JP2019004915A - Myoelectric pattern identification method using time series information and myoelectric artificial hand - Google Patents

Abstract

To provide an algorithm for dramatically improving an identification rate of a myoelectric pattern while minimizing the number of myoelectric electrodes and a learning time for controlling a multiple-degree-of-freedom myoelectric artificial hand.SOLUTION: A myoelectric artificial hand includes a measurement part for measuring a myoelectric pattern at a constant time interval, a myoelectric identification part for identifying a myoelectric pattern from the measured myoelectricity, and a buffer for storing a predetermined number of the myoelectric patterns identified by the myoelectric identification part in time series. When the predetermined number of the myoelectric patterns stored in the buffer are referred to and an occupancy rate of a certain level or higher is found, control according to an operation command corresponding to the myoelectric pattern in which the occupancy rate of a certain level or higher is found is executed. The present invention can achieve a high-speed learning and a high identification rate in spite of a small number of electrode channels, and is effective for practical control of a multiple-degree-of-freedom myoelectric artificial hand.SELECTED DRAWING: Figure 1

Description

  The present invention relates to the identification of a large number of myoelectric patterns, which are necessary when controlling a multi-degree-of-freedom myoelectric hand. The multi-degree-of-freedom myoelectric prostheses that are currently on the market do not require calibration, but there are only a few types of myoelectric patterns that can be identified. In the research stage, many EMG patterns have been identified, but the number of electrodes required for calibration, the length of time, and the accuracy of identification have become problems, and they have not been put to practical use. In this application, we propose a myoelectric identification algorithm that has a high identification accuracy with a small number of electrodes and a small calibration time.

[Background of prosthetic hand]
Prosthetic hands are roughly classified into decorative prostheses, active prostheses, and electric prostheses. The main purpose of the decorative prosthesis is to reproduce the appearance, and the active prosthesis is to reproduce the functionality (Non-patent Document 1). Therefore, the decorative prosthetic hand cannot be moved actively, and the active prosthetic hand is driven by transmitting the movement of the remaining part of the user as power with a wire or the like, so that the gripping force is weak (non-patent) Reference 2). On the other hand, as its name suggests, an electric prosthesis is driven by electric power and a motor. Electric prosthetic hands have fewer restrictions on wire handling than active prosthetic hands, and the direction and amount of movement can be designed to some degree of freedom (Non-Patent Document 1), and various prosthetic hands such as those with multiple degrees of freedom have been produced. . As an electric prosthesis, a myoelectric prosthesis is generally controlled by using a surface myoelectric signal that measures myoelectricity on the skin surface.

[Required specifications of myoelectric prosthetic hand and target elements of this study]
The elements that are said to be the most important in using a myoelectric hand include weight, certainty of operation, ease of wearing, gripping force, and freedom. In particular, it is difficult to achieve both freedom, reliability of operation and ease of installation.
Since the myoelectric prosthetic hand is used as a substitute for the hand and has the function of grasping an object, the certainty of the operation is very important. In particular, in terms of operational certainty, it is desirable that the control be performed with a certainty of at least 80% for practical use.

  As for the ease of wearing, there are ones related to physical attachment to the living body of the prosthetic hand and ones related to calibration. Since prosthetic hands are used on a daily basis, wearing them is required to be simple and immediate. For physical joints, many myoelectric prostheses can be installed simply by placing an electromyograph inside the socket and inserting the stump into the socket. As for calibration, it is the mainstream that can be controlled only by setting the threshold of myoelectric potential amplitude because it is necessary to set each time it is worn and the number of setting parameters is minimal.

