WO2011007569A1 - 分類推定システムおよび分類推定プログラム - Google Patents

分類推定システムおよび分類推定プログラム Download PDF

Info

Publication number
WO2011007569A1
WO2011007569A1 PCT/JP2010/004585 JP2010004585W WO2011007569A1 WO 2011007569 A1 WO2011007569 A1 WO 2011007569A1 JP 2010004585 W JP2010004585 W JP 2010004585W WO 2011007569 A1 WO2011007569 A1 WO 2011007569A1
Authority
WO
WIPO (PCT)
Prior art keywords
value
input
information
element group
values
Prior art date
Application number
PCT/JP2010/004585
Other languages
English (en)
French (fr)
Inventor
森田昌彦
川田浩史
Original Assignee
国立大学法人筑波大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 国立大学法人筑波大学 filed Critical 国立大学法人筑波大学
Priority to JP2011522736A priority Critical patent/JP5545574B2/ja
Publication of WO2011007569A1 publication Critical patent/WO2011007569A1/ja
Priority to US13/350,711 priority patent/US9392954B2/en

Links

Images

Classifications

    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/24Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316Modalities, i.e. specific diagnostic methods
    • A61B5/389Electromyography [EMG]
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/24Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
    • A61B5/316Modalities, i.e. specific diagnostic methods
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/68Arrangements of detecting, measuring or recording means, e.g. sensors, in relation to patient
    • A61B5/6801Arrangements of detecting, measuring or recording means, e.g. sensors, in relation to patient specially adapted to be attached to or worn on the body surface
    • A61B5/6813Specially adapted to be attached to a specific body part
    • A61B5/6824Arm or wrist
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/72Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235Details of waveform analysis
    • AHUMAN NECESSITIES
    • A61MEDICAL OR VETERINARY SCIENCE; HYGIENE
    • A61BDIAGNOSIS; SURGERY; IDENTIFICATION
    • A61B5/00Measuring for diagnostic purposes; Identification of persons
    • A61B5/72Signal processing specially adapted for physiological signals or for diagnostic purposes
    • A61B5/7235Details of waveform analysis
    • A61B5/7264Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems
    • A61B5/7267Classification of physiological signals or data, e.g. using neural networks, statistical classifiers, expert systems or fuzzy systems involving training the classification device
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F3/00Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
    • G06F3/01Input arrangements or combined input and output arrangements for interaction between user and computer
    • G06F3/011Arrangements for interaction with the human body, e.g. for user immersion in virtual reality
    • G06F3/015Input arrangements based on nervous system activity detection, e.g. brain waves [EEG] detection, electromyograms [EMG] detection, electrodermal response detection
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F3/00Input arrangements for transferring data to be processed into a form capable of being handled by the computer; Output arrangements for transferring data from processing unit to output unit, e.g. interface arrangements
    • G06F3/01Input arrangements or combined input and output arrangements for interaction between user and computer
    • G06F3/017Gesture based interaction, e.g. based on a set of recognized hand gestures
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V40/00Recognition of biometric, human-related or animal-related patterns in image or video data
    • G06V40/20Movements or behaviour, e.g. gesture recognition
    • G06V40/28Recognition of hand or arm movements, e.g. recognition of deaf sign language
    • GPHYSICS
    • G16INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16HHEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients

