JP6652252B2 - Myoelectric signal processing method, apparatus and program - Google Patents

Description

  The present invention relates to a technique for processing a surface electromyogram signal (hereinafter, referred to as an electromyogram signal), and in particular, dispersion of a myoelectric signal before rectification and smoothing in which signal intensity-dependent noise is superimposed on an electromyography signal after rectification and smoothing. And a technique for generating an artificial myoelectric signal based on the estimated distribution.

  Myoelectric signals are useful for quantitative evaluation of muscle activity, and have been applied to control of myoelectric prosthetic hands and to exercise analysis during rehabilitation and training. In those applications, a myoelectric signal is modeled and used to extract a feature amount from the myoelectric signal (for example, see Non-Patent Document 1).

  In recent years, with the progress of research on biological signals including myoelectric signals, the existence of signal intensity-dependent noise has become apparent (for example, see Non-Patent Document 2). In other words, signal strength-dependent noise corresponding to the strength of the muscular force is superimposed on the myoelectric signal so that the arm vibrates when the force is applied to the arm.

N. Hogan and RW Mann, “Myoelectric signal processing: Optimal estimation applied to electromyography-part i: Derivation of the optimal myoprocessor,” IEEE Trans. Biomed. Eng., Vol. BME-27, no.7, pp.382- 395, 1980. C. M. Harris and D. M. Wolpert, “Signal-dependent noise determines motor planning,” Nature, vol. 394, no.6695, pp. 780-784, 1998.

  In the conventional myoelectric signal model, the variance of the myoelectric signal is assumed to be constant. However, since the signal strength-dependent noise according to the muscular strength is superimposed on the myoelectric signal, the variance of the myoelectric signal is actually increased. Is not constant but varies according to the signal strength dependent noise. That is, the conventional myoelectric signal model does not sufficiently reflect the presence of signal intensity-dependent noise.

  On the other hand, the signal intensity-dependent noise superimposed on the myoelectric signal can be estimated by analyzing the myoelectric signal, but the sampling frequency of the measured myoelectric signal is 1000 to 2000 [Hz], and the signal bandwidth is wide. Since the amount of information is large, the measured myoelectric signal is actually down-sampled at about 5 [Hz] in an electromyograph and converted into a rectified smoothed signal with a reduced amount of information. There is a need to estimate signal strength-dependent noise from the rectified and smoothed signal obtained from such an electromyograph and to restore the measured myoelectric signal. It is difficult to estimate the signal-intensity-dependent noise from the noise and to restore the measured myoelectric signal.

  Furthermore, in the field of myoelectric prosthetic hand control, there is a need for generating an artificial myoelectric signal in consideration of signal intensity-dependent noise. Support vector machines and neural networks are used to control the myoelectric prosthesis, and it is necessary to train the myoelectric signal pattern of the motion to be identified in advance. Machine learning with muscle strength is performed. For this reason, if a large muscular strength is exerted in real life, erroneous identification of the identification target operation may occur due to the influence of signal intensity-dependent noise, and the myoelectric hand may malfunction. In order to prevent such misidentification or malfunction, the myoelectric signal at the muscle activity corresponding to the muscular strength of various sizes considered to be exerted in real life is generated in consideration of the signal intensity-dependent noise. It is desired to generate artificially and perform machine learning using this artificial myoelectric signal.

  In view of the above problems, the present invention estimates the distribution of the variance of the myoelectric signal before rectification and smoothing in which signal intensity-dependent noise is superimposed from the myoelectric signal after rectification and smoothing, and further based on the estimated distribution. An object is to generate an artificial myoelectric signal.

In the myoelectric signal processing method according to one aspect of the present invention, the time window processing section averages the rectified and smoothed signal obtained by rectifying and smoothing the myoelectric signal on which the signal intensity-dependent noise is superimposed for each predetermined time length. Calculating, and a variance average estimating unit calculating an average value of the variance of the myoelectric signal from the reciprocal of the average of the rectified and smoothed signal and the filter gain related to the rectification and smoothing of the myoelectric signal; dependent noise variance estimation unit, and a step of calculating an estimated value of the myoelectric signal of the variance of the mean of the estimates and Sujiden signal variance of the distribution variance from the parameters defining the signal strength dependent noise, the muscle wire The distribution of the variance of the signal has a property conjugate to the distribution of the myoelectric signal and satisfies the non-negativeness of the variance .

Similarly, the myoelectric signal processing device includes a time window processing section that calculates an average of a rectified and smoothed signal obtained by rectifying and smoothing the myoelectric signal on which the signal intensity-dependent noise is superimposed for each predetermined time length; A variance average estimator for calculating an average of the variance of the myoelectric signal from the reciprocal of the average of the signal and the filter gain relating to the rectification and smoothing of the myoelectric signal; an average of the variance of the myoelectric signal; A signal strength-dependent noise variance estimating unit that calculates an estimated value of the variance of the signal strength-dependent noise from a parameter that defines the distribution of the variance of the signal , wherein the distribution of the variance of the myoelectric signal is the distribution of the myoelectric signal. It is a distribution that has properties that are conjugate to that, and that satisfies the non-negative nature of variance .

Similarly, the myoelectric signal processing program is a means for calculating an average for each predetermined time length of the rectified and smoothed signal obtained by rectifying and smoothing the myoelectric signal on which the signal intensity-dependent noise is superimposed, the average of the rectified and smoothed signal and defining means, and the average of the variance of the myoelectric signal distribution of the variance of the estimate and Sujiden signal to calculate an estimate of the mean from the reciprocal of filter gain of the variance of the myoelectric signal according to the rectified and smoothed myoelectric signal The computer functions as a means for calculating an estimated value of the variance of the signal intensity-dependent noise from the parameters to be performed, and the distribution of the variance of the myoelectric signal has a property conjugate to the distribution of the myoelectric signal, and It is a distribution that satisfies non-negativeness .

  According to the myoelectric signal processing method, the apparatus and the program, from the rectified smoothed signal obtained by rectifying and smoothing the myoelectric signal on which the signal intensity-dependent noise is superimposed, the estimated value of the average of the variance of the myoelectric signal and the Estimates of the variance are calculated, and using these estimates, the distribution of the variance of the myoelectric signal before rectification and smoothing on which the signal intensity-dependent noise is superimposed can be estimated.