[Control method for multi-degree-of-freedom myoelectric hand]
As the degree of freedom increases, it is necessary to improve the degree of freedom of control in order to move fingers with multiple degrees of freedom. Non-patent document 4). In the command method, two electromyographs are placed in the palm flexor muscle group and the dorsiflexor muscle group, and the myoelectric pattern is divided into four categories: weakness, palm flexion, dorsiflexion, and antagonism. Input to perform a specific action. For example, the extensor and flexor muscles are momentarily antagonized twice, and the forearm prolapses when force is applied to the flexor muscles. This type of operation system is used because myoelectric waves show various waveforms depending on the individual, and if they are controlled based on how force is applied or how much force is applied, the number of amputees that can be used is limited. Therefore, it is common to use myoelectric On / Off that can be clearly distinguished and used reliably for each channel of the electromyograph. However, surface myoelectric potentials measure muscle potentials on the skin surface, which causes crosstalk where signals from nearby muscles mix. Therefore, it is difficult to calculate the potential of each muscle from the surface electromyogram, and even if many electromyographs are placed on the skin surface, similar signals are measured in nearby electromyographs. Myoelectricity does not increase the degree of freedom of control simply by increasing the number of EMG channels. With this background, many myoelectric prostheses often place electromyographs at two locations in the flexor and extensor groups that can be contracted separately. Therefore, it is possible to distinguish 2 bits, that is, 4 states by contracting the flexor muscles and extensors separately. However, due to the above-mentioned crosstalk problem, the increase in electromyographs beyond this is effective. This is the limit for the number of states. These four states are often assigned to rest, forward rotation, reverse rotation, and mode change. As the number of operations performed with a prosthetic hand increases, the number of mode changes increases and the operation becomes complicated, so training for about 4 weeks is required to use it freely (Non-patent Document 5). Therefore, a new control method suitable for increasing the degree of freedom of the artificial hand is required (Non-patent Document 6).

[Previous research on control of multi-degree-of-freedom myoelectric hand]
At the research level, attempts have been made to estimate the user's finger posture directly from surface EMG. In many cases, the movement of the prosthetic hand is realized by classifying the myoelectric pattern into multiple classes according to the type of finger movement and learning, and identifying the class to which the measured myoelectric pattern belongs.

  Ajiboye et al. Used fuzzy theory to measure and learn EMG 8 times for each EMG pattern, and measured about 97 EMGs of palm dorsal flexion, ulnar displacement, grip, and rest in healthy subjects. Successful identification with% accuracy (Non-Patent Document 7). Even the amputee has an accuracy of about 96%, and the identification accuracy can withstand practical use. However, since it is recognized that ease of wearing is important for practical use of the prosthetic hand, it may be problematic to measure and learn each of the five patterns of myoelectricity eight times. Therefore, it is necessary to prepare a method that can be easily prepared before use and maintain a high identification rate. Young et al. Estimated 3 degrees of freedom with an accuracy of about 93%, and succeeded in achieving an accuracy of about 89% even in complex motions that moved multiple degrees of freedom simultaneously. In Young's method, the EMG pattern was learned at a high speed of 3 seconds per action, and the electromyogram measurement was four times half that of Ajiboye et al. (Non-patent Document 8). However, the number of electrodes used is six, which is larger than that of two common myoelectric prosthetic hands, which may make it difficult to align the electrodes. Therefore, a method has been used in which the classifier is composed of two stages, the first stage classifier performs rough classification with a low classification rate, and a filter algorithm is developed that improves the classification rate in the second stage (non-patented). Reference 9). This method improved the recognition rate by 5-15%. The number of myoelectric patterns to be identified is 15, which is higher than that of existing myoelectric prostheses, and the classification rate exceeds 90%. However, the number of electrodes is still 12, which is larger than the existing myoelectric prosthesis.

