JP2913611B2 - Learning type electromyogram pattern recognition robot hand controller - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network for recognizing a pattern obtained by performing a high-speed Fourier transform on a myoelectric signal detected from a skin surface electrode and analyzing a frequency by a neural network, and using a finger bending angle recognition value. The present invention relates to a learning-type electromyographic pattern recognition robot hand control device that performs a control operation.

[0002]

2. Description of the Related Art The idea of controlling a robot hand or an artificial hand as an artificial hand by means of a myoelectric signal has been proposed by N.I. Published by Winer in 1948 (Cybernetics, The
M. I. T. Press, 1948), then
Various devices have been developed in practice. In view of myoelectric prosthesis, it is a method of controlling the operation of the prosthesis by the myoelectric potential of the stump of the cut portion, and usually detects the myoelectric potential of about 100 to 1000 microvolts from the stump via several pairs of skin surface electrodes. Amplifies with an amplifier, performs electrical pre-processing (rectification, smoothing, etc.), and recognizes and identifies the processed signal to control the operation of the fingers or indirect joints of the prosthetic hand. Is widely spread.

[0003] Conventional myoelectric prostheses are broadly classified into two types for the purpose of operation control, regardless of the research level or the practical level.

[0004] The first purpose is to operate the forearm hand tool as an object to be controlled from the detected electromyogram.
It is intended to open and close the movement of fingers of a prosthetic hand (hand tool) by limiting the function to one degree of freedom.

[0005] Two types of control methods with one degree of freedom, an on / off control method and a proportional control method, are representative.

In the on / off control method, an in-phase noise signal is removed from myoelectricity detected from two electrodes of a residual flexor and an extensor at a stump, and a difference signal is obtained from the rectified and smoothed two signals. To open and close the prosthetic hand (edited by the Japanese Society of Orthopedic Surgery, Japanese Society of Rehabilitation Medicine, Checkpoint for Prosthetic Orthotics, Medical Shoin, 1987, pp. 59-68).

The proportional control method utilizes the neurophysiologically known fact that a substantially proportional relationship holds between the integrated myoelectric signal and muscle tone, and adjusts the opening / closing speed and grip force of the artificial hand. It is a method of performing. The proportional control method uses pulse width modulation. When the myoelectric signal is weak, the motor for opening and closing the prosthesis is turned on for a short time, and when the muscle contraction increases, the on time is long and the pulse width is large. That is, the motor rotation speed is further increased, and the opening / closing speed of the artificial hand is increased.

Both the on / off control system and the proportional control system switch by setting a threshold value with respect to the rectified and smoothed myoelectric signal. Is set, there is a problem that the wearer cannot move according to the wearer's intention, because the wearer cannot cope with the physical condition of the wearer or a change in the level of myoelectricity with the passage of time after the disconnection.

Normally, in the training process up to the wearing of the myoelectric prosthesis, the wrist joint of the phantom limb is bent,
Bring out the healthy upper limb and amputated limb forward, perform the above exercise, and let the person who will wear the EMG detector observe the EMG that matches the level recognized by the EMG identification circuit. Train. However, even if trained, the intended wearer who cannot generate the myoelectricity of the two channels of flexor and extensor muscles separately from each other is not a target for handling myoelectric prosthesis, which can be said to be a life and death problem for a forearm amputee. There are drawbacks.

The second method focuses on how many degrees of freedom are to be controlled simultaneously from detected myoelectric signals, and aims at controlling the upper arm prosthesis for high-rank amputees such as scapula-thoracic amputation. Things.

One of the myoelectric identification methods for multi-degree-of-freedom control is to smooth and rectify myoelectric signals detected from multiple electrodes,
There is an example in which a time-averaged signal is input to a perceptron at a fixed time interval and learning is performed by giving an operation as a teacher signal (Ryuji Suzuki, Tatsumi Suematsu, learning and discrimination of myoelectric current pattern using Link-8) , Medical Electronics and Biotechnology, Vol. 7, No. 1
No. 1969, pp. 47-48). In this example, the skin surface electrode was mounted just above the three different muscles to identify the seven types of movements. However, due to the characteristic that the perceptron cannot perform pattern separation that cannot be separated linearly,
When the mounting position of the electrode or the type of operation increases, there is a problem that operation cannot be practically identified.