Definitions

  • the present invention relates to a classification estimation system and a classification estimation program for estimating classification information including a motion intention of a subject via a selective desensitization neural network.
  • the machine control is not limited to devices that assist human actions such as nursing / welfare devices (self-helper, self-helper) such as prosthetic hands and prosthetic limbs, and power assist devices such as robot suits. It is expected to be applied to more general control devices such as human interface devices (see Non-Patent Document 1, etc.).
  • the biological information used for the machine control includes, for example, central nervous system information such as information on changes in brain waves and cerebral blood flow, the muscles generated when the muscles contract, and the surroundings of the muscles Peripheral nervous system information such as myoelectric signal (myoelectric signal, myoelectric potential signal, electromyogram) which is information of changes in myoelectric potential (myoelectricity, myoelectric action potential) that is the potential difference between the skin and (For example, refer nonpatent literature 1 etc.).
  • central nervous system information such as information on changes in brain waves and cerebral blood flow, the muscles generated when the muscles contract, and the surroundings of the muscles
  • Peripheral nervous system information such as myoelectric signal (myoelectric signal, myoelectric potential signal, electromyogram) which is information of changes in myoelectric potential (myoelectricity, myoelectric action potential) that is the potential difference between the skin and (For example, refer nonpatent literature 1 etc.).
  • the myoelectric potential signal is observed 30 to 100 [ms] earlier than the contraction of the muscle (see, for example, Non-Patent Document 2). That is, the myoelectric potential signal can be measured with little time delay with respect to the muscle contraction.
  • the myoelectric signal measuring device has high responsiveness (high sensitivity and high speed response), high reliability of measurement values, and can be procured at a relatively low cost. For this reason, it is realistic to use the myoelectric potential signal as biological information for machine control, and it is particularly promising.
  • a method for measuring the myoelectric potential signal a method of directly measuring a specific myoelectric potential signal by inserting a needle electrode into a specific myofiber of a subject (subject, subject), and the skin of the subject
  • a method is generally known in which an electrode (surface electrode) is attached to a wire and a total of myoelectric signals generated by a plurality of muscle fibers, that is, a so-called surface myo-Gram (EMG) is measured.
  • EMG surface myo-Gram
  • the surface myoelectric potential signal is the sum of the myoelectric potential signals and becomes more non-linear
  • the correspondence with the motion (muscle activity) of the subject is complicated. That is, it is known that it is difficult to identify the movement based on the surface myoelectric potential signal. For this reason, when the subject's movement is identified from the measurement result of the surface electromyogram signal, a neural network (neural network) that simplifies a function (nonlinear function) representing a correspondence relationship between an input and an output by a distributed expression.
  • a neural network neural network
  • a function nononlinear function
  • Non-Patent Document 1 three (3ch, 3 channels, 3 channels) are provided by a multilayer perceptron (MLP: Multi-Layer Perceptron, Multi-Layered Perceptron) using an error back propagation method (BP: Back-Propagation, BP method).
  • MLP Multi-Layer Perceptron, Multi-Layered Perceptron
  • BP Back-Propagation, BP method.
  • Electrodes surface electrode and ground electrode
  • 10 types of forearm motions as outputs can be identified with an average discrimination rate of 90%. It is described.
  • Non-Patent Document 1 surface myoelectric potential signals measured by attaching 6-ch electrodes to the subject's forearm using R-LLGMN (Recurrent Log-Linearized Gaussian Mixture Network) are input, and 6 types of outputs are provided. It has been described that the forearm motion of can be identified with an average identification rate of 85 [%].
  • the multilayer perceptron is a so-called layered neural network in which input / output and functions are expressed in a distributed manner by elements (neurons) included in the input layer, the intermediate layer, and the output layer.
  • elements neurons
  • each element of the input layer and the intermediate layer and each element of the intermediate layer and the output layer are coupled with a coupling load, and each element of the intermediate layer and the output layer is coupled to the input layer through the coupling load.
  • the function is expressed in a distributed manner by being calculated as the sum of the inputs from each element of the intermediate layer.
  • the error back propagation method is a learning algorithm used for learning the multilayer perceptron.
  • the output by the multilayer perceptron is the learning data. It is a learning method that corrects the coupling weight between each layer to match the output.
  • the multilayer perceptron using the error back propagation method is described in, for example, Non-Patent Document 3 and the like, and thus detailed description thereof is omitted.
  • LLGMN is a layered neural network containing a Gaussian Mixture Model (GMM) as an example of a statistical model and a Hidden Markov Model (HMM) as an example of a probability model.
  • GMM Gaussian Mixture Model
  • HMM Hidden Markov Model
  • the Gaussian mixture model is a statistical model expressed by a linear combination of Gaussian distributions (normal distributions).
  • the Markov model is a probability model in which a symbol appears in accordance with a stochastic process in which the appearance probability of a certain symbol is generally determined by the immediately preceding n symbols, that is, a so-called Markov process (a stochastic process having Markov properties).
  • the hidden Markov model is a Markov model that assumes that the system is a Markov process with unknown parameters, and estimates the unknown parameters from observable information.
  • the R-LLGMN is a LLGMN having a so-called recurrent coupling that can take into account the output from the previous time between the first intermediate layer and the second intermediate layer.
  • the R-LLGMN is described in, for example, Japanese Patent Application Laid-Open No. 2005-11037, Non-Patent Document 4, and the like, and thus detailed description thereof is omitted.
  • Non-Patent Document 2 discloses seven types of forearm motions (basic postures) based on surface myoelectric potential signals measured by attaching a 4ch electrode to a subject's forearm using a support vector machine (SVM). , Wrist flexion, wrist extension, grasping, opening, wrist gyration, wrist gyration), that is, a technique of outputting a discrimination result by inputting a surface myoelectric potential signal.
  • Non-Patent Document 2 describes that seven types of forearm movements could be identified with an average identification rate of 93.30 [%].
  • the support vector machine is a two-class pattern classifier using a plurality of linear input elements (linear threshold elements) as elements of the input layer. That is, it is a layered neural network that outputs an output value indicating whether or not the sum of values of a plurality of linear input elements multiplied by a coupling load (synaptic load) exceeds a preset threshold value.
  • the element value (output value) of y is y
  • the transposed vector of coupling load vector omega and omega T if the threshold has been is h
  • the output value y is indicated by the following equation (1).
  • sign ( ⁇ T x ⁇ h) is a sign function that becomes +1 when ( ⁇ T x ⁇ h)> 0 and becomes ⁇ 1 when ( ⁇ T x ⁇ h) ⁇ 0. That is, it is a binary function that becomes +1 when the inner product of the vectors x and ⁇ exceeds the threshold value h, and ⁇ 1 when it does not exceed the threshold value h.
  • the support vector machine has a maximum distance from each learning point (training sample), that is, a pair of input feature vector x and output value y, that is, a distance (margin) from the input feature vector x.
  • the parameter of the linear input element is learned, that is, the combined load vector ⁇ is updated based on the standard of margin maximization for obtaining a separation plane (hyperplane).
  • Non-Patent Documents 1 and 2 In the technologies of Non-Patent Documents 1 and 2 and the like, the basic posture of the forearm, wrist flexion, wrist extension, grasping, opening, wrist rotation, wrist rotation, and the like are identified, but the same movement is performed. There is a problem that only the identification is performed when the is always operated at a constant speed. For this reason, the techniques of Non-Patent Documents 1 and 2 have a problem that the same operation cannot be identified at different speeds.
  • the surface myoelectric potential signal has a large individual difference, and in order to perform motion identification with high accuracy, learning data is learned for each subject, and a binding load suitable for each subject is set.
  • Non-Patent Documents 1 and 2 perform motion identification based on a statistical method, and thus require a large amount of learning data, which imposes a heavy burden on the subject and is not practical. It was.
  • Non-Patent Documents 1 and 2 when trying to identify more motions or identifying motion speeds, it is complicated to extract the types of increased motions and speed characteristics. It is necessary to execute feature extraction processing. For this reason, the techniques of Non-Patent Documents 1 and 2 have a problem that the more complicated the operation is identified, the more difficult it is to execute in real time. Therefore, the techniques of Non-Patent Documents 1 and 2 have a problem that a large number of operations cannot be discriminated in real time with a small amount of learning data (learning amount and number of samples) with a high identification rate along with speed. It was.
  • the present invention has a first technical problem to identify the movement of the subject with a small learning amount. Moreover, this invention makes it a 2nd technical subject to estimate the classification information containing the motion intention of a subject with a small learning amount.
  • the classification estimation system provides: A first input element group composed of a plurality of input elements to which the value of the first input variable based on the first information is input, and a plurality of inputs to which the value of the second input variable based on the second information is input; An input element group having a second input element group constituted by elements; A value of each input element of the first input element group and a value of each first output sensitivity calculated so that some values become zero based on the value of each input element of the second input element group; The first multiplexing including at least the integration of each of the first multiplexing output sensitivity values calculated so that some values become zero based on the values of the elements of the preset multiplexing element group
  • the first intermediate element is configured by a plurality of intermediate elements to which the value of the first intermediate variable calculated as the value of each input element of the first input element group that has been desensitized and has some values set to zero is input.
  • An intermediate element group having an intermediate element group; A plurality of values of output variables calculated based on the value of each intermediate element of the first intermediate element group and the value of each coupling load set in advance according to the importance of the value of each intermediate element are input.
  • the invention according to claim 2 is the classification estimation system according to claim 1, A value of each input element of the second input element group and a value of each second output sensitivity calculated so that a part of the values becomes zero based on the value of each input element of the first input element group; Second desensitization including at least integration of each second multiplexing output sensitivity value calculated so that some values become zero based on the values of each element of the multiplexing element group.
  • a second intermediate element group constituted by a plurality of intermediate elements to which the value of the second intermediate variable calculated as the value of each input element of the second input element group that has been made partly zeroed is input
  • the intermediate element group having The value of the output variable calculated based on the value of each intermediate element of each intermediate element group and the value of each coupling load set in advance according to the importance of the value of each intermediate element is input.
  • An output element group It is provided with.
  • the invention according to claim 3 is the classification estimation system according to claim 1 or 2, Based on the difference between the classification information estimated by the classification information estimation means and the actual classification information from which the input first information and the second information are extracted, A connection weight learning means for learning the value of each of the connection weights by updating; It is provided with.
  • the invention according to claim 4 is the classification estimation system according to any one of claims 1 to 3, A value of each element of the multiplexing element group is calculated based on the first information and the second information.
  • the invention according to claim 5 is the classification estimation system according to any one of claims 1 to 4,
  • the first measurement position arranged on the subject when the classification information is the motion intention of the subject and the first information and the second information are the biological signals of the subject accompanying the motion intention
  • a measurement member comprising: a first measurement unit that measures the biological signal in; and a second measurement unit that measures the biological signal at a second measurement position different from the first measurement position disposed on the subject; It is provided with.
  • the invention according to claim 6 is the classification estimation system according to claim 5,
  • the myoelectric potential signal in the subject's skin of the myoelectric potential signal that is information on the change in electric potential based on the muscle activity that is the activity of the muscle group of the subject accompanying the action intention is the surface myoelectric potential signal
  • the biological signal is the surface myoelectric potential signal
  • the first measurement unit that measures the surface myoelectric potential signal at the first measurement position disposed on the skin, and the second measurement disposed on the skin.
  • the second measurement unit that measures the surface myoelectric potential signal at a position, and the measurement member A first integrated value that is an integrated value of the surface myoelectric potential signal from the first measuring unit is calculated, and a second integrated value that is an integrated value of the surface myoelectric potential signal from the second measuring unit is calculated.
  • Integral value calculating means A first integral average value, which is an average value of the first integral values within a preset frame time, is calculated for each preset shift time, and an average value of the second integral values within the frame time is calculated.
  • An integral average value calculating means for calculating a second integral average value for each shift time; When the first information is the first integral average value and the second information is the second integral average value, the value of the first input variable based on the first integral average value is the shift time.
  • the first input element group that is input every time, and the second input element group in which the value of the second input variable based on the second integral average value is input every shift time. Group, It is provided with.
  • the invention according to claim 7 is the classification estimation system according to claim 5 or 6,
  • the classification information estimating means for estimating the classification information having type information which is information on the type of movement specified by the action intention and force information which is information on the movement force; It is provided with.
  • the invention according to claim 8 is the classification estimation system according to claim 7,
  • the classification information estimating means for estimating the classification information which is speed information that is information on the speed of the motion, and the speed information based on the force information; It is provided with.
  • the invention according to claim 9 is the classification estimation system according to claim 8, Of the output variable values, a first output element group composed of a plurality of output elements to which a value of the first output variable corresponding to the first operation intention is input, and a second output element corresponding to the second operation intention.
  • a second output element group composed of a plurality of output elements to which values of two output variables are input, and the output element group, Correspondence relationship between the type information and speed information of the first action intention and the value of the first output variable, the type information of the second action intention, the speed information and the value of the second output variable
  • the correspondence relationship information storage means for storing the correspondence relationship information for specifying the correspondence relationship; Based on the correspondence relationship information stored in the correspondence relationship information storage means, the calculated value of the first output variable, and the calculated value of the second output variable, the input first
  • the classification information estimating means for estimating the type information and the speed information of the classification information according to the information and the second information;
  • the classification information display means for displaying the type information and the speed information of the classification information estimated by the classification information estimation means; It is provided with.
  • the classification estimation program includes: Computer A first input element group composed of a plurality of input elements to which the value of the first input variable based on the first information is input, and a plurality of inputs to which the value of the second input variable based on the second information is input; An input value storage means for storing the input value of the first input variable and the value of the second input variable for an input element group including a second input element group configured by elements; A value of each input element of the first input element group and a value of each first output sensitivity calculated so that some values become zero based on the value of each input element of the second input element group; The first multiplexing including at least the integration of each of the first multiplexing output sensitivity values calculated so that some values become zero based on the values of the elements of the preset multiplexing element group
  • the first intermediate element is configured by a plurality of intermediate elements to which the value of the first intermediate variable calculated as the value of each input element of the first input element group that has been desensitized and has some
  • An intermediate value calculating means for calculating the value of the first intermediate variable for an intermediate element group having an intermediate element group; A plurality of values of output variables calculated based on the value of each intermediate element of the first intermediate element group and the value of each coupling load set in advance according to the importance of the value of each intermediate element are input.
  • Output value calculation means for calculating the value of the output variable for an output element group constituted by the output elements of Correspondence information storage means for storing correspondence information specifying a correspondence relation between a plurality of classification information as classification targets and the value of the output variable corresponding to each classification information, The classification according to the inputted first information and second information based on the correspondence information stored in advance in the correspondence information storage means and the calculated value of the output variable Classification information estimation means for estimating information, Classification information display means for displaying the classification information estimated by the classification information estimation means; It is made to function as.
  • the classification information is input by inputting the value of each input variable based on the first information and the second information to the selective desensitization neural network capable of multiple desensitization.
  • the classification information is estimated via the selective desensitization neural network that can be estimated and has high learning ability (generalization ability), and therefore the classification information is not estimated via the selective desensitization neural network capable of multiple desensitization Compared to the above, classification information can be estimated with a small learning amount.
  • the values of two input variables are mutually modified in order to identify the movement of the subject through the mutual modification model of the selective desensitization neural network capable of multiple desensitization
  • the value of two intermediate variables can be calculated by multiple desensitization
  • the value of one intermediate variable can be calculated by multiplying the values of two input variables (multiple desensitization of only one of them)
  • the learning ability of the selective desensitization neural network capable of multiple desensitization can be increased, and the classification information can be estimated with a higher identification rate.
  • the value of the output variable can be calculated based on the learned value of each coupling load, and the classification information can be estimated with a higher identification rate.
  • the value of each element of the multiplexing element group calculated based on the first information and the second information is calculated based on the first information or the second information. Compared to the value of each element in the multiplexed element group, the value of each input element in the first input element group or the value of each input element in the second input element group is likely to be different, and is desensitized.
  • the learning ability of the selective desensitization neural network that allows multiple desensitization can be increased, and classification information can be estimated with a higher identification rate. can do.
  • it is possible to estimate the movement intention of the subject by inputting the value of each input variable based on the biological signal to the selective desensitization neural network capable of multiple desensitization.
  • the movement of the subject can be recognized.
  • the movement intention of the subject is determined by inputting the value of the input variable based on the integrated average value of the surface myoelectric potential signal calculated every shift time to the selective desensitization neural network. It can be estimated, and the influence of noise on the surface myoelectric potential signal can be reduced as compared with the case where the surface myoelectric potential signal or the value of the input variable based on the integral value is directly input as the biological signal.
  • the change of the surface myoelectric potential signal accompanying the movement of the subject is observed about 30 [ms] to 100 [ms] earlier than the muscle contraction accompanying the movement of the subject.
  • the movement of the subject can be identified in real time by setting the shift time to a time shorter than 30 [ms].
  • the motion type and force can be estimated.
  • the motion speed can be estimated when the motion intention of the subject is estimated.
  • the value of the first output variable corresponding to the first action intention is compared with the value of the second output variable corresponding to the second action intention, The type and speed can be estimated.
  • FIG. 1 is a diagram illustrating the entire operation identification system according to the first embodiment of the present invention.
  • FIG. 2 is an explanatory diagram of six types of forearm motions of the subject identified by the motion identification system according to the first embodiment of the present invention.
  • FIG. 3 is an explanatory diagram showing the function of each device constituting the operation identification system according to the first embodiment of the present invention in a block diagram (functional block diagram).
  • FIG. 4 is an explanatory diagram for the integral average value of the first embodiment of the present invention.
  • FIG. 5 is an explanatory diagram of the selective desensitization neural network according to the first embodiment of the present invention.
  • FIG. 6 is an explanatory diagram of an operation identification start image according to the first embodiment of the present invention.
  • FIG. 7 is an explanatory diagram of the correspondence information according to the first embodiment of the present invention, in which the vertical axis is voltage [mV], the horizontal axis is time [s], and the integrated average value when six types of operations are actually executed sequentially. It is a graph which shows the total value of the error removal value, and the range which identified operation
  • FIG. 8 is an explanatory diagram of an operation identification image according to the first embodiment of the present invention.
  • FIG. 9 is a flowchart of the operation identification process of the operation identification program according to the first embodiment of the present invention.
  • FIG. 10 is a flowchart of input value generation processing of the operation identification program according to the first embodiment of the present invention.
  • 11 is an explanatory diagram of the experimental results of Experimental Example 1, and the six operations “operation A” to “operation F” in one set of all 18 operations with the vertical axis as the identification value and the horizontal axis as time [s].
  • 11A is a graph of identification values when the subject actually executes at intervals of 2 seconds
  • FIG. 11A is a graph of correct values of identification values of actual six actions
  • FIG. 11B is identification of each frame of actual six actions.
  • FIG. 11C is a graph after the actual measurement values of the identification values for each frame of the actual six operations are smoothed.
  • FIG. 12 is an explanatory diagram of the experimental results of Experimental Example 1.
  • FIG. 13 is an explanatory diagram of the experimental results of Experimental Example 1.
  • Six types of operations are performed when six sets of all 18 operations are executed with the vertical axis indicating the operation speed identification rate [%] and the horizontal axis indicating the types of operation. It is a line graph of identification rate [%] according to speed of three speeds.
  • FIG. 14 is an explanatory diagram of the experimental results of Experimental Example 1.
  • FIG. 15 is an explanatory diagram of the experimental results of Experimental Example 1, and is an explanatory diagram illustrating the estimated delay from the start of the operation by the transition of the identification rate with the passage of time.
  • Six types of operations according to the elapsed time from the start of the operation It is explanatory drawing of the table
  • 16 is a graph showing the transition of the three-speed identification rate of the six types of motion according to the elapsed time from the start of the operation based on the values in the table of FIG. 15 with the vertical axis representing the identification rate [%] and the horizontal axis representing the time [ms]. It is.
  • the value of the output sensitivity g 1 1,2 ⁇ g 30 1,2 , g j 1,2 ... first output Sensitivity value, g 1 2,1 to g 30 2,1 , g j 2,1, ... second output sensitivity value, g 1 1,2,8 to g 30 1,2,8 , g k 1,2 , 8 ...
  • First multiplexing output sensitivity value g 1 2,1,8 to g 30 2,1,8 , g k 2,1,8 ...
  • Second multiplexing output sensitivity value N i (p) , N a (p), N d (p), (x 1 i , x 2 i ,..., X 30 i ) ...
  • second intermediate variable x 1 ⁇ , ⁇ to x 30 ⁇ , ⁇ , x i ⁇ , ⁇ , x j ⁇ , ⁇ ... values of intermediate elements, x 1 1,2 to x 30 1,2 , x i 1,2, x j 1,2 ...
  • FIG. 1 is an overall explanatory diagram of an operation identification system according to a first embodiment of the present invention.
  • FIG. 2 is an explanatory diagram of six types of forearm motions of the subject identified by the motion identification system according to the first embodiment of the present invention.
  • a motion identification system motion intention estimation system, event estimation system, classification estimation system S according to the first embodiment of the present invention is associated with a motion intention (event) of a human being as an example of a subject.
  • An identification client personal computer computer, personal computer, user-operable terminal
  • PC as an example of a motion identification device main body (motion intention estimation device main body, event estimation device main body, classification estimation device main body) for identifying a subject's motion;
  • the surface myoelectric potential measuring device U1 of Example 1 has the myoelectric potential at six measurement positions (first measurement position, second measurement position,..., Sixth measurement position) preset on the forearm of the subject. It has six myoelectric sensors (six measuring units, first measuring unit, second measuring unit,..., Sixth measuring unit) SN1 to SN6 for measuring signals.
  • the elbow of the subject's right arm corresponding to the six types of muscles of the subject's lateral carpal flexor, deep finger flexor, ulnar carpal flexor, finger extensor, long lateral carpal flexor and brachial radius Six locations on the nearby skin are set as measurement positions.
  • the wet electrode is an electrode in which a conductive paste is previously incorporated on the electrode side, and is more stable without being affected by the skin condition than a so-called dry electrode in which the paste is not incorporated. Measurement is possible. That is, the surface myoelectric potential measuring device U1 of Example 1 uses the six pairs of wet electrodes (SN1 to SN6) to be applied to the six skins, so-called myoelectric signals corresponding to the six types of muscles, Surface myoelectric potential signals (six types of biological signals, first information, second information,..., Sixth information) are measured.
  • the identification client personal computer PC is configured by a so-called computer device, and includes a computer main body H1, a display H2, input devices such as a keyboard H3 and a mouse H4, and an HD drive (hard disk not shown). Drive) and the like.
  • a motion identification device motion intention estimation device, event estimation device, classification estimation device of the first embodiment by the surface myoelectric potential measurement device U1 having each of the myoelectric sensors SN1 to SN6 and the measurement client personal computer PC ( U1 + PC) is configured.
  • FIG. 3 is an explanatory diagram showing the function of each device constituting the operation identification system according to the first embodiment of the present invention in a block diagram (functional block diagram).
  • the surface myoelectric potential measuring device U1 and the control unit of the computer main body H1 of the identification client personal computer PC perform I / O (input / output) for input / output of signals to / from the outside, adjustment of input / output signal levels, and the like.
  • the ROM of the control unit C stores an action identification program (an action intention estimation program, an event estimation program, and a classification estimation program) AP1.
  • the control unit of the surface myoelectric potential measuring device U1 realizes the following functions to calculate the output signals to be input to the identification client personal computer PC according to the output signals of the six myoelectric sensors SN1 to SN6.
  • the integrated value calculating means C1 is an integrated value (first integrated value s 1 , second integrated value s 2) of the six types of surface myoelectric potential signals from the six myoelectric sensors SN1 to SN6.
  • a sixth integral value s 6 a so-called integrated surface myoelectric signal (integrated electromyogram, integral electromyogram, electromyogram integral value, IEMG: Integrated Electro Myo-Gram) (s 1 to s 6 ). Calculate.
  • the integrated value calculation means C1 performs full-wave rectification smoothing on the six types of surface myoelectric potential signals acquired at a sampling frequency of 3 [kHz], thereby providing six types of integrated surface myoelectric potential signals. (S 1 to s 6 ) is calculated. Specifically, for each of the six types of surface myoelectric potential signals, full wave rectification, that is, after converting an AC wave including both positive and negative waves into a DC wave in which the negative wave is a positive and negative symmetrical wave, smoothing, That is, six types of integrated surface myoelectric potential signals (s 1 to s 6 ) are calculated by performing low-pass filtering at a preset cutoff frequency. In Example 1, the cut-off frequency is set to 2.4 [Hz].
  • Integral Value Transmitting Means uses the six types of integral surface myoelectric potential signals (s 1 to s 6 ) calculated by the integral value calculating means C1 as the computer main body of the identification client personal computer PC. Send to H1.
  • the controller C of the surface myoelectric potential measuring device U1 identifies the movement of the subject based on the output signal of the surface myoelectric potential measuring device U1, that is, the integrated surface myoelectric potential signals (s 1 to s 6 ).
  • the operation identification program AP1 has the following function realizing means.
  • the integrated value receiving means C3 of Embodiment 1 uses the values of the three integrated surface myoelectric potential signals (s 1 to s 6 ) at 1 [ms] and the six types of integrated surface myoelectric potential signals (s 1 to 1 ). s 6 ) Receive for each.
  • C4 Integral Value Storage Unit Integral value storage unit C4 stores the six types of integrated surface myoelectric potential signals (s 1 to s 6 ) received by the integral value receiving unit C3.
  • the integrated value receiving means C3 uses the values of the integrated surface myoelectric signal (s 1 to s 6 ) received at three points every 1 [ms] as the six types of integrated surface myoelectric potential signals (s 1 to s 6 ) Store for each.
  • C5 Integral average value calculation means
  • the integral average value calculation means C5 is an integral average value (first value) that is an average value of the six types of integrated surface myoelectric potential signals (s 1 to s 6 ) within a preset frame time.
  • the sixth integral average value v 6 , and AIEM Average Integrated Electro Myo-Gram) are calculated for each preset shift time.
  • FIG. 4 is an explanatory diagram for the integral average value of the first embodiment of the present invention.
  • the frame time is set to 15 [ms]
  • the shift time is set to 10 [ms]
  • 3 points are received every 1 [ms].
  • the first integrated value s 1 at the Q point is received within one frame and the first integrated average value v 1 is calculated for each frame while shifting by P points, the first q-th value is calculated.
  • AIEMG 1 (p) ⁇ IEMG 1 (1 + P ⁇ (p ⁇ 1)) + IEMG 1 (2 + P ⁇ (p ⁇ 1)) + ... + IEMG 1 (Q + P ⁇ (p ⁇ 1)) ⁇ / Q Formula (2-1)
  • AIEMG i (p) ⁇ IEMG i (1 + P ⁇ (p ⁇ 1)) + IEMG i (2 + P ⁇ (p ⁇ 1)) + ... + IEMG i (Q + P ⁇ (p ⁇ 1)) ⁇ / Q Formula (2-2)
  • Minimum value storage means The minimum value storage means C6 obtains the minimum value of the six types of integral average values (v 1 to v 6 ) calculated by the integral average value calculation means C5 as the six types of integral average values. Store for each of (v 1 to v 6 ).
  • the minimum value of each of the six types from the first to the p-th frame is ⁇ i.
  • the error removing unit C7 removes errors from the six types of integrated average values (v 1 to v 6 ) calculated by the integrated average value calculating unit C5.
  • the error removal unit C7 regards the six types of minimum values ⁇ i stored in the minimum value storage unit C6 as errors for each subject, so-called noise (individual differences, individual differences). Then, the error is removed by subtracting the six types of integral average values v i from the value AIEMG i (p) for each frame. Specifically, when the error removal value of the six types of integrated average values v i is AIEMG ⁇ i (p), the error removal value AIEMG ⁇ i (p) is expressed by the following equation (4).
  • AIEMG ⁇ i (p) AIEMG i (p) ⁇ i (4)
  • the total value calculation unit C8 calculates the total value of the six types of integrated average values (v 1 to v 6 ) from which errors have been removed by the error removal unit C7.
  • the total value calculation means C8 according to the first embodiment calculates the total value of the error removal values AIEMMG ⁇ i (p) for each frame of the six types of integrated average values v i . Specifically, the total value for each frame of six integrated average value v i when the AIEMG all (p), the sum AIEMG all (p) is illustrated by the following equation (5) .
  • AIEMG all (p) AIEMG ⁇ 1 (p) + AIEMG ⁇ 2 (p) + ... + AIEMG ⁇ 6 (p) Formula (5)
  • C9 Total value storage means
  • the total value storage means C9 is the total value AIEMG all (p-1) one frame before the six types of integrated average values (v 1 to v 6 ) calculated by the total value calculation means C8. ) Is memorized.
  • the difference value calculation means C10 according to the first embodiment includes a total value (first total value) AIEMG all (p) calculated by the total value calculation means C8 and one frame before stored in the total value storage means C9.
  • AIEMG dif (p) AIEMG all (p) -AIEMG all (p-1)
  • the maximum value storage means C11 includes six types of maximum values from the first to p-th frame among the six types of integrated average values v i of error removal values AIEMG ⁇ i (p). The maximum value from the first to the p-th frame of the total value AIEMG all (p) and the maximum value from the first to the p-th frame of the difference value AIEMG dif (p) are stored. To do.
  • the maximum value storage means C11 sets the maximum value of the error removal value AIEMG ⁇ i (p) as AIEMG ⁇ imax , the maximum value of the total value AIEMG all (p) as AIEMG amax , and the difference value AIEMG dif
  • the maximum value of (p) is AIEMG dmax
  • eight kinds of maximum values AIEMG ⁇ imax , AIEMG amax , and AIEMG dmax represented by the following formulas (7-1) to (7-3) are stored.
  • AIEMG ⁇ imax max [AIEMG ⁇ i (p)] (7)
  • AIEMG amax max [AIEMG all (p)] Formula (7-2)
  • AIEMG dmax max [AIEMG dif (p)] Formula (7-3)
  • the input value generating unit C12 includes a normalizing unit C12A and a binarizing unit C12B, and an input value to be input to a selective desensitization neural network (SDNN) described later. (Input variable value) is generated.
  • SDNN selective desensitization neural network
  • normalization means normalizing means C12A includes the six integrated average value v i of the error neutralization value AIEMGipushiron i (p), and the total value AIEMG all (p), the difference value AIEMG dif (p) and Is divided by each maximum value AIEMG ⁇ imax , AIEMG amax , and AIEMG dmax to normalize, thereby generating eight types of input values for each shift time, that is, for each frame.
  • Binarization means The binarization means C12B binarizes the eight types of input values N i (p), N a (p), N d (p) calculated by the normalization means C12A. Convert to a representation.
  • the binarization means C12B according to the first embodiment uses the eight types of input values N i (p), N a (p), and N d (p) as 15 + 1's and 15 -1's. In other words, each value is converted into a value that can be represented by a vector expression (binary pattern) having 15 + 1's and 15 -1's as vector components.
  • FIG. 5 is an explanatory diagram of the selective desensitization neural network according to the first embodiment of the present invention.
  • C13 Input value storage means
  • the input value storage means (input value input means) C13 is the eight types of input values N i (p), N a (p), N d (generated by the input value generation means C12. p) is stored.
  • the input value storage means C13 according to the first embodiment converts the value converted into a value that can be represented by a vector expression having 15 + 1's and 15 -1's as vector components by the binarization means C12B. Eight types of input values N i (p), N a (p), and N d (p) are stored.
  • the eight types of input values N i (p), N a (p), and N d (p) that change with each shift time are used as the selectively desensitizing neural network N shown in FIG.
  • the selective desensitization neural network N according to the first embodiment includes an input layer (input element group) Na configured by 240 elements (input elements).
  • first input element group G 1 1 second input element group G 2 1 ,..., Differential input element group G as a seventh input element group are used.
  • 7 1 a total input element group (multiple element group) G 8 1 ) as an eighth input element group.
  • the input layer Na according to the first embodiment has eight types of input variables (first input variable x 1 , second input variable x 2 ,..., Seventh input variable x 7 , eighth input variable x 8 ).
  • a vector representation in which the respective input values N i (p), N a (p), and N d (p) are binarized is stored for each shift time.
  • the values (x 1 i to x 30 i ) of 30 elements of each input element group G i 1 are input every shift time.
  • FIG. 6 is an explanatory diagram of an operation identification start image according to the first embodiment of the present invention.
  • C14 Action identification start determining means
  • the action identification start determining means C14 has action identification start image display means C14A for displaying the action identification start image 1 shown in FIG. 6, and determines whether or not to start the action identification of the subject. Determine. Whether the action identification start determining means C14 of Example 1 starts the action identification of the subject by determining whether or not the identification start button 1a of the action identification start image 1 shown in FIG. 6 is selected. Determine whether or not.
  • the intermediate value calculating means C15 has a selective desensitizing means C15A and a double desensitizing means C15B, and is determined to start the action identification of the subject by the action identification start determining means C14.
  • An intermediate value (value of an intermediate variable) based on the types of input values N i (p), N a (p), and N d (p) is calculated for each shift time.
  • the selective desensitization neural network N includes an intermediate layer (intermediate element group) configured by 1260 elements (intermediate elements) coupled to 240 elements of the input layer Na. Nb.
  • an intermediate layer Nb of Example 1 42 types of intermediate element groups (first intermediate element groups G 1 , 2 2 , second intermediate element groups G 2 , 1 2 ,..., 42nd intermediate element groups G 7, 6 2 ).
  • each intermediate variable x ⁇ , ⁇ is expressed by the following equation (12).
  • the output sensitivity value g j ⁇ , ⁇ is set to 0, and the product ⁇ (g j ⁇ , ⁇ ) ⁇ (x i ⁇ ) ⁇ becomes 0, that is, g j ⁇ , ⁇
  • the product type modification G ⁇ 1 (G ⁇ 1 ) is further subjected to product type context modification by the total input element group G 8 1 corresponding to the normalized total value N a (p). .
  • the calculation is performed in advance based on the element values of the total input element group G 8 1 (values of the elements of the multiplexing element group) x 1 8 to x 30 8.
  • the total output sensitivity value (multiplex output sensitivity value) is g 1 ⁇ , ⁇ , 8 to g 30 ⁇ , ⁇ , 8
  • ⁇ (i 1, 2,..., 30).
  • the values x i ⁇ , ⁇ of the elements of the intermediate element group G ⁇ , ⁇ 2 are expressed by the following formula (13).
  • x i ⁇ , ⁇ (g k ⁇ , ⁇ , 8 ) ⁇ (g j ⁇ , ⁇ ) ⁇ (x i ⁇ ) (13)
  • the values g j ⁇ , ⁇ , g k ⁇ , ⁇ , 8 of the output sensitivities are represented by the following equations (14-1), (14-2), (14-1) ′, (14-2) ′.
