Crosspoint Switch Based EMG Frontend. for Pattern Recognition Myoelectric Control. RUDHRAM GAJENDRAN B.E., Manipal University, India, 2011 THESIS

Transcription

1 Crosspoint Switch Based EMG Frontend for Pattern Recognition Myoelectric Control BY RUDHRAM GAJENDRAN B.E., Manipal University, India, 2011 THESIS Submitted as partial fulfillment of the requirements for the degree of Master of Science in Bioengineering in the Graduate College of the University of Illinois at Chicago, 2013 Chicago, Illinois Defense Committee: James Patton, Chair Thomas J Royston Levi J Hargrove, Advisor, Rehabilitation Institute of Chicago

2 ACKNOWLEDGEMENTS I would like to thank my thesis advisor Dr. Levi J Hargrove from the Center for Bionic Medicine (CBM) for his support and guidance which helped me think and derive solutions to the challenges I faced over the time of my project. I would also like to thank my advisor at UIC, Dr. James Patton for his advice right from the start of my masters regarding coursework and research which helped me identify my areas of interests and define my career goals. I would also like to extend my appreciation to Dr. Thomas J Royston, department head of Bioengineering at UIC, for supporting the research collaboration between UIC and RIC which was crucial for the successful completion of my thesis. I wish to extend special regards to Dr. Dennis Tkach (currently at VHA Inc., Texas) for his support and advice during the first half of my project. All the weekly meetings, brainstorming sessions and useful criticism have culminated into a research document representing my learning and understanding of the field. I appreciate all the time and efforts of the electronics team at CBM in helping me through system development and debugging. It has been a great learning experience working with them. Last but not the least; I would like to thank Dr. Todd Kuiken, director of CBM for giving me the opportunity to work at CBM. ii

3 TABLE OF CONTENTS CHAPTER PAGE ACKNOWLEDGEMENTS...ii TABLE OF CONTENTS... iii LIST OF TABLES... v LIST OF FIGURES... vi LIST OF ABBREVIATIONS... viii ABSTRACT... ix 1. INTRODUCTION BACKGROUND Control of Upper-Limb Prostheses EMG physiology EMG Signal Detection and Signal Processing Noise consideration Pattern Recognition CONCEPTUAL OVERVIEW Specific Aims Specific Aim 1 Prototype signal acquisition system Specific Aim 2 Evaluate device performance in pattern recognition SYSTEM DESIGN Hardware Crosspoint Switch (XPS) Amplifier High Pass Filter iii

4 4.1.4 Microcontroller PIC Band Pass Filter CAPS Prototype System EXPERIMENTS System Validation Pattern Recognition RESULTS System validation Classifier Performance DISCUSSION CONCLUSION AND FUTURE WORK REFERENCES VITA iv

5 LIST OF TABLES TABLE PAGE Table 1. Electrical Characteristics v

6 LIST OF FIGURES FIGURE PAGE Figure 1. Block diagram of a typical powered prosthesis... 5 Figure 2. Motor Unit... 9 Figure 3. Muscle fiber action potential Figure 4. Composition of the EMG signal Figure 5. Functional block diagram of a typical EMG acquisition system Figure 6. Bipolar electrode configuration (28) Figure 7. Band pass filter Figure 8. Pattern Recognition in TMR subjects. Different hand movements result in different patterns of muscular activity in the pectoral region. (Courtesy: Rehabilitation Institute of Chicago) Figure 9. Stages of a pattern recognition controller Figure 10. Crosspoint Switch Figure 11. Application of the crosspoint switch in routing EMG channels and acquiring multiple bipolar channels Figure 12. Bipolar combinations from four electrodes on the forearm Figure 13. Block diagram of the crosspoint switch based EMG acquisition system Figure 14. Functional diagram of INA Figure 15. CAPS Environment Figure channel prototype Figure 17. Input test signals for validation of crosspoint switch Figure 18. Electrode placement for pattern recognition experiments vi

7 Figure 19. Testing crosspoint switch function. Input signals were switched at 2Hz to verify switch operation Figure 20. (A) Input signals (blue) plotted over reconstructed output signals of channel 1 (red) for visual inspection. (B) Quantization error between the input and the output signals Figure 21. Cross-correlation coefficients between input and respective reconstructed signals. 39 Figure 22. Average classification accuracy and standard error for increasing number number of movements and number of channels vii

8 LIST OF ABBREVIATIONS EMG MEP DOF PR CNS PNS BCI EEG SMR SSVEP LIFE MUAP ADC IA CMRR XPS CAN CAPS LDA Electromyography Myoelectric Prosthesis Degree of Freedom Pattern Recognition Central Nervous System Peripheral Nervous System Brain Computer Interface Electroencephalography Sensorimotor Rhythm Steady-State Visual Evoked Potential Longitudinal Intrafascicular Electrodes Motor Unit Action Potential Analog to Digital Converter Instrumentation Amplifier Common Mode Rejection Ratio Crosspoint Switch Controller Area Network Control Algorithm for Prosthetic Systems Linear Discriminant Analysis viii