  In the myoelectric signal processing method, the first random number generator generates a white normal random number, and the shaping filter shapes the frequency characteristic of the white normal random number into a frequency characteristic similar to that of the myoelectric signal. A second random number generator for generating a random number according to a distribution of the variance of the myoelectric signal defined by the estimated value of the average of the variance of the myoelectric signal and the estimated value of the variance of the signal intensity-dependent noise; Generating an artificial myoelectric signal from the shaped random number and a random number according to the distribution of the variance of the myoelectric signal.

  Similarly, the myoelectric signal processing device includes a first random number generator that generates white normal random numbers, a shaping filter that shapes the frequency characteristics of the white normal random numbers into frequency characteristics similar to those of the myoelectric signal, A random number generator that generates random numbers according to the distribution of the variance of the myoelectric signal defined by the average value of the variance and the estimated value of the variance of the signal intensity-dependent noise, and the shaped random number and the myoelectric signal An artificial myoelectric signal generation unit that generates an artificial myoelectric signal from random numbers in accordance with the distribution of signal variances.

  Similarly, the myoelectric signal processing program includes means for generating white normal random numbers, means for shaping the frequency characteristics of the white normal random numbers into frequency characteristics similar to those of the myoelectric signal, an estimated value of the variance of the myoelectric signal and Means for generating a random number according to the distribution of the myoelectric signal defined by the estimated value of the variance of the signal intensity-dependent noise, and an artificial myoelectric signal from the shaped random number and the random number according to the distribution of the myoelectric signal variance The computer may function as a means for generating a signal.

  According to the myoelectric signal processing method, apparatus and program, an artificial myoelectric signal can be generated in consideration of signal intensity-dependent noise from a shaped white normal random number and a random number according to a distribution of variance of the myoelectric signal.

In the myoelectric signal processing method, the variance average estimating section calculates the rectified smoothed signal at the maximum voluntary muscle contraction, the specified muscular activity and the reciprocal of the filter gain related to the rectified smoothing of the myoelectric signal with the muscular activity. Calculating an estimated value of the average of the variance of the myoelectric signal, and a signal intensity-dependent noise variance estimating section calculates a parameter defining a distribution of the variance of the myoelectric signal and a variance of the myoelectric signal at the muscle activity. Calculating an estimate of the variance of the signal intensity dependent noise superimposed on the myoelectric signal at that muscle activity from the average estimate; and wherein the second random number generator generates the myoelectric signal at that muscle activity. The distribution of the variance of the myoelectric signal at the muscle activity defined by the estimated value of the average of the variance of the muscular activity and the estimated value of the variance of the signal intensity-dependent noise superimposed on the myoelectric signal at the myoactivity Generating a random number. There.

Similarly, in the myoelectric signal processing apparatus, the variance average estimating section calculates the muscular rectified smoothed signal at the maximum voluntary muscle contraction, the designated muscular activity, and the reciprocal of the filter gain related to rectified and smoothed myoelectric signal. The signal strength-dependent noise variance estimator calculates a mean value of the average of the variance of the myoelectric signal at the activity level, and a parameter defining the distribution of the variance of the myoelectric signal and the myoelectric signal at the myoactivity level. A second random number generator calculates the estimated value of the variance of the signal intensity-dependent noise superimposed on the myoelectric signal at the muscle activity from the estimated value of the average of the variance of the muscle activity. Of the average of the variance of the myoelectric signal at the same time and the variance of the myoelectric signal at the relevant myoactivity defined by the estimated value of the variance of the signal intensity-dependent noise superimposed on the myoelectric signal at the relevant myoactivity Generates random numbers according to the distribution of It may be.

Similarly, the myoelectric signal processing program calculates the myoelectric signal at the myoelectric activity from the reciprocal of the rectified and smoothed signal at the maximum voluntary muscle contraction, the designated myoactivity and the filter gain related to the myoelectric signal rectifying and smoothing. Means for calculating an estimate of the average of the variance of the signal, a parameter defining the distribution of the variance of the myoelectric signal, and the EMG at the muscle activity from the estimate of the average of the variance of the myoelectric signal at the muscle activity. Means for calculating an estimate of the variance of the signal intensity dependent noise superimposed on the signal, and an estimate of the average of the variance of the myoelectric signal at that muscle activity and being superimposed on the myoelectric signal at that muscle activity The computer may function as means for generating a random number according to the distribution of the variance of the myoelectric signal at the muscle activity defined by the estimated value of the variance of the signal intensity-dependent noise.

  According to the myoelectric signal processing method, apparatus, and program, it is possible to designate an arbitrary myoactivity and generate an artificial myoelectric signal at the myoactivity.

  In the myoelectric signal processing method, apparatus and program, for example, the distribution of the variance of the myoelectric signal may be an inverse gamma distribution, and the parameter may be a shape parameter of the inverse gamma distribution.

  According to the present invention, it is possible to estimate the distribution of the variance of the myoelectric signal before rectification and smoothing on which the signal intensity dependent noise is superimposed, from the myoelectric signal after rectification and smoothing. Further, an artificial myoelectric signal can be generated based on the estimated distribution.

Schematic diagram of the myoelectric signal model used in the present invention 1 is a schematic diagram of a system including a myoelectric signal processing device according to an embodiment of the present invention. Block diagram of myoelectric signal processing device Graph showing the average error rate for estimating the average of the variance of the myoelectric signal at each sampling frequency Graph showing the average error rate for estimating the average of the variance of the myoelectric signal in each time window Graph showing the average error rate for estimating the variance of signal strength-dependent noise at each sampling frequency Graph showing the posterior distribution of the variance of the myoelectric signal at each load in a certain trial Graph showing the average of the variance of the myoelectric signal and the estimated value of the variance of the signal strength-dependent noise at each muscle strength in all trials Graph showing measured myoelectric signals at each load and artificial myoelectric signals generated based on them Graph showing the average absolute error rate of the average amplitude of the measured myoelectric signal and artificial myoelectric signal at each load Graph showing correlation coefficient of power spectral density between measured myoelectric signal and artificial myoelectric signal at each load Graph showing the average absolute error rate of the parameter related to the distribution of the electromyographic signal variance at each load

  Hereinafter, embodiments will be described in detail with reference to the drawings as appropriate. However, an unnecessary detailed description may be omitted. For example, detailed description of well-known matters and redundant description of substantially the same configuration may be omitted. This is to prevent the following description from being unnecessarily redundant and to facilitate the understanding of those skilled in the art.

