HAND PROSTHESIS CONTROL

Software Tool for EMG Signal Analysis

Tomasz Suchodolski and Andrzej Wolczowski

Institute of Informatics, Automation and Robotics, Wroclaw University of Technology

Wybrzeze Wyspianskiego 27, 50-370 Wroclaw, Poland

Keywords: Hand Prosthesis, EMG Signals, Decision Tree, Pattern Recognition.

Abstract: The paper discusses the problem of the decision process of controlling the bio-prosthesis of the hand that is

treated as the human intention recognition by means of the analysis of the electromyography (EMG) signals

from the hand muscles. The number of movements, which is indispensable for the dexterity of the

prosthesis, makes the recognition not entirely reliable. The approach presented herein includes three

methods: the decision tree, neuron networks, and genetic algorithms in order to enhance the reliability of the

EMG signal recognition. Simultaneously, the paper presents the software designed for the needs of the

research and adapted to processing the EMG signals in compliance with these methods.

1 INTRODUCTION

The dexterity of the human hand is the result of both

its kinematic structure and the control of it by the

central nervous system. The construction of multi-

joint anthropomorphic mechanic structure that can

copy even very complicated movements of the

human hand poses at present no problem. However,

the question of the reliable bio-control of such a

dexterous prosthesis remains unsolved.

The usual practice is to make use of myopoten-

tials – which are electrical signals that accompany

the activity of the muscles (Wolczowski, 2001). This

is feasible since, after the amputation of the hand,

there remain in the stamp a significant number of the

muscles that earlier controlled the movement of the

fingers (Su Y, et al., 2007; Mohammadreza, 2007).

Through the tensing of these muscles, the person

with a disability may express their intentions to

control their prosthesis as they do with their healthy

hand. The intention recognition calls for efficient

analysis of EMG signals. Taking into account a large

number of possible movements that characterize a

dexterous hand, the chances of error occurrence are

high. The minimization of this error occurrence

constitutes the current challenge in devising the

person-prosthesis interface. The experiments on the

recognition of the movement intended by a person

with an amputated hand can be replaced with the

analysis of the relationship that holds between the

movements of the healthy hand and the myopoten-

tials that occur then, taking into consideration the

limitations that arise due to the lack of some of the

muscles in the amputated hand.

The solution presented here consists in the

decomposition of the recognition, which leads to the

multi-stage process that can be presented as the

decision tree (DT).

In order to reduce the occurrence of recognition

errors irrespective of the number of the processed

data we have opted for replacing the usual decision

nodes with Neural Networks of the multilayer

perceptron (MLP) type (enodes). In order to create

an MLP population, we have made use of the

Genetic Algorithm (GA), which enhanced the

synthesis of optimal MLPs.

For the needs of the research “EMG Analysis”

application has been developed, which allows for the

program transformation of the registered signals. Its

operation is described in detail in subsequent

chapters. Initially the EMG signal in the form of

numerical data is transmitted to the SQL database.

Signal acquisition and the manner in which the

accumulated data are being refined are described in

Chapter two. Chapter three describes methods of

EMG signal analysis; extraction of characteristic

features and concept for signal classification.

Chapter four describes the experiment that has been

carried out and sums up the results.

321

Suchodolski T. and Wolczowski A. (2010).

HAND PROSTHESIS CONTROL - Software Tool for EMG Signal Analysis.

In Proceedings of the 7th International Conference on Informatics in Control, Automation and Robotics, pages 321-326

DOI: 10.5220/0002957403210326

 SciTePress

2 EMG SIGNALS ACQUISITION

In order to classify the EMG signal (or to determine

the class of the intended movement) one has to make

a number of transformations of the obtained signal,

which boils down to its multistage analysis. Each of

the said stages affects the quality of the whole

process that is the reliability of movement

recognition. That is why it is so important for each

of the stages to be optimised as to their reliability

and calculability. Main stages of the process that are

being carried out by "EMG Analysis" application are

presented below.

2.1 Signal Acquisition

In order to obtain the data the EMG signal was

measured by means of 8-channel measuring system

(Krysztoforski and Wolczowski, 2005). The system

is made up of: (a) – computer that register the results

of the measurements; it is fitted with (b) - A/D

converters card (Ni 4477); and (c) - 8 active

differential electrodes. The electrodes are fitted with

two metal contacts (De Luca, 2006) that are directly

integrated with the circuit amplifying the measured

signals (potentials difference).

Figure 1: The differential measurement method. The EMG

signal is represented by 's' and the noise signals by 'z'.

