With their prehensile mechanism, aesthetics, and simple interface, myoelectric prosthetic hands are useful tools for upper limb amputees. The use of a myoelectrically controlled prosthesis requires the clinical team to be able to set up the prosthesis correctly, and the user to be taught to exploit the device effectively. Reliable prosthetic components and their successive clinical fitting are the two of the four factors that Heckathorne1 described as the explanation for the recent growth of powered component usage, although supporting evidence for this is sparse.
The reliability of the control interface is a key feature in the everyday use of a device. Therefore, the strategy by which the clinical team approaches the assessment and fitting of the myoelectric controller to allow it to be reliable and durable is an important part of the process. The characteristics and signal processing of myoelectric signal and myoelectric controller designs are described by Lovely2 and Childress and Weir,3 respectively. The myoelectric sensor has been studied for many years and can be a useful control interface for prosthetic limbs, but it is also known to be very sensitive to external interference, and the signal can unexpectedly change. Previous research by Roy et al.4 and De Luca5 describes the factors that influence the myoelectric electrodes. Based on this knowledge, the methodology known as “Robust Engineering”6,7 was applied to setting up of myoelectric controllers. It aimed to investigate the relation of different factors considered while installing a myoelectric controller for the generation of simple ON/OFF control signals. Eight factors which influence the fitting of surface myoelectric sensors, including sensor position and activation threshold, were selected, and a multifactorial experiment was conducted on a single able-bodied subject.
Robustness can be defined as designing a product in such a way that the level of its performance under various customer usage conditions is the same as under nominal conditions. Robust Engineering methods, also known as Taguchi Methods, are intended as cost-effective methods to improve the performance of a product by reducing the variability in customer usage conditions. The method has shown success in many industries from major automobile to biochemical companies.
The method of Robust Engineering was applied to assess the relationship of the factors that affect the stability of a myoelectric controller. Robust Engineering provides a framework to create an experimental design and data analysis for examining the key features that influence the target system's stability. The system is modeled to have a linear relation between input signal and output response. Added to this simple model are two factors—control factors and noise factors—that may influence the stability of the system. The control factors are nominal system parameters that contribute to the operation of the system. The noise factors are different forms of uncontrollable factors that impede successful operation, such as various usage conditions, deterioration and wear, and individual difference. Fractional factorial experiments are designed to investigate the relationship between the constants selected for the control factors under the influence of the noise factor on operation. To run the experiment with optimum combinations of numerous control factors, orthogonal array is applied to estimate their effect on variability. Orthogonal array is an inspection device, which was generated from an orthogonal Latin square. In our case with eight control factors, the L18 orthogonal array was chosen. Eighteen experimental runs with a combination of systematically varied parameter levels were conducted. Finally, the collected experimental data were analyzed to investigate the variation of the relationship between the input and the output.
The purpose of this activity was to demonstrate that the technique of Robust Engineering could uncover the factors of importance and the optimum parameter constant in the setting up of an electromyography (EMG) control channel. Once it was shown that it could do so with existing controllers where the requirements for setting up the channel are known, it could then be confidently applied to controllers for different individuals, or on a newer controller where the flaws were as yet undiscovered.
For this investigation, the basic function of the myoelectric controller was defined as the relation between the frequency of the intended prosthesis movement, M, and the frequency of the controller's output, y. In the ideal system, the output and input match, so the coefficient of the relationship between M and y is 1. If the controller cannot react as quickly to a change in input, the relationship moves away from the ideal and this value drops. For the purpose of this trial, the EMG controller was taken as one with switch thresholds (so-called digital twin).
Eight control factors were identified: 1) electrode contact pressure, 2) displacement of the electrode toward long and 3) short axes, 4) EMG sensor orientation, 5) passband frequency, 6) cutoff frequency of the low-pass filter, 7) envelope size, and 8) activation threshold (see Table 1 for description of the control factors and the selected levels).
Three items at two levels were selected as noise factors: with or without moisture (tap water) on the skin at the electrode site, severe or light muscle fatigue, and arm posture (hanging down or in front reaching). To reduce the number of experimental runs, a compounded noise factor with four levels was created by using an orthogonal array on the three noise factors.