9 ABSTRACT Control of myoelectric prosthesis can be achieved in two ways, direct control involving measuring the amplitude of the myoelectric signal or through pattern recognition (PR). PR has the potential to provide a more intuitive control for multi-functional myoelectric prosthesis compared to direct control. Accuracy of PR systems has been shown to improve with increasing number of EMG channels. However, increasing the number of channels requires placing additional electrodes which is not feasible due to space constraints on the residual limb or incorporating multiple hardware components which poses a drawback of increased weight, cost and complexity of the prosthesis. This thesis presents the concept and design of a novel EMG acquisition system to acquire larger number of channels without increasing the number of electrodes placed or the complexity of the signal acquisition system within the prosthetic device. A prototype system was developed and tested to validate performance. Experiments were performed on able-bodied subjects to evaluate system performance in EMG pattern recognition. Subjects were requested to perform nine different hand movements while EMG data was collected into training and test groups. Test results indicate a 15% improvement in classification accuracy with the new system when compared to conventional systems. ix

10 1. INTRODUCTION An estimated 1.7 million people suffer from loss of limbs in the United States (1). Prosthesis is the most effective approach for rehabilitation of individuals that have undergone amputation(s) (2). Prosthesis technology has advanced greatly since the earliest recorded prosthetic leg dating back to 300 B.C. The leg was made of copper and wood and was excavated in Capri, Italy in 1858 (3). Over time, technology has evolved from passive objects to electrically powered devices with multiple degrees of freedom and innovative control strategies. Currently available upper-limb prosthetic devices are 1) Passive limbs; these offer little improvement in functionality and are used primarily for support and aesthetic appeal, 2) Body powered limbs; these have a terminal device (usually a hook or a hand) attached by means of a harness and steel cable to the healthy shoulder to facilitate in opening and closing of the device by shrugging and 3) Electrically powered devices use battery powered motors and actuators to drive joints and are often controlled by electromyographic signals (EMG) generated by the user. This type of prosthetic device is called Myoelectric Prosthesis (MEP) and the EMG result from muscular contractions of the residual limb. These signals can be voluntarily elicited by the user thus making myoelectric prosthesis more intuitive and less stressful to use, especially when using physiologically appropriate muscle contractions (e.g. biceps contractions to control elbow flexion). Benefits of using a MEP are: the user is freed of straps/harnesses, the signal is easy to acquire non-invasively using surface electrodes, the effort required to generate good signals is small and the signals have chronic reliability (4). 1

11 The most common and clinically successful method to control a MEP is to use an estimate of the amplitude of the EMG signals recorded from a pair of the patient s residual agonist/antagonist muscles, and translate it to the appropriate direction of movement of the prosthetic device (5). For example, the EMG amplitude measured from electrodes placed over the flexor compartment of the forearm would be used to control hand closing and the EMG amplitude measured from electrodes placed over the extensor compartment of the forearm would be used to control hand opening. This is termed direct control and is very effective and intuitive for controlling a single degree of freedom (DOF). Multiple degree(s) of freedom can be achieved by using physical switches or muscle co-contraction based mode switches to change functionality. There are limitations in the direct control system; first, the number of signal recording sites is small for amputees thus limiting the degrees of freedom achievable. Second, the use of an external switch to change function is slow and counter-intuitive resulting in frustration of the user and eventual rejection of the prosthesis. Third, since muscular signals are obtained using non-invasive surface electrodes, the recorded signal may contain contributions from neighboring sources in the form of Crosstalk which may cause the wrong movement (2,5). Pattern Recognition (PR) is an alternative control technique that identifies patterns of EMG over multiple channels from different muscle groups and maps them to the respective function/movement of the prosthetic device. PR systems rely on the assumption that the muscle activation pattern generated for a movement is repeatable, unique for that particular movement and can be generated voluntarily by the amputee (6). Many modern PR systems 2

12 have been investigated with high classification accuracies (4-8). PR systems have the following advantages over direct control devices, (i) more intuitive to use because the PR system learns the natural contraction patterns generated by the user for different movements, (ii) more than two degrees of freedom can be achieved, (iii) precise electrode placement is not a critical point for PR as these systems record EMG patterns over multiple electrodes and only require the generated patterns to be consistent and unique for different movements, and (iv) PR systems do not require the signals to be free from cross-talk because these systems do not consider it a contamination but rather an additional component in the complex EMG signal that represents the pattern (2,5,8). Accuracy of a PR classifier is related to several variables including the number of EMG channels and the number and type of movements (2). An increase in the number of active EMG channels has been shown to increase the classification accuracy (2,5,8-12). However, increasing the number of EMG channels calls for an increase in the number of electrodes used to record the EMG signals. Due to limitations in the number of suitable electrode placement sites on the amputee s residual limb and space constraints within the amputee s socket, it is necessary to find a way to obtain maximum amount of useful EMG signals using as few electrodes as possible. The work presented in this thesis is the concept, design and development of a novel EMG acquisition system that can obtain additional EMG signal patterns using only few electrodes placed on the residual limb. The thesis is structured to provide the limited reader a 3

13 complete understanding of the design and development of the EMG acquisition system. Following the introduction, Chapter 2 provides a background on the current state in upper-limb prosthetic control. It also covers the fundamentals of EMG signal generation, acquisition and processing techniques. Chapter 3 explains the concept behind the project and the specific aims. The hardware description is presented in Chapter 4. Experimental protocol to demonstrate proof-of-concept is described in Chapter 5. Chapter 6 provides the results of system validation and experiments. Chapter 7 discusses the implications of the results obtained and Chapter 8 concludes with system limitations and remarks on the future scope/expansion of the project. 4