  • each output sensitivity value g j ⁇ , ⁇ , g k ⁇ , ⁇ , 8 becomes 0 and is desensitized.
  • ⁇ (i) and ⁇ ′ (i) are so-called random number functions, and are permutations obtained by randomly replacing the permutations of (1, 2,..., 30), for example, (21, 15,. 4), i-th value (random number from 1 to 30) such as (17, 6,..., 8).
  • the random number functions ⁇ (i) and ⁇ ′ (i) are not used, there is a strong correlation between the group to be modified G ⁇ 1 and the group to be modified G ⁇ or the total input element group G 8 1.
  • the random number functions ⁇ (i) and ⁇ ′ (i) are used to artificially weaken the correlation between the element groups G ⁇ 1 , G v , and G 8 1 . .
  • Expression (13) ′ in which Expressions (14-1), (14-2), (14-1) ′, and (14-2) ′ are substituted into Expression (13) is shown below.
  • x i ⁇ , ⁇ ⁇ (1 + x ⁇ ′ (i) 8 ) / 2 ⁇ ⁇ ⁇ (1 + x ⁇ (i) v ) / 2 ⁇ ⁇ (x i ⁇ )
  • (13) ′ That is, the value x i mu of said intermediate element group G mu, [nu 2 of the device of Example 1, [nu is the value x i mu of the modified group G mu 1 of the device, the value of the output sensitivity g j mu , ⁇ , and then further selectively desensitized (double desensitization, multiple desensitization) by the total output sensitivity value g k ⁇ , ⁇ , 8 . That is, an appropriate number of the modified group G ⁇ 1 , specifically, a value x i ⁇ of 15 to
  • 42 types of output sensitivity values (first output sensitivity value g j 1,2 , second output sensitivity value g j 2,1 ,..., 42nd output sensitivity value g j 7,6 ) 42 types of gain vectors (first gain vector G 1,2 1 , second gain vector G 2,1 1 ,... 42nd gain vector G 7,8 1 ) are formed.
  • Total output sensitivity value (first total output sensitivity value (first multiplexing output sensitivity value) g k 1,2,8 , second total output sensitivity value (second multiplexing output sensitivity value) ) G k 2,1,8 ,...
  • the 42nd total output sensitivity value (the 42nd multiplexing output sensitivity value) g k 7,6,8 ) is used to obtain 42 types of total gain vectors (first total gain vector G 1, 2, 8 1, second summing gain vector G 2,1,8 1, ... 42 total Gain vector G 7,6,8 1) is constructed.
  • the 42 types of gain vectors G ⁇ , ⁇ 1 and the 42 types of total gain vectors G corresponding to the 42 types of intermediate element groups G ⁇ , ⁇ 2 are used.
  • the values of the 42 kinds of intermediate variables x ⁇ , ⁇ of the first embodiment is a vector representation of a 30-dimensional vector (ternary pattern having three values of ⁇ 1, 0, +1) as vector components. ), For example, (+1, 0, 0, -1, +1,..., 0).
  • the selective desensitizing means C15A is an output sensitivity calculating means C15A1 for calculating the 42 types of output sensitivity values g j ⁇ , ⁇ according to the equations (14-1) and (14-2). As shown by the solid line in FIG. 5, an appropriate number of element values x i ⁇ of the modified group G ⁇ 1 are selectively desensitized by the 42 types of output sensitivity values g j ⁇ and ⁇ . To do.
  • Double desensitization means Double desensitization means (multiple desensitization means)
  • C15B is a value of the total 42 types of output sensitivity g k ⁇ according to the formulas (14-1) ′ and (14-2) ′. , ⁇ , 8 for summing output sensitivity computing means C15B1, and as shown by the dotted lines in FIG. 5 and the equations (13) and (13) ′, the 42 types of summing output sensitivity values g k.
  • the element values x i ⁇ , ⁇ ( ⁇ ⁇ ⁇ , ⁇ ) of the 42 types of intermediate element groups G ⁇ , ⁇ 2 calculated by the double desensitizing means C15B.
  • Example 1 by storing every 30) shift time, the 30 of the intermediate element groups G mu, [nu 2 It can be considered that any one of the three values of -1, 0, and +1 is input (held) to each element at every shift time.
  • the selective desensitization neural network N includes an output layer (output element group) configured by 48 elements (output elements) coupled to 1260 elements of the intermediate layer Nb. Nc.
  • the output layer Nc of Example 1 has six types of output element groups (first output element group G 1 3 , second output element group G 2 3 ,..., Sixth output element group G 6 3 ).
  • the output layer Nc according to the first embodiment includes six output variables (first output variables y 1 , “1” corresponding to “wrist flexion”) that are set in advance according to the six types of motion except “basic posture”.
  • the six types of output element groups G ⁇ 3 are set in advance corresponding to the values of the second output variable y 2 corresponding to “wrist extension”,...,
  • the sixth output variable y 6 corresponding to “wrist pronation” Has been.
  • the value of each coupling load according to each value x j ⁇ , ⁇ is ⁇ ⁇ ij ⁇ , ⁇ , and a threshold value set in advance according to the element value y i ⁇ of the output element group G i 3 H i ⁇ , +1 is output when a value greater than 0 is input, and h is a binary function that outputs 0 when a value less than 0 is input, the output element group G i 3
  • the value y i ⁇ of the element is expressed by the following equation (16).
  • Example 1 48 elements of the output layer Nc are coupled to 1260 elements of the intermediate layer with a coupling load.
  • 48 values y i gamma of elements of the output layer Nc is 1260 pieces of element values x j mu, the value of [nu and its coupling load omega .gamma.ij mu, product of ⁇ ⁇ ( ⁇ ⁇ ij ⁇ , ⁇ ) ⁇ (x j ⁇ , ⁇ ) ⁇ , which is a value calculated based on the sum (binary values 0 and +1).
  • the values of the six types of output variables y ⁇ of the first embodiment is a vector representation (binary pattern) of an 8-dimensional vector having binary values of 0 and +1, for example ( 0, + 1, + 1,0, + 1,0,0,0), etc.
  • Example 1 the value omega .gamma.ij mu connection weights of the 60480 type learned by connection weight learning means C23 to be described later, stores ⁇ and the 48 types of threshold h i gamma.
  • Example 1 it can be considered that one of binary values 0 and +1 is input (held) to each of the eight elements of each output element group G ⁇ 3 for each shift time.
  • Correspondence relation information storage means includes six types of actions (a plurality of classification information and a plurality of events to be classified) excluding “basic posture”, and 6 measured in advance at each of the actions.
  • the correspondence relationship information T according to the first embodiment includes the element values (ideal values) py 1 ⁇ , py 2 ⁇ ,... Of the six types of output element groups G ⁇ 3 corresponding to the actual six types of operations measured in advance.
  • FIG. 7 is an explanatory diagram of the correspondence information according to the first embodiment of the present invention.
  • It is a graph which shows the total value of the error removal value, and the range which identified operation
  • Example 1 in order to store the correspondence relationship information T, first, as shown in FIG. 7, the subject performs the six types of motions in the order of “motion A” to “motion F”.
  • the error when actually executed at intervals of 2 seconds in the order of “wrist flexion”, “wrist extension”, “gripping the hand”, “opening the hand”, “outside the wrist”, “inside the wrist” The total value AIEMG all (p) of the six types of integrated average values (v 1 to v 6 ) removed is calculated.
  • a threshold value for motion identification based on the total value AIEMG all (p) is set.
  • a preset total threshold is set to th and a preset ratio of the total value AIEMG all (p) to the maximum value AIEMG amax is ⁇ (0 ⁇ ⁇ 1)
  • a value ( ⁇ ⁇ AIEMG amax ) obtained by multiplying the total threshold th and the maximum value AIEMG amax by the ratio ⁇ is set as a threshold for action identification.
  • 0.2 [mV] is set in advance as the total threshold th.
  • a previously measured operation is identified.
  • M m ′ (p) H (AIEMG all (p) -th) Xh (AIEMG all (p) ⁇ ⁇ AIEMG amax ) (17)
  • operation A is identified by the first discrimination function M 1 (p) in the first 2 seconds (0 [s] to 2 [s]), and the next 2 seconds (2 In [s] to 4 [s]), the process of identifying “motion B” with the second discrimination function M 2 (p) is repeated, and in the last 2 seconds (10 [s] to 12 [s]) “Operation F” is identified by the sixth identification function M 5 (p).
  • the number of frames (number of continuous frames) obtained by each discriminant function M m ′ (p), (m ′ 1, 2,... .
  • the six motions of the subject at intervals of 2 seconds are repeated a plurality of times, and the motion speed is normalized based on the maximum number of continuous frames.
  • N sp (p) W m '/ W m'max ... formula (18)
  • the normalized operating speed N sp (p) (0 ⁇ N sp (p) ⁇ 1)
  • Each of the values is converted into a value that can be represented by a vector expression (binary pattern) having four + 1's and four zeros as vector components.
  • N sp (p) (0,..., 0, +1,..., +1) Equation (19-1)
  • N sp (p) (0, 0, 0, +1,..., +1, 0)
  • N sp (p) (0, 0, +1,..., +1, 0)
  • N sp (p) (0, 0, +1,..., +1, 0, 0)
  • N sp (p) (0, + 1,..., + 1,0,0,0) (formula (19-4))
  • 0.8 ⁇ N sp (p) ⁇ 1: N sp (p) (+ 1,..., +1, 0,..., 0) Expression (19-5)
  • C20 Motion Identification Unit
  • the motion identification unit (motion intention estimation unit, event estimation unit, classification information estimation unit) C20 includes a motion type identification unit C20A and a motion speed identification unit C20B, and the correspondence relationship information storage unit C19.
  • the movement of the subject is identified based on the correspondence information T stored in advance and the values of the six types of output variables y ⁇ (estimation of movement, estimation of event, estimation of classification information) To do).
  • C20A Motion type identification means
  • the motion type identification means (type information estimation means) C20A has the above-mentioned seven types of motions of the subject ("basic posture”, “wrist flexion”, “wrist extension”, “hold the hand”). , “Open the hand”, “outside the wrist”, and “inside the wrist”) (type information that is the type of each action is estimated).
  • the first output element group G 1 3 is continuously identification frame number Ps or more (0, + 1, + 1, 0, + 1, 0, 0, 0) when the like, identified as "wrist flex" To do.
  • Motion speed identification means The motion speed identification means (speed information estimation means, force information estimation means) C20B identifies the speed of the performed motion when the six types of motions other than the “basic posture” are performed. (Estimate speed information of each motion).
  • the speed of the operation is identified by comparing the value obtained by digitizing the pattern with the value obtained by digitizing the five patterns shown in the equations (19-1) to (19-5).
  • the speed of the operation is identified, but force is generated according to the change in the speed.
  • the speed of the operation speed information, “fast”, “normal”, “slow” information
  • the action force force information, “strong”, “ordinary”, and “weak” information
  • the motion force even when the actual motion speed becomes zero, it is possible to identify the motion force.
  • the elements of the third output element group G 3 3 Since the average value d 3 as the firing pattern changes from a value in the range of 5.5 ⁇ d 3 ⁇ 6.5 to a value in the range of 2.5 ⁇ d 3 ⁇ 3.5, the power of the operation is “ It becomes possible to identify the change from “weak” to “strong”.
  • a surface myoelectric potential signal is measured by attaching the six myoelectric sensors SN1 to SN6 to the six measurement positions of the subject's forearm for a subject who has lost his hand, as in Example 1. Since the surface myoelectric potential signal changes according to the subject's intention to move, and the average value d ⁇ changes, it is possible to estimate the type and force of the action he wanted to perform with the hand lost from the subject's intention to move. It becomes possible.
  • speed information is about three types of “fast”, “normal”, and “slow”
  • power of the motion force information
  • FIG. 8 is an explanatory diagram of an operation identification image according to the first embodiment of the present invention.
  • C21 Motion identification image display means
  • the motion identification image display means (motion intention display means, event display means, classification information display means) C21 is the motion of the subject identified by the motion identification means C20 (the estimated subject's
  • a motion identification image (motion intention estimation image, event estimation image, classification information estimation image) 2 shown in FIG. 8 is displayed with the motion intention, the estimated event, and the estimated classification information.
  • the motion identification image 2 according to the first embodiment includes a display unit 2 a that displays a start time, a type, and a speed of the identified subject's motion, and an end of motion identification for ending the subject's motion identification. And a button 2b.
  • the operation identification image display means C21 of Example 1 displays the start time, type and speed of the operation identified on the display unit 2a.
  • the bond weight learning means C23 includes a threshold learning means C23A and an initial value setting means C23B, and includes seven types of movements (estimated movements of the subject) identified by the movement identification means C20. Update the combined load based on a difference between a pre-measured actual motion (actual motion intention, actual event, actual classification information) and intention, estimated event, estimated classification information) As a result, each of the connection weights is learned.
  • the joint load learning means C23 of the first embodiment determines the 60480 kinds of joint load values ⁇ ⁇ ij ⁇ and ⁇ as follows after the action identification end judging means C22 has determined that the subject's action identification is finished. Update by equation (20-1).
  • Threshold learning means The threshold learning means C23A determines the 48 types of thresholds h i ⁇ by the following formula (20-2) after the action identification end judging means C22 has determined that the action identification of the subject is to be finished.
  • the threshold values h i ⁇ are learned by updating according to the above.
  • the initial value setting means C23B sets the initial values of the 60480 types of coupling load values ⁇ ⁇ ij ⁇ and ⁇ and the 48 types of threshold values h i ⁇ .
  • the coupling load values ⁇ ⁇ ij ⁇ and ⁇ and the threshold value h i ⁇ are learned by the equations (20-1) and (20-2) so that the output value of the frame becomes the ideal value py i ⁇ . .
  • the values ⁇ 1ij ⁇ , ⁇ and the eight threshold values h i 1 can be learned as shown in the following equations (20-1-1) and (20-2-1).
  • the obtained 60480 types of coupling load values ⁇ ⁇ ij ⁇ and ⁇ and the 48 types of threshold values h i ⁇ are obtained as described above.
  • the initial value the coupling weight value omega .gamma.ij mu set as, the ⁇ and the threshold h i gamma, are stored in the connection weight memory means C17A.
  • FIG. 9 is a flowchart of the operation identification process of the operation identification program according to the first embodiment of the present invention.
  • the processing of each ST (step) in the flowchart of FIG. 9 is performed according to a program stored in the ROM or the like of the control unit C. In addition, this process is executed by multitasking in parallel with other various processes of the control unit C.
  • the flowchart shown in FIG. 9 is started when the operation identification program AP1 is started after the surface myoelectric potential measuring device U1 and the identification client personal computer PC are turned on.
  • the start time, type, and speed of the identified operation are displayed (described) on the display unit 2a of the operation identification image 2. Then, the process proceeds to ST13.
  • the subject's upper arm is identified as the “basic posture”, and the start time of the “basic posture” is displayed on the display unit 2a of the motion identification image 2 (described). Then, the process proceeds to ST13.
  • ST13 it is determined whether or not to end the process of identifying the subject's motion by determining whether or not the identification end button 2b of the motion identification image 2 has been selected. If yes (Y), the process proceeds to ST14, and, if no (N), the process returns to ST4.
  • FIG. 10 is a flowchart of input value generation processing of the operation identification program according to the first embodiment of the present invention.
  • the processing of each ST (step) in the flowchart of FIG. 10 is performed according to a program stored in the ROM or the like of the control unit C. Further, this process is executed in a multitasking manner in parallel with other various processes of the control unit C, similarly to the operation identification process.
  • the flowchart shown in FIG. 10 is started when the operation identification program AP1 is started after the surface myoelectric potential measuring device U1 and the identification client personal computer PC are turned on.
  • the control unit of the surface myoelectric potential measurement device U1 uses an integrated surface myoelectric potential signal (an integrated value of the six types of surface myoelectric signals from the six myoelectric sensors SN1 to SN6 shown in FIG. s 1 to s 6 ) are calculated and transmitted to the identification client personal computer PC. Then, the process proceeds to ST102. In ST102, the identification client personal computer PC receives and stores the integrated surface myoelectric potential signals (s 1 to s 6 ). Then, the process proceeds to ST103. In ST103, as shown in equations (2-1), (2-2), and FIG. 4, the integrated average value (v 1- ) that is the average value of the six types of integrated surface myoelectric potential signals (s 1 -s 6 ).
  • the minimum values ( ⁇ 1 to ⁇ 6 ) of the determined integrated average values (v 1 to v 6 ) are stored. Then, the process proceeds to ST106.
  • ST 108 it stores the maximum value AIEMGipushiron imax of the determined error removal value AIEMG ⁇ i (p), respectively. Then, the process proceeds to ST109.
  • the difference value AIEMG dif (p) for each frame of the total value AIEMG all (p) is calculated. Then, the process proceeds to ST113.
  • ST113 as shown in Equation (7-3), it is determined whether or not the calculated difference value AIEMG dif (p) is the maximum value AIEMG dmax up to the current frame (p-th frame). If yes (Y), the process proceeds to ST114, and, if no (N), the process proceeds to ST115.
  • the maximum value AIEMG dmax of the determined total value difference value AIEMG dif (p) is stored. Then, the process proceeds to ST115.
  • ST115 the following processes (1) and (2) are executed, and the process returns to ST101.
  • the operation identification system S executes the same operation at a different speed as compared with the techniques of Non-Patent Documents 1 and 2 that perform only the identification when the same operation is always executed at a constant speed. It is possible to identify when it is made.
  • the motion identification system S of the first embodiment identifies the motion of the subject by a selective desensitization neural network N (for example, see Japanese Unexamined Patent Application Publication No. 2009-64216) with high learning ability (generalization ability). is doing. For this reason, it can be expected that the motion of the subject is identified with a smaller learning amount than the techniques of Non-Patent Documents 1 and 2 that perform motion identification based on a statistical method.
  • the selective desensitization neural network N performs noise based on the six types of surface myoelectric potential signals from the six types of myoelectric sensors SN1 to SN6 by the input value generation processing according to the first example. Eight types of input values N i (p), N a (p), and N d (p) are removed.
  • the input value generation process has a small amount of calculation and can sufficiently generate the input values N i (p), N a (p), and N d (p) for each shift time (10 [ms]).
  • the operation identification process can be executed without delay.
  • connection load values ⁇ ⁇ ij ⁇ and ⁇ for calculating the output values approach the same value every time they are learned ( (See formulas (20-1) and (20-2)). Such a situation is referred to as “average connection weight by one-to-many correspondence”, which causes a significant reduction in learning efficiency and identification accuracy.
  • the total value N a (p) is not only the same input value (input values N 1 (p) to N 4 (p) of 1ch to 4ch) but also different input values.
  • the total value N a (p) that is highly likely to be different from the six types of input values N i (p) is used as the input value.
  • the total value N a (p) is used only for double desensitization of the other seven types of input values N i (p) and N a (p). Therefore, the operation identification system S of the first embodiment is expected to further reduce the learning amount and improve the identification accuracy as compared with the selective desensitization neural network in which the selective desensitization is executed only once (single desensitization). it can.
  • the double desensitization is performed as in the operation identification system S of the first embodiment, it is confirmed by experiments that the identification accuracy is improved by about 20% compared to the case where the double desensitization is not performed. Has been.
  • Example 1 was prepared.
  • the operation identification system S of Experimental Example 1 is produced with the same configuration as the operation identification system S of Example 1, and the input value generation process (see FIG. 10) and the operation identification process (see FIG. 10) of Example 1. 9) is executed.
  • the initial values of the 60480 types of bond load values ⁇ ⁇ ij ⁇ and ⁇ and the 48 types of threshold values h i ⁇ the subject's 6 movements for 2 seconds are “fast”, “normal”, and “slow”.
  • Experimental Example 1 “Personal-EMG” manufactured by Osaka Electronics Co., Ltd. (limited company) was used as the surface myoelectric potential measuring device U1.
  • the connection weight learning process (see ST14 in FIG. 9) was not executed, and the subject's movement was identified with the initial values.
  • the identification rate of the type of operation when any one of the six types of operations is actually executed, the total number of frames in which the actually executed operations are correctly detected is represented by the number of frames in which any of the six types of operations is detected. The value divided by the total number of.
  • the value when the identification value d ⁇ (p ⁇ ) is 4 or more away from the identification value corresponding to the actual 18 movements is set to 0 [%], and the value when it completely matches is 100 [ %] Is the average value of the entire number of identification frames Ps.
  • DR sp (m) [ ⁇ 4- (pd ⁇ (1) -d ⁇ (1)) ⁇ / 4 ⁇ 100 + ⁇ 4- (pd ⁇ (2) ⁇ d ⁇ (2)) ⁇ / 4 ⁇ 100 + ... + ⁇ 4- (pd ⁇ (Ps ) -d ⁇ (Ps)) ⁇ / 4 ⁇ 100 ] / (F m + f m ′′ ) (22)
  • FIG. 11 is an explanatory diagram of the experimental results of Experimental Example 1.
  • 11A is a graph of identification values when the subject actually executes at intervals of 2 seconds
  • FIG. 11A is a graph of correct values of identification values of actual six actions
  • FIG. 11B is identification of each frame of actual six actions.
  • FIG. 11C is a graph after the actual measurement values of the identification values for each frame of the actual six operations are smoothed.
  • FIG. 12 is an explanatory diagram of the experimental results of Experimental Example 1.
  • the discrimination rate of 6 types of motions when 6 sets of all 18 motions are executed with the vertical axis as the discrimination rate [%] and the horizontal axis as the motion type. %] And a three-speed discrimination rate [%] for each operation.
  • the identification rate DR (m) of the six types of operations in Experimental Example 1 when 6 sets of all 18 operations are executed is that “Operation C” is about 90%. (DR (3) ⁇ 90 [%]) and 5 operations (“Operation A”, “Operation B”, “Operation D” to “Operation F”) other than “Operation C” are almost 100 [%].
  • the average value (98.42 [%]) of the identification rate DR (m) is not a value indicating that identification is erroneously performed once or twice when six types of operations are performed 100 times. It is a value indicating that a correct operation is identified for most of the time from the start to the end of the operation identification process, but is instantaneously erroneously identified (see FIG. 11). For this reason, it can be seen that the identification of the type of operation is sufficiently practical.
  • the average value (76.15 [%]) of the three-speed identification rates DR sp (m) of the six types of actions is the average value (98.42 [%] of the discrimination rates DR (m) of the six types of actions. )), It is found that it is lower by 20 [%] or more. This is because the three-speed discrimination rate DR sp (m) of the six types of motions is the actual value (d ⁇ (1) to d ⁇ (15)) of the discrimination value and the ideal value (pd ⁇ (1) to pd).
  • the average value d ⁇ is also the ideal value (pd The value is away from ⁇ (1) to pd ⁇ (15)). For this reason, the smoothing process cannot completely match the actual measurement values (d ⁇ (1) to d ⁇ (15)) with the ideal values (pd ⁇ (1) to pd ⁇ (15)). For this reason, it is considered that the average value (76.15 [%]) of the identification rate DR sp (m) has decreased by 20 [%] or more.
  • FIG. 13 is an explanatory diagram of experimental results of Experimental Example 1.
  • the average values of the identification rates DR sp (m) for the three speeds of the six types of operation are 74.12 [%] for “fast”, 78.30 [%] for “normal”, and “slow”. Was 76.05 [%].
  • FIG. 14 is an explanatory diagram of the experimental results of Experimental Example 1. From the 6-ch myoelectric sensor when six types of operations are actually sequentially executed with the vertical axis representing voltage [mV] and the horizontal axis representing time [s]. It is an example of the graph of surface myoelectric potential.
  • the surface myoelectric potential from the 6ch myoelectric sensor when “operation F” is executed has no conspicuous value other than the maximum value of 3ch being 0.5 [mV]. It can be seen that the voltage value [mV] is relatively smaller than the surface myoelectric potential from the 6-ch myoelectric sensor when the other five operations are executed.
  • FIG. 15 is an explanatory diagram of the experimental results of Experimental Example 1, and is an explanatory diagram showing the estimated delay from the start of the operation by the transition of the identification rate with the passage of time. It is explanatory drawing of the table
  • FIG. 16 is a graph showing the transition of the three-speed identification rate of the six types of motion according to the elapsed time from the start of the operation based on the values in the table of FIG. 15 with the vertical axis representing the identification rate [%] and the horizontal axis representing the time [ms]. It is. Further, as shown in FIGS.
  • the three-speed identification rate DR sp (m) of the six types of operations in Experimental Example 1 when six sets are executed is the rise of each operation, that is, each operation. 10 [ms], 30 [ms], 50 [ms], 100 [ms], and 150 [ms] after the start of reception of the accompanying surface myoelectric potential signal. I can see it going. That is, it can be seen that the identification rate DR sp (m) increases as the shift time for 1, 3, 5, 10, 15 frames elapses.
  • the average value of the identification rate DR sp (m) at each timing of the three speeds of the six types of operations is 11.19 [%] after 10 [ms] and 30 [ms] after the start of each operation. It was 38.16 [%] after 26.31 [%], 50 [ms], 58.16 [%] after 100 [ms], and 63.40 [%] after 150 [ms]. Therefore, from 100 [ms] to 150 [ms] later, the average value of the identification rate DR sp (m) gradually changes, and the increase in the average value of the identification rate DR sp (m) starts to converge. (See FIG. 14).
  • the achievement level [%] with respect to the upper limit of 76.15 [%] is 76.37 [%] ((58.16) / (76.15) ⁇ 100 [ms] after the start of each operation). 100 ⁇ 76.37 [%]), and after 150 [ms] was 83.26 [%] ((63.40) / (76.15) ⁇ 100 ⁇ 83.26 [%]). For this reason, it can be seen that after 100 [ms] from the start of each operation, the three speeds of the six types of operations are almost identified.
  • DR (m) the identification rate
  • the achievement level [%] with respect to the upper limit of 98.42 [%] is 81.45 [%] ((80.16) / (98.42) ⁇ 100 [ms] after the start of each operation). 100 ⁇ 81.45 [%]) and 89.42 [%] after 150 [ms] ((88.09) / (98.42) ⁇ 100 ⁇ 89.42 [%]).
  • the identification can be performed with an identification rate of about 100 [ms] from the start of the operation, about 80% if only the type of operation is identified, and about 60% if the speed of the operation is also identified simultaneously. . That is, in Experimental Example 1, it can be identified with an identification rate of about 100 [ms] from the start of operation, about 80% for identifying 6 operations, and about 60% for identifying 18 operations with 6 operations and 3 speeds. . Therefore, it can be seen that the motion identification system S of Experimental Example 1 obtained a high identification rate with a small learning amount of 36 sample data (2 sets of 18 motions) of the subject.
  • the surface myoelectric potential signal is observed 30 [ms] to 100 [ms] earlier than the muscle contraction accompanying the actual movement of the subject. Therefore, in the motion identification system S of Experimental Example 1, it is almost simultaneous with the start of muscle contraction due to the actual motion, that is, in real time, about 80% if 6 motions are identified, and 18 motions of 6 motions and 3 speeds. Then, it can be seen that it can be identified with an identification rate of about 60%.
  • the identification rate increases as time elapses from the start of each operation. Therefore, when the identification result changes before and after a certain timing, the identification result after the change is changed from the identification result before the change. It is more likely that the answer is correct. For this reason, for example, the motion identification system S of Experimental Example 1 continues the motion identification process even after once identifying 18 motions before muscle contraction, and the identification result immediately after the start of muscle contraction is the identification before contraction. If the method of discriminating that the identification result before contraction is wrong and correcting the identification result immediately after the start of contraction is adopted when there is a change from the result, it is possible to identify 18 actions while maintaining real-time characteristics. It is also possible to further improve the rate. Furthermore, in Experimental Example 1, the connection weight learning process (see ST14 in FIG.
  • connection weight learning process is executed and the initial value is updated, the 18 actions of the subject can be identified in real time and at a higher identification rate. Is possible.
  • the motion identification system S according to the first embodiment operates based on the statistical method that requires a large amount of learning in advance for learning of the initial value of the coupling weight for increasing the identification rate, that is, the initial learning.
  • the learning amount number of samples to be provided
  • the motion identification system S according to the first embodiment reduces the amount of learning necessary in advance for initial learning (relearning) that is performed each time the subject changes, compared to the techniques of Non-Patent Documents 1 and 2 and the like. It is possible to reduce the burden on the subject in advance.
  • Example 1 the movement of the subject is identified by the selective desensitization neural network N having a high learning ability (generalization ability).
  • the eight types of input values N i (p), N a (p), and N d (p) in Example 1 are noises from the output signals (surface myoelectric potential signals) of the six myoelectric sensors SN1 to SN6. (Error, individual difference, individual difference) are removed for calculation (see ST104 to ST106 in FIG. 10, equation (4), etc.). Therefore, the operation identification system S according to the first embodiment is easy to set a coupling load suitable for everyone.
  • the operation identification system S of the first embodiment can identify with a high identification rate of 98.42 [%] or more if 6 operations are identified. It is possible to maintain a high identification rate that is 20% or more, and is practical in terms of identification time. Furthermore, if 18 actions of 6 actions and 3 speeds are discriminated, a high discrimination rate of 76.15 [%] or more can be maintained without additional learning, and a real-time property within 100 [ms] is required. Even so, it can be identified with an identification rate of 60% or more. Therefore, in the motion identification system S of the first embodiment, even if a large number of motions and speeds are identified, the necessary learning amount does not increase rapidly. Furthermore, the movement of the subject can be identified with a small learning amount as compared with the necessary techniques such as Non-Patent Documents 1 and 2.
  • the motion identification system S according to the first embodiment does not need to execute a complicated feature extraction process or the like for extracting features of the increased motion type or speed even if the motion type or speed to be identified increases. . Therefore, the motion identification system S according to the first embodiment is more complicated than the technologies such as Non-Patent Documents 1 and 2 in which the feature extraction process for identifying a large number of motions and speeds may be complicated and impair real-time performance. Even if the type and speed of the movement increase, the movement of the subject can be identified without impairing the real-time property. From these results, the motion identification system S of the first embodiment (1) performs a large number of motions (2) along with the estimation of speed (3) with a high identification rate, (4) with a small number of samples (5 ) Can be identified in real time.
  • values of m, n, m ′, and n ′ are not limited to this, and can be changed to arbitrary values.
  • the number of myoelectric sensors is 5 or less, or 7 or more
  • the number of elements of each layer Na to Nb is applied to an arbitrary number of selective desensitization neural networks N, or the operation of a subject of 5 or less or 7 or more Can also be identified.
  • the number of myoelectric sensors is increased to identify a larger number of movements
  • the selective desensitization neural network N is expanded and identified, the increase in the amount of calculation is moderate as the number of myoelectric sensors increases. Therefore, since the generalization ability is high and the redundancy is high (because the resistance to addition of redundant variables is strong), each operation can be identified with a small learning amount and a high identification rate.
  • the control unit of the surface myoelectric potential measuring device U1 calculates the integrated surface myoelectric potential signals (s 1 to s 6 ) of the surface myoelectric potential signals from the six myoelectric sensors SN1 to SN6. Although it is transmitted to the computer main body H1 of the identification client personal computer PC, the surface of the six myoelectric sensors SN1 to SN6 is controlled by the control unit of the computer main body H1 of the identification client personal computer PC. It is also possible to calculate an integrated surface myoelectric signal (s 1 to s 6 ) of the myoelectric signal. That is, it is possible to include the function of the surface myoelectric potential measuring device U1 in the identification client personal computer PC. For example, the surface myoelectric potential measuring device U1 and the identification client personal computer PC are integrally configured. It is also possible to produce an action identification device (U1 + PC).
  • the frame time for each frame is calculated.
  • 15 [ms] and the shift time are 10 [ms]
  • the present invention is not limited to this, and the frame time and the shift time can be arbitrarily set.
  • the method of identifying the operation is not limited to this.
  • the output element group G ⁇ If the 3 match to the 5 patterns in one frame, i.e., the four elements is continuously (adjacently) when it detects the output elements G gamma 3 was ignited, it can also operate identified is there.
  • the operation is identified only when at least two elements are ignited continuously, It is also possible to identify the operation only when all the elements that have been connected are continuous.
  • the present invention is not limited to this. For example, if it is unknown which of a plurality of corresponding actions has been executed, it is identified as “unidentifiable”, or the state of “basic posture” is maintained without identification. It is also possible to do.
  • the start time, type and speed of the identified subject's motion are described in the display section 2a (see FIG. 8) of the motion identification image 2, and the identification client personal computer PC
  • the identification operation / determination speed is reproduced in real time with a prosthetic hand / prosthesis (for example, prosthetic hand) Identify the next action performed by the wearer (wearer, worker, worker) who wears the robot suit, etc. and assists the identified action. It can also be used to operate a power assist motor or the like.
  • a process of determining a specific motor that applies torque to a specific joint for performing the identified operation that is, a process of specifying a motor to be driven for a plurality of motors of the robot suit
  • the wearer (work) is able to use or reproduce the identification operation with a robot suit in real time at a slightly higher speed than the identification speed (for example, about 10 [%] to 40 [%]). It is also possible to perform power assist for the identification operation. Further, for example, it is possible to execute control such as updating an image on the display H2 of the identification client personal computer PC as a virtual reality device in real time according to the type and speed of movement of the identified subject. It is.
  • the five patterns of the correspondence information T are set based on the six types of surface myoelectric signals from the types of myoelectric sensors SN1 to SN6 and 18 operations of 6 operations and 3 speeds. (See formulas (17), (18), (19-1) to (19-5), (20-1), (20-2), etc.), and the setting method of the correspondence information T is limited to this. However, for example, it is possible to set more accurately based on an analog value based on a speed sensor or acceleration sensor signal (a waveform in which a speed / acceleration value changes with time) as speed information of each operation. . In this case, it can be expected that the learning efficiency such as initial learning is improved and the identification rate can be further improved as compared with the case where five patterns of correspondence information T are set based on the surface myoelectric potential signal and 18 actions. I can expect.
  • the identification rate can be further improved as compared with the case where the operation is identified only by the input values N i (p), N a (p), and N d (p) based on the surface myoelectric potential signal. is there.
  • + d ⁇ (Ps) ⁇ / Ps) is identified as “fast” when 2.5 ⁇ d ⁇ ⁇ 3.5, and 3.5 ⁇ D ⁇ ⁇ 5.5 is identified as “normal” and 5.5 ⁇ d ⁇ ⁇ 6.5 is identified as “slow”, but is not limited to this, “fast” and “normal”
  • An arbitrary value can be set for the boundary value (3.5) that divides the identification range “” and the boundary value (5.5) that divides the identification range “normal” and “slow”.
  • the method for identifying the speed of operation is not limited to the three stages of “fast”, “normal”, and “slow”, and for example, it is possible to identify in two stages or four or more stages. .
  • the case of 2.5 ⁇ d ⁇ ⁇ 3.5 is “high speed”
  • the case of 3.5 ⁇ d ⁇ ⁇ 4.5 is “slightly high speed (medium speed)”
  • the average value d ⁇ itself can be identified as a numerical value indicating a speed, or the speed [m / s] calculated according to the average value d ⁇ can be identified.
  • six types of error removal values AIEMG ⁇ i (p ) Norm root of square sum
  • the norm AIEMG all (p) ′ instead of the total value AIEMG all (p) is expressed by the following equation (5) ′.
  • AIEMG dif (p) ′ AIEMG ⁇ 1 (p) -AIEMG ⁇ 2 (p) -... -AIEMG ⁇ 6 (p) Formula (5) ′′
  • the surface myoelectric potential signal as an example of biological information is measured to identify the movement of the subject.
  • the present invention is not limited to this, for example, the brain wave as an example of biological information is measured. It is also possible to identify the movement of the subject.
  • element group for multiple dead the (multiple element group) was a total input element group G 8 1, without being limited thereto, for example, each of the other input element group Any one or more of G 1 1 to G 7 1 and each of the intermediate element groups G 1, 2 2 to G 7,6 2 may be used.
  • a G 1,2 2 ⁇ G 7,6 2 becomes multiple element group and a plurality, such as the Example 1 Not double desensitization but multiple desensitization more than triple desensitization.
  • the element values of the input element group G [ nu] ( x i mu) and the value of the sum for the output sensitivity (g k ⁇ , ⁇ , product of 8) ⁇ (x i ⁇ ) ⁇ (g k ⁇ , ⁇ , 8) ⁇ was selective desensitization by calculating the After that, the product ⁇ (x i ⁇ ) ⁇ (g k ⁇ , ⁇ , 8 ) ⁇ (g j ⁇ , ⁇ ) ⁇ with the value (g j ⁇ , ⁇ ) of the output sensitivity is calculated to perform double desensitization.
  • the product of the sensitivity value (g j ⁇ , ⁇ ) and the total output sensitivity value (g k ⁇ , ⁇ , 8 ) ⁇ (x i ⁇ ) ⁇ (g j ⁇ , ⁇ ) ⁇ (g k ⁇ , ⁇ ) , 8 ) ⁇ can be simultaneously calculated to double desensitize.
  • the surface myoelectric potential signal is measured to identify the motion of the subject.
  • the present invention is not limited to this.
  • the motion type and speed of the limb lost by the disabled Can also be estimated.
  • the six myoelectric sensors SN1 to SN6 are attached to the six measurement positions of the subject's forearm and the surface myoelectric signal is applied. Since the surface myoelectric potential signal changes according to the movement intention of the subject, it is possible to estimate the type and speed of the movement that the subject wanted to perform with the lost hand. Even when there is no actual movement, it is possible to estimate the movement intention of the subject.
  • the subject's motion is identified from the measured surface myoelectric potential signal.
  • the classification information that can be estimated by the classification estimation system of the present invention is not limited to this. For example, Estimate human / animal movement from the captured video signal, estimate the intention of a stationary human / animal / plant from the measured EEG signal, and determine the state of a non-living object from the measured temperature / humidity Estimate, estimate weather and natural phenomena from measured temperature and humidity, estimate real estate prices from collected location conditions, etc., and collect information on economic indicators and monetary policies of each country It is also possible to estimate exchange rates and stock prices. That is, it is possible to estimate every event from the extracted information.
  • the integrated value of each piece of information (video signal, brain wave signal, temperature / humidity, location condition, economic index / monetary policy, etc.) as a waveform signal whose value changes with time.
  • the integral average value (v i ) of (s i ) is calculated for each shift time, and the total value of the integral average value (v i ) and the difference value for each frame of the total value are calculated (formula (5) ), (6)), and by calculating the input values (N i (p), N d (p), N a (p)) based on the integral average value / difference value / total value (expression (7) -1) to Expression (8-3)), as in the case of the surface myoelectric potential signal, an input value (N i (p), N d (p), it is possible to enter the N a (p)), all events (behavior, intention, state, weather and natural phenomena, real estate Case, it is possible to estimate the foreign exchange rates and stock prices, etc.).
  • the motion identification system S and the motion identification program AP1 of the present invention for example, to reproduce the identification motion / identification speed in real time with a prosthetic hand, a prosthetic limb, etc.
  • a prosthetic hand for example, to reproduce the identification motion / identification speed in real time with a prosthetic hand, a prosthetic limb, etc.
  • the power assist that identifies the next action performed by the worker wearing the robot suit and assists the identified action It is useful when it is used for operating a motor or the like, or when it is also used for control of a virtual reality device such as a sensation type game machine or other human interface devices.