As another example of a myoelectric identification method for multi-degree-of-freedom control, a skin surface electrode is mounted immediately above four different muscles and rectified and smoothed, as in the previous example. On the other hand, there are examples in which five types of shoulder motions were identified using a first-order linear identification function (Kazuo Tani, 7 others, myoelectric pattern identification learning for prosthetic hand control, biomechanism 5, (See The University of Tokyo Press, 1980, pp. 88-95.) When the number of learning EMG patterns increases, it takes time to converge the discriminant function, and in many cases linear separation cannot be performed.
If both the learning pattern and the recognition pattern contain noise, there is a problem that the learning does not converge or the motion cannot be recognized.

The two multi-degree-of-freedom control methods have a common problem that multi-channel identification is required for multi-operation identification.

The third of the myoelectric identification method for multi-degree-of-freedom control
As a second prior art, utilizing the fact that myoelectrics from the same site show different frequency characteristics due to operation, a band filter is applied to a 1-channel myoelectric detected from the forearm flexor carpi flexor to perform 10-division. There have been examples of attempts to identify four types of forearm motions using first and second order linear discriminant functions for the frequency spectrum of the myoelectric signal (Hisashi Sakakibara, et al. 3).
Control of multifunctional forearm prosthesis using name and myoelectric frequency information,
Biomechanism 4, The University of Tokyo Press, 1978, pp. 131-138). Since a linear discriminant function is used, it is usually necessary to create a myoelectric pattern for learning the discriminant function. There is a problem that the subject must be trained to perform the hand motion by decomposing the hand motion performed by the cooperative motion into a simple motion. However, in this example, the frequency characteristic of the myoelectricity uses the characteristic that the strength of the muscle contraction force is small and the change due to muscle fatigue is small.
There is an advantage that reproducibility in motion identification is good.

[0015]

As described above, both the one-degree-of-freedom control method for controlling the finger opening / closing operation and the multi-degree-of-freedom control method for controlling the upper arm or forearm do not have the conventional learning ability. This is a method of operating according to the method of recognizing and identifying myoelectricity, and a body image (self-acquired by growth) that the wearer has already formed so that myoelectricity suitable for the recognition and identification device can be generated by the wearer. (The ability to unconsciously perform the cooperative movements of the muscles of each part of the body) in accordance with the situation. In the conventional linear discriminant function or perceptron method, which has a learning function in the process of recognizing and identifying myoelectricity and converting it into a hand motion pattern, the number of motion patterns to be discriminated increases. However, the probability that patterns cannot be linearly separated increases, and learning often does not converge when learning EMG signals with known motion patterns. In particular, if noise enters the EMG data, learning will be forever. Does not converge (Kazuo Tani, others 7
Name, EMG pattern discrimination learning for prosthetic hand control, Biomechanism 5, University of Tokyo Press, 1980, p.88-
(See page 95).

Further, even if the learning is converged, if noise such as an artifact of myoelectricity is mixed in, there is a problem that the identification is impossible.

Further, when the number of operation patterns for recognition and identification increases, there is a problem that the number of electrodes to be mounted is too large to be practical.

Further, there is a problem that it is impossible to separate and recognize and identify the movement of each finger other than the opening / closing of the finger movement by the conventional identification of the myoelectric signal.

As means for solving such problems, a myoelectric control learning type robot hand for recognizing a myoelectric pattern subjected to a fast Fourier transform (FFT) by a neural network (back propagation) having a non-linear separation capability of the pattern. Has been proposed by the present applicant (Japanese Patent Application No.
114215), a sufficient effect was not obtained.

The present invention has been made in view of the above, and an object of the present invention is to introduce a neural network which is strong in learning and noise as a recognition means, and to use a teacher signal pattern used for learning the neural network. Learning-type electromyographic pattern recognition that performs continuous and time-sequential detection of the user's finger movements from the user's finger movements and accurately performs dynamic and continuous finger bending movements as continuous finger bending angle recognition An object of the present invention is to provide a robot hand control device.