JP 2010-268404 A JP 2011-03362 A

Polhemus A, Doherty B, Mackiw K, Patel R, Paliwal M: uGrip II: A novel functional hybrid prosthetic hand design.39th Annual Northeast Bioengineering Conference, pp. 303-304, 2013. Smit G, Plettenburg DH, van der Helm FCT: Design and evaluation of two different finger concepts for body-powered prosthetic hand.Journal of Rehabilitation Research & Develop-ment. 50 (9), pp. 1253-1266, 2013. Otto Bock HealthCare GmbH: The Michelange-lo ® Hand in Practice-Therapy and Rehabilitation.646D593 = EN-01-1201. Connolly C: Prosthetic hands from Touch Bionics.Industrial Robot: An International Journal. 35 (4), pp. 290-293, 2008. Kannenberg A, Zacharias B: Difficulty of per-forming activities of daily living with the Mi-chelangelo Multigrip and traditional myoelectric hands.American Academy of Orthotists & Pros-thetists 40th Academy Annual Meeting & Scien-tific Symposium, FPTH14, 2014. Zecca M, Micera S, Carrozza MC, Dario P: Con-trol of multifunctional prosthetic hands by pro-cessing the electromyographic signal.Critical Reviews in Biomedical Engineering. 30 (4-6), pp.459-485, 2002. Ajiboye AB, Weir RF: A heuristic fuzzy logic approach to EMG pattern recognition for multi-functional prosthesis control.IEEE Transactions on Neural Systems and Rehabilitation Engineer-ing. 13 (3), pp. 280-291, 2005. Young AJ, Smith LH, Rouse EJ, Hargrove LJ: Classification of simultaneous movements using surface EMG pattern recognition.IEEE Transac-tions on Biomedical Engineering. 60 (5), pp. 1250-1258, 2013. Huang H, Zhou P, Li G, Kuiken T: Spatial filtering improves EMG classification accuracy following targeted muscle reinnervation. Annals of Bio-medical Engineering. 37 (9), pp. 1849-1857, 2009. Liu L, Liu P, Clancy EA, Scheme E, Englehart KB: Electromyogram whitening for improved classification accuracy in upper limb prosthesis control.IEEE Transactions on Neural Systems and Rehabilitation Engineering. 21 (5), pp. 767-774, 2013. Kato R, Yokoi H, Arai T: Real-time learning method for Adaptable motion-discrimination us-ing surface EMG signal.IEEE/RSJ International Conference on Intelligent Robots and Systems.pp. 2127-2132, 2006. Kita K, Kato R, Yokoi H: A self-organizing ap-proach to generate raining data for EMG signal classification.ROBIO 2008. IEEE International Conference on Robotics and Biomimetics.pp. 1230-1235, 2008. P. R. Cavanagh, P. V. Komi: Electromechanical delay in human skeletal muscle under concentric and eccentric contractions.European Journal of Applied Physiology and Occupational Physiology. 42 (3), pp.159-163, 1979.

In EMG pattern identification, more actions can be realized by increasing the number of classes, but the classification rate decreases as the number of classes increases. In order to improve the recognition rate, it is necessary to increase the number of electromyographic electrodes and the learning time, but there is a risk that practicality may be lost.
In order to solve the problem that the discriminator is not sufficiently learned and discrimination is not stable when the number of myoelectric electrodes and the learning time are limited (Non-Patent Document 10), in the present invention, for multi-degree-of-freedom myoelectric hand control, The purpose is to provide an algorithm that dramatically improves the recognition rate of myoelectric patterns while minimizing the number of EMG electrodes and learning time.

The first invention according to the present application is:
A measurement unit for measuring myoelectricity at regular time intervals;
A myoelectric identification unit for identifying which electromyographic pattern of the electromyogram pattern learned in advance should be classified,
A buffer for storing a predetermined number of myoelectric patterns identified by the myoelectric identification unit in time series,
With reference to the predetermined number of myoelectric patterns stored in the buffer and when an occupation ratio exceeding a certain ratio is recognized in a specific myoelectric pattern, control according to the operation command associated with the myoelectric pattern is performed. It is a myoelectric prosthesis characterized by performing.