Abstract

 入力素子群(Na)と、第1入力素子群の各入力素子の値(x ~x30 )と各第1出力感度の値(g 1,2~g30 1,2)と各第1多重用出力感度の値(g 1,2,8~g30 1,2,8)との積算を含む第1の多重不感化がなされて演算された第1中間変数の値(x 1,2~x30 1,2)が入力される中間素子群(Nb)と、第1中間素子群の各中間素子の値(x 1,2~x30 1,2)と各結合荷重の値(ωγij 1,2)とに基づいて演算された出力変数の値(y γ)が入力される出力素子群(Nc)と、予め記憶された対応関係情報(T)と出力変数の値(y γ)とに基づいて分類情報を推定する分類情報推定手段(C20)と、分類情報表示手段(C21)とを備えた分類推定システム(S)により、少ない学習量で被検体の動作意思を含む分類情報を推定する。

Description

分類推定システムおよび分類推定プログラム
 本発明は、選択的不感化ニューラルネットを介して被検体の動作意思を含む分類情報を推定する分類推定システムおよび分類推定プログラムに関する。
 近年、人間等の生体の生体情報(生体信号)を用いた機械制御等についての研究・開発等が盛んに行われている。前記機械制御については、例えば、義手・義肢等の介護・福祉機器(自助具、自助器)やロボットスーツ等のパワーアシスト装置といった人間の動作を補助する機器に限定されず、バーチャルリアリティ装置やその他のヒューマンインタフェース装置等のより一般的な制御機器等にも応用されることが期待されている(例えば、非特許文献1等参照)。また、前記機械制御に利用される前記生体情報としては、例えば、脳波や脳血流の変化の情報等の中枢神経系の情報と、筋肉が収縮する際に発生する前記筋肉と前記筋肉の周囲の皮膚との間の電位差である筋電位(筋電、筋活動電位)の変化の情報である筋電位信号(筋電信号、筋活動電位信号、筋電図)等の末梢神経系の情報とが挙げられる(例えば、非特許文献1等参照)。
 このうち、前記筋電位信号については、前記筋肉の収縮よりも、30~100[ms]早く観測されることが知られている(例えば、非特許文献2等参照)。すなわち、前記筋電位信号について、前記筋肉の収縮に対して時間的な遅れが少ない測定をすることが可能である。また、前記筋電位信号の測定装置は、応答性が高く(高感度・高速応答性)、測定値の信頼性が高く、比較的安価で調達できる。このため、前記筋電位信号は、生体情報として機械制御に利用することが現実的であり、特に有望視されている。
 ここで、前記筋電位信号の測定方法としては、被験者(被検体、被測定者)の特定の筋繊維に針電極を刺して直接的に特定の筋電位信号を測定する方法と、被験者の肌に電極(表面電極)を貼り付けて、複数の筋繊維が発する筋電位信号の総和、いわゆる、表面筋電位信号(EMG:Electro Myo-Gram)を測定する方法とが一般的に知られている。なお、前記針電極により測定する方法については、被験者の負担が大きいため、被験者の負担が比較的小さい表面電極により測定する方法がより一般的である。
 しかしながら、前記表面筋電位信号は、各筋電位信号の総和であって非線形性が強くなるため、被験者の動作(筋活動)との対応関係が複雑になっている。すなわち、前記表面筋電位信号に基づいて、前記動作の識別をすることが困難であることが知られている。
 このため、前記表面筋電位信号の測定結果から被験者の動作識別をする場合には、入力と出力との対応関係を表す関数(非線形関数)を分散表現により単純化するニューラルネット(ニューラルネットワーク)の技術が利用されている。
 ここで、前記表面筋電位信号に基づく動作識別に関する技術として、例えば、下記の非特許文献1,2に記載の技術が知られている。
 非特許文献1には、誤差逆伝播法(BP:Back Propagation、BP法)を用いた多層パーセプトロン(MLP:Multi-Layer Perceptron,Multi-Layered Perceptron)により、3個(3ch、3チャネル、3チャンネル)の電極(表面電極および接地電極)を被験者の前腕に貼り付けて測定した表面筋電位信号を入力して、出力としての10種類の前腕動作が、90[%]の平均識別率で識別できたことが記載されている。また、非特許文献1には、R-LLGMN(Recurrent Log-Linearized Gaussian Mixture Network)により、6chの電極を被験者の前腕に貼り付けて測定した表面筋電位信号を入力して、出力としての6種類の前腕動作が、85[%]の平均識別率で識別できたことが記載されている。
 ここで、前記多層パーセプトロンとは、入力層、中間層、出力層がそれぞれ有する素子(ニューロン)によって、入力・出力および関数を分散表現する、いわゆる、層状ニューラルネットのことである。前記多層パーセプトロンでは、入力層と中間層との各素子と、中間層と出力層との各素子とが結合荷重をもってそれぞれ結合され、中間層および出力層の各素子が、結合荷重を通じた入力層および中間層の各素子からの入力の総和として演算されることにより、前記関数が分散表現されている。また、誤差逆伝播法とは、前記多層パーセプトロンの学習に使われる学習アルゴリズムであり、ある学習データ、すなわち、一対の入力と出力とが与えられた際に、多層パーセプトロンによる出力が、学習データの出力と一致するように各層の間の結合荷重を修正するという学習法のことである。なお、誤差逆伝播法を用いた多層パーセプトロンについては、例えば、非特許文献3等に記載されているため、詳細な説明を省略する。
 また、LLGMNとは、統計モデルの一例としてのガウス混合モデル(GMM:Gaussian Mixture Model)と、確率モデルの一例としての隠れマルコフモデル(HMM:Hidden Markov Model)とを内包する層状ニューラルネットのことである。なお、ガウス混合モデルとは、ガウス分布(正規分布)の線形結合で表現される統計的モデルのことである。また、マルコフモデルとは、一般にある記号の出現確率が直前のn個の記号によって決定される確率過程、いわゆる、マルコフ過程(マルコフ性をもつ確率過程)に従って記号が出現すると仮定する確率モデルのことをいう。また、隠れマルコフモデルとは、「システムがパラメータ未知のマルコフ過程である」と仮定し、観測可能な情報からその未知のパラメータを推定するマルコフモデルのことをいう。
 また、R-LLGMNとは、第1の中間層と第2の中間層との間に1時刻前の出力を考慮可能な結合、いわゆる、リカレント結合を有するLLGMNのことである。なお、R-LLGMNについては、例えば、特開2005-11037号公報や非特許文献4等に記載されているため、詳細な説明を省略する。
 また、非特許文献2には、サポートベクターマシン(SVM:Support Vector Machine)により、4chの電極を被験者の前腕に貼り付けて測定した表面筋電位信号に基づいて、7種類の前腕動作(基本姿勢、手首屈曲、手首伸展、握る、開く、手首回内、手首回外)を識別する技術、すなわち、表面筋電位信号を入力することにより、識別結果を出力する技術が記載されている。なお、非特許文献2では、7種類の前腕動作が、93.30[%]の平均識別率で識別できたことが記載されている。
 ここで、前記サポートベクターマシンとは、入力層の素子としての複数の線形入力素子(線形しきい素子)を利用した2クラスのパターン識別器である。すなわち、結合荷重(シナプス荷重)を乗算した複数の線形入力素子の値の総和が予め設定された閾値を超えるか否かを示す出力値を出力する層状ニューラルネットである。
 具体的には、入力層のn個の素子の値の集合(入力値)としての入力特徴ベクトルをx(x=(x,x,…,x))とし、出力層の1個の素子の値(出力値)をyとし、入力層の各素子と出力層の素子との結合荷重のベクトル(結合荷重ベクトル)をω(ω=(ω,ω,…,ω))とし、結合荷重ベクトルωの転置ベクトルをωとし、閾値をhとした場合に、出力値yが以下の式(1)により示される。
y=sign(ωx-h)    …式(1)
 =sign(xω+xω+…+xω-h)
 なお、sign(ωx-h)は、(ωx-h)>0の場合に+1になり、(ωx-h)≦0の場合に-1になる符号関数である。すなわち、各ベクトルx,ωの内積が閾値hを超える場合に+1になり、超えない場合に-1になる2値関数である。
 また、前記サポートベクターマシンは、学習データ(訓練サンプル)、すなわち、一対の入力特徴ベクトルxおよび出力値yから、各データ点との距離、すなわち、入力特徴ベクトルxとの距離(マージン)が最大となる分離平面(超平面)を求めるマージン最大化という基準により、前記線形入力素子のパラメータを学習、すなわち、結合荷重ベクトルωを更新している。
横井浩史、他1名、「サイボーグの現状と今後」、計測自動制御学会、計測と制御、2008年4月、第47巻、第4号、p.351-358 吉川雅博、他3名、「筋電位信号を利用した手のリアルタイム動作識別と関節角度推定」、電子情報通信学会、電子情報通信学会論文誌D、2009年1月、第J92-D巻、第1号、p.93-103 「誤差逆伝播法」、「online」、九州工業大学、「2009年5月15日検索」、インターネット<URL:http://www.brain.kyutech.ac.jp/~furukawa/note/bp/bp.html> 福田修、他2名、「EMG信号を利用した全方位型ポインティングデバイス」、電子情報通信学会、電子情報通信学会論文誌D-II、2004年10月、第J87-D-II巻、第10号、p.1996-2003
(従来技術の問題点)
 前記非特許文献1、2等の技術では、前腕の基本姿勢、手首屈曲、手首伸展、握る、開く、手首回内、手首回外等の動作の種類についての識別を行っているが、同一動作を常に一定速度で動作させた場合の識別しか行っていないという問題があった。このため、前記非特許文献1、2等の技術では、同一の動作を異なる速度で識別できないという問題があった。
 また、前記表面筋電位信号は、個人差が大きく、動作識別を精度良く実行するためには、被験者ごとに学習データを学習し、被験者ごとに適合する結合荷重を設定したり、不特定多数の被験者の多数の学習データを学習し、万人に適合する結合荷重を設定したりする必要があった。すなわち、前記非特許文献1、2等の技術は、統計的手法に基づいて動作識別を実行するため、多数の学習データを必要とし、被験者への負担が大きく、実用的ではないという問題があった。
 さらに、前記非特許文献1、2等の技術では、さらに多くの動作を識別しようとしたり、動作の速度まで識別しようとしたりすると、増加した動作の種類や速度の特徴を抽出するための複雑な特徴抽出処理を実行する必要がある。このため、前記非特許文献1、2等の技術では、複雑な動作識別になるほど、リアルタイムで実行することが困難になるという問題があった。
 よって、前記非特許文献1、2等の技術では、少ない学習データ(学習量、サンプル数)で、多数の動作を、速度と共に、高い識別率で、リアルタイムに識別することができないという問題があった。
 本発明は、前述の事情に鑑み、少ない学習量で被検体の動作を識別することを第1の技術的課題とする。また、本発明は、少ない学習量で被検体の動作意思を含む分類情報を推定することを第2の技術的課題とする。
 前記技術的課題を解決するために、請求項1記載の発明の分類推定システムは、
 第1の情報に基づく第1入力変数の値が入力される複数の入力素子により構成された第1入力素子群と、第2の情報に基づく第2入力変数の値が入力される複数の入力素子により構成された第2入力素子群と、を有する入力素子群と、
 前記第1入力素子群の各入力素子の値と、前記第2入力素子群の各入力素子の値に基づいて一部の値がゼロになるように演算された各第1出力感度の値と、予め設定された多重用素子群の各素子の値に基づいて一部の値がゼロになるように演算された各第1多重用出力感度の値と、の積算を少なくとも含む第1の多重不感化がなされて一部の値がゼロにされた前記第1入力素子群の各入力素子の値として演算された第1中間変数の値が入力される複数の中間素子により構成された第1中間素子群、を有する中間素子群と、
 前記第1中間素子群の各中間素子の値と前記各中間素子の値の重視度合いに応じて予め設定された各結合荷重の値とに基づいて演算された出力変数の値が入力される複数の出力素子により構成された出力素子群と、
 分類対象としての複数の分類情報と、分類情報ごとに対応する前記出力変数の値と、の対応関係を特定する対応関係情報を記憶する対応関係情報記憶手段と、
 前記対応関係情報記憶手段に予め記憶された前記対応関係情報と、演算された前記出力変数の値と、に基づいて、入力された前記第1の情報および前記第2の情報に応じた前記分類情報を推定する分類情報推定手段と、
 前記分類情報推定手段によって推定された前記分類情報を表示する分類情報表示手段と、
 を備えたことを特徴とする。
 請求項2に記載の発明は、請求項1に記載の分類推定システムにおいて、
 前記第2入力素子群の各入力素子の値と、前記第1入力素子群の各入力素子の値に基づいて一部の値がゼロになるように演算された各第2出力感度の値と、前記多重用素子群の各素子の値に基づいて一部の値がゼロになるように演算された各第2多重用出力感度の値と、の積算を少なくとも含む第2の多重不感化がなされて一部の値がゼロにされた前記第2入力素子群の各入力素子の値として演算された第2中間変数の値が入力される複数の中間素子により構成された第2中間素子群、を有する前記中間素子群と、
 前記各中間素子群の各中間素子の値と前記各中間素子の値の重視度合いに応じて予め設定された各結合荷重の値とに基づいて演算された前記出力変数の値が入力される前記出力素子群と、
 を備えたことを特徴とする。
 請求項3に記載の発明は、請求項1または2に記載の分類推定システムにおいて、
 前記分類情報推定手段により推定された前記分類情報と、入力された前記第1の情報および前記第2の情報が抽出された実際の分類情報との相違に基づいて、前記各結合荷重の値を更新することにより、前記各結合荷重の値を学習する結合荷重学習手段と、
 を備えたことを特徴とする。
 請求項4に記載の発明は、請求項1ないし3のいずれかに記載の分類推定システムにおいて、
 前記多重用素子群の各素子の値が、前記第1の情報と、前記第2の情報と、に基づいて演算された
 ことを特徴とする。
 請求項5に記載の発明は、請求項1ないし4のいずれかに記載の分類推定システムにおいて、
 前記分類情報を被検体の動作意思とし、前記第1の情報および前記第2の情報を前記動作意思に伴う前記被検体の生体信号とした場合に、前記被検体に配置された第1測定位置における前記生体信号を測定する第1測定部と、前記被検体に配置された前記第1測定位置と異なる第2測定位置における前記生体信号を測定する第2測定部と、を有する測定部材と、
 を備えたことを特徴とする。
 請求項6に記載の発明は、請求項5に記載の分類推定システムにおいて、
 前記動作意思に伴う前記被検体の筋群の活動である筋活動に基づく電位の変化の情報である筋電位信号のうちの前記被検体の皮膚における前記筋電位信号を表面筋電位信号とし、前記生体信号を前記表面筋電位信号とした場合に、前記皮膚に配置された前記第1測定位置における前記表面筋電位信号を測定する前記第1測定部と、前記皮膚に配置された前記第2測定位置における前記表面筋電位信号を測定する前記第2測定部と、を有する前記測定部材と、
 前記第1測定部からの前記表面筋電位信号の積分値である第1積分値を演算すると共に、前記第2測定部からの前記表面筋電位信号の積分値である第2積分値を演算する積分値演算手段と、
 予め設定されたフレーム時間内の前記第1積分値の平均値である第1積分平均値を予め設定されたシフト時間ごとに演算すると共に、前記フレーム時間内の前記第2積分値の平均値である第2積分平均値を前記シフト時間ごとに演算する積分平均値演算手段と、
 前記第1の情報を前記第1積分平均値とし、前記第2の情報を前記第2積分平均値とした場合に、前記第1積分平均値に基づく前記第1入力変数の値が前記シフト時間ごとに入力される前記第1入力素子群と、前記第2積分平均値に基づく前記第2入力変数の値が前記シフト時間ごとに入力される前記第2入力素子群と、を有する前記入力素子群と、
 を備えたことを特徴とする。
 請求項7に記載の発明は、請求項5または6に記載の分類推定システムにおいて、
 前記動作意思によって特定される動作の種類の情報である種類情報と、前記動作の力の情報である力情報と、を有する前記分類情報を推定する前記分類情報推定手段、
 を備えたことを特徴とする。
 請求項8に記載の発明は、請求項7に記載の分類推定システムにおいて、
 前記動作の速度の情報である速度情報であって、前記力情報に基づく前記速度情報、を有する前記分類情報を推定する前記分類情報推定手段、
 を備えたことを特徴とする。
 請求項9に記載の発明は、請求項8に記載の分類推定システムにおいて、
 前記出力変数の値のうち、第1の動作意思に応じた第1出力変数の値が入力される複数の出力素子により構成された第1出力素子群と、第2の動作意思に応じた第2出力変数の値が入力される複数の出力素子により構成された第2出力素子群と、を有する前記出力素子群と、
 前記第1の動作意思の前記種類情報および前記速度情報と前記第1出力変数の値との対応関係と、前記第2の動作意思の前記種類情報および前記速度情報と前記第2出力変数の値との対応関係と、を特定する前記対応関係情報を記憶する前記対応関係情報記憶手段と、
 前記対応関係情報記憶手段に記憶された前記対応関係情報と、演算された前記第1出力変数の値と、演算された前記第2出力変数の値と、に基づいて、入力された前記第1の情報および前記第2の情報に応じた前記分類情報の前記種類情報および前記速度情報を推定する前記分類情報推定手段と、
 前記分類情報推定手段によって推定された前記分類情報の前記種類情報および前記速度情報を表示する前記分類情報表示手段と、
 を備えたことを特徴とする。
 前記技術的課題を解決するために、請求項10記載の発明の分類推定プログラムは、
 コンピュータを、
 第1の情報に基づく第1入力変数の値が入力される複数の入力素子により構成された第1入力素子群と、第2の情報に基づく第2入力変数の値が入力される複数の入力素子により構成された第2入力素子群と、を有する入力素子群について、入力された前記第1入力変数の値および前記第2入力変数の値を記憶する入力値記憶手段、
 前記第1入力素子群の各入力素子の値と、前記第2入力素子群の各入力素子の値に基づいて一部の値がゼロになるように演算された各第1出力感度の値と、予め設定された多重用素子群の各素子の値に基づいて一部の値がゼロになるように演算された各第1多重用出力感度の値と、の積算を少なくとも含む第1の多重不感化がなされて一部の値がゼロにされた前記第1入力素子群の各入力素子の値として演算された第1中間変数の値が入力される複数の中間素子により構成された第1中間素子群、を有する中間素子群について、前記第1中間変数の値を演算する中間値演算手段、
 前記第1中間素子群の各中間素子の値と前記各中間素子の値の重視度合いに応じて予め設定された各結合荷重の値とに基づいて演算された出力変数の値が入力される複数の出力素子により構成された出力素子群について、前記出力変数の値を演算する出力値演算手段、
 分類対象としての複数の分類情報と、分類情報ごとに対応する前記出力変数の値と、の対応関係を特定する対応関係情報を記憶する対応関係情報記憶手段、
 前記対応関係情報記憶手段に予め記憶された前記対応関係情報と、演算された前記出力変数の値と、に基づいて、入力された前記第1の情報および前記第2の情報に応じた前記分類情報を推定する分類情報推定手段、
 前記分類情報推定手段によって推定された前記分類情報を表示する分類情報表示手段、
 として機能させることを特徴とする。
 請求項1、10に記載の発明によれば、第1の情報および第2の情報に基づく各入力変数の値を多重不感化が可能な選択的不感化ニューラルネットに入力することによって分類情報を推定でき、学習能力(汎化能力)が高い前記選択的不感化ニューラルネットを介して分類情報を推定するため、多重不感化が可能な選択的不感化ニューラルネットを介して分類情報を推定しない場合に比べ、少ない学習量で分類情報を推定することができる。
 請求項2に記載の発明によれば、多重不感化が可能な選択的不感化ニューラルネットの相互修飾モデルを介して被検体の動作を識別するため、2個の入力変数の値を相互修飾(互いに多重不感化)して2個の中間変数の値を演算でき、2個の入力変数の値を積型修飾(一方のみを多重不感化)して1個の中間変数の値を演算する場合に比べ、多重不感化が可能な選択的不感化ニューラルネットの学習能力を高くでき、より高い識別率で分類情報を推定することができる。
 請求項3に記載の発明によれば、学習した各結合荷重の値に基づいて、出力変数の値を演算でき、より高い識別率で分類情報を推定することができる。
 請求項4に記載の発明によれば、第1の情報および第2の情報に基づいて演算された多重用素子群の各素子の値が、第1の情報または第2の情報に基づいて演算された多重用素子群の各素子の値に比べて、第1入力素子群の各入力素子の値または第2入力素子群の各入力素子の値と異なる値になる可能性が高く、不感化される入力素子の値が偏り難くなり、多重不感化が効率良く実行されるため、多重不感化が可能な選択的不感化ニューラルネットの学習能力を高くでき、より高い識別率で分類情報を推定することができる。
 請求項5に記載の発明によれば、生体信号に基づく各入力変数の値を多重不感化が可能な選択的不感化ニューラルネットに入力することによって被検体の動作意思を推定できる。また、請求項5に記載の発明によれば、被検体が前記動作意思に応じた動作を実行可能な場合には、被検体の動作を認識することができる。
 請求項6に記載の発明によれば、シフト時間ごとに演算される表面筋電位信号の積分平均値に基づく入力変数の値を選択的不感化ニューラルネットに入力することによって被検体の動作意思を推定することができ、生体信号として表面筋電位信号やその積分値に基づく入力変数の値を直接入力する場合に比べ、表面筋電位信号のノイズの影響を低減できる。また、請求項6に記載の発明によれば、被検体の動作に伴う表面筋電位信号の変化が、被検体の動作に伴う筋肉の収縮より30[ms]~100[ms]程度早く観測されるため、シフト時間を、30[ms]より短い時間に設定することにより、リアルタイムで被検体の動作を識別することができる。
 請求項7に記載の発明によれば、被検体の動作意思を推定した際に、動作の種類と力とを推定できる。
 請求項8に記載の発明によれば、被検体の動作意思を推定した際に、動作の速度を推定できる。
 請求項9に記載の発明によれば、第1の動作意思に対応する第1出力変数の値と、第2の動作意思に対応する第2出力変数の値と、を比較して、動作の種類と速度とを推定できる。
図1は本発明の実施例1の動作識別システムの全体説明図である。 図2は本発明の実施例1の動作識別システムが識別する被験者の6種類の前腕動作の説明図である。 図3は本発明の実施例1の動作識別システムを構成する各装置の機能をブロック図(機能ブロック図)で示した説明図である。 図4は本発明の実施例1の積分平均値についての説明図である。 図5は本発明の実施例1の選択的不感化ニューラルネットの説明図である。 図6は本発明の実施例1の動作識別開始画像の説明図である。 図7は本発明の実施例1の対応関係情報の説明図であり、縦軸を電圧[mV]とし横軸を時間[s]とし6種類の動作を実際に順次実行した際の積分平均値の誤差除去値の合計値と動作識別した範囲とを示すグラフである。 図8は本発明の実施例1の動作識別画像の説明図である。 図9は本発明の実施例1の動作識別プログラムの動作識別処理のフローチャートである。 図10は本発明の実施例1の動作識別プログラムの入力値生成処理のフローチャートである。 図11は実験例1の実験結果の説明図であり、縦軸を識別値とし横軸を時間[s]として全18動作の1セットのうちの「動作A」から「動作F」の6動作を被験者が2秒間隔で実際に実行した際の識別値のグラフであり、図11Aは実際の6動作の識別値の正解値のグラフであり、図11Bは実際の6動作のフレームごとの識別値の実測値の平滑化前のグラフであり、図11Cは実際の6動作のフレームごとの識別値の実測値の平滑化後のグラフである。 図12は実験例1の実験結果の説明図であり、縦軸を識別率[%]とし横軸を動作の種類として全18動作の6セットを実行した際の6種類の動作の識別率[%]および各動作の3速度の識別率[%]の折れ線グラフである。 図13は実験例1の実験結果の説明図であり、縦軸を動作の速度の識別率[%]とし横軸を動作の種類として全18動作の6セットを実行した際の6種類の動作の3速度の速度別の識別率[%]の折れ線グラフである。 図14は実験例1の実験結果の説明図であり、縦軸を電圧[mV]とし横軸を時間[s]として6種類の動作を実際に順次実行した際の6chの筋電センサからの表面筋電位のグラフの一例である。 図15は実験例1の実験結果の説明図であり、動作の立ち上がりからの推定遅れを時間の経過に伴う識別率の推移によって示す説明図であり、動作開始からの経過時間による6種類の動作の3速度の識別率の推移をセット別で示した表の説明図である。 図16は縦軸を識別率[%]とし横軸を時間[ms]として図15の表の値に基づく動作開始からの経過時間による6種類の動作の3速度の識別率の推移を示すグラフである。
ωγij μ,ν…結合荷重の値、AP1…分類推定プログラム、C1…積分値演算手段、C5…積分平均値演算手段、C13…入力値記憶手段、C15…中間値演算手段、C17…出力値演算手段、C19…対応関係情報記憶手段、C20…分類情報推定手段、C21…分類情報表示手段、C23…結合荷重学習手段、G ~G ,G ,Na…入力素子群、G …第1入力素子群、G …第2入力素子群、G …多重用素子群、G1,2 ~G7,6 ,Gμ,ν ,Nb…中間素子群、G1,2 …第1中間素子群、G2,1 …第2中間素子群、G ~G ,Gγ ,Nc…出力素子群、G …第1出力素子群、G …第2出力素子群、g μ,ν~g30 μ,ν,g μ,ν,g μ,ν,8~g30 μ,ν,8,g μ,ν,8…出力感度の値、g 1,2~g30 1,2,g 1,2…第1出力感度の値、g 2,1~g30 2,1,g 2,1…第2出力感度の値、g 1,2,8~g30 1,2,8,g 1,2,8…第1多重用出力感度の値、g 2,1,8~g30 2,1,8,g 2,1,8…第2多重用出力感度の値、N(p),N(p),N(p),(x ,x ,…,x30 )…入力変数の値、N(p),(x ,x ,…,x30 )…第1入力変数の値、N(p),(x ,x ,…,x30 )…第2入力変数の値、PC…コンピュータ、S…分類推定システム、SN1~SN6…測定部、SN1…第1測定部、SN2…第2測定部、s…第1積分値、s…第2積分値、T…対応関係情報、U1…測定部材、v…第1積分平均値、v…第2積分平均値、x~x,x…入力変数、x…第1入力変数、x…第2入力変数、x μ~x30 μ,x ~x30 ,x ,x …第1入力素子群の各入力素子の値、x ~x30 ,x ,x …第2入力素子群の各入力素子の値、x ~x30 ,x …多重用素子群の各素子の値、x1,2~x7,6,xμ,ν…中間変数、x1,2…第1中間変数、x2,1…第2中間変数、x μ,ν~x30 μ,ν,x μ,ν,x μ,ν…中間素子の値、x 1,2~x30 1,2,x 1,2,x 1,2…第1中間素子群の各中間素子の値、x 2,1~x30 2,1,x 2,1,x 2,1…第2中間素子群の各中間素子の値、(x μ,ν,x μ,ν,…,x30 μ,ν)…中間変数の値、(x 1,2,x 1,2,…,x30 1,2)…第1中間変数の値、(x 2,1,x 2,1,…,x30 2,1)…第2中間変数の値、y~y,yγ…出力変数、y γ…出力変数の値。
 次に図面を参照しながら、本発明の実施の形態の具体例(以下、実施例と記載する)を説明するが、本発明は以下の実施例に限定されるものではない。
 なお、以下の図面を使用した説明において、理解の容易のために説明に必要な部材以外の図示は適宜省略されている。
 図1は本発明の実施例1の動作識別システムの全体説明図である。
 図2は本発明の実施例1の動作識別システムが識別する被験者の6種類の前腕動作の説明図である。
 図1において、本発明の実施例1の動作識別システム(動作意思推定システム、事象推定システム、分類推定システム)Sは、被検体の一例としての人間、いわゆる、被験者の動作意思(事象)に伴う筋活動(筋群の活動)に基づく筋電位信号(生体信号)を測定する表面筋電位測定装置(測定部材、筋電位信号測定部材)U1と、前記表面筋電位測定装置U1に接続されて前記被験者の動作を識別する動作識別装置本体(動作意思推定装置本体、事象推定装置本体、分類推定装置本体)の一例としての識別用クライアントパソコン(コンピュータ、パーソナルコンピュータ、ユーザが操作可能な端末)PCとを有する。
 なお、実施例1では、図2に示す、前記被験者の前腕の「基本姿勢」(「動作なし」、何も動作識別されていない状態)を含む7種類の動作、すなわち、「手首屈曲」(「動作A」)、「手首伸展」(「動作B」)、「手を握る」(「動作C」)、「手を開く」(「動作D」)、「手首回外」(「動作E」)、「手首回内」(「動作F」)の各動作を識別する。
 実施例1の前記表面筋電位測定装置U1は、前記被験者の前腕に予め設定された6箇所の測定位置(第1測定位置、第2測定位置、…、第6測定位置)において、前記筋電位信号を測定する6個の筋電センサ(6個の測定部、第1測定部、第2測定部、…、第6測定部)SN1~SN6を有する。具体的には、前記被験者の橈側手根屈筋、深指屈筋、尺側手根屈筋、指伸筋、長橈側手根屈筋および腕橈骨筋の6種類の筋肉に対応する前記被験者の右腕の肘付近の皮膚の6箇所が、測定位置として設定されている。また、前記被験者の6箇所の皮膚に貼り付けられた6対(6ch)の湿式電極が、前記6個の筋電センサSN1~SN6として構成されている。なお、前記湿式電極とは、あらかじめ導電性のペーストが電極側に組み込まれた電極のことであり、前記ペーストが組み込まれていない、いわゆる、乾式電極に比べ、皮膚の状態に左右されずに安定した測定が可能となっている。
 すなわち、実施例1の前記表面筋電位測定装置U1は、前記6箇所の皮膚に貼り付ける前記6対の湿式電極(SN1~SN6)により、前記6種類の筋肉に応じた筋電位信号、いわゆる、表面筋電位信号(6種類の生体信号、第1の情報、第2の情報、…、第6の情報)を測定する。
 また、実施例1の前記識別用クライアントパソコンPCは、いわゆる、コンピュータ装置により構成されており、コンピュータ本体H1と、ディスプレイH2と、キーボードH3やマウスH4等の入力装置と、図示しないHDドライブ(ハードディスクドライブ)等とにより構成されている。
 前記各筋電センサSN1~SN6を有する前記表面筋電位測定装置U1と、前記測定用クライアントパソコンPCとにより、実施例1の動作識別装置(動作意思推定装置、事象推定装置、分類推定装置)(U1+PC)が構成されている。
(実施例1の制御部の説明)
 図3は本発明の実施例1の動作識別システムを構成する各装置の機能をブロック図(機能ブロック図)で示した説明図である。
 図3において、前記表面筋電位測定装置U1および前記識別用クライアントパソコンPCのコンピュータ本体H1の制御部は、外部との信号の入出力および入出力信号レベルの調節等を行うI/O(入出力インターフェース)、必要な起動処理を行うためのプログラムおよびデータ等が記憶されたROM(リードオンリーメモリ、記録媒体)、必要なデータ及びプログラムを一時的に記憶するためのRAM(ランダムアクセスメモリ、記録媒体)、前記ROM等に記憶された起動プログラムに応じた処理を行うCPU(中央演算処理装置)、ならびにクロック発振器等を有しており、ROM及びRAM等に記憶されたプログラムを実行することにより種々の機能を実現することができる。実施例1の前記制御部CのROMには、動作識別プログラム(動作意思推定プログラム、事象推定プログラム、分類推定プログラム)AP1が記憶されている。
(表面筋電位測定装置U1の制御部に接続された信号入力要素)
 前記表面筋電位測定装置U1の制御部には、次の信号出力要素SN1~SN6等の出力信号が入力されている。
SN1~SN6:筋電センサ
 6個の筋電センサSN1~SN6は、6箇所の測定位置において測定された前記6種類の筋肉の表面筋電位信号を前記制御部に入力する。
(表面筋電位測定装置U1の制御部の機能)
 前記表面筋電位測定装置U1の制御部は、前記6個の筋電センサSN1~SN6の出力信号に応じて、前記識別用クライアントパソコンPCに入力する出力信号を演算するために、次の機能実現手段を有している。
C1:積分値演算手段
 積分値演算手段C1は、前記6個の筋電センサSN1~SN6からの前記6種類の表面筋電位信号の積分値(第1積分値s、第2積分値s、…、第6積分値s)、いわゆる、積分表面筋電位信号(積分筋電位、積分筋電図、筋電図積分値、IEMG:Integrated Electro Myo-Gram)(s~s)を演算する。
 実施例1の前記積分値演算手段C1は、3[kHz]のサンプリング周波数で取得された前記6種類の表面筋電位信号をそれぞれ全波整流平滑化することにより、6種類の積分表面筋電位信号(s~s)を演算する。具体的には、6種類の表面筋電位信号について、それぞれ、全波整流、すなわち、正負両波を含む交流波を負波が正負対称の正波となる直流波に変換した後、平滑化、すなわち、予め設定された遮断周波数でローパスフィルタリングを実行することにより、6種類の積分表面筋電位信号(s~s)を演算する。なお、実施例1では、前記遮断周波数が2.4[Hz]に設定されている。
C2:積分値送信手段
 積分値送信手段C2は、前記積分値演算手段C1により演算された前記6種類の積分表面筋電位信号(s~s)を、前記識別用クライアントパソコンPCのコンピュータ本体H1に送信する。実施例1の前記積分値送信手段C2は、3[kHz]のサンプリング周波数に基づいて、1[ms]に3点((3×10)×10-3=3)の積分表面筋電位信号(s~s)の値を、前記6種類の積分表面筋電位信号(s~s)それぞれについて前記識別用クライアントパソコンPCのコンピュータ本体H1に送信する。
(識別用クライアントパソコンPCのコンピュータ本体H1の制御部の機能)
 また、前記表面筋電位測定装置U1の制御部Cは、前記表面筋電位測定装置U1の出力信号、すなわち、積分表面筋電位信号(s~s)に基づいて、前記被験者の動作を識別する前記動作識別プログラムAP1において、次の機能実現手段を有している。
C3:積分値受信手段
 積分値受信手段C3は、前記積分値送信手段C2により送信された前記6種類の積分表面筋電位信号(s~s)を受信する。実施例1の前記積分値受信手段C3は、1[ms]に3点の前記積分表面筋電位信号(s~s)の値を、前記6種類の積分表面筋電位信号(s~s)それぞれについて受信する。
C4:積分値記憶手段
 積分値記憶手段C4は、前記積分値受信手段C3により受信された前記6種類の積分表面筋電位信号(s~s)を記憶する。実施例1の前記積分値受信手段C3は、1[ms]に3点ずつ受信する前記積分表面筋電位信号(s~s)の値を、前記6種類の積分表面筋電位信号(s~s)それぞれについて記憶する。
C5:積分平均値演算手段
 積分平均値演算手段C5は、予め設定されたフレーム時間内の前記6種類の積分表面筋電位信号(s~s)の平均値である積分平均値(第1積分平均値v、第2積分平均値v、…、第6積分平均値v、AIEMG:Average Integrated Electro Myo-Gram)を予め設定されたシフト時間ごとに演算する。
 図4は本発明の実施例1の積分平均値についての説明図である。
 実施例1の前記積分平均値演算手段C5では、図4に示すように、前記フレーム時間を15[ms]とし、前記シフト時間を10[ms]として、1[ms]に3点ずつ受信した前記6種類の積分表面筋電位信号(s~s)の平均値(v~v)をそれぞれ演算することにより、前記6種類の積分平均値(v~v)を演算する。すなわち、前記積分平均値演算手段C5は、15[ms]の1フレーム内に含まれる45点(3×15=45)の積分表面筋電位信号(s~s)の値の平均値(v~v)を、30点(3×10=30)ずつシフトさせながら、前記6種類の積分表面筋電位信号(s~s)それぞれについて演算してノイズを除去する。
 ここで、1フレーム内にQ点の第1積分値sを受信し、且つ、P点ずつシフトさせながらフレームごとの第1積分平均値vを演算する際に、q点目の第1積分値sの各値をIEMG(q)(q=1,2,…)とし、p番目のフレームの第1積分平均値vの値をAIEMG(p)(p=1,2,…)とした場合に、p番目のフレームの第1積分平均値vの値AIEMG(p)が、以下の式(2-1)によって示される。なお、実施例1では、Q=45,P=30となるように予め設定されている。
AIEMG(p)={IEMG(1+P×(p-1))
           +IEMG(2+P×(p-1))
           +…
           +IEMG(Q+P×(p-1))}
          /Q  …式(2-1)
 同様にして、p番目のフレームのその他の各積分平均値(v~v)の値(AIEMG(p)~AIEMG(p))についても、前記式(2-1)と同様に示すことができる。よって、iを各chの番号とした場合に(i=1,2,…,6)、p番目のフレームの各積分平均値vの値AIEMG(p)が、以下の式(2-2)によって示される。
AIEMG(p)={IEMG(1+P×(p-1))
           +IEMG(2+P×(p-1))
           +…
           +IEMG(Q+P×(p-1))}
          /Q  …式(2-2)
C6:最小値記憶手段
 最小値記憶手段C6は、前記積分平均値演算手段C5により演算された前記6種類の積分平均値(v~v)の最小値を、前記6種類の積分平均値(v~v)それぞれについて記憶する。なお、実施例1では、前記6種類の積分平均値vのフレームごとの値AIEMG(p)のうち、1番目からp番目のフレームまでの6種類それぞれの最小値をεとした場合に、前記最小値εが、以下の式(3)によって示される。
ε=min[AIEMG(p)]  …式(3)
C7:誤差除去手段
 誤差除去手段C7は、前記積分平均値演算手段C5により演算された前記6種類の積分平均値(v~v)から誤差を除去する。実施例1の前記誤差除去手段C7は、前記最小値記憶手段C6に記憶された前記6種類の最小値εを、前記被験者ごとの誤差、いわゆる、ノイズ(個人差、個体差)とみなして、前記6種類の積分平均値vのフレームごとの値AIEMG(p)から減算することにより、誤差を除去する。具体的には、前記6種類の積分平均値vの誤差除去値をAIEMGε(p)とした場合に、前記誤差除去値AIEMGε(p)が、以下の式(4)によって示される。
AIEMGε(p)=AIEMG(p)-ε  …式(4)
C8:合計値演算手段
 合計値演算手段C8は、前記誤差除去手段C7により誤差が除去された前記6種類の積分平均値(v~v)の合計値を演算する。実施例1の前記合計値演算手段C8は、前記6種類の積分平均値vのフレームごとの誤差除去値AIEMGε(p)の合計値を演算する。具体的には、前記6種類の積分平均値vのフレームごとの合計値をAIEMGall(p)とした場合に、前記合計値AIEMGall(p)が、以下の式(5)によって示される。
AIEMGall(p)=AIEMGε(p)
           +AIEMGε(p)
           +…
           +AIEMGε(p)  …式(5)
C9:合計値記憶手段
 合計値記憶手段C9は、前記合計値演算手段C8により演算された前記6種類の積分平均値(v~v)の1フレーム前の合計値AIEMGall(p-1)を記憶する。
C10:差分値演算手段
 差分値演算手段C10は、前記合計値演算手段C8により演算された前記6種類の積分平均値(v~v)の合計値AIEMGall(p)のフレームごとの差分値を演算する。実施例1の前記差分値演算手段C10は、前記合計値演算手段C8により演算された合計値(第1合計値)AIEMGall(p)と、前記合計値記憶手段C9に記憶された1フレーム前の合計値(第2合計値)AIEMGall(p-1)との差分値を演算する。具体的には、前記差分値をAIEMGdif(p)とした場合に、前記差分値AIEMGdif(p)が、以下の式(6)によって示される。
AIEMGdif(p)=AIEMGall(p)
           -AIEMGall(p-1)  …式(6)
C11:最大値記憶手段
 最大値記憶手段C11は、前記6種類の積分平均値vの誤差除去値AIEMGε(p)のうち、1番目からp番目のフレームまでの6種類それぞれの最大値と、前記合計値AIEMGall(p)のうち、1番目からp番目のフレームまでの最大値と、前記差分値AIEMGdif(p)のうち、1番目からp番目のフレームまでの最大値とを記憶する。実施例1の前記最大値記憶手段C11は、前記誤差除去値AIEMGε(p)の最大値をAIEMGεimaxとし、前記合計値AIEMGall(p)の最大値をAIEMGamaxとし、前記差分値AIEMGdif(p)の最大値をAIEMGdmaxとした場合に、以下の式(7-1)~(7-3)によって示される8種類の最大値AIEMGεimax,AIEMGamax,AIEMGdmaxを記憶する。
AIEMGεimax=max[AIEMGε(p)]  …式(7-1)
AIEMGamax=max[AIEMGall(p)]  …式(7-2)
AIEMGdmax=max[AIEMGdif(p)]  …式(7-3)
C12:入力値生成手段
 入力値生成手段C12は、正規化手段C12Aと、2値化手段C12Bとを有し、後述する選択的不感化ニューラルネット(SDNN:Selective Desensitization Neural Network)に入力する入力値(入力変数の値)を生成する。
C12A:正規化手段
 正規化手段C12Aは、前記6種類の積分平均値vの誤差除去値AIEMGε(p)と、前記合計値AIEMGall(p)と、前記差分値AIEMGdif(p)とをそれぞれの最大値AIEMGεimax,AIEMGamax,AIEMGdmaxで除算して正規化(normalize)することにより、シフト時間ごとに、すなわち、フレームごとに8種類の入力値を生成する。具体的には、8種類の正規化値(入力値)をN(p)(i=1,2,…,6),N(p),N(p)とした場合に、前記8種類の正規化値N(p),N(p),N(p)が、以下の式(8-1)~(8-3)によって示される。
(p)=AIEMGε(p)/AIEMGεimax  …式(8-1)
(p)=ln(AIEMGall(p)+1)
      /ln(AIEMGamax+1)  …式(8-2)
(p)=AIEMGdif(p)/AIEMGdmax  …式(8-3)
 なお、前記式(8-2)に示すように、正規化された合計値N(p)については、前記合計値AIEMGall(p)と前記最大値AIEMGamaxとの差が大きくなり過ぎる可能性がある。このため、実施例1の正規化された合計値N(p)は、自然対数(ln:natural logarithm)の値が0以上となる、ln(AIEMGall(p)+1)とln(AIEMGamax+1)とを除算した値をとって、とり得る値を0~1の範囲に正規化している。また、前記式(8-3)に示す、正規化された差分値N(p)については、AIEMGdif(p)<0,N(p)<0となる場合には、N(p)=0とみなすものとする。
C12B:2値化手段
 2値化手段C12Bは、前記正規化手段C12Aにより演算された前記8種類の入力値N(p),N(p),N(p)を2値化された表現に変換する。実施例1の前記2値化手段C12Bは、前記8種類の入力値N(p),N(p),N(p)を、15個の+1と、15個の-1とを用いて表現された値、すなわち、15個の+1と、15個の-1とをベクトル成分とするベクトル表現(2値パターン)で示すことが可能な値にそれぞれ変換する。具体的には、正規化された入力値N(p),N(p),N(p)は、0から1までの値をとり得る(0≦N(p)≦1,0≦N(p)≦1,0≦N(p)≦1)。このため、まず、0の場合には、1番目から15番目までの15個のベクトル成分が+1となり、16番目から30番目までの15個のベクトル成分が-1となるように変換する。すなわち、以下の式(9-1)に示すように変換する。
0=(+1,+1,…,+1,-1,-1,…,-1)  …式(9-1)
 次に、1/15(0.0666…)だけ大きな値になる度に、連続する15個の+1を右側に1つずつシフトさせる。例えば、0.1(1/15<0.1<2/15)の場合には、以下の式(9-2)に示すように右側シフトさせる。
0.1=(-1,+1,+1,…,+1,-1,-1,…,-1)
  …式(9-2)
 そして、1の場合には(15/15=1)、1番目から15番目までの15個のベクトル成分が-1となり、16番目から30番目までの15個のベクトル成分が+1となるように変換する。すなわち、以下の式(9-3)に示すように変換する。
1=(-1,-1,…,-1,+1,+1,…,+1)  …式(9-3)
 図5は本発明の実施例1の選択的不感化ニューラルネットの説明図である。
C13:入力値記憶手段
 入力値記憶手段(入力値入力手段)C13は、前記入力値生成手段C12で生成された前記8種類の入力値N(p),N(p),N(p)を記憶する。実施例1の入力値記憶手段C13は、前記2値化手段C12Bにより、15個の+1と、15個の-1とをベクトル成分とするベクトル表現で示すことが可能な値に変換された前記8種類の入力値N(p),N(p),N(p)を記憶する。
 ここで、実施例1では、シフト時間ごとに変化する前記8種類の入力値N(p),N(p),N(p)を、図5に示す選択的不感化ニューラルネットNに入力することにより、前記被験者の動作を識別する。図5において、実施例1の前記選択的不感化ニューラルネットNは、240個の素子(入力素子)により構成された入力層(入力素子群)Naを有する。
 実施例1の前記入力層Naでは、8種類の入力素子群(第1入力素子群G 、第2入力素子群G 、…、第7入力素子群としての差分用入力素子群G 、第8入力素子群としての合計用入力素子群(多重用素子群)G )を有する。また、前記8種類の入力素子群G (i=1,2,…,8)は、それぞれ30個の素子を有する(8×30=240)。
 なお、実施例1の前記入力層Naは、8種類の入力変数(第1入力変数x、第2入力変数x、…、第7入力変数x、第8入力変数x)の値、すなわち、前記8種類の入力値N(p),N(p),N(p)に対応して、前記8種類の入力素子群G が予め設定されている。
 よって、各入力素子群G の30個の素子の値をx ,x ,…,x30 とした場合に、各入力変数xが、以下の式(10)によって示される。
=(x ,x ,…,x30 )  …式(10)
 また、前記8種類の入力値N(p),N(p),N(p)をベクトル成分とする8次元の入力ベクトルをxとした場合に、前記入力ベクトルxが、以下の式(11)によって示される。
x=(x,…,x,x
 =(N(p),…,N(p),N(p))  …式(11)
 なお、実施例1の前記各入力素子群G の30個の素子には、+1または-1の2値(x ~x30 )を入力(保持)することが可能である。このため、実施例1の前記入力値記憶手段C13では、前記各入力値N(p),N(p),N(p)が2値化されたベクトル表現をシフト時間ごとに記憶する(更新する)ことにより、前記各入力素子群G の30個の素子の値(x ~x30 )をシフト時間ごとに入力したものとみなすことができる。なお、実施例1では、入力値N(p)として、例えば、(+1,+1,…,+1,-1,-1,…,-1)等の30個の素子の値を有する8種類の入力素子群の値、すなわち、合計で240(30×8=240)個の値(-1,+1)が記憶される。
 図6は本発明の実施例1の動作識別開始画像の説明図である。
C14:動作識別開始判別手段
 動作識別開始判別手段C14は、図6に示す動作識別開始画像1を表示する動作識別開始画像表示手段C14Aを有し、前記被験者の動作識別を開始するか否かを判別する。実施例1の前記動作識別開始判別手段C14は、図6に示す、前記動作識別開始画像1の識別開始釦1aが選択されたか否かを判別することにより、前記被験者の動作識別を開始するか否かを判別する。
C15:中間値演算手段
 中間値演算手段C15は、選択的不感化手段C15Aと、二重不感化手段C15Bとを有し、前記動作識別開始判別手段C14によって前記被験者の動作識別を開始すると判別された場合に、前記入力層Naに入力された前記8種類の入力値N(p),N(p),N(p)、すなわち、前記入力値記憶手段C13に記憶された前記8種類の入力値N(p),N(p),N(p)に基づく中間値(中間変数の値)をシフト時間ごとに演算する。
 図5において、実施例1の前記選択的不感化ニューラルネットNは、前記入力層Naの240個の素子と結合された1260個の素子(中間素子)により構成された中間層(中間素子群)Nbを有する。
 実施例1の前記中間層Nbでは、42種類の中間素子群(第1中間素子群G1,2 、第2中間素子群G2,1 、…、第42中間素子群G7,6 )を有する。また、前記42種類の中間素子群Gμ,ν (μ≠ν,μ=1,2,…,7,ν=1,2,…,7)は、それぞれ30個の素子を有する(42×30=1260)。
 なお、実施例1の前記中間層Nbでは、正規化された合計値N(p)を除く7種類の入力値N(p),N(p)に対応する7種類の入力素子群(G ~G )のうちの1つを被修飾群Gμ (μ=1,2,…,7)とし、前記被修飾群Gμ を除く前記6種類の各入力素子群(G ~G )のうちの1つを修飾群Gν (ν≠μ,ν=1,2,…,7)とした場合に、前記修飾群Gν により積型文脈修飾(積型修飾)された前記被修飾群Gμ に基づく合計42種類の中間素子群Gμ,ν が予め設定されている。
 ここで、各中間素子群Gμ,ν (μ≠ν,μ=1,2,…,7,ν=1,2,…,7)の30個の素子の値をx μ,ν,x μ,ν,…,x30 μ,νとした場合に、各中間変数xμ,νが、以下の式(12)によって示される。
μ,ν=(x μ,ν,x μ,ν,…,x30 μ,ν)  …式(12)
 なお、「積型文脈修飾」とは、前記被修飾群Gμの素子の値x μ~x30 μを前記修飾群Gνの素子の値x ν~x30 νに基づいて変化させる方法である。具体的には、図5の実線に示すように、前記修飾群Gνの素子の値x ν~x30 νに基づいて予め演算された出力感度(ゲイン)の値をg μ,ν~g30 μ,νとした場合に、前記被修飾群Gμの素子の値x μ(i=1,2,…,30)と前記出力感度の値g μ,ν(j=1,2,…,30)との積{(g μ,ν)×(x μ)}を演算することをいう。
 このとき、特に、前記出力感度の値g μ,νを0にして、前記積{(g μ,ν)×(x μ)}が0になること、すなわち、g μ,ν=0,{(g μ,ν)×(x μ)}=0となる場合を「不感化」という。また、前記被修飾群Gμの一部の素子の値x μ(i=1,2,…,30)をランダムに不感化することを「選択的不感化」という。
 なお、修飾群Gν による被修飾群Gμ の積型修飾については、以降、Gμ (Gν )と記載する。
 さらに、実施例1では、正規化された合計値N(p)に対応する合計用入力素子群G により、前記積型修飾Gμ (Gν )をさらに積型文脈修飾する。具体的には、図5の点線に示すように、前記合計用入力素子群G の素子の値(多重用素子群の各素子の値)x ~x30 に基づいて予め演算された合計用出力感度の値(多重用出力感度の値)をg μ,ν,8~g30 μ,ν,8とした場合に、前記積{(g μ,ν)×(x μ)}と前記合計用出力感度の値g μ,ν,8(k=1,2,…,30)との積を前記中間素子群Gμ,ν の素子の値x μ,ν(i=1,2,…,30)とする。このため、実施例1では、前記中間素子群Gμ,ν の素子の値x μ,νが、以下の式(13)によって示される。
μ,ν=(g μ,ν,8)×(g μ,ν)×(x μ)  …式(13)
 なお、実施例1では、前記各出力感度の値g μ,ν,g μ,ν,8が、以下の式(14-1),(14-2),(14-1)′,(14-2)′によって示される。
μ,ν=(1+x ν)/2  …式(14-1)
j=σ(i)  …式(14-2)
μ,ν,8=(1+x )/2  …式(14-1)′
k=σ′(i)  …式(14-2)′
 すなわち、各入力素子群Gν ,G の素子の値x ν,x が+1の場合には、各出力感度の値g μ,ν,g μ,ν,8も+1となるが、-1の場合には、各出力感度の値g μ,ν,g μ,ν,8が0となり、不感化される。
 ここで、σ(i),σ′(i)は、いわゆる、乱数関数であって、(1,2,…,30)の順列をランダムに置換した順列、例えば、(21,15,…,4),(17,6,…,8)等のi番目の値(1から30までの乱数値)である。
 なお、前記乱数関数σ(i),σ′(i)を使用しなければ、前記被修飾群Gμ と、前記修飾群Gνまたは前記合計用入力素子群G との相関が強い場合、前記被修飾群Gμ の素子のほとんどが選択的不感化されたり、前記素子のほとんどが選択的不感化されなかったり、x μ=+1となる素子ばかり選択的不感化されたり、x μ=-1となる素子ばかり選択的不感化されたりする可能性が高く、動作識別の性能が著しく低下する。このため、実施例1では、前記乱数関数σ(i),σ′(i)を使用して、前記各素子群Gμ ,Gν,G の相関を擬似的に弱くしている。
 ここで、前記式(13)に、前記式(14-1),(14-2),(14-1)′,(14-2)′を代入した式(13)′を以下に示す。
μ,ν={(1+xσ′(i) )/2}
     ×{(1+xσ(i) ν)/2}×(x μ)  …式(13)′
 すなわち、実施例1の前記中間素子群Gμ,ν の素子の値x μ,νは、前記被修飾群Gμ の素子の値x μが、前記出力感度の値g μ,νによって選択的不感化された後、前記合計用出力感度の値g μ,ν,8によってさらに選択的不感化(二重不感化、多重不感化)された値である。すなわち、前記被修飾群Gμ の適切な数、具体的には、15個以上30個以下の素子の値x μが不感化された値である。
 実施例1では、42種類の出力感度の値(第1出力感度の値g 1,2、第2出力感度の値g 2,1、…第42出力感度の値g 7,6)の集合により、42種類のゲインベクトル(第1ゲインベクトルG1,2 、第2ゲインベクトルG2,1 、…第42ゲインベクトルG7,8 )が構成されており、42種類の合計用出力感度の値(第1合計用出力感度の値(第1多重用出力感度の値)g 1,2,8、第2合計用出力感度の値(第2多重用出力感度の値)g 2,1,8、…第42合計用出力感度の値(第42多重用出力感度の値)g 7,6,8)の集合により、42種類の合計用ゲインベクトル(第1合計用ゲインベクトルG1,2,8 、第2合計用ゲインベクトルG2,1,8 、…第42合計用ゲインベクトルG7,6,8 )が構成されている。
 このため、実施例1の前記中間値演算手段C15では、42種類の中間素子群Gμ,ν に対応する前記42種類のゲインベクトルGμ,ν と前記42種類の合計用ゲインベクトルGμ,ν,8 とが演算された後、42種類の中間素子群Gμ,ν の素子の値x μ,νが2重不感化によって演算されることにより、42種類の中間変数xμ,ν(μ≠ν,μ=1,2,…,7,ν=1,2,…,7)の値(中間値)が演算される。
 よって、実施例1の前記42種類の中間変数xμ,νの値、すなわち、前記中間値は、-1,0,+1の3値をベクトル成分とする30次元ベクトルのベクトル表現(3値パターン)、例えば、(+1,0,0,-1,+1,…,0)等で示すことができる。
C15A:選択的不感化手段
 選択的不感化手段C15Aは、前記式(14-1),(14-2)によって前記42種類の出力感度の値g μ,νを演算する出力感度演算手段C15A1を有し、図5の実線で示すように、前記42種類の出力感度の値g μ,νにより前記被修飾群Gμ の適切な数の素子の値x μを選択的不感化する。
C15B:二重不感化手段
 二重不感化手段(多重不感化手段)C15Bは、前記式(14-1)′,(14-2)′によって前記42種類の合計用出力感度の値g μ,ν,8を演算する合計用出力感度演算手段C15B1を有し、図5の点線、前記式(13),(13)′で示すように、前記42種類の合計用出力感度の値g μ,ν,8により前記積型修飾Gμ (Gν )の適切な数の素子の値、すなわち、前記積{(g μ,ν)×(x μ)}を選択的不感化することにより、前記被修飾群Gμ の適切な数の素子の値x μを二重不感化する。
C16:中間値記憶手段
 中間値記憶手段(中間値入力手段)C16は、前記中間値演算手段C15で演算された前記42種類の中間値、すなわち、前記42種類の中間変数xμ,ν(μ≠ν,μ=1,2,…,7,ν=1,2,…,7)の値を記憶する。実施例1の中間値記憶手段C16では、前記二重不感化手段C15Bによって演算された前記42種類の中間素子群Gμ,ν の素子の値x μ,ν(μ≠ν,μ=1,2,…,7,ν=1,2,…,7,i=1,2,…,30)をシフト時間ごとに記憶することにより、前記各中間素子群Gμ,ν の30個の素子に、-1,0,+1の3値のいずれかがシフト時間ごとに入力(保持)されたものとみなすことができる。なお、実施例1では、前記中間値として、例えば、(+1,0,0,-1,+1,…,0)等の30個の素子の値を有する42種類の中間素子群の値、すなわち、合計で1260(30×42=1260)個の値(-1,0,+1)が記憶される。
C17:出力値演算手段
 出力値演算手段C17は、結合荷重記憶手段C17Aを有し、前記中間層Nbに入力された前記42種類の中間値、すなわち、前記42種類の中間変数xμ,ν(μ≠ν,μ=1,2,…,7,ν=1,2,…,7)の値に基づく出力値(出力変数の値)をシフト時間ごとに演算する。
 図5において、実施例1の前記選択的不感化ニューラルネットNは、前記中間層Nbの1260個の素子と結合された48個の素子(出力素子)により構成された出力層(出力素子群)Ncを有する。
 実施例1の前記出力層Ncは、6種類の出力素子群(第1出力素子群G 、第2出力素子群G 、…、第6出力素子群G )を有する。また、前記6種類の出力素子群Gγ (γ=1,2,…,6)は、それぞれ8個の素子を有する(6×8=48)。
 なお、実施例1の前記出力層Ncは、「基本姿勢」を除く6種類の動作に応じて予め設定された6種類の出力変数(「手首屈曲」に対応する第1出力変数y、「手首伸展」に対応する第2出力変数y、…、「手首回内」に対応する第6出力変数y)の値に対応して、前記6種類の出力素子群Gγ が予め設定されている。
 よって、各出力素子群Gγ の8個の素子の値をy γ,y γ,…,y γとした場合に、各出力変数yγが、以下の式(15)によって示される。
γ=(y γ,y γ,…,y γ)  …式(15)
 実施例1では、6種類の出力素子群Gγ の素子の値y γ(i=1,2,…,8)が、前記中間層Nbの1260個の素子の値x μ,ν(j=1,2,…,30)と、前記各値x μ,νの重視度合いに応じて予め設定された1260個の各結合荷重の値とに基づいて演算される。
 ここで、前記各値x μ,νに応じた前記各結合荷重の値をωγij μ,νとし、前記出力素子群G の素子の値y γに応じて予め設定された閾値をh γとし、0より大きい値が入力されると+1を出力し且つ0以下の値が入力されると0を出力する2値関数をhとした場合に、前記出力素子群G の素子の値y γが、以下の数1の式(16)によって示される。
Figure JPOXMLDOC01-appb-M000001