Three-pole dry myoelectric sensors with differential amplifier (Personal-EMG, Oisaka Electronic Device Ltd., Japan) were used as test instrument in the experiment (sensor size: 11.9 × 19.0 mm2, CMR 104 dB). The system is capable of simultaneous recording and monitoring of two myoelectric signal channels with notch filter of 60 Hz, (local power line frequency). The myoelectric signals were sampled at 3 kHz with 12-bit AD converter and recorded. The signals were processed offline with a low-pass IIR filter, envelope algorithm, and a discriminant function with activation thresholds.
Informed consent was obtained from an able-bodied 22-year-old male subject. The subject had no experience with myoelectric control in advance. His left arm, which is his nondominant side, was used as the signal source. The myoelectric sensors were placed on the skin surface above the extensor digitorum and flexor digitorum superficialis and fixed with adhesive tapes and wrapped with an elastic band. Before each trial, each time the sensor was placed on the skin, the signals were confirmed to be distinct. Additionally, the maximum voluntary contraction was recorded before and after each test.
The speed of the movement (frequencies of 45, 60, 75, and 90 changes per minute) were presented as beep tones from a digital metronome, the subject having to react each time he heard the tone. The myoelectric signals were visually fed back to the subject through a monitor throughout the test. The trials were conducted with 15-second recording with a minimum of 1-minute rest between runs.
The protocol for the trial is described in Figure 1. Each trial started with a time keeper counting down 5 seconds. The subject was asked to move through the four states in turn—flexion-neutral-extension-neutral—each time he heard the tone. The conditions of the myoelectric sensor were only changed on the flexor side.
With the 18 experimental runs for each four movement frequency and four levels of the compounded noise factor, a total of 288 experimental trial data were collected. Following the trial data analysis, 32 trial data were additionally collected to confirm the derived optimal combination of the control factor levels.
The maximum activation cycle times computed from each output response gives the response frequencies. Then, signal-to-noise (SN) ratio and sensitivity were computed with the functions provided in the zero-point proportional equation for a dynamic problem. Dynamic problem are characterized by the presence of signal factor M and the response variable y, and the relation between y and M is given by the following equation.
where sensitivity β is the slope of the best-fit line between M and y. The SN ratio,η, is obtained by the next equation, where σ2 is the variance of the unpredictable and undesirable part of y, and β2 is the variance of the predicable and desirable part except for the range of M.
The logarithm of the results for each factor level was plotted in decibels. Figure 2 shows the effect of the factors on the SN ratio, and Figure 3 depicts the sensitivity. SN ratio is the reciprocal of error variance, whereas sensitivity is the coefficient of the input-output relation, thus, higher the SN ratio value indicates better performance.
After the initial tests, a confirmation test was performed. Data were collected for the initial configuration (A1B1C1D1E1F3G2H3) and the optimized configuration (A1B1C1D1E2F2G1H2). The letters refer the factors on Figures 2, 3. The results of the initial configuration were a SN ratio −10.4 and sensitivity 0.7, and optimized configuration SN ratio −8.0 and sensitivity 0.5.
As Figure 2 shows, the factors D and B, which are the sensor orientation and displacement in the short-axis direction have the best (highest) and the worst (lowest) signal to noise ratio. Whereas, the factors A and G (contact pressure and the size of the envelope window) had the smallest difference among the factor levels for both SN ratio and sensitivity. These results show that sensor orientation requires the greatest care and fine tuning during the setting up of the myoelectric sensor. Similarly, contact pressure, unless it becomes zero, requires much less attention from a signal acquisition perspective.
The impact of sensor displacement in short-axis direction, resulted in a “V” shaped curve, this result was unexpected. This is likely to be related to factors that were not considered in this test, e.g., electrode dimension, sensor initial relative position on the target muscle, and subcutaneous muscle movement. Furthermore, the results of the envelope window size need additional testing to determine whether this factor has interaction effect with others, or if the result is dependent on the narrow range of selected level values. Regarding the other control factors, the range of SN ratios are too similar and no major priority among the factors can be determined. However, from the control factor values the highest SN ratio can be chosen to improve the robustness of the sensor. If the indicated differences are small, then other criteria must be used, such as sensitivity or economic reasons for assembling the controller.