  The second invention according to the present application is characterized in that when the buffer is a ring buffer and the number of myoelectric patterns to be stored reaches the predetermined number, the old myoelectric pattern is discarded. It is a myoelectric prosthesis as described in 1 invention.

  Furthermore, the third invention according to the present application is to perform control according to the specific myoelectric pattern when the occupation ratio is 50% or more. It is.

  In addition, according to a fourth aspect of the present application, in the case where an action that can be expressed simultaneously in the myoelectric prosthesis is set as an n pattern, the occupation ratio is 100 / (n + 1)% or more when the occupation ratio is 100 / (n + 1)% or more. The myoelectric prosthetic hand according to the second or third aspect of the invention is characterized in that control is performed simultaneously in accordance with an operation command corresponding to each of a plurality of myoelectric patterns occupying the same.

  According to the present invention, it is possible to achieve high-speed learning and a high identification rate with few electrode channels, which is effective for practical control of a multi-degree-of-freedom myoelectric hand.

(A) is a conventional algorithm, (b) is a conceptual diagram showing the algorithm of the present invention. It is a conceptual diagram which shows the structure of the neural network used by this invention. It is a flowchart which shows the algorithm of this invention. It is a schematic diagram which shows the measurement site | part of an electromyograph. (A) is an electromyograph, (b) is a photograph showing a state in which a wet electrode is attached to the electromyograph of (a). It is a conceptual diagram which shows the PC screen in experiment. It is a graph which shows the comparison of the recognition rate at the time of mounting the stabilization algorithm of myoelectric identification, and non-mounting. (A)-(d) is a graph which shows the comparison of the identification in the case where the proposal algorithm is not mounted and the case where it is not mounted.

  Hereinafter, specific embodiments to which the present invention is applied will be described in detail.

  The present invention is a filter algorithm that acts as a first-layer discriminator using the myoelectric discrimination algorithm (Non-patent Document 11) published by Kato et al. In 2007 and further improving the discrimination result. . Kato et al.'S EMG identification algorithm uses backpropagation (BP) neural network to identify EMG patterns. In order to improve this discrimination rate, we propose an algorithm for filtering BP neural networks by focusing on features that are mutually contradictory to events in the output time series. In the BP neural network, the connection strength between the nodes constituting the neural network is learned by backpropagation. In the present invention, a hierarchized filter algorithm has been developed that dramatically improves the identification rate without changing the structure of the BP neural network. First, Fig. 1 (a) shows a schematic diagram of a conventional myoelectric pattern identification algorithm. In contrast, the algorithm proposed in the present invention is shown schematically in FIG.

  In the present invention, the time series of the identified myoelectric pattern is stored in a buffer, and the discrimination rate is improved by performing the second stage identification based on the time series data. To identify the first stage, an algorithm based on a BP neural network that has been used conventionally is used (Non-patent Document 12). In this algorithm, the EMG waveform is first measured and stored in the vector emg. Vector emg contains time series data measured from 3 channel electromyographs. Next, a feature vector X indicating the frequency distribution of the vector emg is extracted from the vector emg by the feature extraction function GFE.

This feature vector X is input as a feature value to the input layer of the BP neural network. As a result, one of the nodes in the output layer corresponding to different hand postures fires and is identified (Fig. 2). The mechanism for performing this myoelectric pattern identification process is schematically shown as a myoelectric identifier in FIGS. 1 (a) and 1 (b).
If the output value at the output layer of the BP neural network is vector L and the number of EMG patterns to be identified is M
It is expressed as The output of the final BP neural net y, the ID of the myoelectric patterns m _i, the ID is the set of vectors L when the myoelectric patterns of m _i is identified as the output y and O _i