  …式(16)
 すなわち、実施例1では、前記出力層Ncの48個の素子が、それぞれ結合荷重をもって前記中間層の1260個の素子と結合している。このため、前記出力層Ncの48個の素子の値y γが、1260個の素子の値x μ,νとその結合荷重の値ωγij μ,νの積{(ωγij μ,ν)×(x μ,ν)}の総和に基づいて演算された値(0,+1の2値)になっている。
 この結果、実施例1の前記6種類の出力変数yγの値、すなわち、前記出力値は、0,+1の2値をベクトル成分とする8次元ベクトルのベクトル表現(2値パターン)、例えば(0,+1,+1,0,+1,0,0,0)等で示すことができる。
C17A:結合荷重記憶手段
 結合荷重記憶手段C17Aは、48種類の閾値h γを記憶する閾値記憶手段C17A1を有し、前記出力層Ncの48個の素子それぞれについて、前記中間層の1260種類の素子の値x μ,νに応じた1260個の結合荷重の値ωγij μ,νを記憶する。すなわち、実施例1の前記結合荷重記憶手段C17Aは、合計60480種類の結合荷重の値ωγij μ,νを記憶する(1260×48=60480)。なお、実施例1では、後述する結合荷重学習手段C23によって学習した前記60480種類の結合荷重の値ωγij μ,νおよび前記48種類の閾値h γを記憶する。
C18:出力値記憶手段
 出力値記憶手段(出力値入力手段)C18は、前記出力値演算手段C17で演算された前記6種類の出力値、すなわち、前記6種類の出力変数yγ(γ=1,2,…,6)の値を記憶する。実施例1の出力値記憶手段C18では、前記6種類の出力素子群Gγ の素子の値y γ(i=1,2,…,8)をシフト時間ごとに記憶することにより、前記各出力素子群Gγ の8個の素子に、0,+1の2値のいずれかがシフト時間ごとに入力(保持)されたものとみなすことができる。なお、実施例1では、前記出力値として、例えば、(0,+1,+1,0,+1,0,0,0)等の8個の素子の値を有する6種類の出力素子群の値、すなわち、合計で48(8×6=48)個の値(0,+1)が記憶される。
C19:対応関係情報記憶手段
 対応関係情報記憶手段C19は、「基本姿勢」を除く6種類の動作(分類対象である複数の分類情報、複数の事象)と、前記各動作時に予め測定された6種類の出力変数yγ(γ=1,2,…,6)の値(分類情報ごとに対応する出力変数の値、事象ごとに対応する出力変数の値)との対応関係を特定する対応関係情報Tを記憶する。実施例1の前記対応関係情報Tは、予め測定された実際の6種類の動作に応じた6種類の出力素子群Gγ の素子の値(理想値)py γ,py γ,…,py γについての情報(動作正解情報、動作正解データ、教師信号、辞書情報)である。具体的には、連続する(隣接する)4個の素子が+1、すなわち、発火している(+1,…,+1,0,…,0),(0,+1,…,+1,0,0,0),(0,0,+1,…,+1,0,0),(0,0,0,+1,…,+1,0),(0,…,0,+1,…,+1)の5パターンが予め設定されている。
 図7は本発明の実施例1の対応関係情報の説明図であり、縦軸を電圧[mV]とし横軸を時間[s]とし6種類の動作を実際に順次実行した際の積分平均値の誤差除去値の合計値と動作識別した範囲とを示すグラフである。
 ここで、実施例1では、前記対応関係情報Tを記憶するために、まず、図7に示すように、前記被験者が前記6種類の動作を、「動作A」から「動作F」の順、すなわち、「手首屈曲」、「手首伸展」、「手を握る」、「手を開く」、「手首回外」、「手首回内」の順に、2秒間隔で実際に実行した際の誤差が除去された前記6種類の積分平均値(v~v)の合計値AIEMGall(p)を演算する。次に、前記被験者の2秒間隔の動作は、誤差を伴うため、前記合計値AIEMGall(p)による動作識別のための閾値を設定する。
 具体的には、予め設定された合計用閾値をthとし、前記合計値AIEMGall(p)の最大値AIEMGamaxに対する予め設定された割合をα(0<α<1)とした場合に、前記合計用閾値thと前記最大値AIEMGamaxに前記割合αを乗算した値(α×AIEMGamax)を動作識別のための閾値とする。なお、実施例1では、前記合計用閾値thとして、0.2[mV]が予め設定されている。
 次に、前記合計値AIEMGall(p)と前記各閾値th,(α×AIEMGamax)とに基づいて、予め測定された動作を識別する。具体的には、「動作A」から「動作F」の6種類の各動作を識別する際に+1を出力し且つ識別しない際に0を出力する6種類の各識別関数をM(p),M(p),…,M(p)とした場合に、前記各動作の各識別関数Mm′(p),(m′=1,2,…,6)の値を、以下の式(17)によって演算する。
m′(p)
=h(AIEMGall(p)-th)
 ×h(AIEMGall(p)-α×AIEMGamax)  …式(17)
 すなわち、図7に示すように、最初の2秒間(0[s]~2[s])で「動作A」を第1の識別関数M(p)で識別し、次の2秒間(2[s]~4[s])で「動作B」を第2の識別関数M(p)で識別する処理を繰り返してゆき、最後の2秒間(10[s]~12[s])で「動作F」を第6の識別関数M(p)で識別する。
 次に、前記各識別関数Mm′(p),(m′=1,2,…,6)が+1を継続して得たフレーム数(継続フレーム数)を、擬似的な動作速度とする。次に、前記被験者の2秒間隔の6動作を複数回繰り返し、前記継続フレーム数の最大値に基づいて、動作速度の正規化を行う。具体的には、前記継続フレーム数をWm′とし、前記継続フレーム数Wm′の最大値をWm′maxとし、正規化された動作速度をNsp(p)とした場合に、正規化された動作速度Nsp(p)が、以下の式(18)に示される。
sp(p)=Wm′/Wm′max  …式(18)
 そして、前記入力値N(p),N(p),N(p)と同様に、正規化された動作速度Nsp(p)(0≦Nsp(p)≦1)を、4個の+1と、4個の0とをベクトル成分とするベクトル表現(2値パターン)で示すことが可能な値にそれぞれ変換する。具体的には、前記正規化された動作速度Nsp(p)と、前記理想値py γ(i=1,2,…,8)の5パターンのベクトル表現とが、以下の式(19-1)~(19-5)によって対応付けられる。
0≦Nsp(p)<0.2の場合:
sp(p)=(0,…,0,+1,…,+1)  …式(19-1)
0.2≦Nsp(p)<0.4の場合:
sp(p)=(0,0,0,+1,…,+1,0)  …式(19-2)
0.4≦Nsp(p)<0.6の場合:
sp(p)=(0,0,+1,…,+1,0,0)  …式(19-3)
0.6≦Nsp(p)<0.8の場合:
sp(p)=(0,+1,…,+1,0,0,0)  …式(19-4)
0.8≦Nsp(p)≦1の場合:
sp(p)=(+1,…,+1,0,…,0)  …式(19-5)
C20:動作識別手段
 動作識別手段(動作意思推定手段、事象推定手段、分類情報推定手段)C20は、動作種類識別手段C20Aと、動作速度識別手段C20Bとを有し、前記対応関係情報記憶手段C19に予め記憶された前記対応関係情報Tと、前記6種類の出力変数yγの値とに基づいて、前記被験者の動作を識別する(動作意思を推定する、事象を推定する、分類情報を推定する)。
C20A:動作種類識別手段
 動作種類識別手段(種類情報推定手段)C20Aは、前記被験者の動作が前記7種類の動作(「基本姿勢」、「手首屈曲」、「手首伸展」、「手を握る」、「手を開く」、「手首回外」、「手首回内」)のいずれであるかを識別する(各動作の種類の情報である種類情報を推定する)。
 実施例1の前記動作種類識別手段C20Aは、前記対応関係情報Tである前記式(19-1)~(19-5)に示す5パターンが、4個の素子が連続して発火しているため、3個以上の素子が予め設定された識別フレーム数Ps以上継続して発火した前記出力素子群Gγ (γ=1,2,…,6)がある場合には、前記出力素子群Gγ (γ=1,2,…,6)に対応する動作であると識別する。例えば、前記第1出力素子群G が識別フレーム数Ps以上継続して(0,+1,+1,0,+1,0,0,0)等となる場合には、「手首屈曲」と識別する。また、3個以上の素子が識別フレーム数Ps以上継続して発火した前記出力素子群Gγ (γ=1,2,…,6)がない間は、「基本姿勢」であると識別する。なお、実施例1では、3個以上の素子が識別フレーム数Ps以上継続した前記出力素子群Gγ が複数存在する場合には、対応する複数の動作を同時実行したと識別する。よって、前記動作種類識別手段C20Aは、6種類の出力素子群Gγ について、3個以上の素子の発火が継続したフレーム数をpγ(γ=1,2,…,6,0≦pγ≦Ps)とした場合に、pγ=Psとなった出力素子群Gγ に応じた動作を実行したと識別し、pγ=Psとなった出力素子群Gγ がない間は、「基本姿勢」であると識別する。
C20B:動作速度識別手段
 動作速度識別手段(速度情報推定手段、力情報推定手段)C20Bは、「基本姿勢」を除く前記6種類の動作が実行された場合に、実行された動作の速度を識別する(各動作の速度情報を推定する)。実施例1の前記動作速度識別手段C20Bは、前記動作種類識別手段C20Aによって識別された特定の動作に対応する前記出力素子群Gγ (γ=1,2,…,6)の素子の発火パターンを数値化した値と、前記式(19-1)~(19-5)に示す5パターンを数値化した値とを比較して前記動作の速度を識別する。具体的には、まず、特定された前記出力素子群Gγ の1から8番目の素子に、1から8の重み付けをし、1から8番目の素子の値y γ(i=1,2,…,8)と、1から8の重み付けとを乗算した値を、発火素子数で除算した値(識別値)を演算する。
 例えば、「手首屈曲」と識別された際の前記第1出力素子群G が(0,+1,+1,0,+1,0,0,0)である場合には、前記識別値が3.3となる({(0×1)+(+1×2)+(+1×3)+(0×4)+(+1×5)+(0×6)+(0×7)+(0×8)}/3=(2+3+5)/3=10/3≒3.3)。また、前記式(19-1)~(19-5)に示す5パターンの前記識別値を演算すると、2.5((1+2+3+4)/4=2.5)から6.5((5+6+7+8)/4=6.5)となる。
 そして、前記識別フレーム数Ps分の前記識別値の平均値を演算した結果、前記式(19-1)~(19-5)に示す5パターンの前記識別値(2.5~6.5)のどの範囲に属するかによって、速度を識別する。なお、実施例1では、特定の出力素子群Gγ のフレームごとの識別値をdγ(pγ)(γ=1,2,…,6,pγ=1,2,…,Ps)とし、前記識別値dγ(pγ)の平均値をdγ(γ=1,2,…,6,dγ={dγ(1)+dγ(2)+…+dγ(Ps)}/Ps)とした場合に、2.5≦dγ≦3.5となる場合には「速い」と識別し、3.5<dγ<5.5の場合には「普通」と識別し、5.5≦dγ≦6.5の場合には「遅い」と識別する。
 なお、実施例1では動作の速度を識別するが、速度の変化に応じて力が発生するため、例えば、動作の速度(速度情報、「速い」、「普通」、「遅い」の各情報)に応じた動作の力(力情報、「強い」、「普通」、「弱い」の各情報)を推定することも可能である。
 具体的には、速度の識別の場合と同様に、特定の動作に対応する前記出力素子群Gγ (γ=1,2,…,6)の素子の発火パターンとしての前記平均値dγについて、2.5≦dγ≦3.5となる場合には「強い」と識別し、3.5<dγ<5.5の場合には「普通」と識別し、5.5≦dγ≦6.5の場合には「弱い」と識別することにより、動作の力を識別することが可能である。
 この場合、実際の動作の速度が0となる場合でも、動作の力を識別することが可能となる。例えば、被験者が「手を握った(動作C)」状態において、被験者が、握る力を「弱く」から「強く」に変化させた場合には、前記第3出力素子群G の素子の発火パターンとしての前記平均値dが、5.5≦d≦6.5の範囲の値から2.5≦d≦3.5の範囲の値に変化するため、動作の力が「弱い」から「強い」に変化したことを識別することが可能となる。また、例えば、障害者の失った手足の動作意思についての動作の種類と力とを推定することも可能となる。例えば、手を失った被験者に対し、実施例1と同様に、前記被験者の前腕の前記6箇所の測定位置に前記6個の筋電センサSN1~SN6を貼り付けて表面筋電位信号を測定すると、表面筋電位信号が前記被験者の動作意思に応じて変化し、前記平均値dγが変化するため、被験者の動作意思から失った手で行いたかった動作の種類と力とを推定することが可能となる。
 さらに、実施例1のように、動作の速度(速度情報)が「速い」、「普通」、「遅い」の3種類程度であれば、動作の力(力情報)が「強い」、「普通」、「弱い」の3種類のいずれかを動作の速度と同時に推定することも可能である。
 図8は本発明の実施例1の動作識別画像の説明図である。
C21:動作識別画像表示手段
 動作識別画像表示手段(動作意思表示手段、事象表示手段、分類情報表示手段)C21は、前記動作識別手段C20により識別された前記被験者の動作(推定された被検体の動作意思、推定された事象、推定された分類情報)を図8に示す動作識別画像(動作意思推定画像、事象推定画像、分類情報推定画像)2を表示する。図8において、実施例1の前記動作識別画像2は、識別された前記被験者の動作の開始時期、種類および速度を表示する表示部2aと、前記被験者の動作識別を終了するための動作識別終了釦2bとを有する。実施例1の前記動作識別画像表示手段C21は、前記表示部2aに識別された前記動作の開始時期、種類および速度を表示する。
C22:動作識別終了判別手段
 動作識別終了判別手段C22は、前記被験者の動作識別を終了するか否かを判別する。実施例1の前記動作識別終了判別手段C22は、前記識別終了釦2bが選択されたか否かを判別することにより、前記被験者の動作識別を終了するか否かを判別する。
C23:結合荷重学習手段
 結合荷重学習手段C23は、閾値学習手段C23Aと、初期値設定手段C23Bとを有し、前記動作識別手段C20により識別された7種類の動作(推定された被検体の動作意思、推定された事象、推定された分類情報)と、予め測定された実際の動作(実際の動作意思、実際の事象、実際の分類情報)との相違に基づいて、前記結合荷重を更新することにより、前記各結合荷重を学習する。実施例1の前記結合荷重学習手段C23は、前記動作識別終了判別手段C22によって前記被験者の動作識別を終了すると判別された後に、前記60480種類の結合荷重の値ωγij μ,νを、以下の式(20-1)によって更新する。なお、予め設定された正の値の定数(正定数)をcとする(0<c<1、例えば、c=0.3)。
ωγij μ,ν←ωγij μ,ν+c(py γ-y γ)×(x μ,ν)  …式(20-1)
C23A:閾値学習手段
 閾値学習手段C23Aは、前記動作識別終了判別手段C22によって前記被験者の動作識別を終了すると判別された後に、前記48種類の閾値h γを、以下の式(20-2)によって更新することにより、前記各閾値h γを学習する。
γ←h γ-c(py γ-y γ)  …式(20-2)
C23B:初期値設定手段
 初期値設定手段C23Bは、前記60480種類の結合荷重の値ωγij μ,νおよび前記48種類の閾値h γの初期値を設定する。実施例1の前記初期値設定手段C23Aは、前記対応関係情報記憶手段C19における前記理想値py γの設定の際に、測定された前記6種類の積分平均値(v~v)を前記選択的不感化ニューラルネットNに入力して演算された実測値、すなわち、前記中間変数xμ,ν(μ≠ν,μ=1,2,…,7,ν=1,2,…,7)の値と、前記6種類の出力変数yγ(γ=1,2,…,6)の値とに基づいて、前記初期値を設定する。具体的には、図7に示す前記被験者の2秒間隔の6動作において、前記各識別関数Mm′(p),(m′=1,2,…,6)が+1を継続していたフレームの出力値が、前記理想値py γとなるように、前記式(20-1),(20-2)によって前記結合荷重の値ωγij μ,νおよび前記閾値h γを学習する。
例えば、最初の2秒間(0[s]~2[s])の「動作A」が「速い」であった場合、第1出力素子群G の8個の素子の値については、(y ,y ,…,y )=(py ,py ,…,py )=(0,…,0,+1,…,+1)が成立し、他の出力素子群Gγ (γ=2,3,…,6)の8個の素子の値については、(y γ,y γ,…,y γ)=(0,0,…,0)が成立するものとする。
 この場合において、前記式(13)′によって前記中間層Nbの1260種類の素子の値x μ,νが二重不感化されて得られた際に、ωγij μ,ν=h γ=0,c=0.3とすると、第1出力素子群G の8個の素子の実測値について、(y ,y ,…,y )=(0,0,…,0)が成立するため、前記式(20-1),(20-2)に代入すると、第1出力素子群G の8個の素子に対する10080種類(1260×8)の結合荷重の値ω1ij μ,νおよび8種類の閾値h が、以下の式(20-1-1),(20-2-1)に示すように学習できる。
ω1ij μ,ν←0+0.3×(py -0)×(x μ,ν)  …式(20-1-1)
←0-0.3×(py -0)  …式(20-2-1)
 よって、例えば、第1出力素子群G の1番目の素子に応じた1260種類の結合荷重の値ω11j μ,νについては、py =0であるため更新されず、前記1番目の素子に応じた閾値h についても更新されない(ω11j μ,ν=0,h =0)。また、例えば、8番目の素子に応じた1260種類の結合荷重の値ω18j μ,νについては、py =1であるため、x μ,ν=1であれば0.3に更新され、前記8番目の素子に応じた閾値h については-0.3に更新される(ω18j μ,ν=0.3,h =-0.3)。
 また、以降の2秒毎に実行される各動作(「動作B」から「動作F」)についても、同様に前記結合荷重の値ωγij μ,νおよび前記閾値h γを学習できる。
 そして、動作速度の正規化のために試行(測定)した回数だけ学習を繰り返した結果、得られた前記60480種類の結合荷重の値ωγij μ,νおよび前記48種類の閾値h γが前記初期値として設定される。なお、前記初期値として設定された前記結合荷重の値ωγij μ,νおよび前記閾値h γは、前記結合荷重記憶手段C17Aに記憶される。
(実施例1のフローチャートの説明)
 次に、実施例1の制御部Cの関数近似プログラムAP1の処理の流れをフローチャートを使用して説明する。なお、前記対応関係情報Tと、前記60480種類の結合荷重の値ωγij μ,νおよび前記48種類の閾値h γの初期値とを設定する処理については、前記被験者の2秒間隔の6動作を測定した際に、前記式(17),(18),(19-1)~(19-5),(20-1),(20-2)によって、理想値および初期値を演算して記憶するだけであるため、フローチャートによる詳細な説明を省略する。
(実施例1の動作識別処理のフローチャートの説明)
 図9は本発明の実施例1の動作識別プログラムの動作識別処理のフローチャートである。
 図9のフローチャートの各ST(ステップ)の処理は、前記制御部CのROM等に記憶されたプログラムに従って行われる。また、この処理は前記制御部Cの他の各種処理と並行してマルチタスクで実行される。
 図9に示すフローチャートは前記表面筋電位測定装置U1および前記識別用クライアントパソコンPCが電源オンした後、前記動作識別プログラムAP1が起動した場合に開始される。
 図9のST1において、図6に示す動作識別開始画像1をディスプレイH2に表示する。そして、ST2に移る。
 ST2において、動作識別開始画像1の識別開始釦1aが選択されたか否かを判別することにより、図2に示す被験者の7種類の動作を識別する処理を開始するか否かを判別する。イエス(Y)の場合はST3に移り、ノー(N)の場合はST2を繰り返す。
 ST3において、以下の(1)~(3)の処理を実行し、ST4に移る。
(1)6種類の継続フレーム数pγ(γ=1,2,…6)、(6×Ps)種類の識別値dγ(pγ)(γ=1,2,…,6,pγ=1,2,…,Ps)、6種類の平均値dγ(γ=1,2,…,6)に0をセットする(pγ=dγ(pγ)=dγ=0)。
(2)60480種類の結合荷重の値ωγij μ,ν(γ=1,2,…6,i=1,2,…,8,j=1,2,…,8,μ≠ν,μ=1,2,…,7,ν=1,2,…,7)および48種類の閾値h γ(γ=1,2,…6,i=1,2,…,8)の初期値をセットする。
(3)図8に示す動作識別画像2をディスプレイH2に表示する。
 ST4において、後述する図10に示す入力値生成処理を実行して、8種類の入力変数x(i=1,2,…,8)の値(入力値)を生成する。すなわち、8種類の入力素子群G (i=1,2,…,8)の30個の素子、すなわち、合計で240(8×30=240)個の素子に-1,+1のいずれかを入力する(240個の値(-1,+1)を記憶する)。そして、ST5に移る。
 ST5において、式(13)′に示す二重不感化を実行して、42種類の中間変数xμ,ν(μ≠ν,μ=1,2,…,7,ν=1,2,…,7)の値(中間値)を演算する。すなわち、42種類の中間素子群Gμ,ν (μ≠ν,μ=1,2,…,7,ν=1,2,…,7)の30個の素子、すなわち、合計で1260(42×30=1260)個の素子に-1,0,+1のいずれかを入力する(1260個の値(-1,0,+1)を記憶する)。そして、ST6に移る。
 ST6において、式(16)に示す、中間層Nbの1260個の素子の値x μ,ν(j=1,2,…,8,μ≠ν,μ=1,2,…,7,ν=1,2,…,7)と結合荷重の値ωγij μ,νとの積の総和を閾値h γでフィルタリングする演算を実行して、6種類の出力変数yγの値(出力値)を演算する。すなわち、6種類の出力素子群Gγ (γ=1,2,…6)の8個の素子、すなわち、合計で48(6×8=48)個の素子に0,+1のいずれかを入力する(48個の値(0,+1)を記憶する)。そして、ST7に移る。
 ST7において、3個以上の素子が発火した出力素子群Gγ (出力変数yγの値、出力値、発火パターン)が検出されたか否かを判別する。イエス(Y)の場合はST8に移り、ノー(N)の場合はST12に移る。
 ST8において、以下の(1),(2)の処理を実行し、ST9に移る。
(1)継続フレーム数pγに+1を加算する。(pγ=pγ+1)
(2)識別値dγ(pγ)を演算する。
 ST9において、継続フレーム数pγが識別フレーム数Psになったか否かを判別する。イエス(Y)の場合はST10に移り、ノー(N)の場合はST13に移る。
 ST10において、以下の(1)~(3)の処理を実行し、ST11に移る。
(1)継続フレーム数pγが識別フレーム数Psになった出力素子群Gγ に応じた動作の種類を識別する。
(2)識別された動作に応じた出力素子群Gγ の識別値dγ(pγ)の平均値dγを演算して(dγ={dγ(1)+dγ(2)+…+dγ(2)}/Ps)、動作の速度を識別する。すなわち、2.5≦dγ≦3.5ならば「速い」と識別し、3.5<dγ<5.5ならば「普通」と識別し、5.5≦dγ≦6.5ならば「遅い」と識別する。
(3)識別フレーム数Psに達した継続フレーム数pγを0にリセットする(pγ=0)。
 ST11において、識別された動作の開始時期・種類・速度を動作識別画像2の表示部2aに表示する(記載する)。そして、ST13に移る。
 ST12において、被験者の上腕が「基本姿勢」であると識別し、「基本姿勢」の開始時期を動作識別画像2の表示部2aに表示する(記載する)。そして、ST13に移る。
 ST13において、動作識別画像2の識別終了釦2bが選択されたか否かを判別することにより、被験者の動作を識別する処理を終了するか否かを判別する。イエス(Y)の場合はST14に移り、ノー(N)の場合はST4に戻る。
 ST14において、式(20-1),(20-2)に示す結合荷重学習処理を実行して、60480種類の結合荷重の値ωγij μ,νと48種類の閾値h γを学習する(更新する)。そして、ST1に戻る。
(実施例1の入力値生成処理のフローチャートの説明)
 図10は本発明の実施例1の動作識別プログラムの入力値生成処理のフローチャートである。
 図10のフローチャートの各ST(ステップ)の処理は、前記制御部CのROM等に記憶されたプログラムに従って行われる。また、この処理は前記動作識別処理と同様に前記制御部Cの他の各種処理と並行してマルチタスクで実行される。
 図10に示すフローチャートは前記表面筋電位測定装置U1および前記識別用クライアントパソコンPCが電源オンした後、前記動作識別プログラムAP1が起動した場合に開始される。
 図10のST101において、表面筋電位測定装置U1の制御部で、図1に示す6個の筋電センサSN1~SN6からの6種類の表面筋電位信号の積分値である積分表面筋電位信号(s~s)を演算し、識別用クライアントパソコンPCに送信する。そして、ST102に移る。
 ST102において、識別用クライアントパソコンPCにおいて、積分表面筋電位信号(s~s)を受信して記憶する。そして、ST103に移る。
 ST103において、式(2-1),(2-2)、図4に示すように、6種類の積分表面筋電位信号(s~s)の平均値である積分平均値(v~v)のフレームごとの値AIEMG(p)(i=1,2,…,6,p=1,2,…)を演算する。そして、ST104に移る。
 ST104において、式(3)に示すように、演算された6種類の積分平均値vのフレームごとの値AIEMG(p)が、現フレーム(p番目のフレーム)までの最小値ε(i=1,2,…,6)であったか否かをそれぞれ判別する。イエス(Y)の場合はST105に移り、ノー(N)の場合はST106に移る。
 ST105において、判別された積分平均値(v~v)の最小値(ε~ε)をそれぞれ記憶する。そして、ST106に移る。
 ST106において、式(4)に示すように、6種類の積分平均値vのフレームごとの値AIEMG(p)から誤差としての6種類の最小値εをそれぞれ減算することにより、誤差除去値AIEMGε(p)(i=1,2,…,6,p=1,2,…)を演算する。そして、ST107に移る。
 ST107において、式(7-1)に示すように、演算された6種類の誤差除去値AIEMGεi(p)が、現フレーム(p番目のフレーム)までの最大値AIEMGεimaxであったか否かをそれぞれ判別する。イエス(Y)の場合はST108に移り、ノー(N)の場合はST109に移る。
 ST108において、判別された誤差除去値AIEMGεi(p)の最大値AIEMGεimaxをそれぞれ記憶する。そして、ST109に移る。
 ST109において、式(5)に示すように、6種類の誤差除去値AIEMGεi(p)の合計値AIEMGall(p)(p=1,2,…)を演算し、次のフレームの差分値AIEMGdif(p+1)を演算するための1フレーム前の合計値AIEMGall(p)として用いるために一時記憶する。そして、ST110に移る。
 ST110において、式(7-2)に示すように、演算された合計値AIEMGall(p)が、現フレーム(p番目のフレーム)までの最大値AIEMGamaxであったか否かを判別する。イエス(Y)の場合はST111に移り、ノー(N)の場合はST112に移る。
 ST111において、判別された合計値AIEMGall(p)の最大値AIEMGamaxを記憶する。そして、ST112に移る。
 ST112において、式(6)に示すように、合計値AIEMGall(p)のフレームごとの差分値AIEMGdif(p)を演算する。そして、ST113に移る。
 ST113において、式(7-3)に示すように、演算された差分値AIEMGdif(p)が、現フレーム(p番目のフレーム)までの最大値AIEMGdmaxであったか否かを判別する。イエス(Y)の場合はST114に移り、ノー(N)の場合はST115に移る。
 ST114において、判別された合計値差分値AIEMGdif(p)の最大値AIEMGdmaxを記憶する。そして、ST115に移る。
 ST115において、以下の(1),(2)の処理を実行し、ST101に戻る。
(1)式(8-1)~(8-3)に示す正規化により、8種類の入力値N(p)(i=1,2,…,6),N(p),N(p)を演算する(生成する)。
(2)式(9-1)~(9-3)に示すように、8種類の入力値N(p),N(p),N(p)を2値化して記憶することにより、240(8×30=240)個の素子に-1,+1のいずれかを入力する(240個の値(-1,+1)を記憶する)。そして、ST101に戻る。
(実施例1の作用)
 前記構成を備えた実施例1の前記動作識別システムSは、被験者の右腕に6個の筋電センサSN1~SN6を貼り付けた後(図1参照)、動作識別開始画像1の識別開始釦1a(図6参照)を選択することにより、被験者の7種類の動作(図2参照)についての識別を開始する(図9のST2参照)。
 また、実施例1の前記動作識別処理(図9参照)では、6個の筋電センサSN1~SN6からの6種類の表面筋電位信号が、シフト時間ごとに選択的不感化ニューラルネットN(図5参照)の入力値N(p)(i=1,2,…,8),N(p),N(p)に変換され、前記選択的不感化ニューラルネットNの入力層Naに入力される(図9のST4、図10のST101~ST115参照)。この結果、前記選択的不感化ニューラルネットNの入力層Naでは、8種類の入力素子群G (i=1,2,…,8)の30個の素子、合計240個の素子に-1,+1の2値のいずれかがシフト時間ごとに入力される。
 また、前記選択的不感化ニューラルネットNの中間層Nbでは、式(13)′に示す二重不感化により、42種類の中間素子群Gμ,ν (μ≠ν,μ=1,2,…,7,ν=1,2,…,7)の30個の素子、合計1260個の素子に-1,0,+1の3値のいずれかが入力される(図9のST5参照)。また、前記選択的不感化ニューラルネットNの出力層Ncでは、式(16)により、6種類の出力素子群Gγ (γ=1,2,…6)の8個の素子、合計で48個の素子に0,+1の2値のいずれかが入力される(図9のST6参照)。そして、3個以上の素子が識別フレーム数Ps以上継続して発火した出力素子群Gγ が存在する場合には被験者の動作の種類(「手首屈曲」、「手首伸展」、「手を握る」、「手を開く」、「手首回外」、「手首回内」)および速度(「速い」、「普通」、「遅い」)を識別すると共に、存在しない場合には被験者の上腕が「基本姿勢」であると識別する(図9のST7~ST12参照)。この結果、識別された動作の開始時期・種類・速度が動作識別画像2の表示部2a(図8参照)に表示される(図9のST11参照)。
 よって、実施例1の前記動作識別システムSは、同一動作を常に一定速度で実行させた場合の識別しか行っていない前記非特許文献1、2等の技術に比べ、同一動作を異なる速度で実行させた場合の識別をすることができる。
 なお、実施例1の前記動作識別システムSは、学習能力(汎化能力)が高い選択的不感化ニューラルネットN(例えば、特開2009-64216号公報等参照)により、前記被験者の動作を識別している。このため、統計的手法に基づいて動作識別を実行する前記非特許文献1、2等の技術に比べ、少ない学習量で被検体の動作を識別することが期待できる。
 また、実施例1の前記選択的不感化ニューラルネットNには、実施例1の前記入力値生成処理により、6種類の筋電センサSN1~SN6からの6種類の表面筋電位信号に基づいてノイズ等が除去された8種類の入力値N(p),N(p),N(p)が入力される。なお、前記入力値生成処理は、計算量が少なく、シフト時間(10[ms])ごとに入力値N(p),N(p),N(p)を十分に生成できるため、遅延なく前記動作識別処理を実行できる。
 ここで、6chの筋電センサSN1~SN6に応じた6種類の入力値N(p)については、0から1の範囲の値に正規化されているため、異なる種類や異なる速度の動作を実行した場合でも、複数の入力値N(p)が同じ値となる場合がある。例えば、1ch~4chの入力値N(p)~N(p)について、N(p)=N(p)=N(p)=N(p)が成立する可能性がある。この場合、同じ値となる2種類の入力値N(p)どうしで相互修飾(互いに選択的不感化)しても、対応する12種類の中間素子群Gμ,ν (μ=1,2,3,4、ν=1,2,3,4,μ≠ν)で同じ中間値が複数得られる結果となる可能性が高くなる。すなわち、異なる入力値に対して同じ3値パターン(例えば、(+1,0,0,-1,+1,…,0))となる可能性が高くなる。
 そして、異なる入力値に対して同じ3値パターンが多数得られた場合、出力値を演算するための結合荷重の値ωγij μ,νも学習する度に同じ値に近付いてゆくことになる(式(20-1),(20-2)参照)。このような状況を、「1対多対応による結合荷重の平均化」といい、学習効率や識別精度を大きく低下させる原因となる。
 これに対して、前記合計値N(p)は、同じ値となる入力値(1ch~4chの入力値N(p)~N(p))だけでなく、異なる値となる入力値(5ch,6chの入力値N(p)~N(p))も含む6種類の入力値N(p)~N(p)の合計値を正規化した値である。このため、前記合計値N(p)は、同じ値となる入力値(1ch~4chの入力値N(p)~N(p))とは異なる値となる可能性が高くなる。
 すなわち、実施例1では、前記6種類の入力値N(p)と値が異なる可能性が高い前記合計値N(p)を入力値として採用している。
 また、前記合計値N(p)については、他の7種類の入力値N(p),N(p)の二重不感化のためにのみ使用されている。よって、実施例1の前記動作識別システムSは、選択的不感化が1回のみ実行(一重不感化)される選択的不感化ニューラルネットに比べ、さらなる学習量の低減や識別精度の向上が期待できる。なお、実施例1の前記動作識別システムSように、二重不感化を行った場合には、二重不感化を行わない場合に比べ、識別精度が2割程度向上することが実験等により確認されている。
(実験例)
 ここで、実施例1の前記動作識別システムSが、前記入力値生成処理および前記動作識別処理を実行することにより、少ない学習量で、高い識別率を得られるかを調べるために、次の実験例1を準備した。
(実験例1)
 実験例1の動作識別システムSは、実施例1の前記動作識別システムSと同等の構成で作製されており、実施例1の前記入力値生成処理(図10参照)および前記動作識別処理(図9参照)が実行される。なお、前記60480種類の結合荷重の値ωγij μ,νおよび前記48種類の閾値h γの初期値については、前記被験者の2秒間の6動作を、「速い」、「普通」、「遅い」の3速度で、それぞれ2回(2セット)実行して設定した。すなわち、36(6×3×2)個のサンプルデータで学習した。
 また、実験例1の動作識別システムSでは、実際の動作識別として、3速度の6動作、すなわち、全18動作を、6回(6セット)実行した際の動作を識別した。
 なお、実験例1では、表面筋電位測定装置U1として、(有限会社)追坂電子機器製の「Personal-EMG」を使用した。