  The inventors provide the accompanying drawings and the following description so that those skilled in the art can fully understand the present invention, and they are intended to limit the subject matter described in the claims. is not.

≫Estimation of posterior distribution of EMG signal variance≫
(EMG signal model)
FIG. 1 is a schematic diagram of a myoelectric signal model used in the present invention. The EMG signal model, a myoelectric signal x _t at time t, expressed on the basis of the white Gaussian noise W and variance sigma _t ² through the shaping filter H. Here, σ _t ² is the value of the random variable σ ² at time t, and the distribution of σ ² is determined by muscular strength f σ ￣ ² (in this specification, the average of σ (sigma bar) is referred to as “σ ￣” for convenience). And signal strength dependent noise ε. Further, the distribution of σ ² can be estimated using the signal y _t obtained by rectifying and smoothing x _t .

  First, the relationship between f and σ￣ can be expressed as follows.

Here, k and a are constants and are experimentally obtained.

Next, sigma ² is represented as the sum of Shiguma ² and noise epsilon.

Here, the noise ε is a random variable having an average of 0.

Thus, sigma ² average E [σ ^2] and variance Var [σ ^2] may,

It is expressed as

x _t is because it is expressed by the product of the white Gaussian noise W and sigma _t through a shaping filter H, average 0 follows a normal distribution with variance sigma _t ^2.

y _t is obtained by rectifying and smoothing by using the N-order low-pass filter x _t.

_{Here, a 0, a 1, ···} , a N-1 and _{b 0, b 1, ···,} b N is the filter coefficient.

(Dispersion distribution estimation)
Next, a method for estimating the distribution of the sigma ² from x _t. First, when the myoelectric signal _xt is measured, the posterior distribution P (σ ² | _xt ) of the variance can be expressed as follows using Bayes' theorem.

Here, the likelihood P (x _t | σ ² ) is given by the following equation (5).

It is. In addition, an inverse gamma distribution IG (α, β) which is a prior distribution conjugate to a normal distribution is assumed for the prior distribution P (σ ² ) in consideration of σ ² > 0.

Here, α and β are parameters for determining the inverse gamma distribution. At this time, the numerator on the right side of the equation (7) can be expanded as follows.

Here, κ (α, β) is a collection of constant terms,

Becomes In equation (7), P (x _t ) must be a constant, and P (σ ² | x _t ) must be a probability density function. Therefore, κ (α, β) and P (x _t ) are canceled and P (σ ² | _{x t)} is IG (α + 1/2, β + x t 2/2) to match. In this case, sigma ² the posterior mean _{^{E [σ 2 | x t]}} , the posterior variance Var [σ ² | _{x t]} is

Becomes At this time, according to equations (3) and (4), E [σ ² | x _t ] and Var [σ ² | x _t ] are estimated values of σ￣ ² and Var [ε] under the condition where x _t is observed. to match each value, it can be determined Var [epsilon] from Shiguma ² by setting the α advance.

(Σ ² of the estimate)
Next, a method for estimating the Shiguma ² from the signal y _t rectified smoothed. Since y _t can be expressed by equation (6), the expected value of y _t is as follows.

Here, _{x t} is E to follow the formula (5) _{[| x} t _|] is

Becomes ε _t is the value of ε at time t. Assuming that E [y _t ] does not depend on time, Expression (14) uses Expression (15).

Becomes Here, the second term on the right side of the equation (16) has an expected value of 0, and includes a load average based on b _i (> 0) of ε _t (≪σ￣), and is therefore sufficiently smaller than the first term and can be ignored. Thus, σ￣ ² is calculated by using equation (16).

Can be estimated. At this time, the fractional coefficient term on the right side corresponds to the reciprocal of the filter gain.

From the above, by using the EMG signal y _t and the shape parameter α, which is rectified and smoothed (13), it can be estimated (17) σ¯ ² and Var [epsilon] based on the formula.

の Generation of artificial myoelectric signal≫
If the posterior distribution of the variance sigma _t ² from the signal y _t rectified smoothed rectified smoothed signal before x _t is estimated, the same statistical properties as the rectified smoothed signal before x _t using the distributions the estimated An artificial myoelectric signal having the following can be generated.

(Reproduction of myoelectric signal)
Time Artificial EMG z _t at t can be expressed on the basis of the normal random number w _'t through a shaping filter H to the variance sigma _t ^2.

w _'t is a normal random number having the same frequency characteristics and EMG signals x _t, generated by the shaping filter H based on the autoregressive model of order M.

Here, w _t is a white normal random number having a mean of 0 and a variance of 1, and v, a _j (j = 1,..., M) are an estimated error variance and parameters in the autoregressive model, respectively. Further, in the estimation of the autoregressive model parameters, by using the myoelectric signal x _t dispersion is normalized to a 1, w _'t mean 0 and normal random number of dispersion 1.

The artificial EMG z _t is defined as follows: the product of (3) (4) and the random number sigma _t generated by distribution based on formula w _'t. That is, it is possible to replace the myoelectric signal x _t to artificial EMG z _t in EMG model of FIG.

From the above, it is possible to generate a parameter k, the distribution of a and sigma ^2, and by pre-set parameters of the shaping filter H, corresponding to exert muscle force f artificial EMG z _t of equation (1). That is, it is possible to restore the rectified smoothed before signal x _t as artificial EMG z _t from the signal y _t rectified smoothed (artificially reproduced).