Is determined. If it cannot be identified, -1 is assigned to y. Or myoelectric patterns m _i is assigned to any orientation of the prosthetic hand is programmed in advance by the experimenter. Generally, a smoothing process such as moving average is performed on the myoelectric waveform (Non-Patent Document 10). However, since the output identified by the neural network has a discrete posture, a simple smoothing process can be used. Can not. In addition, in smoothing, if the time width is too large, the responsiveness decreases, so the time width of smoothing needs to be kept to a minimum. Therefore, we focused on the fact that there is a limited number of movements that may occur simultaneously in the target myoelectric pattern. For example, the electromyographic pattern handled in this example is forearm pronation / extroversion, wrist palm flexion / dorsiflexion, five finger grip / open, thumb flexion, 4/5 finger grip, resting state There are 9 patterns. For example, palm dorsiflexion of the wrist may be performed while holding the five fingers, but palm flexing while dorsiflexing the wrist means that the wrist is in an antagonistic state, and the current myoelectric prosthetic hand performs antagonistic control. Since it is not done, it is not necessary to identify. Based on this, the combined action of the three patterns of forearm pronation or prolapse, wrist palm flexion or dorsiflexion, and finger pattern is the largest. The myoelectric patterns that can be handled in the present invention are not limited to nine patterns, and may be less than nine patterns or more than ten patterns.

Based on the fact that there are three movements that may occur simultaneously when there are nine EMG patterns, we constructed an algorithm as shown in Fig. 3. In this algorithm, the time series data of the myoelectric pattern for a certain time stored in the ring buffer is referenced, and the output is determined based on the ratio of the myoelectric pattern included in the time series. If the number of myoelectric patterns buffered at a certain time is N, the ring buffer buff _{t at} a certain time _t is
It is expressed as The myoelectric pattern ID is stored in y _i which is a component of buff _t . Thereafter, performed again electromyographic measurements in the next control cycle, after t _l s identification is performed of the new myoelectric patterns y, the ring buffer buff is
And updated. The number S _i myoelectric patterns m _i contained in the ring buffer buff _{t + tl} is
It is expressed as Than this, the ratio P _i occupied by the myoelectric patterns m _i contained in the ring buffer buff
It becomes. The proposed algorithm first determines whether the myoelectric pattern is single or contains multiple myoelectric patterns. If more than half of the ring buffer buff is an arbitrary myoelectric pattern M _p , it is determined to be single, and the action associated with the myoelectric pattern M _p is output to the prosthetic hand.
That is, the proposed algorithm first refers to the N myoelectric patterns stored in the ring buffer buff. If more than half are the myoelectric patterns M _p , the control unit issues an operation command associated with the myoelectric pattern M _p , and the prosthetic hand operates according to the operation command.
When the ring buffer buff contains multiple myoelectric patterns M _p1 , M _p2 ..., The number of operations that occur simultaneously is set to 3, so that the myoelectric pattern occupies at least 1/4 of the ring buffer buff. Output all M _p1 , M _p2 …. Here, for generalization, if the operation o that is finally output is described as n, the maximum value of the number of operations that occur simultaneously is described as
It becomes. Here, when there are multiple myoelectric patterns o to be output, they are not actually output at the same time, and y _i to y _k are output in order.
When the number of operations that are simultaneously expressed is n, the threshold occupancy rate of the ring buffer buff serving as a reference for the simultaneous expression is set to 1 / (n + 1), and the ratio of the ring buffer buff to 1 / (n + 1) It is desirable to develop all of the above myoelectric patterns simultaneously. However, the threshold threshold occupancy rate is not limited to 1 / (n + 1), and can be set arbitrarily.
The filter algorithm proposed in this study is defined as a function GTS that outputs the output o to the prosthetic hand as a filter result using the time series buff of the identification result as an argument.
Can be written as

[Smooth EMG pattern identification]
For the algorithm GTS proposed in the present invention, it is assumed that buff is given by Eq. (11) when the buffer length is N = 6 and the maximum number n of simultaneous operations is n = 3.

