

In scenario C was used one feature: the area under the curve, as explained in section 3.4.2.

The results of scenario C with just one feature (area) are presented in table B.9 for each individual. The

global results for all the individuals are presented in table 4.11.

Analyzing the table one see that it were achieved good results just with one feature. For example

individual 1, 5 and 11 have approximately 90% of good identification. The global matrix shows that 61%

of the active zones are correctly classified, which comparing with the use of just one feature in the other

scenarios, is a good result. In this case it was not possible to use more than one feature but as mentioned

50

Table 4.11: Scenario C with 1 feature (AC), global matrix.

PRE POSPRE 61,0390 38,9610POS 38,9610 61,0390

before the better results, with this kind of algorithm, are achieved with more than one feature.

Scenario D

In scenario D was used just one feature: the area. The results achieved for each individual are presented

in table B.10 and the global results are presented next in table 4.12.

Table 4.12: Scenario D with 1 feature (AD), global matrix.

PRE POSPRE 79,2208 20,7792POS 27,2727 72,7273

With just one feature, one see that the results are very good, compared with the other scenarios. It can

be seen that 9 in 11 individuals have more than 50% of a correct classification. And 4 of them approximately

90%. Those results are very good, knowing that just one feature was used. The global matrix shows that

active zones have about 75% of being identified with the correct task. This could mean that the active

zones of the same task could have a similar pattern but shifted.

Comparison of the 4 scenarios

All the 4 scenarios were attempts to reach the best results so it is important to compare them. As it

was mentioned before the percentages achieved with just one feature are not so good as using more

features. Comparing the use of just one feature in the 4 scenarios, scenario D is the best one because the

percentage of a good classification is very high. Even compared with the other scenarios but using two

features the results of scenario D are similar. So it can be said that scenario D is the best one. This result

let us to conclude that, probably, the different active zones from the same task are shifted as time evolves.

A reason for that shift could be fatigue because the pattern of the individual is the same but due to fatigue

the actives zones are not equal to each other.

With the two features the results are improved, in some cases significatively but it is with the three

51

features that one has the best results and almost all the active zones are correctly classified. This fact

proves that all the features are relevant for this study and the combination of the three identifies the task in

focus.

Another result that can be discussed is the fact that active zones from PRE2000 and POS2000 are

different. Theoretically PRE2000 and POS2000 are two similar tasks in which athletes should produce the

same results. But in fact the results are different, probably due to fatigue. Although the tasks are equal, the

athletes are not in the same conditions. When they perform POS2000 they had already done another task,

the 2000, so fatigue can explain the difference in the active zones from task PRE2000 from POS2000.

4.5.2 Individuals classification

The active zones acquired in the process were classified using the k-NN algorithm to identify the individu-

als. Several scenarios were considered and two features were used, which are the duration of the active

zone and the time of fall, both mentioned before.

The algorithm was used with each feature separately and with both simultaneously. This means that 3

scenarios were used, one with the duration of the active zones, other with the time of fall and another with

these two features together. For the 3 scenarios was applied the algorithm using just the signals from the

3 tasks isolated and then the signals from the 3 tasks simultaneously.

The results, achieved with these analysis, were summarized in several confusion matrices which are

shown in annex C. The rows correspond to the real class of the signal, in this case the individual that the

signals correspond to. The columns refer to the individual to which the signal was classified.

The first 4 tables presented in annex C are related to the duration of the active zone. Comparing the

tables related to the separated tasks one can see that the task which leads to the worst results is the 2000.

This can be explained because this task is longer than the others so the active zones could have different

patterns inside this task. To better explain one can say that, for example, at the beginning of the task

individuals can have active zones with a different pattern from the active zones at end of the same task.

It is necessary to refer that, the fact that the first column of table C.3 has the higher values is due to the

method used in the algorithm, because when we have a tie in the decision process the individual chosen is

the first one in the sequence. So this allows us to conclude, that several situations of this nature happened,

which explains those results. Using the tasks PRE2000 and POS2000 individually this doesn’t happen and

one achieved better results. One see that, in those cases, it is possible to identify several individuals

by their active zones. For example using the active zones from task PRE2000 (table C.1) the individual

4 is completely identified by his active zones and the individual 1 and 2 also have a good percentage

of identification. Using the task POS2000 it is the individual 8 which has a 100% correct identification,

52

although individuals 1, 2 and 6 also have a good percentage. When it is joined the three tasks, as most of

the active zones are from task 2000, the results are very similar to the ones achieved with this task.