 また、実験例1では、各セットの実行後には、結合荷重学習処理(図9のST14参照)を実行せず、前記初期値のまま被験者の動作を識別した。
 また、実験例1では、3フレーム連続で3個以上の素子が発火した出力素子群Gγ (i=1,2,…,6)を検出した場合に、前記出力素子群Gγ に対応する動作の種類・速度を識別すると共に、動作識別後も12フレーム連続で前記出力素子群Gγ の識別値dγ(pγ)(γ=1,2,…,6,pγ=1,2,…,Ps)の演算を継続して、前記識別値dγ(pγ)の全15フレーム分の平均値dγをとって平滑化した。すなわち、実験例1では、動作を識別してから15フレームまでは、0以外の識別値dγ(pγ)が得られたフレームだけで平均値dγを演算した。
 さらに、実験例1では、動作識別後、10フレーム連続で0が出力された場合に、識別された動作が終了したと判別させている。
(動作の種類の識別率DR(m)について)
 ここで、動作の種類の識別率については、6種類の動作のいずれかを実際に実行した際に、実際に実行した動作を正しく検出したフレームの総数を、6動作のいずれかを検出したフレームの総数で除算した値とする。例えば、「手首屈曲」の識別率は、「手首屈曲」を実際に実行した際に、前記第1出力素子群G の3個以上の素子が発火したフレームの総数を、「手首屈曲」を実行する2秒間で3個以上の素子が発火した出力素子群Gγ (i=1,2,…,6)を検出したフレームの総数で除算した値とする。
 このため、特定の動作を実行する2秒間において、前記特定の動作を正しく検出したフレームの総数をf(m=1,2,…,6)とし、前記特定の動作以外の動作のいずれかを検出したフレームの総数をfm″(m≠m″,m″=1,2,…,6)とし、前記特定の動作の識別率をDR(m)(m=1,2,…,6)とした場合に、前記特定の動作の識別率DR(m)[%]は、以下の式(21)によって示すことができる。
DR(m)=f/(f+fm″)×100  …式(21)
(動作の速度の識別率DRsp(m)について)
 また、動作の速度の識別率については、18動作のいずれかを実際に実行した際に、3個以上の素子が発火した出力素子群Gγ (i=1,2,…,6)の識別値dγ(pγ)(γ=1,2,…,6,pγ=1,2,…,Ps)が、前記実際の動作に応じた識別値、すなわち、前記対応関係情報Tである5パターンに応じた識別値、すなわち、(2.5),(3.5),(4.5),(5.5),(6.5)の5値のいずれかに、どの程度近似しているかを示す値の識別フレーム数Ps全体の平均値とする。具体的には、前記識別値dγ(pγ)が、実際の18動作に応じた識別値から4以上離れたときの値を0[%]とし、完全に一致したときの値を100[%]とした値の識別フレーム数Ps全体の平均値である。例えば、「手首屈曲」を「速い」で実際に実行した際の前記対応関係情報Tに基づく識別値が2.5であるため、「手首屈曲」の動作識別率が前記第1出力素子群G の識別値d(1)~d(Ps)が全て3.3ならば(d(1)=3.3)、識別率は80%である({1-(2.5-3.3)/4}×100=80[%])。
 このため、特定の動作を3速度のいずれかによって実行する2秒間において、前記特定の動作に応じたフレームごとの識別値(2.5から6.5の5値のいずれかの正解値)をpdγ(pγ)(γ=1,2,…,6,pγ=1,2,…,Ps)とし、前記特定の動作の速度の識別率をDRsp(m)(m=1,2,…,6)とした場合に、前記特定の動作の速度の識別率DRsp(m)[%]は、以下の式(22)によって示すことができる。DRsp(m)
=[{4-(pdγ(1)-dγ(1))}/4×100
  +{4-(pdγ(2)-dγ(2))}/4×100
  +…
  +{4-(pdγ(Ps)-dγ(Ps))}/4×100
 ]/(f+fm″)  …式(22)
(実験結果)
 ここで、実験例1の実験結果について以下に示す。
 図11は実験例1の実験結果の説明図であり、縦軸を識別値とし横軸を時間[s]として全18動作の1セットのうちの「動作A」から「動作F」の6動作を被験者が2秒間隔で実際に実行した際の識別値のグラフであり、図11Aは実際の6動作の識別値の正解値のグラフであり、図11Bは実際の6動作のフレームごとの識別値の実測値の平滑化前のグラフであり、図11Cは実際の6動作のフレームごとの識別値の実測値の平滑化後のグラフである。
 まず、実験例1の動作の種類の識別、すなわち、グラフの位置・横幅については、図11に示すように、正解値のグラフ(図11A参照)と、実測値のグラフ(図11B、図11C参照)とが、ほぼ識別できていることがわかる。
 また、図11Bに示すように、フレーム単位では、「動作A」が、「動作C」の実施直後に誤って識別されていると共に(図11Bの太線参照)、「動作C」が、「動作A」から「動作B」までの間と「動作B」の実施直後と「動作D」の実施直後とに誤って識別されたことが散見できる(図11Bの点線参照)。しかしながら、図11Cに示すように、誤って識別された「動作A」および「動作C」については、15フレーム以下、すなわち150[ms](10×15=150[ms])以下の瞬間的なものであるため、15フレームで平滑化した後には、消滅していることがわかる。
 また、実験例1の動作の速度の識別、すなわち、グラフの高さにについても、図11に示すように、「動作B」以外の5動作については、ほぼ識別できていることがわかる。
 図12は実験例1の実験結果の説明図であり、縦軸を識別率[%]とし横軸を動作の種類として全18動作の6セットを実行した際の6種類の動作の識別率[%]および各動作の3速度の識別率[%]の折れ線グラフである。
 また、図12に示すように、全18動作の6セットを実行した際の実験例1の6種類の動作の識別率DR(m)は、「動作C」が約90[%]であるが(DR(3)≒90[%])、「動作C」以外の5動作(「動作A」、「動作B」、「動作D」~「動作F」)がほぼ100[%]であることがわかる(DR(m)≒100[%](m=1,2,4,5,6))。なお、6種類の動作の識別率DR(m)の平均値は、98.42[%]であった({DR(1)+…+DR(6)}/6=98.42[%])。
 ここで、前記識別率DR(m)の平均値(98.42[%])は、6種類の動作を100回実行した場合に、1、2回誤って識別することを示す値ではなく、動作識別処理を開始してから終了するまでの殆どの時間、正しい動作を識別するが、瞬間的に誤って識別することを示す値である(図11参照)。このため、動作の種類の識別については、実用に十分耐え得るものであることがわかる。
 また、6種類の動作の3速度の識別率DRsp(m)は、「動作C」については約70[%]、「動作F」については約60[%]であるが(DRsp(3)≒70[%],DRsp(6)≒60[%])、「動作C」および「動作F」以外の4動作(「動作A」、「動作B」、「動作D」、「動作E」)が約80[%]であることがわかる(DRsp(m)≒80[%](m=1,2,4,5))。なお、6種類の動作の3速度の識別率DRsp(m)の平均値は、76.15[%]であった({DRsp(1)+…+DRsp(6)}/6=76.15[%])。
 よって、6種類の動作の3速度の識別率DRsp(m)の平均値(76.15[%])は、6種類の動作の識別率DR(m)の平均値(98.42[%])に比べて、20[%]以上低くなったことがわかる。これは、6種類の動作の3速度の識別率DRsp(m)が、前記識別値の実測値(dγ(1)~dγ(15))と理想値(pdγ(1)~pdγ(15))との近似の度合いを示す値であるため(式(22)参照)、出力素子群Gγ の8個の素子の値に基づく前記実測値(dγ(1)~dγ(15))が、前記理想値(pdγ(1)~pdγ(15))に完全に一致する値とならなかったことが原因と考えられる。
 ここで、実験例1では、前記識別率DRsp(m)を演算する前に、前記識別値(dγ(1)~dγ(15))の全15フレーム分の平均値dγを演算して平滑化する。このため、平滑化された前記実測値(平均値dγ)が前記理想値(pdγ(1)~pdγ(15))から離れた値になり易くなる。
 すなわち、一部の前記実測値(dγ(1)~dγ(15))が前記理想値(pdγ(1)~pdγ(15))と完全一致していても、残りの前記実測値(dγ(1)~dγ(15))が前記理想値(pdγ(1)~pdγ(15))から大きく離れた値であれば、平均値dγも前記理想値(pdγ(1)~pdγ(15))から離れた値となる。このため、前記平滑化処理が、前記実測値(dγ(1)~dγ(15))を前記理想値(pdγ(1)~pdγ(15))に完全一致させられなかった一因となり、前記識別率DRsp(m)の平均値(76.15[%])が20[%]以上低くなったと考えられる。
 図13は実験例1の実験結果の説明図であり、縦軸を動作の速度の識別率[%]とし横軸を動作の種類として全18動作の6セットを実行した際の6種類の動作の3速度の速度別の識別率[%]の折れ線グラフである。
 また、図13に示すように、6種類の動作の3速度の速度別の識別率[%]は、「動作F(速い)」が約40[%]であるが、「動作F(速い)」以外の17動作が約60[%]から約90[%]の範囲内であることがわかる。なお、6種類の動作の3速度の速度別の識別率DRsp(m)の平均値は、「速い」が74.12[%]、「普通」が78.30[%]、「遅い」が76.05[%]であった。
 図14は実験例1の実験結果の説明図であり、縦軸を電圧[mV]とし横軸を時間[s]として6種類の動作を実際に順次実行した際の6chの筋電センサからの表面筋電位のグラフの一例である。
 ここで、図14に示すように、「動作F」を実行した際の6chの筋電センサからの表面筋電位は、3chの最大値が0.5[mV]となる以外に目立つ値がなく、他の5動作を実行した際の6chの筋電センサからの表面筋電位に比べ、相対的に電圧値[mV]が小さいことがわかる。このため、「動作F(速い)」を実行した際の入力値のフレームごとの変化が小さく、識別率[%]が約40[%]と極端に低くなったものと考えられる。
 このため、識別率[%]を向上させるためには、「動作F」を実行した際に表面筋電位が変化する他の測定位置に新たな筋電センサを追加することが効果的であると考えられる。
 図15は実験例1の実験結果の説明図であり、動作の立ち上がりからの推定遅れを時間の経過に伴う識別率の推移によって示す説明図であり、動作開始からの経過時間による6種類の動作の3速度の識別率の推移をセット別で示した表の説明図である。
 図16は縦軸を識別率[%]とし横軸を時間[ms]として図15の表の値に基づく動作開始からの経過時間による6種類の動作の3速度の識別率の推移を示すグラフである。
 また、図15、図16に示すように、6セットを実行した際の実験例1の6種類の動作の3速度の識別率DRsp(m)は、各動作の立ち上がり、すなわち、各動作に伴う表面筋電位信号の受信開始から10[ms]後、30[ms]後、50[ms]後、100[ms]後、150[ms]後の各タイミングを経過する度に、高くなっていくことがわかる。すなわち、前記識別率DRsp(m)は、1,3,5,10,15フレーム分のシフト時間が経過する度に、高くなっていくことがわかる。
 なお、6種類の動作の3速度の各タイミングごとの識別率DRsp(m)の平均値は、各動作の開始から10[ms]後が11.19[%]、30[ms]後が26.31[%]、50[ms]後が38.16[%]、100[ms]後が58.16[%]、150[ms]後が63.40[%]であった。
 よって、100[ms]後から150[ms]後にかけて、識別率DRsp(m)の平均値の変化が緩やかになっており、識別率DRsp(m)の平均値の上昇が収束し始めていることがわかる(図14参照)。なお、上限値となる76.15[%]に対する達成度[%]は、各動作の開始から100[ms]後が76.37[%]((58.16)/(76.15)×100≒76.37[%])、150[ms]後が83.26[%]であった((63.40)/(76.15)×100≒83.26[%])。このため、各動作の開始から100[ms]後には、6種類の動作の3速度の識別をほぼ達成していることがわかる。
 ここで、図15に示す表において、DRsp(m)=0[%]以外の動作は、動作の種類の識別率が100[%]である。このため、6種類の動作の各タイミングごとの識別率DR(m)についても演算することができ、各動作の開始から10[ms]後が17.46[%]、30[ms]後が39.68[%]、50[ms]後が53.97[%]、100[ms]後が80.16[%]、150[ms]後が88.09[%]となる。なお、上限値となる98.42[%]に対する達成度[%]は、各動作の開始から100[ms]後が81.45[%]((80.16)/(98.42)×100≒81.45[%])、150[ms]後が89.42[%]であった((88.09)/(98.42)×100≒89.42[%])。
 よって、実験例1では、動作開始から100[ms]程度で、動作の種類だけの識別ならば8割程度、動作の速度も同時に識別するならば6割程度の識別率で、識別できることがわかる。すなわち、実験例1では、動作開始から100[ms]程度で、6動作の識別ならば8割程度、6動作3速度の18動作の識別ならば6割程度の識別率で、識別できることがわかる。したがって、実験例1の前記動作識別システムSは、被験者の36個のサンプルデータ(18動作を2セット)という少ない学習量で、高い識別率が得られたことがわかる。
 なお、前記表面筋電位信号は、前記被験者の実際の動作に伴う筋肉の収縮より30[ms]~100[ms]早く観測される。したがって、実験例1の前記動作識別システムSでは、実際の動作に伴う筋肉の収縮開始とほぼ同時、すなわち、リアルタイムに、6動作の識別ならば8割程度、6動作3速度の18動作の識別ならば6割程度の識別率で、識別できることがわかる。
 また、実験例1では、各動作の立ち上がりから時間が経過する程、識別率が高くなっていくため、あるタイミングの前後で識別結果が変化した場合、変化前の識別結果より変化後の識別結果のほうが正解である可能性が高いことになる。このため、実験例1の前記動作識別システムSは、例えば、筋肉の収縮前に、18動作を一旦識別した後も動作識別処理を継続し、筋肉の収縮開始直後の識別結果が収縮前の識別結果から変化した場合に、前記収縮前の識別結果が誤りであると判別して、前記収縮開始直後の識別結果に修正をするという方法を採用すれば、リアルタイム性を保持したまま18動作の識別率をさらに向上させることも可能である。
 さらに、実験例1では、前記結合荷重学習処理(図9のST14参照)を実行しておらず、前記60480種類の結合荷重の値ωγij μ,νおよび前記48種類の閾値h γが、初期値のまま被験者の動作を識別している。このため、実施例1のように、前記結合荷重学習処理を実行して、前記初期値を更新してゆけば、前記被験者の18動作を、リアルタイムで、且つ、さらに高い識別率で識別することが可能となる。
 この結果、実施例1の前記動作識別システムSは、識別率を高めるための結合荷重の初期値の学習、いわゆる、初期学習に大量の学習量を予め必要とする前記統計的手法に基づいて動作を識別する非特許文献1、2等の技術に比べ、初期学習に予め必要な学習量(提供するサンプル数)を低減することができる。したがって、実施例1の前記動作識別システムSは、非特許文献1、2等の技術に比べ、前記被験者が交代する度に実行される初期学習(再学習)に予め必要な学習量を低減することができ、被験者に対する事前の負担を低減することができる。
 なお、実施例1では、学習能力(汎化能力)が高い前記選択的不感化ニューラルネットNにより、前記被験者の動作を識別している。また、実施例1の8種類の入力値N(p),N(p),N(p)は、6個の筋電センサSN1~SN6の出力信号(表面筋電位信号)からノイズ(誤差、個人差、個体差)が除去されて演算されている(図10のST104~ST106、式(4)等参照)。よって、実施例1の前記動作識別システムSは、万人に適合する結合荷重を設定し易くなっている。このため、実施例1の前記動作識別システムSは、被験者以外の第三者の動作をサンプルとして初期学習しても、被験者の動作を精度良く識別できる可能性がある。すなわち、前記被験者自身の動作をサンプルとして初期学習しなくても、前記被験者の動作を精度良く識別することが期待できる。
 また、実施例1の前記動作識別システムSは、6動作の識別ならば、98.42[%]以上の高い識別率で識別でき、100[ms]以内というリアルタイム性を要求しても、8割以上の高い識別率を維持することができ、識別時間という点においても実用的である。また、6動作3速度の18動作の識別ならば、追加学習をしなくても、76.15[%]以上の高い識別率を維持することができ、100[ms]以内というリアルタイム性を要求しても、6割以上の識別率で識別できる。
 したがって、実施例1の前記動作識別システムSは、多数の動作や速度を識別しても必要な学習量が急増することがないため、増加した動作の種類や速度の分だけ大量の学習量がさらに必要となる非特許文献1、2等の技術に比べ、少ない学習量で前記被験者の動作を識別することができる。
 さらに、実施例1の前記動作識別システムSは、識別する動作の種類や速度が増加しても、増加した動作の種類や速度の特徴を抽出する複雑な特徴抽出処理等を実行する必要がない。
 したがって、実施例1の前記動作識別システムSは、多数の動作や速度を識別するための前記特徴抽出処理が複雑化してリアルタイム性を損なう可能性ある非特許文献1、2等の技術に比べ、動作の種類や速度が増加してもリアルタイム性を損なわずに被験者の動作を識別することができる。
 これらの結果から、実施例1の前記動作識別システムSは、(1)多数の動作を、(2)速度の推定と共に、(3)高い識別率をもって、(4)少ないサンプル数で、(5)リアルタイムに識別できる。
(変更例)
 以上、本発明の実施例を詳述したが、本発明は、前記実施例に限定されるものではなく、特許請求の範囲に記載された本発明の要旨の範囲内で、種々の変更を行うことが可能である。本発明の変更例(H01)~(H017)を下記に例示する。
(H01)前記実施例では、6個の筋電センサSN1~SN6を使用して(m=6)、選択的不感化ニューラルネットNの入力素子群G および中間素子群Gμ,ν の素子の数をそれぞれ30個とし(n=30)、出力素子群Gγ の素子の数を8個とし(n′=8)として、被験者の6種類の動作の種類を識別したが(m′=6、「基本姿勢」を除く)、m,n,m′,n′の値についてはこれに限定されず、任意の値に変更可能である。例えば、被験者の膝の屈曲・伸展を識別する場合には、2個の筋電センサSN1,SN2を使用して(m=2)、選択的不感化ニューラルネットNの入力素子群G および中間素子群Gμ,ν の素子の数をそれぞれ10個とし(n=10)、出力素子群Gγ の素子の数を6個とし(n′=6)として、被験者の2種類の動作の種類を識別することも可能である(m′=2、「基本姿勢」を除く)。すなわち、筋電センサを5個以下または7個以上使用したり、各層Na~Nbの素子数が任意の数の選択的不感化ニューラルネットNに適用したり、5以下または7以上の被験者の動作を識別したりすることも可能である。なお、筋電センサを増やして、より多数の動作を識別する場合でも、前記選択的不感化ニューラルネットNを拡張して識別すれば、筋電センサの増加に対して、計算量の増加が緩やかになり、汎化能力が高く、冗長性も高いため(冗長変数の追加等への耐性も強いため)、少ない学習量、且つ、高い識別率で各動作を識別することが可能である。
(H02)前記実施例では、前記表面筋電位測定装置U1の制御部が、6個の筋電センサSN1~SN6からの表面筋電位信号の積分表面筋電位信号(s~s)を演算して、前記識別用クライアントパソコンPCのコンピュータ本体H1に送信したが、これに限定されず、識別用クライアントパソコンPCのコンピュータ本体H1の制御部で、6個の筋電センサSN1~SN6からの表面筋電位信号の積分表面筋電位信号(s~s)を演算することも可能である。すなわち、前記識別用クライアントパソコンPCに前記表面筋電位測定装置U1の機能を包含させることも可能であり、例えば、前記表面筋電位測定装置U1と前記識別用クライアントパソコンPCとを一体的に構成した動作識別装置(U1+PC)を作製することも可能である。
(H03)前記実施例では、前記6種類の積分平均値(v~v)を演算するために(式(2-1),(2-2)参照)、1フレームごとのフレーム時間を15[ms]とし、前記シフト時間を10[ms]としたが、これに限定されず、前記フレーム時間および前記シフト時間を任意に設定することが可能である。なお、前記フレーム時間および前記シフト時間を変更した場合には、実験例1で示す動作の種類および速度の識別率DR(m),DRsp(m)(m=1,2,…,6)に影響する可能性がある。また、前記フレーム時間および前記シフト時間を長くし過ぎると、識別結果が実際の動作より遅れてしまう可能性がある。このため、前記識別結果を利用する処理の用途等を考慮して、各用途の処理で要求される識別率やリアルタイム性に応じて前記フレーム時間および前記シフト時間を設定することが好ましい。
(H04)前記実施例では、前記対応関係情報Tの5パターン(式(19-1)~(19-5)参照)に基づいて、前記出力素子群Gγ (γ=1,2,…,6)の3個以上の素子が識別フレーム数Ps以上継続して発火した場合に、対応する動作を識別したが、動作の識別方法はこれに限定されず、例えば、前記出力素子群Gγ が1フレームでも前記5パターンに一致した場合、すなわち、4個の素子が連続して(隣接して)発火した前記出力素子群Gγ を検出した場合に、動作識別することも可能である。また、例えば、前記出力素子群Gγ の3個以上の素子が識別フレーム数Ps以上継続して発火した場合でも、少なくとも2個の素子が連続して発火した場合のみ動作識別したり、発火した素子が全て連続している場合のみ動作識別したりすることも可能である。
(H05)前記実施例の動作識別システムSでは、複数の動作を同時実行した場合、いわゆる、複合動作を実行した場合には、前記複合動作を1度も学習させていないにも関わらず、どの動作とどの動作の組合せであるかを識別できる可能性がある。例えば、手首を屈曲しながら手を握った場合に、「手首屈曲」と「手を握る」組合せであると識別できる可能性があり、前記複合動作を学習させれば、より高い識別率で前記複合動作を識別させることも期待できる。このため、前記実施例では、3個以上の素子が識別フレーム数Ps以上継続した前記出力素子群Gγ が複数存在する場合には、対応する複数の動作を同時実行したと識別させているが、これに限定されず、例えば、対応する複数の動作のどちらを実行したかが不明であるとして、「識別不可」と識別したり、識別せずに「基本姿勢」の状態を維持したりすることも可能である。
(H06)前記実験例1では、18動作(6動作3速度)、2セットの36個のサンプルデータを初期学習するという条件で動作の種類および速度の識別率DR(m),DRsp(m)(m=1,2,…,6)等を測定したが、初期学習するサンプルデータ数はこれに限定されず、例えば、6動作のみの識別であれば、35個以下のサンプルデータを初期学習したり、初期状態から18動作をより高い識別率でしたい場合やより多くの動作や速度を識別したい場合には、37個以上のサンプルデータを初期学習することも可能である。この場合、実用性を考慮して、被験者に対する事前の負担が大きくならない程度にサンプルデータ数を設定することが好ましい。
(H07)前記実施例では、識別された前記被験者の動作の開始時期、種類および速度を、前記動作識別画像2の表示部2a(図8参照)に記載して、前記識別用クライアントパソコンPCのディスプレイH2にリアルタイムで表示したが、これに限定されず、例えば、識別された前記被験者の動作の種類および速度に基づいて、義手・義肢等で識別動作・識別速度をリアルタイムに再現(例えば、義手ならば手を握る等、義肢ならば膝の曲げ伸ばし等)したり、ロボットスーツを着用する着用者(装着者、作業者、作業員)が次に行う動作を識別し、識別した動作を補助するパワーアシスト用のモータ等を作動させるために利用することも可能である。具体的には、識別した動作を行うための特定の関節にトルクを付与する特定の前記モータを判別する処理、すなわち、ロボットスーツが有する複数のモータについて、駆動対象となるモータを特定する処理に利用したり、ロボットスーツで識別動作を識別速度より若干高速(例えば、10[%]~40[%]程度高速)でリアルタイムに再現(例えば、重い荷物を持ち上げる動作等)して装着者(作業者)の識別動作のパワーアシストをしたりすることも可能である。また、例えば、識別された前記被験者の動作の種類および速度に応じて、バーチャルリアリティ装置としての前記識別用クライアントパソコンPCのディスプレイH2上の画像をリアルタイムに更新する等の制御を実行することも可能である。
(H08)前記実施例では、前記被験者の6動作3速度の18動作(「基本姿勢」を含めると19動作)を識別したが、識別するのは動作の種類および速度に限定されず、例えば、前記出力素子群Gγ に動作の角度を識別するための素子を別途追加することにより(例えば、前記出力素子群Gγ に6個の素子を追加して12個の素子で構成することにより)、動作の角度についても識別することを期待できる。また、例えば、前記出力素子群Gγ に動作の加速度を識別するための素子も同様に別途追加するにより、動作の加速度についても識別することを期待できる。この場合、動作の種類(種類情報)、速度(速度情報)だけでなく、角度(角度情報)や加速度(加速度情報)についても同時に識別することが期待できる。
(H09)前記実施例では、対応関係情報Tの5パターンを、種類の筋電センサSN1~SN6からの6種類の表面筋電位信号と6動作3速度の18動作とに基づいて設定しているが(式(17),(18),(19-1)~(19-5),(20-1),(20-2)等参照)、対応関係情報Tの設定方法についてはこれに限定されず、例えば、各動作の速度情報としての速度センサや加速度センサの信号(時間経過により速度・加速度の値が変化する波形)に基づくアナログ値に基づいてより正確に設定することも可能である。この場合、表面筋電位信号と18動作とに基づいて対応関係情報Tの5パターンを設定する場合に比べ、初期学習等の学習効率が向上することが期待でき、識別率をさらに向上させることが期待できる。
(H010)前記実施例では、6種類の筋電センサSN1~SN6からの6種類の表面筋電位信号に基づく8種類の入力値N(p),N(p),N(p)を、選択的不感化ニューラルネットNに入力して、被験者の6動作3速度の18動作を識別したが、入力値については、表面筋電位信号に基づく値に限定されず、例えば、速度センサや加速度センサの信号に基づく入力値も、選択的不感化ニューラルネットNに入力することも可能である。この場合、表面筋電位信号に基づく入力値N(p),N(p),N(p)のみで動作を識別する場合に比べ、識別率をさらに向上させることができる可能性がある。
(H011)前記実施例では、識別された動作に応じた出力素子群Gγ のフレームごとの識別値dγ(pγ)の平均値dγ(γ=1,2,…,6,dγ={dγ(1)+dγ(2)+…+dγ(Ps)}/Ps)が、2.5≦dγ≦3.5となる場合に「速い」と識別し、3.5<dγ<5.5の場合に「普通」と識別し、5.5≦dγ≦6.5の場合に「遅い」と識別したが、これに限定されず、「速い」と「普通」との識別範囲を分ける境界値(3.5)や、「普通」と「遅い」との識別範囲を分ける境界値(5.5)については任意の値を設定できる。また、動作の速度の識別方法については、「速い」、「普通」、「遅い」の3段階で識別することに限定されず、例えば、2段階や4段階以上で識別することも可能である。例えば、2.5≦dγ<3.5の場合を「高速」、3.5≦dγ<4.5の場合を「やや高速(中高速)」、4.5≦dγ<5.5の場合を「中速」、5.5≦dγ≦6.5の場合を「低速」と識別することにより、4段階で識別することも可能である。また、例えば、前記平均値dγ自体を速度を示す数値として識別したり、前記平均値dγに応じて演算された速度[m/s]を識別することも可能である。
(H012)前記実施例では、式(5)に示す6種類の誤差除去値AIEMGεi(p),(p=1,2,…)の合計値AIEMGall(p)に基づく入力値N(p)を、多重不感化のための合計用入力素子群G の素子の値x ~x30 に入力したが、これに限定されず、例えば、6種類の誤差除去値AIEMGεi(p)のノルム(2乗和のルート)に基づく入力値を入力することも可能である。なお、合計値AIEMGall(p)に替わるノルムAIEMGall(p)′については、以下の式(5)′により示される。
AIEMGall(p)′
={(AIEMGε(p))
  +(AIEMGε(p))
  +…
  +(AIEMGε(p))1/2  …式(5)′
また、例えば、予め設定された2種類の誤差除去値AIEMGεi(p)の差分値に基づく入力値を入力することも可能である。さらに、第1の誤差除去値AIEMGε1(p)を残りの誤差除去値AIEMGεi(p)(i=2,3,…,6)で減算した差分値に基づく入力値を入力することも可能である。なお、前記差分値をAIEMGdif(p)′とした場合、合計値AIEMGall(p)に替わる差分値AIEMGdif(p)′については、以下の式(5)″により示される。
AIEMGdif(p)′=AIEMGε(p)
            -AIEMGε(p)
            -…
            -AIEMGε(p)  …式(5)″
(H013)前記実施例では、生体情報の一例としての表面筋電位信号を測定して被験者の動作を識別したが、これに限定されず、例えば、生体情報の一例としての脳波等を測定して被験者の動作を識別することも可能である。
(H014)前記実施例では、多重不感化のための素子群(多重用素子群)を、合計用入力素子群G としたが、これに限定されず、例えば、その他の各入力素子群G ~G や各中間素子群G1,2 ~G7,6 のいずれか1つまたは複数とすることも可能である。なお、複数の素子群G ~G ,G1,2 ~G7,6 を多重用素子群として選択した場合、多重用素子群が複数となり、前記実施例1のような二重不感化でなく、三重不感化以上の多重不感化となる。例えば、二重不感化のための多重用素子群を合計用入力素子群G とし、三重不感化のための多重用素子群を差分用入力素子群G とした場合、入力素子群Gνの素子の値(x μ)が二重不感化された値{(x μ)×(g μ,ν)×(g μ,ν,8)}と、三重不感化のための差分用出力感度の値(gk′ μ,ν,7),(k′=σ″(i)、すなわち、k′は乱数)との積{(x μ)×(g μ,ν)×(g μ,ν,8)×(gk′ μ,ν,7)}を演算することにより、三重不感化が行われ、四重不感化以上の多重不感化も同様に行うことで可能となる。
(H015)前記実施例では、入力素子群Gνの素子の値(x μ)と出力感度の値(g μ,ν)との積{(x μ)×(g μ,ν)}を演算して選択的不感化をした後、合計用出力感度の値(g μ,ν,8)との積{(x μ)×(g μ,ν)×(g μ,ν,8)}を演算して二重不感化したが、式(13)に示すように、演算の順番についてはこれに限定されず、例えば、入力素子群Gνの素子の値(x μ)と合計用出力感度の値(g μ,ν,8)との積{(x μ)×(g μ,ν,8)}を演算して選択的不感化をした後、出力感度の値(g μ,ν)との積{(x μ)×(g μ,ν,8)×(g μ,ν)}を演算して二重不感化をしたり、入力素子群Gνの素子の値(x μ)と、出力感度の値(g μ,ν)と合計用出力感度の値(g μ,ν,8)との積{(g μ,ν)×(g μ,ν,8)}と、の積{(x μ)×(g μ,ν)×(g μ,ν,8)}を演算して二重不感化をしたり、入力素子群Gνの素子の値(x μ)と出力感度の値(g μ,ν)と合計用出力感度の値(g μ,ν,8)との積{(x μ)×(g μ,ν)×(g μ,ν,8)}を同時に演算して二重不感化をしたりすることも可能である。
(H016)前記実施例の動作識別システムSでは、表面筋電位信号を測定して被験者の動作を識別したが、これに限定されず、例えば、障害者の失った手足の動作の種類と速度とを推定することも可能である。具体的には、手を失った被験者に対し、実施例1と同様に、前記被験者の前腕の前記6箇所の測定位置に前記6個の筋電センサSN1~SN6を貼り付けて表面筋電位信号を測定することにより、表面筋電位信号は前記被験者の動作意思に応じて変化するため、被験者が失った手で行いたかった動作の種類と速度とを推定することが可能となり、失った手による実際の動作がない場合でも被験者の動作意思を推定することが可能となる。
(H017)前記実施例の動作識別システムSでは、測定された表面筋電位信号から被験者の動作を識別したが、本願発明の分類推定システムが推定可能な分類情報についてはこれに限定されず、例えば、撮像された映像信号から人間・動物の動作を推定したり、測定された脳波信号から静止した人間・動物・植物の意思を推定したり、測定された温湿度から非生物の物体の状態を推定したり、測定された温湿度等から天候や自然現象を推定したり、収集された立地条件等の情報から不動産価格を推定したり、収集された各国の経済指標・金融政策等の情報から為替や株価等を推定したりすることも可能である。すなわち、抽出された情報からあらゆる事象を推定することが可能である。具体的には、実施例1と同様に、時間ごとに値が変化する波形信号としての前記各情報(映像信号、脳波信号、温度・湿度、立地条件、経済指標・金融政策等)の積分値(s)の積分平均値(v)をシフト時間ごとに演算し、前記積分平均値(v)の合計値と前記合計値のフレームごとの差分値とを演算して(式(5)、式(6)参照)、積分平均値・差分値・合計値に基づく入力値(N(p),N(p),N(p))を演算することにより(式(7-1)~式(8-3)参照)、表面筋電位信号と同様に、多重不感化が可能な前記選択的不感化ニューラルネットN(図5参照)に入力値(N(p),N(p),N(p))を入力することができ、あらゆる事象(動作、意思、状態、天候や自然現象、不動産価格、為替や株価等)を推定することが可能となる。
 本発明の前記動作識別システムSおよび前記動作識別プログラムAP1は、例えば、障害者の失った手足による日常の動作を補うために、義手・義肢等で識別動作・識別速度をリアルタイムに再現する場合や、筋肉への負担が大きい作業員(作業者、肉体労働者)等の支援のために、ロボットスーツを着用する前記作業員が次に行う動作を識別し、識別した動作を補助するパワーアシスト用のモータ等を作動させるために利用する場合や、体感型ゲームマシン等のバーチャルリアリティ装置やその他のヒューマンインタフェース装置等の制御にも利用する場合に有用である。