(Generation of artificial myoelectric signal at any muscle activity)
Further, a myoelectric signal at an arbitrary muscle activity can be artificially generated. For example, the maximum voluntary signals rectified smoothed during contraction _{y max,} the muscle activity degree (% MVC) and _{r t,}
r _t = y _t / y _max (20)
It is expressed as Therefore, the expected value of _{y t} E _{[y t]} is,
_{E [y t] = E [} r t] · y max (21)
And substituting equation (21) into equation (17),

Becomes That gives a signal y _max rectified smoothed at the maximum voluntary contraction, by specifying any muscle activity r _t, (13) equation (22) based on the equation of muscle activity r _t it is possible to estimate the posterior distribution of the variance of the myoelectric signal x _t. Then, it is possible to produce an artificial myoelectric signal z _t of muscle activity r _t on the basis of the equation (19).

In the case where the myoelectric signals of a plurality of channels are measured, the time of the muscle activity of each channel is calculated from the myoelectric signals of the channel l of interest and the channel m of the main operating muscle related to the operation during a certain motion of the subject. calculate the average lambda _l and lambda _m, using the ratio λ _l / λ _m as a value _{r t} × λ _l / λ _m multiplied muscle activity _{r t} (22) in the formula E _{[r t]} You may. For example, considering the hand-opening operation, the pattern of the myoelectric signal of each channel is considered to be the same regardless of whether the hand is opened lightly or openly. By correcting the designated muscle activity by multiplying the ratio of the time average of the muscle activity of the channel of interest to the time average, the desired muscle activity at the desired muscle activity can be maintained without breaking the characteristic myoelectric pattern corresponding to the motion. An artificial myoelectric signal can be generated.

≪Myoelectric signal processing device≫
Next, a myoelectric signal processing device according to the present invention will be described. FIG. 2 is a schematic diagram of a system including a myoelectric signal processing device according to an embodiment of the present invention. The system 100 includes an electromyograph 10 and an electromyography signal processing device 20.

  The electromyograph 10 includes an arbitrary number (four in the present embodiment) of myoelectric sensors 11 and a data collection device 12. The data collection device 12 and the myoelectric sensor 11 are connected by a parallel cable or a serial cable. Further, these cables can be detached from the data collection device 12.

  The myoelectric sensor 11 is for measuring a surface myoelectric potential signal (myoelectric signal) of an arbitrary body part of the subject. The myoelectric signal is a waveform of the compound action potential of the muscle, and the compound action potential is the sum of the action potentials generated from each muscle fiber constituting the muscle and propagated to the body surface by volume conduction. Say. Complex action potentials vary with the degree of muscle contraction. For example, when the muscle exerts a small force, the number of contracting muscle fibers is small, and the number of action potentials generated from the muscle fibers is also small. As the force is increased from there, the number of contracting muscle fibers gradually increases, and the number of action potentials also increases. Thereby, the composite action potential observed as a myoelectric signal increases. As described above, there is a correlation between how to output the force and the myoelectric signal.

  The surface signal is captured as a potential difference between two surface electrodes. Specifically, each myoelectric sensor 11 has two surface electrodes 101 (a plus electrode and a minus electrode). These electrodes 101 are attached to the body surface of the subject at intervals of about several cm along the running direction of the muscle fiber at the target site. Note that any one of the myoelectric sensors 11 further includes a reference electrode (not shown).

The data collection device 12 includes an unillustrated amplifier, a filter, an AD converter, a CPU, a communication device, and the like, and collects a myoelectric signal (x _t in a myoelectric signal model) measured by the myoelectric sensor 11, and The collected signal is rectified and smoothed to generate a rectified and smoothed signal (y _t in the myoelectric signal model) and transmit it to the myoelectric signal processing device 20. Wireless communication such as Wi-Fi (registered trademark), Bluetooth (registered trademark), ZigBee (registered trademark), or specific low-power wireless is used for communication between the data collection device 12 and the myoelectric signal processing device 20. be able to. As described above, by rectifying and smoothing the measured myoelectric signal in the data collection device 12 and reducing the data amount, the myoelectric signal of many channels can be transmitted to the myoelectric signal processing device 20 through a relatively low-speed communication line. Can be.

The myoelectric signal processing device 20 is a device that performs the above-described myoelectric signal processing based on the myoelectric signal model of FIG. That is, the myoelectric signal processing device 20 receives the rectified and smoothed signal y _t from the electromyograph 10 and performs the above-described myoelectric signal processing, and after the dispersion of the myoelectric signal x _t measured by the myoelectric sensor 11, The distribution is estimated, and an artificial myoelectric signal is generated based on the estimated distribution. The artificial myoelectric signal generated by the myoelectric signal processing device 20 is displayed in real time on a monitor screen of the myoelectric signal processing device 20 or another device, or is transmitted to another device as a signal for machine learning of myoelectric prosthetic hand control. Output.

  As the myoelectric signal processing device 20, a general-purpose computer (a desktop computer, a notebook computer, or the like), a smartphone, a tablet terminal, or the like can be used. By installing a program or application for realizing the above-described myoelectric signal processing in these computers and terminals, the computers and terminals can be used as the myoelectric signal processing device 20.

FIG. 3 is a block diagram of the myoelectric signal processing device 20. The myoelectric signal processing device 20 includes a time window processing unit 21, a variance average estimating unit 22, a signal intensity dependent noise variance estimating unit 23, a white Gaussian noise generator 24, a shaping filter 25, an inverse gamma distribution random number generator 26, and an artificial muscle. An electric signal generation unit 27 is provided. Among them, the rectified smoothed signal y _t myoelectric sensors 11 EMG measured is x _t is the block required to estimate the posterior distribution of the variance of the time window processing section 21, the dispersion average estimator 22 and the signal strength There are three dependent noise variance estimators 23, and an artificial myoelectric signal can be generated by adding the remaining four blocks.

The time window processing unit 21 calculates an average E [y _t ] of the rectified and smoothed signal y _t for each predetermined time length (time window). Here, y _t is a signal obtained by rectifying and smoothing the myoelectric signal x _t on which the signal intensity-dependent noise ε is superimposed, and the time window processing unit 21 substantially executes the calculation of Expression (14).