That kind of construction makes it possible to

eliminate most of the disturbances that occur during

the measurement. The value of the reduction is

determined by coefficient CMRR of the amplifier. In

this system the coefficient amounts to 84dB.

The signals from the electrodes are transformed

into a digital form that is convenient for further

analysis on the A/D converters card and they are

stored in the SQL database of the measuring

computer. These data are labelled for their further

identification (measurement number, class of

problem – i.e. posture). Data acquisition carried out

in this way enables further program transformations

of the obtained EMG signal by means of consecutive

application modules of “EMG Analysis”.

2.2 Preparing the Data for the Analysis

In order to reduce the level of noise in the obtained

EMG signal, the signal undergoes an initial

processing that consists in averaging of the signal

course in the domain of time, independently of each

channel (1), according to the rule:

]1[)1(][*][ 







nynxny



(1)

y[n] - current output sample from averaging system;

x[n] - current EMG sample.

The selection window placed in the form shown

in Figure 2 marked as “alpha coefficient” enables

the operator to select a proper degree of the signal

averaging depending on its quality (

0.1,5.0



Figure 2: Actual signal (upper part of the figure), the

filtered signal. The alpha parameter equals 0,9.

The signal that has been processed as described

above may be used for identifying characteristic

features i.e. such features that differentiate the

analysed signals one from another (in consequence

the type of the hand movement).

3 EMG SIGNAL ANALYSIS

The process of EMG signal recognition is made up

of two main stages: that of determining the features

of the measured signal and that of signal

classification on the basis of the determined features

(Wolczowski and Suchodolski, 2007).

3.1 Features Extraction

In order to determine characteristic features of the

signal two methods have been applied: Fast Fourier

Transform and Wavelet Transform.

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

322

To carry the task out, the signal amplitude had to

be analysed (Wojtczak, et al., 2008). A low-ampli-

tude signal is difficult to classify due to the partici-

pation of the constant constituent (in the form of

noise) and also due to the fact that the low value of

the signal amplitude may imply both that the muscle

(the hand) is at rest or its insignificant participation

in a given class of movement. The analysis has been

carried out according to the following scheme:

 Division of the input data vector that consists of

7.000 samples/channel (7s measurement) into 40

element windows (40 ms), and calculation of the

sum of the contents each window (E

sum

). In this way

the information of the energy of the recorded

movement is obtained; the minimum E

min

and the

maximum E

max

energy values in the set;

 Finding out the energy threshold E

edg

(activate-

ing energy) in accordance with rule (2). In this way a

group of windows is singled out that may contain

important information in relation to the signal for the

calculations that are being carried out (level of the

signal amplitude);

 The Windows that fulfil the energy condition

egd

> E

sum

are then used to single out 256 element-

windows (necessary for the spectrum analysis of the

signal). This is done through the enlargement of the

space of the selected 40-element windows by a

further 216 elements (216 ms).





minmaxmin

* EEkEE

egd



(2)

Figure 3: Signal usefulness analysis.

This stage is represented in Figure 3. In order to

optimize this process, the application carries out a

simulation of calculations simultaneously for a few

values of the coefficient k (k(0.1,…0.4)). The

operator can for each movement class select by

means of the checkbox different values of his

coefficient, which affects the number of the newly

created 256-element windows.

The Windows determined in such a manner

become the input vector for the feature extractor. As

a result of the transformations for Fourier Transform

we obtain the representation of the signal in the

frequency domain (Figure 4).

Figure 4: Exemplary FFT course (before and after

averaging).

The spectrum pictures of the signal thus obtained

are further subjected to the process of averaging of

harmonics for the purpose of enhancing the degree

of resemblance (data generalization) according to

(3). The averaged values thus become the vector of

signal features.

][









IPE

IPSi

IPSIPE

1,...,1,0  Nk

(3)

In order to determine the features by means of

Wavelet Transform, Mallat algorithm has been used

(Mallat, 1989), which is used to obtain, among

others, the vector of signal details which, after

transformation is treated as the feature vector. The

signal is filtered in accordance with:







mknm

clnc )()(



(4)







mknm

chnd )()(



(5)

According to the size of the windows (256

elements) that constitute the input vector that

undergoes transformations, 6 levels of signal

filtration had to be applied; each of the Mallat levels

reduces by half the size of the input vector. Each of

the Mallat levels is represented by number values

that are the sum of absolute magnitudes of the detail

vectors that arose earlier. The number of the features

that thus arose (6 features) is smaller than the

HAND PROSTHESIS CONTROL - Software Tool for EMG Signal Analysis

323

number of the assumed movement classes to be

recognised. Thus a further 6 features had to be

created in order for them to represent the dynamics

of the changes in the recorded movement. This was

done by summing up the real values respectively for

each of the Mallat vectors that arose. In this way it

was possible to obtain 12 values that represent the

information contained in the signal as well as their

approximations at each of the obtained levels of

filtration.