[0021]

In order to achieve the above object, the present invention provides a skin surface electrode, an amplifier for amplifying a myoelectric signal detected from the skin surface electrode, and the amplifier amplified by the amplifier. Filter means for removing low-frequency components and high-frequency components of the myoelectric signal, and high-speed Fourier transform for performing high-speed Fourier transform and frequency analysis of the myoelectric signal from which the low-frequency and high-frequency components have been removed by the filter means Means, a pattern conversion means for converting the signal component level for each bandwidth of the myoelectric signal frequency-analyzed by the Fourier conversion means into a numerical matrix pattern, and a plurality of units and the plurality of units. A neural network having a learning function consisting of connected weighting links is provided. The numerical matrix pattern is used as an input, and the finger bending angle data corresponding to the pattern is input. And a finger bending angle teacher value recognizing unit that recognizes the finger bending angle teacher value in a time series manner. The finger bending angle teacher value recognized by the finger bending angle teacher value recognition unit is used as teacher data. A teacher signal generating means for outputting to the recognizing means, and a finger bending angle of the hand provided with the skin surface electrode is detected as the finger bending angle teacher value, and this detection signal is detected by the finger bending angle of the teacher signal generating means. Finger bending angle detecting means for outputting to the teacher value recognizing section, robot hand controlling means for outputting an operation command signal based on the finger bending angle data outputted from the recognizing means, and output from the robot hand controlling means. And a robot hand that performs an operation based on the operation command signal.

[0022]

According to the learning-type electromyogram pattern recognition robot hand control device of the present invention, a normal person or a forearm amputee attaches the myoelectric generated in the movement of the hand intended by himself to the skin attached to the muscle involved in the movement. It has the ability to detect by the surface electrode, perform fast Fourier transform of the detected myoelectric signal, and separate the frequency-analyzed pattern from those that cannot be linearly separated according to the attribute of each pattern. Train the neural network several times with supervised learning. In the learning process, a finger bending angle teacher value recognizing unit that continuously recognizes the user's finger bending angle teacher value in a time-series manner is provided. Since the finger bending angle teacher value is time-averaged and presented to the neural network, a plurality of finger bending angles can be continuously recognized in time series from the electromyogram pattern, and the corresponding finger motion pattern is recognized. Identify and accurately control the robot hand. In addition, since the finger bending angle teacher value recognition unit takes in the actual finger bending angle detected by the finger bending angle detecting means as the finger bending angle teacher value, more accurate control is possible.

[0023]

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram showing the overall configuration of a learning type electromyographic pattern recognition robot hand control device according to a first embodiment of the present invention. In the figure, a user A who is a robot hand operation candidate wears skin surface electrodes 1 and 3 near the total finger extensor muscles of the wrist and skin surface electrodes 5 and 7 near the superficial finger flexors of the same wrist, respectively. A glove with a finger bending angle recognition sensor (for example, a product name "Data Glove",
Light is passed through an optical fiber attached to the surface of a glove, manufactured by VPL, and a finger joint bending angle which recognizes the bending angle of the finger joint from the attenuation of light accompanying the bending of the fiber is fitted. Here, in the case of a healthy person, it is put on the operating hand, and in the case of a amputee, it is put on the remaining hand.