Dispersion average estimation unit 22, rectification average E of the smoothed signal [y _t] and calculate an estimate of the average Shiguma ² of the variance of the EMG from the reciprocal of filter gain in accordance with the rectified and smoothed myoelectric signal x _t I do. That is, the variance-average estimating unit 22 substantially executes the calculation of Expression (17).

Also, when generating the artificial myoelectric signal z _t in any muscle activity r _t, the dispersion average estimation unit 22, the maximum voluntary rectified smoothed signal y _max during _contraction, the muscle activity r _t and designated calculating an estimate of the average Shiguma ² of the variance of the EMG signals from the reciprocal of filter gain in accordance with the rectified and smoothed by the muscle activity r _t of the EMG. In this case, the variance-average estimating unit 22 substantially executes the calculation of the expression (22). The dispersion average estimation unit 22, to the designated muscle activity r _t may be used as it is, as described above, the time of the muscle activity of the target channel for the time average of the muscle activity of the main operation muscle A value corrected by multiplying by the average ratio may be used.

In order to measure the myoelectric signal during the maximum voluntary muscle contraction, it is necessary to have the subject demonstrate muscle strength at a muscle activity level of 100 (100% MVC), but in some cases it is not possible because of the burden on the subject. is there. In such a case, since the rectified smoothed signal at the time of the maximum voluntary muscle contraction cannot be obtained, the rectified smoothed signal y _max at the time of the maximum voluntary muscle contraction input to the variance average estimating unit 22 is, for example, an actual measurement value of another subject. Can be used as the assumed value. Of course, when a myoelectric signal at the time of the maximum voluntary muscle contraction of the subject can be measured, it is needless to say that an actual measurement value may be used instead of an assumed value as y _max .

Signal intensity dependent noise variance estimation unit 23, calculates an estimate of the variance Var of the parameters from the signal intensity dependent noise according to the distribution of the variance of the average Shiguma ² estimates and Sujiden signal variance [epsilon] of the EMG I do. Here, assuming an inverse gamma distribution as the distribution of the variance of the myoelectric signal, the shape parameter α of the inverse gamma distribution can be adopted as the parameter. That is, the signal intensity dependent noise variance estimating unit 23 substantially executes the calculation of the expression (13).

Also, when generating the artificial EMG z _t in any muscle activity r _t, the signal strength dependent noise variance estimation unit 23, according to the distribution of the variance of the electromyographic signal parameter and muscle activity r _t calculate the estimate of the variance Var [epsilon] of the signal strength dependent noise from the estimated value of the average Shiguma ² of the variance of the myoelectric signal is superimposed on the EMG in muscle activity r _t.

White Gaussian noise generator 24 generates a white Gaussian random numbers w _t. Here, the average of w _t is 0 and the variance is 1.

Shaping filter 25, the frequency characteristics of the white Gaussian random numbers w _t to produce an electromyographic signal x _t the same random number w _'t is shaped in the frequency characteristic. Here, the average of w _'t 0, the dispersion is 1. That is, the shaping filter 25 substantially executes the operation of the expression (18).

Inverse gamma distribution random number generator 26, in accordance with the distribution of the variance of the myoelectric signal defined by the estimated value of the variance Var [epsilon] of the estimate of the mean Shiguma ² of the variance of the myoelectric signal and the signal strength dependent noise Generate a random number σ _t . Here, it is assumed the inverse gamma distribution as a distribution of the variance of the myoelectric signal, using the inverse gamma distribution random number as the random sigma _t.

Also, when generating the artificial EMG z _t in any muscle activity r _t, inverse gamma distribution random number generator 26, the average Shiguma ² of estimation of the variance of the EMG in muscle activity r _t in accordance with the distribution of the variance of the electromyographic signals in value and muscle activity r variance of the signal intensity depends noise superimposed on the myoelectric signal at _t Var [epsilon] muscle activity r _t defined by the estimated value of Generate a random number σ _t .

  The shape parameter α and the scale parameter β, which are parameters that define the inverse gamma distribution, maximize the marginal likelihood of a myoelectric signal sequence obtained in advance based on, for example, the second type of maximum likelihood estimation. You can ask for it.

Artificial electromyogram signal generator 27 generates an artificial EMG z _t from random sigma _t in accordance with the distribution of the variance of the random number w _'t and electromyographic signals after shaping in shaping filter 25. That is, the artificial myoelectric signal generation unit 27 substantially executes the calculation of Expression (19).

≪Demonstration experiment≫
(Dispersion distribution estimation accuracy)
An experiment was performed using artificial myoelectric signals generated artificially to confirm the variance distribution estimation accuracy by the myoelectric signal processing device 20. The generation of the artificial myoelectric signal was performed in the following procedure, so that a signal in which noise was superimposed on the variance was pseudo-generated.
(1) Generate a discrete sequence {σ _t ² ; t = 1,..., T} of length T with random numbers according to an inverse gamma distribution IG (α ₀ , β ₀ ).
(2) At each t, a random number according to the normal distribution N (0, σ _t ² ) is generated one sample at a time, and this is defined as a sequence {x _t }.
(3) {x _t } is regarded as a myoelectric signal measured at the sampling frequency F _s [Hz].
At this time, the average of the variance of {x _t } and the true value of the variance of the variance are as follows.

Dispersion distribution estimation accuracy verification was performed by comparing the σ¯ estimated by these true values and myoelectric signal processing device 20 ² and Var [epsilon].

Estimation of Shiguma _2, using _{{x t}} rectified smoothed signal as _{{y t},} the first L of _{{y t}} was downsampling _F down [Hz] signal [ms] went. At this time, when L is fixed to 1000 [ms] and F _{down is changed} to 500 [Hz], 200 [Hz], 100 [Hz], 50 [Hz], 20 [Hz], and 10 [Hz], And F _down is fixed at 1000 [Hz], and L is changed to 1000 [ms], 500 [ms], 200 [ms], 100 [ms], 50 [ms], 20 [ms], and 10 [ms]. Under each condition, the% error between the true values σ￣ ₀ ² and σ￣ ₂ was calculated. As a comparison method, the maximum likelihood estimated value of the variance calculated from {x _t } under the same conditions was used.