Tables C.5 to C.8 are concerned to the time of fall. Using task 2000 one obtains the same behavior,

explained before, for the other feature. Concerning the other two tasks one see again that some individuals

are well classified. Using the active zones from task PRE2000 one sees that individual 6 is better identified

than the others, while using task POS2000 it is individual 4 that achieves best results. The joining of the

three tasks lead us to the same conclusions presented in the previous paragraph.

Comparing the results achieved with the two features, one sees that the individuals are better classified

by the duration of the strokes. So this feature is more related to the individuality of each signal than the

other feature. It is expected however that, when the 2 features are considered, one obtains better results.

Tables C.9 to C.12 show the results achieved using the 2 features simultaneously. Indeed one see

that using the 2 features we have better results and more individuals are correctly classified. And also

considering the task 2000 the results are improved. Using the active zones from task PRE2000 one sees

that 5 individuals have percentages of a correct identification higher than 50%. But it can be seen that

the individuals are better identified by the active zones from task POS2000. In this one 7 individuals have

percentages higher or equal to 50%. With task 2000, in this case it can be seen an improvement because,

comparing with the use of just one feature, in this case one has a higher number of active zones correctly

identified. Consequently using all the tasks together it is possible to see that the results are better than

using just one feature. Although one has higher percentages when the identification is done with active

zones taken from tasks PRE2000 and POS2000.

53

Chapter 5

Conclusions

In this chapter, it is presented the main conclusions of this thesis, as well as suggestions for future work.

The goal of this thesis was to analyze the EMG signal of athletes performing rowing. The purpose of the

analysis was to identify and characterize fatigue. It was also studied, trough the use of pattern recognition

techniques, the individuality of the signals.

The first analysis made was using global parameters of the active zones. It was calculated the dominant

frequency of the active zone as well as the mean power. The results obtained with these two features didn’t

lead us to any specific conclusion. The behavior of these features is different from individual to individual

and task to task.

The analysis made next was using time features. Those features, dominant frequency and mean power,

were calculated for each instant of time and not only for the global active zone. The results achieved with

the dominant frequency show a concentration in lower frequencies. Observing the graphics of the mean

power, it was seen a pattern with this feature. In order to see if each individual had a typical active zone,

mean active zone was calculated. This allowed us to conclude that the individuals have different active

zones. In some of them the mean power starts decreasing sooner than in others, others have longer active

zones.

Another approach was the classification of the active zones. Using pattern recognition techniques, the

active zones were classified according to the individual who performed it and according to the task.

Using the k-NN algorithm was made the classification of the active zones by its task, 4 scenarios were

considered. The difference between the scenarios was the way the area, above the curve of the mean

power, was calculated. The results achieved allow us to conclude that tasks PRE2000 and POS2000 are

different from each other, even though the nature of the two tasks is similar. Fatigue could be an explanation

55

for this fact, because task POS2000 is performed when the athlete had already made a significant effort.

The active zones were also used to identify individuals, using the k-NN algorithm. The results achieved

let us conclude that each individual is unique and has a unique way to perform the activity. So it is possible

to identify the individuals by its active zones, using the right features.

For future work, one would suggest the study of the other channels (muscles). Since all the work was

done for the first channel it is also important to see the behavior of active zones for the other muscles. The

techniques should be slightly different because the signals achieved in the other channels have a different

behavior. It is also suggested a more exhaustive study of task 2000. One suggests the division of task

2000 in smaller signals to easily analyze the signal and study the different parts of it. Another approach

could be the use of other features to improve the classification of tasks and individuals.

56

Appendices

57

Appendix A

Mean Active Zones

59

(a) Individual 1 (b) Individual 2

(c) Individual 3 (d) Individual 4

(e) Individual 5 (f) Individual 6

Figure A.1: Mean active zones of all individuals and tasks for channel 1.

60

(g) Individual 7 (h) Individual 8

(i) Individual 9 (j) Individual 10

(k) Individual 11

Figure A.1: Mean active zones of all individuals and tasks for channel 1.

61

Appendix B

Tasks identification

63

Table B.1: Scenario A1 with 1 feature.(a) Individual 1.