Claims (10)

  1.  第1の情報に基づく第1入力変数の値が入力される複数の入力素子により構成された第1入力素子群と、第2の情報に基づく第2入力変数の値が入力される複数の入力素子により構成された第2入力素子群と、を有する入力素子群と、
     前記第1入力素子群の各入力素子の値と、前記第2入力素子群の各入力素子の値に基づいて一部の値がゼロになるように演算された各第1出力感度の値と、予め設定された多重用素子群の各素子の値に基づいて一部の値がゼロになるように演算された各第1多重用出力感度の値と、の積算を少なくとも含む第1の多重不感化がなされて一部の値がゼロにされた前記第1入力素子群の各入力素子の値として演算された第1中間変数の値が入力される複数の中間素子により構成された第1中間素子群、を有する中間素子群と、
     前記第1中間素子群の各中間素子の値と前記各中間素子の値の重視度合いに応じて予め設定された各結合荷重の値とに基づいて演算された出力変数の値が入力される複数の出力素子により構成された出力素子群と、
     分類対象としての複数の分類情報と、分類情報ごとに対応する前記出力変数の値と、の対応関係を特定する対応関係情報を記憶する対応関係情報記憶手段と、
     前記対応関係情報記憶手段に予め記憶された前記対応関係情報と、演算された前記出力変数の値と、に基づいて、入力された前記第1の情報および前記第2の情報に応じた前記分類情報を推定する分類情報推定手段と、
     前記分類情報推定手段によって推定された前記分類情報を表示する分類情報表示手段と、
     を備えたことを特徴とする分類推定システム。
  2.  前記第2入力素子群の各入力素子の値と、前記第1入力素子群の各入力素子の値に基づいて一部の値がゼロになるように演算された各第2出力感度の値と、前記多重用素子群の各素子の値に基づいて一部の値がゼロになるように演算された各第2多重用出力感度の値と、の積算を少なくとも含む第2の多重不感化がなされて一部の値がゼロにされた前記第2入力素子群の各入力素子の値として演算された第2中間変数の値が入力される複数の中間素子により構成された第2中間素子群、を有する前記中間素子群と、
     前記各中間素子群の各中間素子の値と前記各中間素子の値の重視度合いに応じて予め設定された各結合荷重の値とに基づいて演算された前記出力変数の値が入力される前記出力素子群と、
     を備えたことを特徴とする請求項1に記載の分類推定システム。
  3.  前記分類情報推定手段により推定された前記分類情報と、入力された前記第1の情報および前記第2の情報が抽出された実際の分類情報との相違に基づいて、前記各結合荷重の値を更新することにより、前記各結合荷重の値を学習する結合荷重学習手段と、
     を備えたことを特徴とする請求項1または2に記載の分類推定システム。
  4.  前記多重用素子群の各素子の値が、前記第1の情報と、前記第2の情報と、に基づいて演算された
     ことを特徴とする請求項1ないし3のいずれかに記載の分類推定システム。
  5.  前記分類情報を被検体の動作意思とし、前記第1の情報および前記第2の情報を前記動作意思に伴う前記被検体の生体信号とした場合に、前記被検体に配置された第1測定位置における前記生体信号を測定する第1測定部と、前記被検体に配置された前記第1測定位置と異なる第2測定位置における前記生体信号を測定する第2測定部と、を有する測定部材と、
     を備えたことを特徴とする請求項1ないし4のいずれかに記載の分類推定システム。
  6.  前記動作意思に伴う前記被検体の筋群の活動である筋活動に基づく電位の変化の情報である筋電位信号のうちの前記被検体の皮膚における前記筋電位信号を表面筋電位信号とし、前記生体信号を前記表面筋電位信号とした場合に、前記皮膚に配置された前記第1測定位置における前記表面筋電位信号を測定する前記第1測定部と、前記皮膚に配置された前記第2測定位置における前記表面筋電位信号を測定する前記第2測定部と、を有する前記測定部材と、
     前記第1測定部からの前記表面筋電位信号の積分値である第1積分値を演算すると共に、前記第2測定部からの前記表面筋電位信号の積分値である第2積分値を演算する積分値演算手段と、
     予め設定されたフレーム時間内の前記第1積分値の平均値である第1積分平均値を予め設定されたシフト時間ごとに演算すると共に、前記フレーム時間内の前記第2積分値の平均値である第2積分平均値を前記シフト時間ごとに演算する積分平均値演算手段と、
     前記第1の情報を前記第1積分平均値とし、前記第2の情報を前記第2積分平均値とした場合に、前記第1積分平均値に基づく前記第1入力変数の値が前記シフト時間ごとに入力される前記第1入力素子群と、前記第2積分平均値に基づく前記第2入力変数の値が前記シフト時間ごとに入力される前記第2入力素子群と、を有する前記入力素子群と、
     を備えたことを特徴とする請求項5に記載の分類推定システム。
  7.  前記動作意思によって特定される動作の種類の情報である種類情報と、前記動作の力の情報である力情報と、を有する前記分類情報を推定する前記分類情報推定手段、
     を備えたことを特徴とする請求項5または6に記載の分類推定システム。
  8.  前記動作の速度の情報である速度情報であって、前記力情報に基づく前記速度情報、を有する前記分類情報を推定する前記分類情報推定手段、
     を備えたことを特徴とする請求項7に記載の分類推定システム。
  9.  前記出力変数の値のうち、第1の動作意思に応じた第1出力変数の値が入力される複数の出力素子により構成された第1出力素子群と、第2の動作意思に応じた第2出力変数の値が入力される複数の出力素子により構成された第2出力素子群と、を有する前記出力素子群と、
     前記第1の動作意思の前記種類情報および前記速度情報と前記第1出力変数の値との対応関係と、前記第2の動作意思の前記種類情報および前記速度情報と前記第2出力変数の値との対応関係と、を特定する前記対応関係情報を記憶する前記対応関係情報記憶手段と、
     前記対応関係情報記憶手段に記憶された前記対応関係情報と、演算された前記第1出力変数の値と、演算された前記第2出力変数の値と、に基づいて、入力された前記第1の情報および前記第2の情報に応じた前記分類情報の前記種類情報および前記速度情報を推定する前記分類情報推定手段と、
     前記分類情報推定手段によって推定された前記分類情報の前記種類情報および前記速度情報を表示する前記分類情報表示手段と、
     を備えたことを特徴とする請求項8に記載の分類推定システム。
  10.  コンピュータを、
     第1の情報に基づく第1入力変数の値が入力される複数の入力素子により構成された第1入力素子群と、第2の情報に基づく第2入力変数の値が入力される複数の入力素子により構成された第2入力素子群と、を有する入力素子群について、入力された前記第1入力変数の値および前記第2入力変数の値を記憶する入力値記憶手段、
     前記第1入力素子群の各入力素子の値と、前記第2入力素子群の各入力素子の値に基づいて一部の値がゼロになるように演算された各第1出力感度の値と、予め設定された多重用素子群の各素子の値に基づいて一部の値がゼロになるように演算された各第1多重用出力感度の値と、の積算を少なくとも含む第1の多重不感化がなされて一部の値がゼロにされた前記第1入力素子群の各入力素子の値として演算された第1中間変数の値が入力される複数の中間素子により構成された第1中間素子群、を有する中間素子群について、前記第1中間変数の値を演算する中間値演算手段、
     前記第1中間素子群の各中間素子の値と前記各中間素子の値の重視度合いに応じて予め設定された各結合荷重の値とに基づいて演算された出力変数の値が入力される複数の出力素子により構成された出力素子群について、前記出力変数の値を演算する出力値演算手段、
     分類対象としての複数の分類情報と、分類情報ごとに対応する前記出力変数の値と、の対応関係を特定する対応関係情報を記憶する対応関係情報記憶手段、
     前記対応関係情報記憶手段に予め記憶された前記対応関係情報と、演算された前記出力変数の値と、に基づいて、入力された前記第1の情報および前記第2の情報に応じた前記分類情報を推定する分類情報推定手段、
     前記分類情報推定手段によって推定された前記分類情報を表示する分類情報表示手段、
     として機能させることを特徴とする分類推定プログラム。
PCT/JP2010/004585 2009-07-15 2010-07-14 分類推定システムおよび分類推定プログラム WO2011007569A1 (ja)