On the system of the example, each element in buff is identified and updated in about 20ms. In EMG prosthetic hand operation, it is considered unlikely that the user has intentionally output a myoelectric pattern that differs only in one element discontinuously as shown in Equation (11) within 20 ms. Must be removed as a misclassification. To solve this problem, algorithm GTS first calculates S _i and P _i by Eq. (7).

To calculate. Next, using equation (9)
It can be seen that the myoelectric pattern mixed into the buff can be removed discontinuously.

[Support for multiple operations]
When multiple myoelectric patterns are output simultaneously, such as bending 4 or 5 fingers while palm flexing the wrist, many of those myoelectric patterns appear.
It becomes. From Equation (7), S _i and P _i are

It becomes. Next, using equation (9)
It can be seen that this is also effective when multiple myoelectric patterns are output simultaneously.

[Scope of application]
The algorithm proposed in the present invention is effective when the myoelectricity is constantly exerted. However, it is difficult to apply in the case where myoelectricity is instantaneously exhibited. For example, when a spike-like myoelectric signal is output in a short time, such as a tapping operation, buff sets the weak myoelectric pattern to m ₀
It becomes. At this time, S _i and P _i are

Thus, the output o to the prosthesis is m _{0 according} to Equation (9).

[Purpose of experiment]
The myoelectric pattern identification algorithm was evaluated by experiments. We verified how much the EMG pattern identification with the proposed algorithm could improve the accuracy compared to the conventional EMG pattern identification.

[Subject and experiment setup]
This experiment was conducted with 2 healthy people in their 20s who had experience with the operation of myoelectric prostheses. The electromyograph was attached to the forearm at three locations (extension group, flexor group, and long thumb flexor) (Fig. 4), and the reference electrode was the elbow. The subject sits in a chair so that the PC monitor can be seen in a comfortable posture. The specifications for electromyography are shown in Table 1. Wet surface electromyograph (Fig. 5) was used for measurement. The buffer length N of the myoelectric pattern ring buffer buff was set to 15. In addition, it takes about 10 ms for the ring buffer to be updated after EMG measurement and EMG pattern identification.

[BP Neural Network Learning Procedures and Conditions]
The learning process of the myoelectric pattern of BP neural network presses the key of the number of the action that the subject wants to correspond while outputting the myoelectric pattern. Correspondence between movement and number is as follows: 0: Rest, 1: Forearm gyrus, 2: Forearm pronation, 3: Wrist palmar flexion, 4: Wrist dorsiflexion, 5: Five finger grip, 6: Five finger open, 7: Mother Finger flexion, 8: 4/5 finger flexion. Through this work, a specific myoelectric pattern and a specific prosthetic hand movement are labeled, and a neural network is learned by BP. All of the pairs of myoelectric patterns taught in the past and their recognition actions were memorized, and consideration was given so that the recognition rate of the last taught myoelectric pattern was not significantly increased.

  The teaching conditions were that the myoelectric pattern to be assigned to each action was taught once, and then additional learning was allowed once for the myoelectric pattern considered to be inadequate. The time taken to teach one myoelectric pattern varies depending on the similarity between the newly taught myoelectric pattern and the already learned myoelectric pattern, and the first teaching without any other learning data is 100 to 300 ms. The second and subsequent teachings are about 100 to 1000 ms because BP learning continues until a certain recognition rate is secured for all other learning data.

  The screen shown in Fig. 6 is displayed on the PC monitor. The target posture is displayed on the left and the posture identified from the myoelectric pattern is displayed on the right. The subject is given a task to control the electromyographic pattern so that the target posture matches the identified posture.