PRE POSPRE 85,7143 14,2857POS 28,5714 71,4286

(b) Individual 2.

PRE POSPRE 57,1429 42,8571POS 28,5714 71,4286

(c) Individual 3.

PRE POSPRE 28,5714 71,4286POS 57,1429 42,8571

(d) Individual 4.

PRE POSPRE 100 0POS 0 100

(e) Individual 5.

PRE POSPRE 71,4286 28,5714POS 28,5714 71,4286

(f) Individual 6.

PRE POSPRE 42,8571 57,1429POS 28,5714 71,4286

(g) Individual 7.

PRE POSPRE 57,1429 42,8571POS 71,4286 28,5714

(h) Individual 8.


(i) Individual 9.

PRE POSPRE 57,1429 42,8571POS 42,8571 57,1429

(j) Individual 10.

PRE POSPRE 57,1429 42,8571POS 71,4286 28,5714

(k) Individual 11.

PRE POSPRE 71,4286 28,5714POS 42,8571 57,1429

64

Table B.2: Scenario A1 with 2 features.

(a) Individual 1.

PRE POSPRE 100 0POS 14,2857 85,7143

(b) Individual 2.

PRE POSPRE 57,1429 42,8571POS 57,1429 42,8571

(c) Individual 3.

PRE POSPRE 42,8571 57,1429POS 57,1429 42,8571

(d) Individual 4.


(e) Individual 5.

PRE POSPRE 57,1429 42,8571POS 42,8571 57,1429

(f) Individual 6.

PRE POSPRE 57,1429 42,8571POS 14,2857 85,7143

(g) Individual 7.

PRE POSPRE 71,4286 28,5714POS 28,5714 71,4286

(h) Individual 8.


(i) Individual 9.

PRE POSPRE 71,4286 28,5714POS 28,5714 71,4286

(j) Individual 10.

PRE POSPRE 57,1429 42,8571POS 71,4286 28,5714

(k) Individual 11.

PRE POSPRE 100 0POS 14,2857 85,7143

65


(a) Individual 1.

PRE POSPRE 85,7143 14,2857POS 14,2857 85,7143

(b) Individual 2.

PRE POSPRE 100 0POS 14,2857 85,7143

(c) Individual 3.

PRE POSPRE 71,4286 28,5714POS 14,2857 85,7143

(d) Individual 4.


(e) Individual 5.


(f) Individual 6.


(g) Individual 7.


(h) Individual 8.


(i) Individual 9.

PRE POSPRE 85,7143 14,2857POS 14,2857 85,7143

(j) Individual 10.


(k) Individual 11.


66

Table B.4: Scenario A2 with 1 feature.

(a) Individual 1.

PRE POSPRE 71,4286 28,5714POS 14,2857 85,7143

(b) Individual 2.

PRE POSPRE 42,8571 57,1429POS 57,1429 42,8571

(c) Individual 3.

PRE POSPRE 71,4286 28,5714POS 42,8571 57,1429

(d) Individual 4.

PRE POSPRE 28,5714 71,4286POS 71,4286 28,5714

(e) Individual 5.

PRE POSPRE 71,4286 28,5714POS 14,2857 85,7143

(f) Individual 6.

PRE POSPRE 28,5714 71,4286POS 57,1429 42,8571

(g) Individual 7.

PRE POSPRE 71,4286 28,5714POS 42,8571 57,1429

(h) Individual 8.

PRE POSPRE 57,1429 42,8571POS 42,8571 57,1429

(i) Individual 9.

PRE POSPRE 42,8571 57,1429POS 71,4286 28,5714

(j) Individual 10.

PRE POSPRE 57,1429 42,8571POS 57,1429 42,8571

(k) Individual 11.

PRE POSPRE 71,4286 28,5714POS 14,2857 85,7143

67


(a) Individual 1.


(b) Individual 2.

PRE POSPRE 28,5714 71,4286POS 57,1429 42,8571

(c) Individual 3.

PRE POSPRE 57,1429 42,8571POS 71,4286 28,5714

(d) Individual 4.


(e) Individual 5.

PRE POSPRE 57,1429 42,8571POS 14,2857 85,7143

(f) Individual 6.

PRE POSPRE 71,4286 28,5714POS 14,2857 85,7143

(g) Individual 7.