14 2. BACKGROUND 2.1. CONTROL OF UPPER-LIMB PROSTHESES 8% (41,000) of all amputees in the United States are major upper limb amputees and 32% (500,000) are minor upper limb amputees (distal level amputation). 80% of the people with upper limb amputation are fitted with prosthetic arms; 64% of which are body powered devices and 72% of the rest are electrically powered prosthetic devices (13). Powered prostheses can be controlled in multiple ways such as force sensors, linear potentiometers or more commonly biological signals generated by the user (10). There are two main challenges in control of bio-signal powered prosthetic devices, i) capabilities of the mechanical devices in performing multiple degree of freedom movements and ii) signal processing necessary to translate the user s intent to control commands. Recent technological advancement has led to the development of mechanical arms capable of multiple degrees of freedom and several hand grasps, e.g. OttoBock DynamicArm, touch bionics i-limb, be-bionics etc. However, to control these multi-functional MEPs, we need to accurately extract and decode the signals that represent the user s intent. Figure 1 shows the different stages of a typical bio-signal powered prosthetic device. Figure 1. Block diagram of a typical powered prosthesis 5

15 Critical stages are the signal extraction and signal conditioning which involves successful extraction of the user s movement intent. Once detected accurately, the prosthetic arm can be programmed to perform the desired function. Several signal sources exist that contain information regarding the desired movement. Some of the commonly investigated signals with its advantages and disadvantages are presented in the following paragraphs. The central nervous system (CNS) and the peripheral nervous system (PNS) are commonly tapped to record electrical signals to interface with hybrid bionic devices. Brain Computer Interface (BCI) is a technology that interfaces with the CNS to record the electrical activity of the brain (Electroencephalography, EEG) and use it to control an external device such as a wheelchair, or a prosthetic hand. BCI systems provide the users an alternative communication and control channel that does not depend on the brain s normal output channels of peripheral nerves and muscles (14,15). EEG can be recorded either non-invasively by placing electrodes on the scalp or invasively by surgically implanting micro-electrodes in the brain. Wolpaw and MacFarland demonstrated multidimensional control of a cursor on a computer screen using Sensorimotor Rhythm (SMR) based BCI (16). A study by Muller-Putz and Pfurtscheller demonstrated that it is possible to asynchronously control a neuroprostheses using steady-state visual evoked potentials (SSVEP) with an accuracy of upto 88% (15). Although acceptable results have been obtained in laboratory setting, noninvasive BCI systems are far from being clinically viable solution due to the long training period involved, noisy signals, slow communication speed, electrode and skin problems with long recording time and the requirement for controlled attention. Invasive BCIs on the other hand offer improved 6

16 classification, success rates and overall performance. However, these systems have only been tested on small groups because of unwillingness of subjects to participate in invasive BCI research (17). BCI is an active area of research and has the potential to successfully provide rehabilitation for highly disabled individuals (e.g. complete spinal cord injury); the current systems however, do not provide clinically acceptable results for amputees. Recent developments in technology have enabled us to interface with the PNS for neural signal recording to control artificial prosthesis. Cuff electrodes, which are extra-neural electrodes that surround the epineurium (outermost layer of dense connective tissue surrounding a peripheral nerve), provide relatively low-risk, reliable chronic interface to the PNS (18). Cuff electrodes have the advantage that they minimize mechanical damage to the electrodes and have been observed to be effective for extended periods of time. An important limitation of these however, is the small number of contacts to the nerve that limits the resolution of the signals recorded (19). Intrafascicular electrodes are placed inside the nerve fascicles to improve selectivity of the recordings. Longitudinal Intrafascicular Electrode(s) (LIFE) are small sized electrodes that are inserted into the nerve fascicle to record signals from upper limb amputees to decode motor intent of the user (18,20). Dhillon et al. suggested that LIFEs could provide amputees with prosthetic devices that feel more natural to use than the current systems (21). Both, cuff and intrafascicular electrodes pose a risk of infection due to the invasive application. While these methods seem promising, they are still in the research phase and are not ready for clinical implementation. 7

17 Electromyographic (EMG) signals representing the neuromuscular activity of the contracting muscle(s) are the most commonly used signals for control of upper limb prosthesis (4). EMG can be recorded non-invasively by placing electrodes on the surface of the skin over the muscle belly (surface EMG) or invasively by inserting fine wire or needle electrodes into the muscle using guide wires. The main advantage of surface EMG is easy setup and non-invasive access to physiological signals. Intramuscular electrodes inserted into the muscle provide more reliable and clean signals but since prosthetic systems require long term signal recording, these electrodes pose a risk of infection, tissue damage or pain and discomfort (18,20). This project will focus on using non-invasive EMG. Clinical advancements in prosthetic control include Targeted Muscle Reinnervation, a surgical procedure conceived by Dr. Todd Kuiken at the Rehabilitation Institute of Chicago which involves surgically rewiring the residual nerve ends at the site of amputation to new muscles in the chest, thighs or laterals. High level amputees with no muscle sites benefit from TMR surgery such that, the new muscle sites act as effectors of the signals that were intended for muscles of the lost limb. As a result thoughts corresponding to movement of the lost limb result in EMG at the new muscle sites (pectorals, laterals or quadriceps) that can be used to control a prosthetic arm (22) EMG PHYSIOLOGY EMG is an electrophysiological signal that is generated by electrochemical activity within the body (23). Skeletal muscle is composed of specialized cells called muscle fibers that are 8