In the estimation accuracy verification of Var [ε], under each condition in which L is fixed to 1000 [ms] and F _down is changed to 1000 [Hz] and 50 [Hz], the% error from Var [ε ₀ ] is calculated. did. The variance based on the maximum likelihood estimation was compared with the variance. Here, the variance of the variance based on the maximum likelihood estimation refers to the variance of the K variance maximum likelihood estimation values calculated by dividing the {x _t } of L [ms] into K and calculating each division. Note that the prior distribution parameters were determined using the empirical Bayes method.

In each experiment, and _{T = 1000, F s = 1000} , was used a second-order Butterworth filter having a cut-off frequency of 1 [Hz] in the smoothing process. The% error is an average value when the true value is changed in 20 ways (α = 15, β = 0.5, 1, 1.5,..., 10).

Figure 4 is a graph showing average error rate according to the estimation of the average Shiguma ₂ of the variance of myoelectric signals at each sampling frequency. Figure 5 is a graph showing average error rate according to the average Shiguma ₂ of estimation of the variance of the electromyographic signals in each time window. From the graph of FIG. 4, F _down ≦ 200, and from the graph of FIG. 5, L ≦ 100, there is a significant difference between the present invention (estimation based on the myoelectric signal model of FIG. 1) and the conventional technique (maximum likelihood estimation). Differences were noted. FIG. 6 is a graph showing the average error rate for estimating the variance Var [ε] of the signal strength dependent noise at each sampling frequency. As a result of performing a multiple comparison test based on the Bonferroni method, a significant difference was confirmed between the present invention and the prior art.

Referring to the graph of FIG. 4, when F _down is high, the error rate is 5% or less for both the present invention and the conventional technology, and the estimation accuracy is high. However, as the F _down decreases, the accuracy of the conventional technology deteriorates. The invention maintains high accuracy. Referring to the graph of FIG. 5, if the window width L is shortened, the accuracy of both the present invention and the conventional technology deteriorates. However, even when L = 10 [ms], the estimation error of the present invention is about 5%, and Significantly smaller than the prior art. Therefore, according to the present invention, it can be said that can be estimated Shiguma ² with fewer samples.

  Referring to the graph of FIG. 6, the estimation error of Var [ε] of the present invention is significantly smaller than that of the related art. This is because the conventional maximum likelihood method assumes that the variance is constant in each division, so that noise superimposed independently at each time cannot be sufficiently expressed. On the other hand, in the present invention, it is possible to estimate the variance of the variance with a good accuracy of 10% or less by estimating the posterior distribution of the population by assuming the variance value independent of each time.

  As described above, according to the myoelectric signal processing device 20 according to the embodiment of the present invention, the distribution of the variance of the myoelectric signal on which the signal intensity-dependent noise is superimposed can be estimated with high accuracy.

Next, in order to confirm the effectiveness of the present invention with respect to actual data, an experiment of estimating the variance posterior distribution P (σ ² | x) was performed using myoelectric signals. In the measurement of the myoelectric signal, an Ag / AgCl electrode was attached to the skin surface on the biceps of five subjects (mean age: 23.3 ± 0.8, right-handed), and the upper arm was dropped in a sitting position and palms were applied. With the forearm bent horizontally and forward with the elbow facing upward, the elbow was placed on the wrist, the load was placed on the wrist, and the myoelectric signal when the posture was maintained for 10 seconds was acquired at a sampling frequency of 1000 [Hz]. At this time, the load was changed to 500 [g], 1000 [g], 1500 [g], and 2000 [g], and measurement was performed for 5 trials each. Note that a multi-telemeter system WEB-5000 (high-frequency cutoff frequency: 100 [Hz], low-frequency cutoff frequency: 5.3 [Hz]) manufactured by Nihon Kohden Corp. was used for the measurement.

For the analysis of the myoelectric signal, data of the latter 5 seconds out of the measured 10 seconds was used. At this time, the weight of the forearm and the hand is 0.022, the center of mass is 0.318, the length ratio from the elbow joint is 0.318, and the moment arm length of the biceps relative to the rotation center of the elbow joint is 0.03 [m]. Assuming, the muscular strength of the biceps was calculated. Then, the change of the distribution posterior distribution P (σ ² | x) with the increase in the load and the change of E [σ ² | x] and Var [σ ² | x] with the change of the muscular strength were obtained.

  The prior distribution parameters were set for each subject based on the empirical Bayes method using one trial of the myoelectric signal under a load of 2000 [g] measured in advance.

  FIG. 7 is a graph showing the posterior distribution of the variance of the myoelectric signal at each load in one trial. FIG. 8 is a graph showing the average of the variance of the myoelectric signal and the estimated value of the variance of the signal strength-dependent noise at each muscle strength in all trials. FIG. 8 also shows test results based on the Steel-Dwass method.

Referring to the graph of FIG. 7, as the load increases, P (σ ² | x) moves to the right and spreads. This means that the posterior average and posterior variance tend to increase as the load increases. Referring to the graph of FIG. 8, it can be seen that the muscular strength can be expressed by the equation (1) since the muscular strength and the posterior average E [σ ² | x] tend to increase monotonically. At this time, when the index a in the equation (1) was obtained, a = 1.18, 0.56, 0.81, 0.74, and 0.65 in the order of the subject. On the other hand, when the index a is obtained using the variance σ ² obtained based on the maximum likelihood estimation as in Non-patent Document 1, a = 1.17, 0.55, 0.78, 0.73, 0.65 It became. From this, it can be seen that the posterior average E [σ ² | x] that can be calculated in the present invention plays the same role as the variance obtained by the maximum likelihood method in the method of Non-Patent Document 1.

On the other hand, the increase of the posterior variance Var [σ ² | x] according to the muscle strength is apparent from the increase of the posterior mean E [σ ² | x] and the equation (13). Var [σ ² | x] is a point proportional to the square of the posterior average E [σ ² | x]. In the past, it has been reported that the variance of muscle strength during isometric contraction is proportional to the square of the average muscular strength, but the present invention has found that a similar relationship holds for the variance of myoelectric signals. Is shown. This means that the noise is superimposed on the variance of the myoelectric signal due to the signal intensity-dependent noise.