PRE POSPRE 57,1429 42,8571POS 14,2857 85,7143

(h) Individual 8.


(i) Individual 9.

PRE POSPRE 71,4286 28,5714POS 14,2857 85,7143

(j) Individual 10.

PRE POSPRE 85,7143 14,2857POS 57,1429 42,8571

(k) Individual 11.

PRE POSPRE 100 0POS 14,2857 85,7143

68


(a) Individual 1.

PRE POSPRE 85,7143 14,2857POS 14,2857 85,7143

(b) Individual 2.

PRE POSPRE 100 0POS 14,2857 85,7143

(c) Individual 3.

PRE POSPRE 71,4286 28,5714POS 14,2857 85,7143

(d) Individual 4.


(e) Individual 5.


(f) Individual 6.


(g) Individual 7.


(h) Individual 8.


(i) Individual 9.

PRE POSPRE 85,7143 14,2857POS 0 100

(j) Individual 10.


(k) Individual 11.


69

Table B.7: Scenario B with 1 feature.

(a) Individual 1.

PRE POSPRE 85,7143 14,2857POS 14,2857 85,7143

(b) Individual 2.

PRE POSPRE 42,8571 57,1429POS 71,4286 28,5714

(c) Individual 3.

PRE POSPRE 57,1429 42,8571POS 42,8571 57,1429

(d) Individual 4.


(e) Individual 5.

PRE POSPRE 85,7143 14,2857POS 0 100

(f) Individual 6.

PRE POSPRE 42,8571 57,1429POS 71,4286 28,5714

(g) Individual 7.

PRE POSPRE 14,2857 85,7143POS 57,1429 42,8571

(h) Individual 8.

PRE POSPRE 57,1429 42,8571POS 42,8571 57,1429

(i) Individual 9.

PRE POSPRE 71,4286 28,5714POS 28,5714 71,4286

(j) Individual 10.

PRE POSPRE 42,8571 57,1429POS 57,1429 42,8571

(k) Individual 11.

PRE POSPRE 42,8571 57,1429POS 71,4286 28,5714

70

Table B.8: Scenario B with 2 features.

(a) Individual 1.

PRE POSPRE 85,7143 14,2857POS 28,5714 71,4286

(b) Individual 2.

PRE POSPRE 28,5714 71,4286POS 28,5714 71,4286

(c) Individual 3.

PRE POSPRE 57,1429 42,8571POS 57,1429 42,8571

(d) Individual 4.


(e) Individual 5.

PRE POSPRE 85,7143 14,2857POS 0 100

(f) Individual 6.

PRE POSPRE 57,1429 42,8571POS 57,1429 42,8571

(g) Individual 7.

PRE POSPRE 57,1429 42,8571POS 14,2857 85,7143

(h) Individual 8.


(i) Individual 9.

PRE POSPRE 71,4286 28,5714POS 14,2857 85,7143

(j) Individual 10.

PRE POSPRE 57,1429 42,8571POS 71,4286 28,5714

(k) Individual 11.

PRE POSPRE 100 0POS 28,5714 71,4286

71

Table B.9: Scenario C with 1 feature.

(a) Individual 1.

PRE POSPRE 85,7143 14,2857POS 0 100

(b) Individual 2.

PRE POSPRE 57,1429 42,8571POS 71,4286 28,5714

(c) Individual 3.

PRE POSPRE 57,1429 42,8571POS 42,8571 57,1429

(d) Individual 4.

PRE POSPRE 71,4286 28,5714POS 28,5714 71,4286

(e) Individual 5.

PRE POSPRE 85,7143 14,2857POS 0 100

(f) Individual 6.

PRE POSPRE 42,8571 57,1429POS 42,8571 57,1429

(g) Individual 7.

PRE POSPRE 57,1429 42,8571POS 57,1429 42,8571

(h) Individual 8.

PRE POSPRE 42,8571 57,1429POS 42,8571 57,1429

(i) Individual 9.

PRE POSPRE 57,1429 42,8571POS 42,8571 57,1429

(j) Individual 10.

PRE POSPRE 28,5714 71,4286POS 85,7143 14,2857

(k) Individual 11.

PRE POSPRE 85,7143 14,2857POS 14,2857 85,7143

72

Table B.10: Scenario D with 1 feature.