18 capable of contraction and relaxation. The contraction of the muscle is initiated by a biochemical stimulus from the motor neurons that innervate these muscle fibers. Motor neurons are efferent neurons that originate in the spinal cord and innervate muscle fibers to facilitate muscle contraction. Typically, the motor neuron along with its branches innervates more than one muscle fiber and the group collectively forms the smallest functional unit of muscular contraction called a motor unit as shown in figure 2. Figure 2. Motor Unit Stimulus from the motor neuron causes the muscle fiber to depolarize and twitch. This depolarization, accompanied by a movement of ions in and out the cell generates an electrical impulse termed as muscle fiber action potential, figure 3. 9

19 Figure 3. Muscle fiber action potential A motor unit comprises of several muscle fibers and the summation of the action potentials from all of the muscle fibers in a motor unit is called motor unit action potential (MUAP). Depending on the size, a muscle usually contains a few to several thousand motor units. The EMG signal detected at the skin s surface is the superposition of MUAPs from all the active motor units in the detection volume of the electrodes. Figure 4 shows the block diagram of the composition of the EMG. S i is the single fiber action potential of the i th muscle fiber in the motor unit and the summation of all the action potentials in the motor unit gives M i, the motor unit action potential. The superposition of all the MUAPs is the raw EMG signal recorded at the surface (24-26). 10

20 Raw EMG Figure 4. Composition of the EMG signal 2.3. EMG SIGNAL DETECTION AND SIGNAL PROCESSING As discussed previously, EMG may be measured invasively using fine-wire intramuscular electrodes or non-invasively using surface electrodes placed on the skin. When measured at the surface, EMG normally has peak-to-peak amplitude of 1 10 mv and lies in the frequency range of Hz (25). The block diagram of a typical signal acquisition system for obtaining clean and useful EMG signal is shown in Figure 5. 11

21 Figure 5. Functional block diagram of a typical EMG acquisition system The quality of an EMG measurement strongly depends on skin preparation and electrode positioning. Before electrode placement, the skin is usually abraded with a cleansing medium to remove the dirt and oil to ensure low electrode-skin impedance. Skin impedance of < 5kΩ is suggested to ensure good quality EMG signals (24,27). It is important for the impedance to remain consistent over the duration of the measurement. EMG signals are detected via surface electrodes placed longitudinally over the muscle belly on the skin, usually 2.5cms apart. The voltages are measured with respect to a reference electrode placed at a site away from the EMG electrodes over electrically neutral tissue such as the elbow, wrist or other bony surfaces (28). For myoelectric control applications, EMG electrodes are most often placed in a bipolar configuration (29,30) as shown in figure 6. The signals from the two electrodes placed close to each other (referenced to ground) are passed through a differential amplifier which amplifies the difference between the two signals thus removing the noise that is common to both the electrodes. Other configurations, such as 12

22 monopolar or multi-polar, are sometimes used in research studies but have not been implemented clinically (25), (29). Figure 6. Bipolar electrode configuration (28) Differential amplification of the signal is achieved using an Instrumentation amplifier (IA). The ability of the IA to remove the noise is measured in terms of the Common Mode Rejection Ratio (CMRR). CMRR is therefore an important factor in choosing an instrumentation amplifier for the system hardware. After the differential amplification stage, the EMG is relatively free of 50-60Hz electrical noise but may still contain additional noise from DC offsets, electrode shifts, or interference from computers, radio and cellular phones which can be removed by filtering. A Band pass filter is used to remove the low frequency motion artifact, DC offset and other irrelevant frequency components from the EMG, leaving only the desired part of the signal. A lower cutoff frequency of Hz and a higher cutoff frequency of Hz are normally used to design the band pass filter (Figure 7). A band pass filter includes a low pass 13

23 and a high pass filter constructed in series using discrete resistor and capacitor components. The lower and higher cutoff frequencies are set according to equation 1. Where is the cutoff frequency, R and C are the resistor and capacitor values used in the circuit. Figure 7. Band pass filter The low pass region of a band pass filter also acts as an anti-aliasing filter. An anti-alias filter restricts the frequency content of the signal below a certain cutoff so that the signal can be sampled at a desired rate without aliasing. Aliasing is a common signal processing artifact in which improper sampling rate results in loss of essential EMG information. Most clinically implemented acquisition systems use a cutoff frequency for the anti-alias filter of 500 Hz (31). 14

24 After filtering, the signals are amplified again, sampled and converted to digital data that is transmitted to the computer for display, analysis and pattern recognition NOISE CONSIDERATION EMG recorded over the surface is a weak signal and is easily contaminated by noise and disturbances. Reduction of noise in surface EMG is mandatory to extract the information regarding the neural intent of the patient. In order to fully understand the techniques in acquiring clean EMG, it is important to understand the noise and their respective sources. The common noise sources in EMG are, i. Power line interference is the electromagnetic interference that arises primarily from the electrical mains. The main frequency content of this noise is 50/60 Hz depending on the frequency of the supply. Notch filters can remove the Hz noise but it is not recommended because the information density of EMG is high in that frequency band. Other sources of electromagnetic interference include lamps, computers, electric fans, cellular phones or any other electronic equipment in the surrounding area. This noise can never be fully removed but can be reduced significantly with appropriate electrode configuration and signal processing. ii. Motion artifact due to the movement of electrodes at the electrode-skin interface or movement of leads can cause low frequency noise in the recorded EMG signals. Current technology coupled with good circuitry is capable of removing noise due to movement. 15