Reference numerals 9 and 11 denote skin surface electrodes 1, 3, 5, and 7, respectively.
An amplifier that amplifies the myoelectric signals a ₁ and a ₂ detected in
Reference numerals 3 and 15 denote low-cut filters that block low-frequency components from the myoelectric signal b amplified by the amplifiers 9 and 11, and 17, 19, respectively.
Is a high cut filter that cuts off high frequency components from the myoelectric signal c whose low frequency components are cut off by the low cut filters 13 and 15, and 21 and 23 are a high cut filter that cuts off the high frequency components by the high cut filters 17 and 19. A high-speed Fourier transform unit (FFT) 27, which performs high-speed Fourier transform and frequency-analyzes and converts into a myoelectric signal pattern e, 27 is a myo-electrical signal whose frequency is analyzed by the high-speed Fourier transform units 21 and 23. A matrix pattern conversion unit 29 for converting the signal component level of each band of the pattern e into a numerical matrix pattern f, a neural network 29 having a learning function comprising a plurality of units and weighting links connecting these units A numerical matrix pattern f from the matrix pattern conversion unit 27 is input and corresponds to the numerical matrix pattern f. Recognition unit which outputs the data h, - operating data of
Reference numeral 31 denotes a finger bending angle teacher value recognition unit that continuously recognizes the user's finger bending angle teacher value in a time series, and time averages the user's finger bending angle teacher value recognized by the finger bending angle teacher value recognition unit. And an interface unit for presenting to the neural network, generating teacher data for the recognition unit 29 and outputting it as a finger bending angle teacher signal pattern g to continuously recognize the finger bending angle teacher value in a time series. The finger bending angle teacher value presentation interface unit 25 to synchronize the finger bending angle processing in the finger bending angle teacher value presentation interface unit 31 with the processing of the myoelectric signal e by the synchronization signals j, k, m, and n. And a signal control unit 33 for inputting the finger bending angle operation data h from the recognition unit 29 and outputting an operation command signal i corresponding to the finger bending angle operation data h. Department, 5 This robot hand control unit 33
The robot hand reproduces the movement of the finger based on the operation command signal i from the robot hand.

FIG. 2 shows the neural network 3 of the recognition unit 29 in FIG.
7 is a diagram showing a configuration according to FIG. In the figure, reference numeral 39 denotes an input / output pattern storage unit for receiving a pair of input / output patterns of signals f and g from the matrix pattern conversion unit 27 and the finger bending angle teaching value presentation interface unit 31;
41 is an input signal pattern for recognition which is a numerical matrix pattern f supplied from the input / output pattern storage unit 39;
Reference numeral 43 denotes an output pattern output as finger bending angle operation data h, and reference numeral 45 denotes bending angle teacher signal data g supplied from the finger bending angle teacher value presentation interface unit 31.
Is a learning finger bending angle teacher value pattern, 47 is an input layer unit of the neural network 37, 49 is an intermediate layer unit,
Reference numeral 1 denotes an output layer unit, and the weight of the connection by the link between these layers is updated based on the result of the output pattern 43.

FIG. 3 (a) shows the input / output relationship of the units of the neural network 37 in the recognition unit 29, and FIG. 3 (b) shows the structure of the neural cell unit of the neural network 37. The output layer unit 51 fires for the bandwidth-specific level pattern input to the. The output of the output layer unit 51 corresponds to the finger bending angles of the first and second joints of each finger.
It is a joint.

FIGS. 4 to 8 show the high-speed Fourier transform units 21 and
23, which is a high-speed Fourier transform performed by the high-speed Fourier transform units 21 and 23, and
FIG. 4 is a diagram showing a myoelectric signal pattern e subjected to octave frequency analysis. FIG. 4 shows a myoelectric signal pattern e in a state in which all five fingers are bent, and FIG. 5 shows only the index finger bent. FIG. 6 shows a myoelectric signal pattern e with only the middle finger bent, FIG. 7 shows a myoelectric signal pattern e with only the thumb bent, and FIG. A myoelectric signal pattern e in a state where all the fingers are opened is shown.

Next, the operation of the learning-type electromyogram pattern recognition robot hand controller configured as described above will be described based on the learning mode in the neural network 37 of the recognition unit 29 and the hand of the intended hand from the generated electromyogram signal. The operation will be described separately in two modes, a recognition mode for recognizing an operation.

First, the learning mode of the recognition unit 29 will be described. Here, two-channel skin surface electrodes 1, 3
An example will be described in which the myoelectric signals of 5 and 7 are detected to recognize the finger bending angles of a total of 10 joints of the first and second joints of the finger on one side. The user A continuously performs a hand operation intended for the operation for several minutes. At this time, in the case of a healthy person, the hand is actually moved, and in the case of a forearm amputee, the hand movement is imaged in the head as a phantom limb, or the skin surface electrodes 1, 3, The left and right healthy arms 5 and 7 and the cut side which does not actually exist are moved symmetrically on the image. During this operation, the myoelectric signals a ₁ and a ₂ detected by the skin surface electrodes 1, 3, 5 and 7 are amplified by the amplifiers 9 and 11, and are converted into the time-series myoelectric signals b by the low-cut filter 13. 15 and the high-cut filters 17, 19, respectively, the low-frequency component and the high-frequency component are cut off, and the time-series myoelectric signal d is supplied to the high-speed Fourier transform units 21, 23.