(a) Individual 1.

PRE POSPRE 85,7143 14,2857POS 0 100

(b) Individual 2.

PRE POSPRE 85,7143 14,2857POS 42,8571 57,1429

(c) Individual 3.

PRE POSPRE 42,8571 57,1429POS 42,8571 57,1429

(d) Individual 4.

PRE POSPRE 85,7143 14,2857POS 28,5714 71,4286

(e) Individual 5.

PRE POSPRE 100 0POS 14,2857 85,7143

(f) Individual 6.

PRE POSPRE 85,7143 14,2857POS 28,5714 71,4286

(g) Individual 7.

PRE POSPRE 42,8571 57,1429POS 71,4286 28,5714

(h) Individual 8.


(i) Individual 9.

PRE POSPRE 85,7143 14,2857POS 14,2857 85,7143

(j) Individual 10.

PRE POSPRE 71,4286 28,5714POS 57,1429 42,8571

(k) Individual 11.

PRE POSPRE 85,7143 14,2857POS 0 100

73

Appendix C

Individuals Classification - Confusion

Matrix

75

Table C.1: Feature: Active zone duration. Task: PRE20001 2 3 4 5 6 7 8 9 10 11

1 62,5 0 12,5 0 0 25 0 0 0 0 02 0 62,5 0 37,5 0 0 0 0 0 0 03 0 0 37,5 0 0 0 62,5 0 0 0 04 0 0 0 100 0 0 0 0 0 0 05 50 0 25 0 0 0 12,5 0 12,5 0 06 33,333 0 11,111 0 0 44,444 11,111 0 0 0 07 0 0 44,444 0 0 11,111 44,444 0 0 0 08 25 0 12,5 0 0 12,5 50 0 0 0 09 12,5 12,5 0 0 37,5 0 0 0 37,5 0 0

10 0 28,571 57,143 0 0 14,286 0 0 0 0 011 0 62,5 12,5 25 0 0 0 0 0 0 0

Table C.2: Feature: Active zone duration. Task: POS20001 2 3 4 5 6 7 8 9 10 11

1 50 0 0 37,5 0 0 0 0 0 12,5 02 0 88,889 0 0 0 0 0 11,111 0 0 03 0 12,5 37,5 12,5 0 0 0 0 37,5 0 04 25 0 12,5 37,5 0 0 0 0 0 25 05 12,5 37,5 12,5 12,5 0 25 0 0 0 0 06 0 37,5 0 0 0 50 0 12,5 0 0 07 0 62,5 25 0 0 12,5 0 0 0 0 08 0 0 0 0 0 0 0 100 0 0 09 16,667 0 66,667 0 0 0 0 0 0 16,667 0

10 42,857 0 14,286 42,857 0 0 0 0 0 0 011 0 0 50 25 0 12,5 0 0 0 12,5 0

Table C.3: Feature: Active zone duration. Task: 20001 2 3 4 5 6 7 8 9 10 11

1 92,983 1,3158 0 2,193 0,4386 0 0 0 3,0702 0 02 80,745 17,081 0 0 0,6211 0,9317 0 0 0 0 0,62113 94,311 0,2994 4,1916 0,2994 0,2994 0 0 0 0 0 0,59884 89,163 3,9409 0 1,9704 0 1,4778 0 1,4778 0,4926 0 1,47785 91,327 3,0612 0,5102 0 1,0204 0,5102 0 0 1,5306 0 2,04086 90,863 2,5381 0 3,0457 0,5076 0 0 1,0152 0,5076 0 1,52287 84,456 13,99 0 0,5181 0,5181 0,5181 0 0 0 0 08 88,835 5,8252 0 1,9417 0 1,9417 0 0 0,4854 0 0,97099 83,333 1,3889 0 0 2,0833 0 0 0 13,194 0 0

10 93,056 2,3148 0 2,3148 0,463 1,3889 0 0 0,463 0 011 69,231 11,539 1,9231 3,8462 3,2051 3,8462 0 0 0,641 0 5,7692