25 iii. Thermal noise or more commonly white noise is random in nature containing a wide range of frequency components. It is caused due to the random irregular movement of electrons (Brownian movement) within the conducting medium including electrolytes. Thermal noise does not pose a serious problem for recording as the system parameters can be adjusted to acquire signals only in the desired frequency band thereby increasing the signal to noise ratio (25). iv. Chemical noise is generated at the interface of electrode and skin due to reaction of ions at the electrode-electrolyte or electrolyte-skin interface. This is an inherent source of noise that can be minimized by appropriate skin preparation before electrode placement (32). The human body is a conducting medium. Depending on the impedance on the surface of the skin and from skin to ground, potential difference between the body and the ground or between two points on the body induced from the surrounding can be high (24,33). Placing a ground electrode on the body provides a low impedance pathway between the body and the ground for the induced voltages to flow thereby reducing the contribution from body induced noise. Furthermore, an important factor for noise reduction is the matching of impedance at the electrode-skin interface between two recording sites. A difference or fluctuation of 16

26 impedance can result in inflow of common mode current that will induce a voltage which will be amplified greatly resulting in very high common mode noise thus significantly reducing the SNR. Therefore, bipolar configuration is commonly used PATTERN RECOGNITION PR systems utilize multiple EMG sites, feature extraction and multidimensional classifiers to map a pattern of muscular activity of a group of muscles (recorded via spatially arranged electrodes) to the respective movement that generated such a pattern (4-12,34-38). Figure 8 shows the different patterns of muscular activation for different hand movements in a TMR subject. Figure 8. Pattern Recognition in TMR subjects. Different hand movements result in different patterns of muscular activity in the pectoral region. (Courtesy: Rehabilitation Institute of Chicago) 17

27 The controller for a PR based system consists of the following stages, (Fig. 9) Data windowing Feature extraction Classification Figure 9. Stages of a pattern recognition controller Data windowing is the first step of a PR algorithm which involves extraction of a small segment of data from the recorded EMG. The classification decision is made by analyzing individual data windows. Increasing the window length stabilizes the features for classification and therefore improves accuracy; but it results in increased processing time for the controller that causes high user perceived delays. A study by Smith et al. determined the optimum window length for PR to be between 150 and 250 ms (39). Feature extraction is performed to increase the density of relevant information in the EMG data stream to ensure distinction between classes. Selection of feature(s) plays a major role in classification accuracy. In fact, performance of a classifier is more closely related to the choice of features than the classifier itself (5,7,40). Several Time Domain (TD) and Timefrequency methods have been investigated to determine the ideal set of features to represent 18

28 the signal. TD features such as mean absolute value (MAV), Willison amplitude (WAMP), zero crossing (ZC), variance of EMG, auto-regressive (AR) model are quick and easy to calculate and are therefore widely used in research. Time-frequency representations (TFR) presents a more accurate description of the signal in both time and frequency domains but is computationally heavier and therefore requires more processing time (31,41). A study comparing the feature sets observed that a combination of TD and AR feature sets outperforms all the other feature sets (4,5,39). The classifier recognizes the patterns embedded in the extracted feature set and classifies the signal into the respective movement class. Commonly used classifiers include Bayesian pattern classifier, Artificial Neural Network (ANN), Fuzzy logic, Gaussian Mixture Models (GMM), Linear Discriminant Analysis (LDA), Support Vector Machines (SVM) and Hidden Markov Model (HMM). While the performances of most classifiers for EMG signals have been observed to be similar, with appropriate signal representation linear classifiers have been reported to perform better than non-linear classifiers (8,31). defined as, Performance of a classifier is measured in terms of its classification error (or accuracy), Classification error is affected by several factors such as window length, feature selection, number of EMG channels and number and type of movements attempted to classify. 19

29 Generally, increasing the number of channels reduces classification error and increasing the number of movements increases the classification error (2). Electrodes are usually longitudinally oriented with respect to the muscle fibers, spaced approximately 2cm apart and are fed into a differential amplifier to obtain bipolar channels. Young et al. observed that including the transverse channel in addition to the longitudinal EMG channel significantly improved the classification accuracy (9). Increasing the number of channels either requires placing additional electrodes as mentioned previously or attaching multiple parallel leads to the existing electrodes and incorporating additional amplifiers in the signal acquisition system. Including additional hardware components and lead wires adds weight, cost, and complexity to the system and increases the risk of wire breaking which is a common cause of failure of myoelectric prosthesis (35,42). Furthermore, it is not clinically feasible to incorporate numerous electrodes in the prosthetic socket. There are fewer active muscle sites on a residual limb and the electrode placement sites must not be over the scarred tissues. Therefore, a system that can acquire multiple bipolar channels without increasing the number of electrodes placed or the number of leads attached to the electrodes would be more efficient and feasible to implement clinically. A simple approach to increase the number of channels is to uniquely combine different electrodes on the skin to obtain all the bipolar combinations. The additional channels obtained can potentially improve classification accuracy. This project is the realization of a concept to obtain all possible bipolar combinations from monopolar electrodes without using additional leads or electrodes. 20