[Experimental procedure]
In the experimental procedure, an electromyograph is first attached to the subject and the subject's myoelectric pattern is learned by the BP neural network. The subject is then assigned an experimental task and data is collected. Since the myoelectric pattern discrimination system uses supervised learning using BP neural network, the system first teaches the subject's myoelectric pattern. Next, the target electromyographic pattern of the hand and the identification result are displayed on the PC monitor (Fig. 6), and the task of matching the identification result to the target myoelectric pattern is performed. The task repeatedly outputs the target myoelectric pattern displayed on the PC monitor for 4 seconds and then weakens for 6 seconds. Each target EMG pattern is displayed 5 times during one task. At this time, each with and without the proposed algorithm is performed for about 7 minutes, and the degree of coincidence of the identified myoelectric pattern with respect to the target myoelectric pattern is evaluated as the discrimination rate.

[Experimental result]
The proposed algorithm is evaluated by comparing the proposed system with the developed algorithm with the conventional system without it. First, Fig. 7 shows the identification rates of both the conventional system and the proposed system. The classification rate of the developed algorithm was 82.5% when all 9 patterns were identified, and 92.9% when 7 patterns were excluded from the forearm pronation and pronation. When the F test was performed with the trial with and without the developed algorithm, the trial for all patterns had a one-sided probability of 0.095, and the trial excluding pronation and pronation was 0.37, both of which exceeded the 5% level. The t test was performed with equal variance. As a result, both trials showed a significant difference (p <0.001) in the high recognition rate of the proposed method. The improvement of the recognition rate was 17.7% for all nine patterns, and 11.5% for seven patterns excluding forearm pronation and pronation. Figure 8 shows the classification time series, and Table 2 shows the classification rate under each condition. Figure 8 shows the EMG pattern identified by the algorithm, logged in CSV format, with 10-20ms between samples. In addition, the correspondence of the myoelectric pattern and the number in Fig. 8 is as follows: 0: Rest, 1: Forearm gyrus, 2: Forearm pronation, 3: Wrist palm flexion, 4: Wrist dorsiflexion, 5: Five finger grip, 6 : 5 finger open, 7: thumb flexion, 8: 4/5 finger flexion.

  Figures 8 (a) to 8 (d) show a comparison of identification when the algorithm of the present invention is installed and when it is not installed. 8A and 8B are graphs of the subject 1, and FIGS. 8C and 8D are graphs of the subject 2. 8 (a) and 8 (c) show the proposed method of this application, and FIGS. 8 (b) and 8 (d) show the conventional method. The sampling interval on the horizontal axis is 10 to 20 ms. The top and bottom are differences between subjects, and ○ indicates each identification result. The change in the target posture is shown by a solid line. The target posture is shown on the display only in the part where the solid line is horizontal. The subjects showed identification results and target myoelectric patterns in a representative section of about 7 minutes of outputting the target myoelectric pattern displayed on the PC monitor. From Fig. 8, it can be seen that the continuity of the myoelectric pattern is maintained and there are fewer misidentifications than the conventional method.

[About identification rate]
As a result of the experiment, it was found that the recognition rate changed slightly depending on whether or not pronation and pronation were included. The classification rate without the pronation and pronation exceeded 90%. The identification rate was 82.5% even when including pronation and pronation. Focusing on subject 1 in Fig. 8 (a), which shows the time series of discrimination, it can be seen that when grasping (myoelectric pattern 5), it was misidentified as prolapse (myoelectric pattern 1). There was a possibility that it was difficult to distinguish between gripping and pronation because there was force in the pronation direction when grasping. Focusing on the subject 2 in FIG. 8C, the pronation (myoelectric pattern 2) is also identified in the pronation (myoelectric pattern 1) and the opening (myoelectric pattern 6). It is suggested that it is hanging.

[Control cycle]
The proposed method buffers 15 EMG discriminator outputs with a control period of 10 ms, but because the decision is made at 1/4 or 1/2 of the buffer as shown in Fig. 3, the delay time of the proposed method is Can be estimated as 40-80 ms. Considering that myoelectricity is observed several tens to hundreds of milliseconds before the start of joint movement (Non-patent Document 13), it is considered to have sufficient responsiveness to be used as a prosthetic hand.