76

Table C.4: Feature: Active zone duration. Task: all1 2 3 4 5 6 7 8 9 10 11

1 93,443 1,2295 0 1,2295 0,4098 0 0 0 3,6885 0 02 80,531 17,994 0 0,295 0,59 0 0 0 0 0 0,593 94,571 0,2857 4 0,2857 0,2857 0 0 0 0 0 0,57144 87,671 5,0228 0 3,653 0 0,9132 0 1,3699 0,9132 0 0,45665 91,509 2,8302 0,4717 0,9434 0,9434 0 0 0 1,8868 0 1,41516 90,654 2,8037 0 4,6729 0,4673 0 0 0 0,4673 0 0,93467 85,714 12,857 0 0,9524 0,4762 0 0 0 0 0 08 87,387 6,3063 0 4,5045 0 0,4505 0 0 0,9009 0 0,45059 84,81 1,2658 0 0 1,8987 0 0 0 12,025 0 0

10 95,217 2,1739 0 1,3043 0,4348 0,4348 0 0 0,4348 0 011 71,512 10,465 1,1628 4,6512 2,907 2,3256 0 0 0,5814 0 6,3953

Table C.5: Feature: Time of fall. Task: PRE20001 2 3 4 5 6 7 8 9 10 11

1 37,5 25 0 12,5 25 0 0 0 0 0 02 75 0 0 0 12,5 0 12,5 0 0 0 03 12,5 25 12,5 12,5 25 0 12,5 0 0 0 04 25 12,5 25 0 0 12,5 25 0 0 0 05 37,5 0 50 0 0 12,5 0 0 0 0 06 11,111 0 0 11,111 0 66,667 11,111 0 0 0 07 0 0 33,333 22,222 11,111 0 33,333 0 0 0 08 62,5 0 25 0 0 0 12,5 0 0 0 09 0 0 0 12,5 0 0 0 0 75 12,5 0

10 0 0 0 0 0 71,429 0 0 0 28,571 011 12,5 25 12,5 25 0 12,5 0 0 12,5 0 0

Table C.6: Feature: Time of fall. Task: POS20001 2 3 4 5 6 7 8 9 10 11

1 50 25 0 0 12,5 12,5 0 0 0 0 02 44,444 22,222 0 0 11,111 0 11,111 11,111 0 0 03 37,5 25 0 12,5 0 12,5 12,5 0 0 0 04 0 0 0 75 0 0 0 0 12,5 12,5 05 25 12,5 0 12,5 37,5 0 0 12,5 0 0 06 37,5 0 0 12,5 12,5 0 0 12,5 25 0 07 12,5 37,5 0 0 25 0 0 25 0 0 08 12,5 0 0 0 12,5 0 12,5 62,5 0 0 09 0 0 0 50 0 0 0 0 0 16,667 33,333

10 0 0 0 57,143 0 14,286 0 0 14,286 0 14,28611 12,5 12,5 0 0 0 0 0 0 25 0 50

77

Table C.7: Feature: Time of fall. Task: 20001 2 3 4 5 6 7 8 9 10 11

1 88,597 5,7018 3,9474 0 0 0 0 0 0,8772 0 0,87722 79,193 18,323 0 0 0 0 1,2422 0,3106 0 0 0,93173 96,407 0,5988 1,7964 0,5988 0,2994 0 0 0 0 0 0,29944 17,734 0 10,837 66,503 0 0 0 0 4,9261 0 05 93,878 3,5714 0 0,5102 0 0 0 0 0,5102 0 1,53066 98,985 0,5076 0 0 0 0 0 0 0 0 0,50767 34,197 47,668 0 0 0 0 17,098 0 0 0 1,03638 59,223 30,097 0,4854 0,4854 0 0 7,767 0 0,4854 0 1,45639 15,278 0 6,25 59,028 0 0 0 0 19,444 0 0

10 79,63 0 5,0926 14,352 0 0 0 0 0,9259 0 011 69,231 0,641 0 6,4103 1,2821 0 0,641 0,641 0,641 0 20,513

Table C.8: Feature: Time of fall. Task: all1 2 3 4 5 6 7 8 9 10 11

1 90,164 4,5082 3,6885 0 0 0 0 0 0,8197 0 0,81972 84,956 12,094 0 0 0 0 1,4749 0,59 0 0 0,8853 96,857 0,2857 1,7143 0,5714 0,2857 0 0 0 0 0 0,28574 22,831 0 10,046 62,557 0 0 0 0 4,5662 0 05 95,283 1,8868 0 0,4717 0 0 0,4717 0 0,4717 0 1,41516 99,065 0,4673 0 0 0 0 0 0 0 0 0,46737 44,286 40,952 0 0 0 0 13,81 0 0 0 0,95248 64,414 24,775 0,4505 0,4505 0 0 7,6577 0 0,4505 0 1,80189 18,354 0 6,3291 56,962 0 0 0 0 18,354 0 0