30 3. CONCEPTUAL OVERVIEW A Crosspoint switch (XPS), sometimes called a switch matrix, is an analog device that connects M inputs to N outputs in the form of an array such that any of the M inputs can be connected to any of the N outputs. The device has M x N intersections, called crosspoints where an input line and output line can be connected as shown below in Figure 10. It is commonly used for network switching, parallel computing and Audio/Video telecommunication purposes where one signal needs to be output to multiple appliances. Compact, low cost, power efficient XPS ICs are now available that can be reconfigured digitally using embedded microprocessors. Figure 10. Crosspoint Switch The idea is to use an XPS for routing EMG signals to obtain all possible bipolar combinations from the monopolar electrodes placed on the skin. Raw EMG signals from four electrodes placed on the forearm form the input to a 4 x 4 Crosspoint switch. The outputs of 21

31 the XPS are combined at the two IAs to obtain bipolar signals which are sampled at the ADC. When triggered, the XPS continuously reconfigures the signals at the output channels thus enabling us to combine two different signals to obtain unique bipolar channel(s) as illustrated in figure 11. By switching and sampling fast enough, all the bipolar combinations can be acquired without loss of data or signal quality. Figure 11. Application of the crosspoint switch in routing EMG channels and acquiring multiple bipolar channels. Number of unique bipolar channels that can be generated from m number of electrodes is given by, With 4 electrodes, the maximum number of bipolar channels that can be obtained is six. Figure 12 shows the different bipolar channels that can be obtained from 4 electrodes placed 22

32 on the forearm. As opposed to conventional systems that generate two bipolar channels from four electrodes, the proposed system can obtain up to 6 channels. Figure 12. Bipolar combinations from four electrodes on the forearm. 3.1 SPECIFIC AIMS This is a preliminary study to demonstrate the proof-of-concept of a system that will allow for the creating of multiple bipolar EMG channels from few electrodes. The goal of this project was to design the proposed system and test its performance on able-bodied subjects in classifying different movements SPECIFIC AIM 1 PROTOTYPE SIGNAL ACQUISITION SYSTEM Design and develop a prototype signal acquisition system incorporating a 4 x 4 crosspoint switch. Validate system performance by, (i) Passing sinusoidal signals of known parameters as inputs and observing the output signals to determine effects of crosspoint switching on signal quality. 23

33 SPECIFIC AIM 2 EVALUATE DEVICE PERFORMANCE IN PATTERN RECOGNITION Conduct experiments on able-bodied subjects to evaluate the performance of the proposed system in pattern recognition. EMG data will be analyzed in the following conditions; (i) Device with crosspoint switch (a) 6 bipolar channels, (b) 4 bipolar channels and (ii) Device without crosspoint switch - 2 bipolar channels. 24

34 4. SYSTEM DESIGN This chapter discusses the design considerations for the acquisition system. It provides a brief description of the different components used in the different stages of the system. 4.1 HARDWARE Figure 13 shows the block diagram of different stages of the proposed system. The hardware was developed with the following goals in mind: Extract millivolt level EMG signals from flexor and extensor muscles of the forearm with minimal distortion. Switch the signals at high speeds without affecting the signal quality and obtain six bipolar EMG channels. Figure 13. Block diagram of the crosspoint switch based EMG acquisition system 25

35 4.1.1 CROSSPOINT SWITCH (XPS) The crosspoint switch is the most crucial component of the system. The desired characteristics of the XPS include, High speed switching ( > 10Khz) Low settling time Low crosstalk SPI/Parallel communication Low power consumption Breadboard compatible for testing M22100 (ST Microelectronics), a 4 x 4 XPS with memory was chosen for this project. Simple design, easy control, high switching frequency of 1.2MHz, low propagation delay (<75 ns), low power consumption (200 mw) and DIP packaging for breadboard use were some of the striking features of the device. The Switch is controlled via parallel communication using an embedded microcontroller. Power to the XPS is provided using a 5V regulated DC power supply AMPLIFIER The instrumentation amplifier is the primary noise limiting component of the signal acquisition system. After researching potential options for IA, the INA 129 instrumentation amplifier (Analog Devices) was selected. Figure 14 shows the functional diagram of the INA 129. The first stage of the IA creates a bipolar signal by removing most of the signals common to the electrodes of the bipolar montage and the second stage amplifies the signal. Features such as 26

36 low offset voltage (50 µv), high CMRR (>120 db), high input impedance and programmable gain of make the INA 129 an excellent choice as a front end IA. Figure 14. Functional diagram of INA 129 The programmable gain of the IA is given by the following equation, Where Rg is the resistor placed between pin 1 and 8 of the IA (figure 14). An adjustable gain of approximately 200 was set using potentiometers in place of Rg across pins 1 and 8. The INA 129 requires a power of +/- 5V that is provided using a regulated power supply. Two INA 129s have been incorporated in the design to create two bipolar channels from the four outputs of the XPS. 27