In the high-speed Fourier transform units 21 and 23, a 1/3 octave frequency analysis of the myoelectric signal d detected in each operation state of each finger is performed and shown in FIGS. 4 to 8 described above. Is converted into such a myoelectric signal pattern e,
In the matrix pattern converter 27, the signal component level of the myoelectric signal pattern e is converted into a numerical matrix pattern f for each bandwidth. FIG. 9 shows an example of a numerical matrix pattern f showing the relationship between the signal component level and the hand movement for each bandwidth. In FIG. 9, one horizontal line next to input is a pattern of one operation. It corresponds to.

The numerical value matrix pattern f from the matrix pattern conversion unit 27 is input to the neural network 37 of the recognition unit 29. At the same time, an operation corresponding to the numerical matrix pattern f is given to the recognizing unit 29 from the finger bending angle teacher value presentation interface unit 31 as a bending angle teacher signal pattern g. Here, the finger bending angle teacher value presentation interface unit 31 captures the actual finger bending angle of the user A detected by the finger bending angle recognition sensor-equipped glove 8 by the finger bending angle teacher value recognition unit.

Here, the correspondence between the timing of the numerical value matrix pattern f input to the recognizing section 29 and the timing of the bending angle teaching signal pattern g will be described with reference to FIGS. Here, the fast Fourier transform units 21 and 23 are used.
Is a frame length of 500 msec for the time-series myoelectric signal d,
Fast Fourier transform is performed at a frame period of 250 msec, while sampling of the finger bending angle teaching value is performed at 32 points per second. The time-series myoelectric signal is filtered at a high cut frequency of 3000 Hz and a low cut frequency of 50 Hz, and the fast Fourier transform units 21 and 23 perform 1/3 octave analysis on 10 bands (center frequencies 63, 80, 100, 125,
160, 200, 250, 315, 400, 500H
Analyze in z).

First, as shown in FIG. 10, under the above-described conditions, the recognizing unit 29 recognizes four patterns per second. The signals f and g of the time-series electromyogram and the teacher are opened so that their observation windows coincide with each other as shown in FIG. As shown in FIG. 10, the time-series myoelectric signal d is obtained by opening the observation window of 500 msec and taking in the fast Fourier transform sections 21 and 23 to obtain the myoelectric signal pattern e as shown in FIGS. Becomes Further, this is converted into the inpu
t Convert to the next horizontal row of numeric patterns.

In FIG. 10, the myoelectric signal f is EMG (1)
, The finger bending angle teacher value signal g corresponding to the time
Corresponds to Teacher (1). Similarly, Teachers (2) and (3) correspond to EMGs (2) and (3), respectively. At this time, since the sampling rate of the finger bending angle teacher value is different from the sampling rate of the recognition unit 29, the finger bending angle teacher value is time-averaged so that both sampling rates match. In this time averaging process, as shown in FIG. 11, a time window having the same length as the time-series myoelectric signal subjected to the fast Fourier transform is opened for the finger bending angle teacher value sampled at 32 points per second. Capture the point signal,
Time-averaging is performed by taking 1/16 of the sum of these 16 signals. The processing of the finger bending angle teacher value is performed by the finger bending angle teacher value presentation interface unit 31 based on the synchronization signals j, k, m, and n of the signal control unit 25, as shown in FIG. Perform while synchronizing.