(n)

m +1

(n)

m +k

(n)

m+1

(n)

m+k

(n)

Figure 5: Signal filtration in compliance with Mallat

algorithm.

Figure 6: Exemplary Wavelet form.

3.2 Data Classification

The classification of the feature vector is of great

importance in the presented approach. Due to a large

number of movement classes and high recognition

reliability (which is indispensable for the dexterity

of the prosthesis) the classification process is a

difficult problem. That is why, for convenience’s

sake, the classification process should be performed

in an extensive manner – through many stages. In

the presented research the method of multi-stage

problem solution has been proposed, which is

performed through the decision tree (DT). The

structure of DT calls for: (a) the determination of its

decision logic (structures and the assignment of

classes to the terminal nodes); (b) determination of

the subsets of features for indirect nodes; (c)

determination of the decision making rules

(algorithms of recognition in each indirect node). To

build the DT based on Neural Network Tree

(NNTree) solution (Zhao, 2001) the C4.5 algorithm

was used, the one worked out by Quinlan (1993); it

makes use of the entropy of information in the

learning subset. Entropy in information theory is

defined as the average amount of information that

corresponds to a sign that symbolises an occurrence

out of a certain set. The selection of the attribute is

determined on the basis of the increase of the

information.

Figure 7: Exemplary NNTree.

The input vector that enables tree induction is the

constituents: Fourier Transform (128 size vectors) or

Wavelet Transform (12 size vectors) previously

determined by information entropy of the set.

At this stage (Figure 8) it is possible to determine

the data size that is to be transformed (single

measurement, measurement grouping; specification

of relevant signals).

Figure 8: Entropy tab page.

Quinlan’s method divides a given node into as

many descending nodes, as many values there are in

the most informative feature (constituent of the

selected transform). The tree structure is made by

recurring computation of the information increase

coefficient for all available attributes at each

recurring selection of the algorithm and the selection

of the attribute for which this value is maximal

(constituent of the selected transform). The recurring

selection of the algorithm that makes up the tree

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

324

structure ends the searching process of the input

vector (FFT or DWT feature vector) provided that in

a given sub-tree created by means of the recurring

procedure a set of attributes represents one decision

class. In this case a terminal mode is created that is

marked with class label. The structure thus created

has an input node, a root, and descendant nodes

some of which (terminals) determine the decision

class.

Figure 9: Exemplary decision tree.

To avoid too large a tree (and excessive

matching), the pruning was used on the basis of the

estimation of the significance of the difference of the

classify-cation error for a given node and its sub-

nodes. The algorithm estimates the probability of the

decrease of the error occurrence and cuts off the

divisions for which this probability does not exceed

a given threshold value or provides identical results

(similar results for the sub-nodes). The boundaries

of the ranges that divide the space of the input vector

(FFT or DWT) into the scopes of values that can be

obtained were selected as a result of experiments so

that their distribution was even for the whole span of

the input. However, this method has a drawback;

namely, new feature vectors that are to be classified

can have values that cross the boundaries of two

neighbouring scopes and represent different classes

of hand movement. Therefore, the simple decision

rules that consisted in comparing the new value with

existing scopes were replaced with the decision

making module i.e. MLP. On the basis of actual

values from the database, n networks were trained (n

– number of scopes in the MLP database {m

...m

}).

Each of them (MLP

..MLP

) is responsible for the

classification of the value from a given range

...m

)

This solution introduces the genetic

algorithm to create the MLP population that will

allow for solving the problem of range classification.

The process is running until the intended

classification error is obtained (the value falls within

the range) or until the maximum iteration number is

obtained (then the best configuration is returned –

the least obtained error).

After the training process, the configuration and

the weights of best MLP modules are directed to

SQL database where they are assigned to given

value ranges to be further used for classification

process. The process consists in computing the

attributes taken earlier from the input vector (EMG

spectrum) in compliance with the decision scheme

i.e. the tree structure. The decision rules are taken

from the truth table. Additionally, the taken value is

computed by means of relevant MLP

module that

simulates a given tree node (Figure 10). Thus we

obtain the classic and neuron tree response. The

activation of one of the descendant nodes brings

about the recurring selection of another rule. The

process is running until the final node (terminal) is

reached. This node determines the decision class i.e.

the type of identified posture.