37 4.1.3 HIGH PASS FILTER A passive high pass filter was incorporated as part of the IA programmable gain circuit with a cutoff frequency of 20Hz to filter out the DC offset that is added to the signals from the skin and XPS. A 22 µf capacitor was placed in series with the resistor Rg of the IA to create the high pass filter with a cutoff frequency of 20 Hz MICROCONTROLLER PIC 32 USB32 Bit Whacker (UBW32) development board consisting of a PIC 32 microcontroller is the control unit for the system. Firmware for the microcontroller was developed to periodically reconfigure the outputs of the switch, sample EMG signals and transmit the digital data to a computer. Any desired channel of the XPS can be opened or closed in software using commands from the PIC 32. The microcontroller was programmed to reconfigure the XPS to three different configurations. The microcontroller comprises of an on board 10-bit ADC. EMG from the XPS is sampled at 3000Hz to maintain signal fidelity and satisfy the sampling theorem. The ADC has an analog voltage range of V. Therefore, an offset of 1.5V was added to the output of the IA. The step size of the ADC is given by, Controller area network (CAN) is the network protocol used for communication between the microcontroller and the host computer. CAN messages are device specific and 28

38 consists of an ID associated with the respective device. Sampled signals are encoded as CAN message with appropriate device ID and transmitted to the computer via USB communication. Data was transmitted at a rate of 1000 sps BAND PASS FILTER Digital band pass filters were implemented in the firmware to filter out the signal outside of the desired EMG band. A sixth order band pass filter with a lower cutoff frequency of 20 Hz and a higher cutoff frequency of 350 Hz was implemented to increase the SNR of the acquired EMG. In addition to the band pass filtering, a fourth order high pass filter with a cutoff frequency of 100 Hz was implemented. This removes all remaining power-line interferences and increases the EMG signal quality. 4.2 CAPS Control Algorithm for Prosthetic Systems (CAPS), a custom software developed at the Center for Bionic Medicine is used for signal acquisition, recording and pattern recognition (figure 15). Some important functions of CAPS in this project include, EMG signals embedded in the CAN message is decoded and displayed on a GUI for visualization and recording. The digital filters are implemented using inbuilt filtering functions. Pattern recognition is performed in CAPS. Window length, feature sets and classifiers can be selected appropriately for classifying the EMG signals into different movement classes. 29

39 Virtual reality can be used to observe the classifier performance in controlling a virtual prosthesis. Unique prosthesis configuration for different users can be created and stored for future use. Figure 15. CAPS Environment 30

40 4.3 PROTOTYPE SYSTEM Figure 16 shows the final version of the prototype EMG acquisition system (protoboard design). Labeled are the different parts as described in the previous section. Several versions of the device were tested on the breadboard before arriving at the final design that was soldered on to a protoboard. The system is powered using a regulated +/- 5V power supply. Table 1 outlines the electrical characteristics of the system. PIC 32 UBW32 Development board Crosspoint Switch (M22100) Instrumentation Amplifier (INA 129) Figure channel prototype 31

41 Table 1. ELECTRICAL CHARACTERISTICS Parameter Power supply Input current Power dissipation Sampling frequency Value +/- 5V 150 ma 750 mw 3000 Hz Switching time (4 channels) 80 µs Data Transfer rate to PC 1000 sps 32

42 5. EXPERIMENTS This chapter is divided into two sections; the first part discusses the testing and validation of the designed system, the second part describes the experiment carried out with able-bodied subjects. The objective of the experiment was to measure the performance of the PR classifier using the proposed system and compare it to the performance of a conventional signal acquisition system that does not use a crosspoint switch. 5.1 SYSTEM VALIDATION Initial tests were done to verify operation of the XPS. Sinusoidal and square inputs were switched at a rate of 2 Hz to visualize switching on CAPS. Following system design, in order to validate the functionality, four low amplitude (20-40 mv) signals from a function generator were input to the system. Signals selected were 100 Hz sinusoid, a 50 Hz triangle, a 50Hz square and a 0V constant DC (figure 17). The signal values were chosen to approximate the scale and frequency of physiological EMG. For this part of the experiment, the signals from the XPS were not passed through the IA and were not filtered. The PIC32 was programmed to reconfigure the XPS such that each of its output channels contains all the 4 input signals. The signals from the XPS were switched and sampled at a rate of 4 khz. The signals at each output channel were separated into its constituent signals, reconstructed and compared to their respective inputs to ensure that no signal distortion is caused due to high speed switching. 33

43 Figure 17. Input test signals for validation of crosspoint switch. 5.2 PATTERN RECOGNITION Three able-bodied subjects participated in an IRB approved experiment to compare the performance of the prototype system (6 EMG channels) to the system without the XPS (2 channels) and observe the trend in classification error with increasing number of channels and movements. To ensure stable electrode contact and low skin impedance, the skin surface was cleaned using alcohol swabs. Two pairs of electrodes were placed longitudinally over the flexors and the extensor muscle groups with an inter-electrode distance of 2.5cms. A ground electrode was placed away from the muscles of interest on the wrist, over the pisiform carpal bone (figure 18). Standard Ag/AgCl disposable pregelled electrodes were used to record surface EMG signals. Custom leads were attached to the electrodes and connected to the input of the XPS. 34