10 80 0 5,2174 13,913 0 0 0 0 0,8696 0 011 70,349 0,5814 0 5,814 1,1628 0 0,5814 1,1628 0,5814 0 19,767

Table C.9: Feature: Active zone duration and time of fall. Task: PRE20001 2 3 4 5 6 7 8 9 10 11

1 50 0 0 0 12,5 0 0 37,5 0 0 02 0 75 0 25 0 0 0 0 0 0 03 0 0 37,5 0 0 0 62,5 0 0 0 04 0 0 0 100 0 0 0 0 0 0 05 50 0 12,5 0 12,5 0 25 0 0 0 06 0 0 11,111 0 0 77,778 11,111 0 0 0 07 0 0 44,444 0 0 11,111 44,444 0 0 0 08 37,5 0 0 0 0 0 25 37,5 0 0 09 0 0 0 0 0 12,5 0 0 75 0 12,5

10 0 0 42,857 0 0 14,286 0 0 0 14,286 28,57111 0 75 0 0 0 0 0 0 0 25 0

78

Table C.10: Feature: Active zone duration and time of fall. Task: POS20001 2 3 4 5 6 7 8 9 10 11

1 87,5 0 12,5 0 0 0 0 0 0 0 02 0 55,556 0 0 0 22,222 11,111 11,111 0 0 03 0 0 75 12,5 0 0 0 0 0 0 12,54 0 0 0 87,5 0 0 0 0 12,5 0 05 0 0 37,5 0 12,5 0 37,5 0 0 12,5 06 0 25 0 12,5 0 50 0 12,5 0 0 07 0 25 37,5 0 37,5 0 0 0 0 0 08 0 0 0 0 0 0 0 100 0 0 09 0 0 0 16,667 0 0 0 0 16,667 16,667 50

10 0 0 0 42,857 0 0 0 0 0 42,857 14,28611 0 0 25 0 0 0 0 0 0 12,5 62,5

Table C.11: Feature: Active zone duration and time of fall. Task: 20001 2 3 4 5 6 7 8 9 10 11

1 62,281 18,421 4,386 0 1,7544 1,3158 3,5088 3,9474 2,193 0,8772 1,31582 20,186 60,559 5,2795 0 2,1739 1,5528 8,0745 0,6211 0 0,6211 0,93173 7,7844 8,6826 76,647 1,1976 0,5988 0,8982 0,8982 0,5988 0 1,7964 0,89824 0,9852 0,4926 9,8522 74,384 0 0 0 0 7,3892 3,9409 2,95575 36,225 34,184 10,204 0 6,6327 3,5714 2,551 1,0204 1,5306 1,0204 3,06126 24,366 38,579 15,736 0 13,706 6,0914 1,0152 0 0,5076 0 07 6,2176 27,979 1,5544 0 0 0 58,031 5,1813 0 0 1,03638 29,126 19,903 2,4272 0,4854 1,9417 0 31,068 13,107 0,4854 0 1,45639 4,1667 0,6944 11,111 47,222 0 1,3889 0 0 34,722 0,6944 0

10 8,3333 6,0185 44,907 21,296 0,9259 0,463 0 0 1,3889 16,667 011 20,513 13,462 15,385 5,7692 5,7692 2,5641 0 1,9231 4,4872 3,2051 26,923

Table C.12: Feature: Active zone duration and time of fall. Task: all1 2 3 4 5 6 7 8 9 10 11