As shown in FIG. 11, for the recognition unit 29 at t (1) in FIG. 10, the finger bending teacher value at t (1) in FIG. It is assumed that the finger bending teacher value of t (2) corresponding to this is given to the recognizing unit 29, respectively. FIG. 9 shows g obtained by converting the finger bending angle teaching value by the finger bending angle teaching value presentation interface unit 31 and the corresponding fast Fourier-transformed numerical matrix EMG pattern f.
Is shown as a learning input / output pattern of the recognizing unit 29, and the next row and column of the teacher is the converted finger bending angle teacher value pattern. In this conversion, a finger bending angle of 0 to 120 degrees is proportionally converted from a numerical value of -0.5 to +0.5.

Learning of the neural network 37 in the recognition section 29 is performed based on the flowchart of FIG. First, the weight between each unit of the neural network 37 of the recognition unit 29 and the offset of each unit are initialized by the number of columns (step 11).
0). Next, the user repeats the hand motion for several minutes (step 120), and recognizes the pair of the input / output pattern of the myoelectric signal f and the finger bending angle teacher value g processed as described above. The data is stored in the storage unit 39 (step 130). When the hand operation is completed (step 14
0), the pair of input / output patterns is extracted from the input / output pattern storage unit 39, and the learning of the neural network 37 is performed by, for example, the back propagation method (see Kaoru Nakano, “Neurocomputer”, Technical Review, 1989, p. 47). Then, the weight of each unit of the neural network 37 is updated, and the offset of each unit is updated (step 160).

More specifically, learning is performed by giving pairs of input / output patterns shown in FIG. 9 to the neural network 37 in order from top to bottom. Learning is repeated from the pair (steps 170 to 1).
90). When the number of times of learning repeatedly performed exceeds a predetermined number of times (step 200), the learning is terminated.

Next, the recognition mode of the recognition unit 29, that is, the mode of controlling the robot hand by mounting the skin surface electrodes 1, 3, 5, 7 will be described with reference to the flowchart of FIG. In this recognition mode, the user A first removes the glove 8 with the finger bending angle recognition sensor. When the user is a healthy person, the hand on the side on which the skin surface electrodes 1, 3, 5, 7 are worn is moved, and when the amputee is a hand, the hand is moved in the head by a phantom limb ( Step 210).

The myoelectric signals detected by the skin surface electrodes 1, 3, 5, 7 during the movement or image of the hand are as follows:
Amplified by the amplifiers 9 and 11 as in the learning mode.
Cut filters 13, 15 and high cut filter 1
At 7 and 19, the low-frequency component and the high-frequency component are cut off, respectively, and the time-series myoelectric signal d is converted to the high-speed Fourier transform unit 2.
1 and 23. The myoelectric signal d is discretized for each time window unit of the fast Fourier transform opened on the time axis under the same conditions as in the learning mode and frequency-analyzed. It is supplied as a pattern e (step 220).

After that, the myoelectric signal pattern e is converted into a numerical matrix pattern f in the matrix pattern conversion section 27 in the same manner as in the learning mode described above, and the neural network 37 of the recognition section 29 is used. (Step 23)
0). Numerical matrix pattern f recognizes the corresponding finger bending angle value as finger bending angle data h based on the pattern already learned in the learning mode (step 2).
40), and outputs the result to the robot hand control unit 33 (step 250). The robot hand control unit 33 outputs an operation command signal i formed on the basis of the finger bending angle data h to the robot hand 35, and the robot hand 35 reproduces the hand motion based on the operation command signal i (step). 26
0). If the use is to be continued, the process returns to step 220 and the same operation is repeated (step 270). Otherwise, the hand operation is terminated (step 280).

FIG. 14 shows the myoelectric pattern of the finger bending angle of the first finger of the index finger when the learned recognizing section 29 recognizes the two-channel electromyographic pattern of the finger bending operation of the unlearned hand. An example of recognition in which the finger bending recognition value and the true value of the finger bending angle measured by the glove (data glove) 8 with the finger bending angle recognition sensor at the same time are shown in time series. The finger bending angle (indicated by a solid line) recognized from the myoelectric pattern according to the above embodiment and the true bending angle (indicated by a broken line) are in good agreement, and the effectiveness of the present invention is well shown.