44 Figure 18. Electrode placement for pattern recognition experiments Data were collected for nine movement classes: wrist flexion, wrist extension, wrist supination, wrist pronation, power grip, hand open, chuck grip, fine pinch and relaxed (no movement) class. Subjects were requested to perform the set of movements presented to them on the computer screen. Each contraction was held with a normal force for three seconds. A resting time of three seconds was provided between different contractions. Four sets of data were collected for training and testing the classifier. The acquired data was grouped into training and testing data. CAPS was used to record the data and perform offline pattern recognition. Acquired data were segmented into 250ms windows with 25ms increments which have been shown to be well-suited for pattern recognition (39). The signals in each window is 35

45 represented using the following TD features, mean relative value, zero crossings, slope changes and waveform vertical length. Selection of features was based on a previous study by Englehart et al. that demonstrated superior performance of time-domain features in classification. An LDA classifier was chosen because of its ease of implementation and training (8). Classifier was tested for four different number of movements (three, five, seven and nine movement classes) and three different number of channels; two, four and six. The first two channels were created by combining the two longitudinally placed electrodes along the flexors and the extensors. The four channel system comprised of the previous two channels plus the transverse pairs formed across the forearm and finally the 6 channel system consisted of all the possible combinations. Classifier error rate was calculated by four folds validation for increasing number of channels and increasing number of movements. 36

46 6. RESULTS 6.1 SYSTEM VALIDATION Figure 19 shows the signals switched at 2Hz. Output signals were visually inspected for glitches and switching errors. Figure 19. Testing crosspoint switch function. Input signals were switched at 2Hz to verify switch operation. Tests with basic signals revealed functional limits of the XPS. Switching time as low as 80 µs was achieved without signal distortion. Higher switching speeds resulted in waveform distortion by the XPS. Signals at the output of the XPS were reconstructed and compared with the input signals. Visual inspection revealed nearly perfect reconstruction indicating that XPS does not affect signal quality (figure 20 (A)). Quantization error due to poor resolution of the 37

47 ADC (or low value of the input signals) was represented by slight mismatch between the reconstructed output and the corresponding input signal (figure 20 (B)). Error was observed to be 3mV, which corresponds to the step size of the ADC. (A) (B) Figure 20. (A) Input signals (blue) plotted over reconstructed output signals of channel 1 (red) for visual inspection. (B) Quantization error between the input and the output signals. 38

48 Quantitative measure of similarity between input and reconstructed output was obtained by means of correlation coefficient. r values of the four input signals compared to their corresponding signals reconstructed from output channel 1 of the XPS was calculated to be 0.97, 0.92, 0.93 and 1.0 respectively (Figure 21). Figure 21. Cross-correlation coefficients between input and respective reconstructed signals The r values indicate a nearly perfect reconstruction of the input signals at the output of the XPS at switching speeds of 4 khz. This validates the function of XPS in switching the signals at high speed without distorting the signals and therefore can be incorporated in the system in synchronization with the IA and ADC to simultaneously switch and sample to obtain all bipolar combinations. 39

49 6.2 CLASSIFIER PERFORMANCE Performance of the classifier for increasing number of EMG channels was evaluated for four different movement sets. Classification error rates calculated by four folds validation averaged over three subjects along with the standard error are shown in Figure 22. The trend observed is consistent with previous studies which state that classification error rate decreases with increasing number of channels and increases for higher number of movement classes (2,9). Reduction in classification error rate when increased from 2 to 4 and 6 channels was observed to be highest in the 5 movement class. Improvement in classification accuracy was observed to be same for the 7 and 9 movement classes (10% and 14% reduction in error when number of channels was increased from 2 to 4 and 6 respectively). An average reduction of 15% was observed in classification error rates when increased from 2 to 6 channels for all movement classes. While the overall classification error was high with more movement classes, there is a notable improvement in performance when increasing the number of channels from 2 to 6 thus validating the purpose of the crosspoint switch based EMG acquisition system. 40

50 Figure 22. Average classification accuracy and standard error for increasing number number of movements and number of channels. 41

51 7. DISCUSSION The goal of this project was to devise a technique to acquire multiple EMG channels without having to place additional electrodes on the skin or attach additional leads. The work presented in this thesis is the proof-of-concept of a novel EMG acquisition system capable of achieving this goal. The novelty in the proposed system is to incorporate a simple crosspoint switch to reconfigure outputs fast enough to acquire all the bipolar combinations within the sample duration while maintaining sufficient signal fidelity for pattern recognition. This is the first study exploring the use of a crosspoint switch in routing biological signals. Increasing the number of EMG channels improves classification accuracy in myoelectric PR until a point beyond which the benefits of adding additional EMG channels are diminishing (8). These studies however, acquired additional EMG channels by using high number of electrodes and leads. Using multiple leads in parallel from the same electrodes or placing additional electrodes over the residual limb requires complex signal acquisition systems with multiple amplifiers for signal conditioning which makes the system more expensive, heavy and bulky. In addition, the presence of loose wires from electrodes adds to the inconvenience in use and increases the risk of broken wires. One other alternative is to acquire single ended EMG from electrodes and combining them digitally to obtain desired bipolar channels. This approach however is susceptible to common-mode noise and therefore requires complex signal processing equipment to maintain signal quality. The proposed method mitigates the need to place additional electrodes or attach additional leads in parallel. 42