1 59,836 18,033 4,0984 0 2,459 1,6393 4,0984 5,7377 2,0492 0,4098 1,63932 17,699 62,537 6,1947 0,885 2,9499 0,885 7,0796 0,59 0 0,295 0,8853 9,1429 10,571 73,143 1,1429 0,5714 0,5714 1,1429 0,8571 0 1,7143 1,14294 3,1963 3,1963 9,1324 69,863 0 0 0 0,4566 7,3059 4,1096 2,73975 36,793 33,491 10,377 0 4,717 2,3585 2,8302 2,3585 1,4151 2,3585 3,30196 25,234 35,981 16,355 0,4673 12,617 5,1402 0,9346 1,4019 0,4673 0,4673 0,93467 9,0476 26,191 3,3333 0 0 0 51,905 8,5714 0 0 0,95248 27,027 18,018 3,1532 0,4505 2,2523 0 29,73 16,216 0,4505 0,9009 1,80189 5,6962 1,8987 11,392 44,304 0,6329 1,2658 0 0 34,177 0,6329 0

10 9,5652 6,087 43,044 21,304 1,7391 0,4348 0 0 1,7391 15,652 0,434811 20,349 12,791 15,116 5,2326 5,2326 4,0698 0 1,7442 4,0698 2,907 28,488

79

Bibliography

[1] Yousef Al-Assaf and Hason Al-Nashash. Myoelectric signal segmentation and classification using

wavelets based neural networks. In Proceedings of the 23rd Annual EMBS International Conference,

2001.

[2] Pedro A. Berger, Francisco A. de O. Nascimento, Jake C. do Carmo, and Adson F. da Rocha. Com-

pression of emg signals with wavelet transform and artificial neural networks. Physiological Measure-

ment, (27):457–465, 2006.

[3] Liyu Cai, Zhizhong Wang, and Haihang Zhang. An emg classification method based on wavelet

transform. In Proceedings of The First Joint BMES/EMBS Conference Serving Humanity, Advancing

Technology, page 565, October 1999.

[4] Hanne Christensen. Emg - analysis. IEEE Engineering in Medicine & Biology Society 11th Annual

International Conference, 1989.

[5] Beau Crawford, Kai Miller, Pradeep Shenoy, and Rajesh Rao. Real-time classification of electromyo-

graphic signals for robotic control. In Proceedings of AAAI, pages 523–528, 2005.

[6] Richard O. Duda, Peter E. Hart, and David G. Stork. Pattern Classification. Wiley, John & Sons,

Incorporated, 2000.

[7] Ana Fred. Nn-classifiers. Modeling and classification of biomedical signals.

[8] Xiao Hu, Zhizhong Wang, and Xiaomei Ren. Classification of surface emg signal using relative wavelet

packet energy. Computer Methods and Programs in Biomedicine, (79):189–195, 2005.

[9] Adham R. Ismail and Shihab S. Asfour. Continuous wavelet transform application to emg signals

during human gait. IEEE, 1998.

[10] Dinesh Kant Kumar, Nemuel D. Pah, and Alan Bradley. Wavelet analysis of surface electromyography

to determine muscle fatigue. IEEE Transactions on neural systems and rehabilitation engineering,

11(4), 2003.

81

[11] F. Laterza and G. Olmo. Analysis of emg signals by means of the matched wavelet transform. Elec-

tronics Letters Online, 33(5), 1997.

[12] Carlo J. De Luca. Myoelectrical manifestations of localized muscular fatigue in humans. Critical

Reviews in Biomedical Engineering, 11(4):251–279, 1984.

[13] Carlo J. De Luca. The use of surface electromyography in biomechanics. Journal of Applied Biome-

chanics, 13(2):135–163, 1997.

[14] Mario. Remo em ergometro - analise cinematica. Master’s thesis.

[15] Thomas Mazzone. Kinesiology of the rowing stroke. NSCA Journal, 10(2), 1988.

[16] Roberto Merletti and Philip A. Parker. Electromyography. Institute of Electrical and Electronics Engi-

neers, 2004.

[17] Larry D. Paarmann and Michael J. Korenberg. Accurate arma signal identification - empirical results.

IEEE, 1992.

[18] M. B. I. Reaz, M. S. Hussain, and F. Mohd-Yasin. Techniques of emg signal analysis: detection,

processing, classification and applications. Biological Procedures Online, 2006.

[19] Peter Wellig and George S. Moschytz. Analysis of wavelet features for myoelectric signal classification.

Signal and Information Processing Laboratory, 1998.

82

Fatigue detection in EMG signals - ULisboa · Neste estudo foram analisadas 3 provas de remo...

Documents

Transcript of Fatigue detection in EMG signals - ULisboa · Neste estudo foram analisadas 3 provas de remo...