FIG. 15 shows an example of recognition showing the correlation between the finger bending recognition value based on the myoelectric pattern and the true value of the finger bending angle under the same conditions as in FIG. According to this, the two have a correlation coefficient of 0.691, which are almost the same, indicating the effectiveness of the present invention.

It should be noted that the present invention is not limited to the above embodiment, and may have the following configuration, for example.

The number of skin surface electrodes is increased from two pairs of two channels, and the signals from each electrode are subjected to fast Fourier transform processing to convert them into band-specific level numerical patterns, and furthermore, a unit of the input layer 47 of the neural network 37. The number of layers is increased so that patterns from many pairs of electrodes are input simultaneously.

The bandwidth and the number of bands in the fast Fourier transform processing are changed.

The present invention is applied to movement of other parts of the human body instead of fingers.

A neural network other than the back propagation type neural network, which has the ability to separate patterns that cannot be linearly separated by pattern attributes, is used.

An apparatus in which the weight and offset of the neural network of the recognition unit suitable for an unspecified number of users are fixed in advance and the function of the learning mode is omitted is used.

FIG. 16 is a block diagram showing the overall configuration of a learning-type electromyographic pattern recognition robot hand control device according to a second embodiment of the present invention. In this embodiment, the first
Instead of the glove 8 with the finger bending angle recognition sensor of the embodiment,
One or more finger imaging cameras (visible light cameras,
The bending angle of the finger is image-processed and recognized by an infrared thermography camera 53, and the recognition signal p is taken into the finger bending angle teacher value presentation interface unit 31. Other configurations and operations are the same as those of the first embodiment.

The field of application of the present invention includes the use of a myoelectric hand for a forearm amputee, the use of a prosthetic hand for an amputee of another part, a welfare field as a power prosthesis such as a prosthesis, and the use of a healthy person as a pilot in outer space. , Inside a nuclear power plant containment vessel, deep sea, fire site,
Control of robot hand of remote control robot for extreme work such as work in culture vessel of toxic microorganisms, operation of robot hand by doctor at remote operation, blood vessel,
Operation of micro robot hand by doctor at the operation of micro organ such as eyeball, Manned operation of robot hand for handling micro device such as LSI device, Unmanned telephone office, Unmanned radio wave relay facility, Todo, Manned operation of a robotic hand of a remotely controlled robot that performs maintenance management of unmanned substations and other facilities that require maintenance management, input devices to aircraft pilot control devices, and manipulation in virtual space created by computer CG Applications can be considered.

[0052]

As described above, according to the present invention, a pattern obtained by subjecting a myoelectric signal from a skin surface electrode to high-speed Fourier transform processing is a pattern that cannot be linearly separated. A neural network that has a learning function that identifies and learns by the attributes of each robot, performs supervised learning for each type of hand movement, and after learning, recognizes the myoelectricity from the robot hand operator wearing the skin surface electrode using the neural network and the robot. Since the hand is controlled, it automatically extracts the pattern characteristics of the neural network, distinguishes similar patterns, and has an electrode directly above the muscle directly involved in hand movement in learning mode due to its high tolerance. Even if the learning pattern is not located, there are many learning patterns, or if noise is mixed in the learning pattern, the learning converges. Even if noise and artifacts are mixed in, When a group moves in coordination, the myoelectric potentials generated from several muscle groups simultaneously interfere nonlinearly even if the electrodes are not located directly above the muscles directly involved in the hand movement. The corresponding finger bending angle of the hand can be recognized from the skin surface electromyogram reaching the skin surface by crosstalk. Therefore, it is possible to control the bending angle of the finger of the robot hand by recognizing the bending angle of the finger from the skin surface myoelectricity, which was not possible in the past. In addition, with regard to changes in the generated myoelectricity due to changes in physical condition and age, the use of the robot hand is interrupted occasionally before or during use, and the learning mode is performed. Further, it can be used in a state suitable for generated myoelectricity. In addition, in this learning mode, the finger bending angle detecting means detects the actual bending angle of the finger and incorporates it as the finger bending angle teacher value, so that more accurate control is possible.

[Brief description of the drawings]

FIG. 1 is a block diagram showing an overall configuration of a learning-type electromyographic pattern recognition robot hand control device according to a first embodiment of the present invention.