  In order for a myoelectric prosthesis to be used practically, it is desirable to control it with a certainty of at least 80%. In conventional multi-degree-of-freedom myoelectric pattern identification, the number of electrodes (Non-Patent Documents 8 and 9) and the length of learning time (Non-Patent Document 7) are problems in order to realize the identification rate. It was. On the other hand, in the present invention, a discrimination rate of 90% or more is realized when 7 patterns excluding pronation and pronation are targeted for discrimination with a small number of electrodes of 3 channels and a short learning time of 1 second or less. . Moreover, even when including pronation and pronation, the recognition rate of 82.5% is realized, so it can be used effectively as a prosthetic hand. In addition, some subjects gave a recognition rate of over 90% for all 9 patterns, and it was suggested that proficiency could be over 90% for all 9 patterns.

  In the present invention, by focusing on the time series of the identification results and performing two-stage identification, practical identification of 7 patterns from 3 electrodes is 90% and 9 patterns is 82.5% without greatly increasing the calculation cost. We built a system that can be identified with a certain rate. Also, the learning time of the classifier is less than a dozen seconds for all 7 patterns, and a control method for a multi-degree-of-freedom prosthesis that can withstand practical use is approaching realization.

  In recent years, there is an increasing need for multi-degree-of-freedom myoelectric prostheses, and there have been a series of research presentations of highly functional multi-degree-of-freedom myoelectric prostheses. However, the difficulty of various problems such as countermeasures against EMG noise and changes in the electrical properties of the skin has hindered substantial improvement in control freedom. Therefore, in order to make a multi-degree-of-freedom robot hand function as a myoelectric prosthetic hand, a methodology capable of absorbing the characteristic changes of the living body in an information processing manner and identifying multiple myoelectric patterns stably is required. In this application, we propose an algorithm that improves the recognition rate of myoelectric patterns. Focusing on the output time series of the classifier, we developed a filter algorithm that dramatically improves the classification rate by being positioned behind the classifier without changing the structure of the classifier itself. In the evaluation experiment, the discrimination rate of the myoelectric pattern was compared between the discriminating system with the developed algorithm and the discriminating system without it. The surface electrode 3 channels, the learning time per class was within 1 second, the number of classification classes was set to 7 and 9 conditions, and the pattern identification output was sampled at 10-20 ms. As a result, the accuracy was improved by 11.5% for 7 classes and 17.7% for 9 classes. The classification rate by the proposed method was 82.5% for 9 classes and 92.9% for 7 classes. Although there are few electrode channels, high-speed learning and a high identification rate have been achieved, and it has been proposed a highly effective methodology for practical control of multi-degree-of-freedom myoelectric hand.

Claims (4)

A measurement unit for measuring myoelectricity at regular time intervals;
A myoelectric identification unit for identifying which electromyographic pattern of the electromyogram pattern learned in advance should be classified,
A buffer for storing a predetermined number of myoelectric patterns identified by the myoelectric identification unit in time series,
With reference to the predetermined number of myoelectric patterns stored in the buffer, if an occupation ratio exceeding a certain ratio is recognized in a specific myoelectric pattern, control is performed according to an operation command corresponding to the myoelectric pattern. A myoelectric prosthesis characterized by that.   The myoelectric prosthesis according to claim 1, wherein when the buffer is a ring buffer and the number of myoelectric patterns to be stored reaches the predetermined number, the old myoelectric pattern is discarded.   The myoelectric prosthetic hand according to claim 2, wherein when the occupation ratio is 50% or more, control according to an operation command corresponding to the specific myoelectric pattern is performed. When the myoelectric pattern that can be expressed simultaneously in the myoelectric prosthetic hand is set as n pattern, each of the plurality of myoelectric patterns that occupy the occupation rate when the occupation rate is 100 / (n + 1)% or more. 4. The myoelectric prosthetic hand according to claim 2, wherein control according to a corresponding operation command is performed simultaneously.