Figure 10: Classification process of the recorded signal.

4 EXPERIMENT AND OBTAINED

RESULTS

The experiments were carried out on healthy persons

by placing each time 6 electrodes on the skin of the

forearm over relevant muscles.

Eleven movement classes were tested: 1. neutral;

2(3). flexion /extension of the hand digits; 4(5)

pronation and supination of the forearm; 6(7)

flexion / extension of the wrist; 8(9). flexion/

extension of the thumb; 10(11). flexion / extension

of the fingers II-V.

For each class of movement the measurement

lasted 7s and was preceded with a 5s break. In that

way the discreet signals were obtained of the size

7.000 samples * 6 channels) per class (sampling

HAND PROSTHESIS CONTROL - Software Tool for EMG Signal Analysis

325

Table 1: Averaged classification errors of the tested movements for FFT input vector (Err

-DT, Err

-NNTree).

Class 1 2 3 4 5 6 7 8 9 10 11

Err

[%] 13 16 14 19 17 12 18 23 19 17 19

Err

[%] 9 14 14 17 9 10 13 21 12 11 14

Table 2: Averaged classification errors of the tested movements for DWT input vector (Err

-DT, Err

-NNTree).

Class 1 2 3 4 5 6 7 8 9 10 11

Err1 [%] 12 14 19 17 12 13 15 18 17 18 15

Err2 [%] 6 11 11 13 13 7 14 16 10 8 8

frequency 1 kHz). The data were located in the SQL

database. The measurement data (1/3 of the signals

selected at random) were used to generate the

decision tree and MLPs. The remaining part of the

set (2/3 of the signals selected at random) was

assigned as a testing set.

Figure 11: Tested movement classes.

The NNTree generated in this way gave an

average recognition value equal to 87% (FFT) when

the DT with classical nodes in turn gave a value

equal to 83% (FFT). The detail results for tested

movements are presented in Table1; were Err1 and

Err2 represent the errors obtained respectively by

DT and by NNTree methods for the FFT feature

vectors. As we can see the best results were for

classes 1, 5, 6 and 10.

In the case of DWT feature vectors an average

recognition values were equal to 89% for NNTree

and 85% for DT method. I this case the best results

were for classes 1, 6, 10 and 11 (see: Table 2).

5 CONCLUSIONS

The tests with the presented “EMG Analysis”

software are in the incipient stage. The conducted

experiments aimed on principle at comparing the

features extraction methods for EMG signals and

their further classification. They prove the

usefulness of the presented multistage analysis

methods for the recognition of the decisions that

control the manipulation and grasping movements of

the artificial hand.

REFERENCES

Wolczowski A., 2001. Smart hand: the concept of sensor

based control. In MMAR’01, 7th IEEE International

Conference on Methods and Models in Automation

and Robotics. Szczecin Univ. of Technol. Publishing.

Su Y., Fisher M. H., Wolczowski A., Bell D. G., Burn D.

J., Gao R. X., 2007. Towards and EMG-controlled

prosthetic hand using a 3-D electromagnetic

positioning system. IEEE Transactions on

Instrumentation and Measurement, (56/1).

Mohammadreza A. O., 2007. Myoelectric control system

– A survey. Biomedical Signal Processing and

Control, vol. 2, issue 4, pp 275-294.

Krysztoforski, K., Wolczowski, A., 2005. Progress of

Robotics - Industrial and medical robotic systems,

A laboratory stand for EMG signal measurements.

WKL. Warszawa.

De Luca, C. J., 2006. Electromyography. Encyclopedia of

Medical Devices and Instrumentation. (John G.

Webster, Ed.) John Wiley Publisher.

Wolczowski, A., Suchodolski T., 2007. Bioprosthesis

Control: Human-Machine Interaction Problem. In

Challenges for Assistive Technology - AAATE 07. IOS

Press.

Wojtczak P., Amaral T. G., Dias O. P., Wolczowski A.,

Kurzynski M., 2008. Hand movement recognition

based on biosignal analysis. Engineering Applications

of Artificial Intelligence, (22/(4/5)).

Mallat, S., 1989. A Theory for Multiresolution Signal

Decomposition, IEEE Trans. Pattern Analysis and

Machine Intelligence, (11).

Zhao, Q. F., 2001. Evolutionary design of neural network

tree – integration of decision tree, neural network and

GA. In IEEE Congress on Evolutionary Computation.

Quinlan, J. R., 1993. C4.5: Programs for Machine

Learning, Morgan Kaufmann. San Mateo, California.

ICINCO 2010 - 7th International Conference on Informatics in Control, Automation and Robotics

326