The sensitivities in Figure 3 show that the all results exist near 0 db. This indicates that the slope of the input-output relation model is approximately 1. The positive to negative decibel changes within the levels, which are seen in C, D, E, F, and H, signify that these are the superior factors for adjusting the setting to the ideal condition. When comparing the variance among the factors, E, F, and H are larger then the others. This shows that the low-pass filter design and the activation threshold value have larger leverage effect for tuning the response; it therefore requires caution when the parameters are selected.
When taken together, the sensor should first be adjusted to the optimal orientation and then the activation threshold should be tweaked to gain a certain tolerate level. If needed to superbly tune the response, displacing the sensor in the long axis has a lower risk of declining the SN ratio while having the appropriate sensitivity for modifying the reaction.
The results from the verification test showed that the optimized configuration as chosen by this method does improve the SN ratio. This indicates that the method succeeds in selecting the critical control factors, and most importantly, that following these recommendations produces the improvement predicted.
Although these results do not produce any surprises, but rather simply confirming existing knowledge, they show that the proposed method is effective and produces a quantitative assessment of the model. The intention of applying the Robust Engineering for this research was limited to test the factors' effect on reducing the trouble shooting of an assembled myoelectric controller; therefore, the test required only one subject. To enforce robustness for long-term use of a myoelectric controller, noise factors related to physical features, e.g., subcutaneous fat and muscle volume, can be added. These factors were not included because of the difficulty of altering within our short experiment period. It will, therefore, be possible to conduct future work, which will extend the method using Robust Engineering to look at other design parameters, e.g., electrode dimensions and signal processing, in conjunction with new myoelectric sensor prototypes. Additionally, by limiting the control and noise factors to those that pertain to the fitting procedure and to characteristics of myoelectric sensors used in prosthetic limb control and signals generated by amputee users, experiments can be conducted to quantify a better understanding of the procedure. Finally, the result of this experiment should be interpreted under the presented conditions only—changing the controller main function, control and noise factors of the target model may result different outcomes.
The author thanks Hirotaka Chinnami for the support on experimental data analysis, and Kiyoshi Gotoh in Department of Sports System Engineering for the advice and experimental apparatus for electromyography measurements.
1. Heckathorne CW. Chapter 11 Components for Electric-Powered System. In: Smith DG, Michael JW, Bowker JH, eds. Atlas of Amputation and Limb Deficiencies Surgical, Prosthetic, and Rehabilitation Principles.
3rd ed. Rosemont, IL: American Academy of Orthopaedic Surgeons; 2004:145–171.
2. Lovely DF. Signals and signal processing
for myoelectric control. In: Muzumdar A, ed. Powered Upper Limb Prostheses
. Berlin, Germany: Springer-Verlag; 2004:35–54.
3. Childress DS, Weir RF. Chapter 12 Control of limb prostheses. In: Smith DG, Michael JW, Bowker JH, eds. Atlas of Amputation and Limb Deficiencies Surgical, Prosthetic, and Rehabilitation Principles.
3rd ed. Rosemont, IL: American Academy of Orthopaedic Surgeons; 2004:173–195.
4. Roy SH, De Luca G, Cheng MS, et al. Electro-mechanical stability of surface EMG sensors. Med Biol Eng Comput
5. De Luca CJ. Electromyography. In: Webster JG, ed. Encyclopedia of Medical Devices and Instrumentation
. Hoboken, NJ: John Wiley & Sons; 2006:98–109.
6. Taguchi G, Chowdhury S, Wu Y. Taguchi's Quality Engineering Handbook
. Hoboken, NJ: John Wiley & Sons; 2005.
7. Taguchi G, Jugulum R, Taguchi S. Computer-Based Robust Engineering essentials for DFSS
. Milwaukee, WI: American Society for Quality; 2004.