  As described above, according to the myoelectric signal processing device 20 according to the embodiment of the present invention, the variance distribution of the myoelectric signal can be estimated, and the noise superimposed on the variance can be estimated.

(Experiment for generating artificial myoelectric signals)
Next, an experiment of generating an artificial myoelectric signal by the myoelectric signal processing device 20 was performed. In the experiment, first measure the myoelectric signal from healthy college one person, dispersion distribution parameters Shiguma ² and Var [epsilon], and estimated the parameters of the shaping filter 25. Then, a comparison was made between the artificial myoelectric signal generated based on the estimated parameters and the measured myoelectric signal.

  The measurement conditions of the myoelectric signal are the same as above. The load was changed to 500 [g], 1000 [g], 1500 [g], and 2000 [g], and each was measured for one trial. For the parameter estimation and comparison, data of the latter 5 seconds of the measured 10 seconds was used.

The post-variance distribution was estimated for each load. However, the prior parameters were set based on the second type of maximum likelihood estimation using a pre-measured myoelectric signal at a load of 2000 [g], and were made common throughout the experiment. The parameters of the shaping filter 25 were determined by the Burg method based on the myoelectric signal at a load of 2000 [g], and the model order was set to M = 20 based on the Bayesian information criterion. The artificial myoelectric signal was generated five times for each load, and the Tanizaki method was used to generate the inverse gamma distribution random number σ _t .

  The comparison was performed for three items: average amplitude, frequency component, and variance distribution. The average amplitude was the average value of the values obtained by rectifying and smoothing the signal, and the average absolute error rate when the value in the measured myoelectric signal was a true value was calculated for each load. In the comparison of the frequency components, the power spectral densities of the measured myoelectric signal and the artificial myoelectric signal were obtained, respectively, and the correlation between them was calculated. As for the comparison of the variance distributions, the parameters α and β are estimated by the second type of maximum likelihood estimation by assuming the distribution of the variance of the myoelectric signal to be an inverse gamma distribution, and the average absolute error rate is calculated similarly to the case of the average amplitude Was calculated.

  FIG. 9 is a graph showing a measured myoelectric signal at each load and an artificial myoelectric signal generated based on the measured myoelectric signal. Referring to the graph of FIG. 9, it can be seen that the amplitude of both the measured myoelectric signal and the artificial myoelectric signal increases as the load increases.

  FIG. 10 is a graph showing the average absolute error rate of the average amplitude of the measured myoelectric signal and the artificial myoelectric signal at each load. Referring to the graph of FIG. 10, it can be confirmed that the artificial myoelectric signal can reproduce the amplitude of the myoelectric signal with an average absolute error rate of about 5% under any load.

  FIG. 11 is a graph showing the correlation coefficient of the power spectral density between the measured myoelectric signal and the artificial myoelectric signal at each load. Referring to the graph of FIG. 11, it can be seen that the correlation coefficient is 0.8 or more for all loads, and that the artificial myoelectric signal can reproduce the myoelectric signal with high accuracy even in the frequency component.

  FIG. 12 is a graph showing the average absolute error rate of the parameter related to the distribution of the variance of the myoelectric signal at each load. Referring to the graph of FIG. 12, the shape parameter α is very small, less than 0.01 [%], although the average error rate tends to increase as the load increases. On the other hand, the average error rate of the scale parameter β is about 5 to 10%, which is larger than the average error rate of α. The reason for this is that, in Equations (13) and (17), α is directly used to derive the posterior distribution, but β is not used in it, and the estimation error for the scale of the posterior distribution has increased. Conceivable. However, since the shape parameter α determines the essential shape of the distribution, while the scale parameter β determines the spread of the distribution, the effect of β on the generation of artificial myoelectric signals is α. We think that it is small and does not matter.

  As described above, the myoelectric signal processing device 20 according to one embodiment of the present invention can generate an artificial myoelectric signal that reproduces the amplitude, frequency, and variance distribution of a measured myoelectric signal.

  As described above, the embodiments have been described as examples of the technology according to the present invention. For that purpose, the accompanying drawings and the detailed description have been provided.

  Accordingly, among the components described in the accompanying drawings and the detailed description, not only those components that are essential for solving the problem, but also those that are not essential for solving the problem in order to exemplify the technology. May also be included. Therefore, it should not be immediately recognized that these non-essential components are essential based on the fact that the non-essential components are described in the accompanying drawings and the detailed description.

Further, since the above-described embodiment is for exemplifying the technology of the present invention, various changes, replacements, additions, omissions, and the like can be made within the scope of the claims or the equivalents thereof. For example, in the above description, the inverse gamma distribution is assumed as the distribution of the myoelectric signal, but in addition to the inverse gamma distribution, the distribution of the myoelectric signal has a property conjugate to the distribution of the myoelectric signal. May be applied.

Reference Signs List 20 myoelectric signal processing device 21 time window processing unit 22 variance average estimation unit 23 signal intensity dependent noise variance estimation unit 24 white Gaussian noise generator (first random number generator)
25 shaping filter 26 inverse gamma distribution random number generator (second random number generator)
27 Artificial myoelectric signal generator

Claims (12)