FIG. 2 is a configuration diagram of a neural network of a recognition unit used in the learning-type electromyogram pattern recognition robot hand control device of FIG. 1;

3 (a) is an explanatory diagram showing the input / output relationship of the neuron unit of the recognition unit in FIG. 2, and FIG. 3 (b) is a configuration diagram of the neuron unit.

FIG. 4 is a myoelectric pattern diagram after a fast Fourier transform process.

FIG. 5 is a myoelectric pattern diagram after a fast Fourier transform process.

FIG. 6 is a myoelectric pattern diagram after a fast Fourier transform process.

FIG. 7 is a myoelectric pattern diagram after a fast Fourier transform process.

FIG. 8 is a myoelectric pattern diagram after the fast Fourier transform processing.

FIG. 9 is an explanatory diagram illustrating an example of a learning pattern of a recognition unit including an input pattern which is a signal component level for each bandwidth of a myoelectric signal subjected to frequency analysis and a finger bending angle teacher value pattern;

FIG. 10 is a timing chart of an input / output pattern in a learning mode of the recognition unit.

FIG. 11 is an explanatory diagram illustrating a relationship between sampling of a finger bending angle teacher value and a time averaging method in a learning mode of a recognition unit.

FIG. 12 is a flowchart illustrating a control operation of the recognition unit in a learning mode.

FIG. 13 is a flowchart illustrating a control operation of the recognition unit in a recognition mode.

FIG. 14 is an explanatory diagram showing an example of recognition in which a finger bending recognition value based on a myoelectric pattern and a true value of a finger bending angle are shown in time series.

FIG. 15 is an explanatory diagram showing an example of recognition showing a correlation between a finger bending recognition value based on a myoelectric pattern and a true value of a finger bending angle.

FIG. 16 is a block diagram showing an overall configuration of a learning-type electromyographic pattern recognition robot hand control device according to a second embodiment of the present invention.

[Explanation of symbols]

1, 3, 5, 7 Skin surface electrode 8 Glove with finger bending angle recognition sensor (finger bending angle detecting means) 9, 11 Amplifier 13, 15 Low cut filter (filter means) 17, 19 High cut filter (filter means) 21 , 23 High-speed Fourier transform section (high-speed Fourier transform section) 27 Matrix pattern transform section (pattern transform section) 29 Recognizing section (recognizing section) 31 Finger bending angle teacher value presentation interface section (teacher signal generating section) 33 Robot hand control unit (robot hand control means) 35 Robot hand

Continuation of front page (58) Field surveyed (Int.Cl. ⁶ , DB name) A61F 2/72 B25J 3/00-3/04

Claims (1)

(57) [Claims] A skin surface electrode; an amplifier for amplifying a myoelectric signal detected from the skin surface electrode; and a filter means for removing a low frequency component and a high frequency component of the myoelectric signal amplified by the amplifier. A high-speed filtering of the myoelectric signal from which the low frequency component and the high frequency component have been removed by the filter means.
High-speed Fourier transform means for performing frequency analysis by performing a linear transform;
Pattern conversion means for converting the signal component level for each bandwidth of the myoelectric signal frequency-analyzed by the Fourier conversion means into a numerical matrix pattern; and a plurality of units and a weight for connecting the plurality of units. A recognizing means comprising a neural network having a learning function consisting of a link, receiving the numerical matrix pattern as input, and outputting finger bending angle data corresponding to the pattern; A teacher signal generating means for outputting a finger bending angle teacher value recognized by the finger bending angle teacher value recognition section as teacher data to the recognition means, comprising: The skin surface electrode was provided
The finger bending angle of the hand is detected as the finger bending angle teacher value.
This detection signal is used as the finger bending angle of the teacher signal generation means.
Finger bending angle detecting means for outputting to the teacher value recognizing section, robot hand controlling means for outputting an operation command signal based on the finger bending angle data outputted from the recognizing means, and output from the robot hand controlling means. A learning-type electromyographic pattern recognition robot hand control device, comprising: a robot hand that performs an operation based on an operation command signal.