A time window processing unit that calculates an average for each predetermined time length for a rectified and smoothed signal obtained by rectifying and smoothing the myoelectric signal on which the signal strength dependent noise is superimposed;
A variance average estimating unit for calculating an average of the variance of the myoelectric signal from the reciprocal of the average of the rectified and smoothed signal and the filter gain related to the rectification and smoothing of the myoelectric signal;
A step of calculating an estimated value of the variance of the signal intensity-dependent noise from the parameter that defines the average value of the variance of the myoelectric signal and the distribution of the variance of the myoelectric signal. Prepared ,
The myoelectric signal processing method , wherein the distribution of the myoelectric signal variance has a property conjugate to the distribution of the myoelectric signal and satisfies the non-negativeness of the variance . A first random number generator generating white normal random numbers;
A shaping filter shapes the frequency characteristic of the white normal random number to a frequency characteristic similar to the myoelectric signal,
A second random number generator for generating random numbers according to the distribution of the variance of the myoelectric signal defined by the estimated value of the average of the variance of the myoelectric signal and the estimated value of the variance of the signal intensity-dependent noise. When,
The myoelectric signal processing method according to claim 1, further comprising the step of: generating an artificial myoelectric signal from the shaped random number and a random number according to a distribution of variance of the myoelectric signal. . The variance average estimating unit calculates the myoelectric signal at the myoactivity from the reciprocal of the rectified and smoothed signal at the time of maximum voluntary muscle contraction, a designated muscle activity and a reciprocal of a filter gain related to the rectification and smoothing of the myoelectric signal. Calculating an estimate of the mean of the variance;
The signal intensity-dependent noise variance estimating unit converts the parameter that defines the distribution of the variance of the myoelectric signal and the estimated value of the average of the variance of the myoelectric signal at the muscular activity into the myoelectric signal at the muscular activity. Calculating an estimate of the variance of the signal strength dependent noise to be superimposed;
The second random number generator defines an average value of the variance of the myoelectric signal at the muscle activity and an estimated value of the variance of the signal intensity-dependent noise superimposed on the myoelectric signal at the muscle activity. Generating a random number in accordance with the distribution of the variance of the myoelectric signal at the muscle activity level to be performed. The distribution of the variance of the myoelectric signal is an inverse gamma distribution,
4. The myoelectric signal processing method according to claim 1, wherein the parameter is a shape parameter of an inverse gamma distribution. A time window processing unit that calculates an average for each predetermined time length with respect to a rectified and smoothed signal obtained by rectifying and smoothing the myoelectric signal on which the signal strength dependent noise is superimposed,
A variance average estimating unit that calculates an average of the variance of the myoelectric signal from the reciprocal of the average of the rectified and smoothed signal and the filter gain related to the rectification and smoothing of the myoelectric signal,
A signal strength-dependent noise variance estimator that calculates an estimated value of the variance of the signal strength-dependent noise from a parameter that defines an average value of the variance of the myoelectric signal and a distribution of the variance of the myoelectric signal,
The myoelectric signal processing device is a distribution of the variance of the myoelectric signal having a property conjugate to the distribution of the myoelectric signal and satisfying the non-negativeness of the variance . A first random number generator for generating white normal random numbers;
A shaping filter for shaping the frequency characteristics of the white normal random numbers into the same frequency characteristics as the myoelectric signal,
A second random number generator that generates a random number according to the distribution of the variance of the myoelectric signal defined by the estimated value of the average of the variance of the myoelectric signal and the estimated value of the variance of the signal intensity-dependent noise;
The myoelectric signal processing device according to claim 5, further comprising: an artificial myoelectric signal generation unit configured to generate an artificial myoelectric signal from the shaped random number and a random number according to a distribution of the variance of the myoelectric signal. The variance average estimating unit calculates the myoelectric signal at the myoactivity from the reciprocal of the rectified and smoothed signal at the time of maximum voluntary muscle contraction, a designated muscle activity and a reciprocal of a filter gain related to the rectification and smoothing of the myoelectric signal. Computes an estimate of the mean of the variance,
The signal intensity-dependent noise variance estimating unit converts the parameter that defines the distribution of the variance of the myoelectric signal and the estimated value of the average of the variance of the myoelectric signal at the muscular activity into the myoelectric signal at the muscular activity. Calculating an estimate of the variance of the signal strength dependent noise to be superimposed,
The second random number generator defines an average value of the variance of the myoelectric signal at the muscle activity and an estimated value of the variance of the signal intensity-dependent noise superimposed on the myoelectric signal at the muscle activity. The myoelectric signal processing device according to claim 6, wherein a random number is generated in accordance with a distribution of a variance of the myoelectric signal at the muscle activity. The distribution of the variance of the myoelectric signal is an inverse gamma distribution,
The myoelectric signal processing device according to claim 5, wherein the parameter is a shape parameter of an inverse gamma distribution. Means for calculating an average for each predetermined time length for a rectified and smoothed signal obtained by rectifying and smoothing the myoelectric signal on which the signal strength dependent noise is superimposed,
Means for calculating an average of the variance of the myoelectric signal from the average of the rectified and smoothed signal and the reciprocal of a filter gain relating to the rectification and smoothing of the myoelectric signal; and estimating the average of the variance of the myoelectric signal. A computer functioning as a means for calculating an estimated value of the variance of the signal strength-dependent noise from a value and a parameter defining a distribution of the variance of the myoelectric signal,
A myoelectric signal processing program , wherein the distribution of the myoelectric signal variance has a property conjugate to the distribution of the myoelectric signal and satisfies the non-negativeness of the variance . Means for generating white normal random numbers,
Means for shaping the frequency characteristics of the white normal random numbers into the same frequency characteristics as the myoelectric signal,
Means for generating a random number according to the distribution of the variance of the myoelectric signal defined by the estimated value of the variance of the myoelectric signal and the estimated value of the variance of the signal intensity-dependent noise; and 10. The myoelectric signal processing program according to claim 9, causing a computer to function as means for generating an artificial myoelectric signal from random numbers according to a distribution of the variance of the myoelectric signal. The rectified smoothed signal at the maximum voluntary muscle contraction, the estimated value of the average of the variance of the myoelectric signal at the myoelectric activity from the reciprocal of the filter gain according to the specified muscular activity and the rectifying and smoothing of the myoelectric signal. Means to calculate,
The parameter defining the distribution of the variance of the myoelectric signal and the estimated value of the average of the variance of the myoelectric signal at the muscular activity, the variance of the signal intensity-dependent noise superimposed on the myoelectric signal at the muscular activity Means for calculating an estimated value, defined by an estimated value of the average of the variance of the myoelectric signal at the muscular activity and an estimated value of the variance of the signal intensity-dependent noise superimposed on the myoelectric signal at the muscular activity. The myoelectric signal processing program according to claim 10, causing a computer to function as a means for generating a random number according to a distribution of variance of the myoelectric signal at the muscle activity level. The distribution of the variance of the myoelectric signal is an inverse gamma distribution,
The myoelectric signal processing program according to claim 9, wherein the parameter is a shape parameter of an inverse gamma distribution.

Images