Priority Applications (2)

Application Number Priority Date Filing Date Title
JP2011522736A JP5545574B2 (ja) 2009-07-15 2010-07-14 分類推定システムおよび分類推定プログラム
US13/350,711 US9392954B2 (en) 2009-07-15 2012-01-13 Classification estimating system and classification estimating program

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
JP2009-167222 2009-07-15
JP2009167222 2009-07-15

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US13/350,711 Continuation US9392954B2 (en) 2009-07-15 2012-01-13 Classification estimating system and classification estimating program

Publications (1)

Publication Number Publication Date
WO2011007569A1 true WO2011007569A1 (ja) 2011-01-20

Family

ID=43449179

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2010/004585 WO2011007569A1 (ja) 2009-07-15 2010-07-14 分類推定システムおよび分類推定プログラム

Country Status (3)

Country Link
US (1) US9392954B2 (ja)
JP (1) JP5545574B2 (ja)
WO (1) WO2011007569A1 (ja)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2016030096A (ja) * 2014-07-29 2016-03-07 株式会社きもちラボ 生体電気ノイズ同定システム及び生体電気ノイズ除去システム
JP2018054765A (ja) * 2016-09-27 2018-04-05 日本電気株式会社 データ処理装置、データ処理方法、およびプログラム
JP2018196603A (ja) * 2017-05-24 2018-12-13 日本電信電話株式会社 行動推定装置、及び行動推定方法
WO2022195657A1 (ja) * 2021-03-15 2022-09-22 日本電信電話株式会社 筋音抽出方法、筋音抽出装置及びプログラム

Families Citing this family (42)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
TW201121335A (en) * 2009-12-02 2011-06-16 Sunplus Core Technology Co Ltd Method and apparatus for adaptively determining compression modes to compress frames
US9907489B2 (en) * 2012-06-14 2018-03-06 Rehabilitation Institute Of Chicago Systems and methods for hierarchical pattern recognition for simultaneous control of multiple-degree of freedom movements for prosthetics
US20140198034A1 (en) 2013-01-14 2014-07-17 Thalmic Labs Inc. Muscle interface device and method for interacting with content displayed on wearable head mounted displays
WO2014186370A1 (en) 2013-05-13 2014-11-20 Thalmic Labs Inc. Systems, articles and methods for wearable electronic devices that accommodate different user forms
US20150057770A1 (en) * 2013-08-23 2015-02-26 Thaimic Labs Inc. Systems, articles, and methods for human-electronics interfaces
US11921471B2 (en) 2013-08-16 2024-03-05 Meta Platforms Technologies, Llc Systems, articles, and methods for wearable devices having secondary power sources in links of a band for providing secondary power in addition to a primary power source
US11426123B2 (en) 2013-08-16 2022-08-30 Meta Platforms Technologies, Llc Systems, articles and methods for signal routing in wearable electronic devices that detect muscle activity of a user using a set of discrete and separately enclosed pod structures
US10188309B2 (en) 2013-11-27 2019-01-29 North Inc. Systems, articles, and methods for electromyography sensors
US10042422B2 (en) 2013-11-12 2018-08-07 Thalmic Labs Inc. Systems, articles, and methods for capacitive electromyography sensors
US20150124566A1 (en) 2013-10-04 2015-05-07 Thalmic Labs Inc. Systems, articles and methods for wearable electronic devices employing contact sensors
US9788789B2 (en) 2013-08-30 2017-10-17 Thalmic Labs Inc. Systems, articles, and methods for stretchable printed circuit boards
US9372535B2 (en) 2013-09-06 2016-06-21 Thalmic Labs Inc. Systems, articles, and methods for electromyography-based human-electronics interfaces
US9483123B2 (en) * 2013-09-23 2016-11-01 Thalmic Labs Inc. Systems, articles, and methods for gesture identification in wearable electromyography devices
DE112013007524T5 (de) * 2013-10-24 2016-08-04 Apple Inc. Armbandgerät-Eingabe mittels Handgelenkbewegung
KR102372544B1 (ko) 2014-02-14 2022-03-10 페이스북 테크놀로지스, 엘엘씨 탄성 전기 케이블 및 이를 사용하는 웨어러블 전자 장치를 위한 시스템, 물품, 및 방법
US10199008B2 (en) 2014-03-27 2019-02-05 North Inc. Systems, devices, and methods for wearable electronic devices as state machines
US9880632B2 (en) 2014-06-19 2018-01-30 Thalmic Labs Inc. Systems, devices, and methods for gesture identification
US9807221B2 (en) 2014-11-28 2017-10-31 Thalmic Labs Inc. Systems, devices, and methods effected in response to establishing and/or terminating a physical communications link
US10078435B2 (en) 2015-04-24 2018-09-18 Thalmic Labs Inc. Systems, methods, and computer program products for interacting with electronically displayed presentation materials
US10945653B2 (en) * 2015-06-12 2021-03-16 Koninklijke Philips N.V. Surface electromyography system, recorder and method
ITUB20159413A1 (it) * 2015-12-23 2017-06-23 Laboratori Archa Srl Metodo e sistema di rilevazione di movimenti
KR102619981B1 (ko) 2016-02-02 2024-01-02 삼성전자주식회사 근전도 신호를 이용한 동작 분류 방법 및 장치
CN110300542A (zh) * 2016-07-25 2019-10-01 开创拉布斯公司 使用可穿戴的自动传感器预测肌肉骨骼位置信息的方法和装置
US11216069B2 (en) 2018-05-08 2022-01-04 Facebook Technologies, Llc Systems and methods for improved speech recognition using neuromuscular information
US20190121306A1 (en) 2017-10-19 2019-04-25 Ctrl-Labs Corporation Systems and methods for identifying biological structures associated with neuromuscular source signals
US11185270B1 (en) * 2017-02-03 2021-11-30 Yongwu Yang Wearable device and method for monitoring muscle tension and other physiological data
WO2018191755A1 (en) * 2017-04-14 2018-10-18 REHABILITATION INSTITUTE OF CHICAGO d/b/a Shirley Ryan AbilityLab Prosthetic virtual reality training interface and related methods
US10481699B2 (en) * 2017-07-27 2019-11-19 Facebook Technologies, Llc Armband for tracking hand motion using electrical impedance measurement
US10984054B2 (en) * 2017-07-27 2021-04-20 Robert Bosch Gmbh Visual analytics system for convolutional neural network based classifiers
US11907423B2 (en) 2019-11-25 2024-02-20 Meta Platforms Technologies, Llc Systems and methods for contextualized interactions with an environment
US11493993B2 (en) 2019-09-04 2022-11-08 Meta Platforms Technologies, Llc Systems, methods, and interfaces for performing inputs based on neuromuscular control
US11150730B1 (en) * 2019-04-30 2021-10-19 Facebook Technologies, Llc Devices, systems, and methods for controlling computing devices via neuromuscular signals of users
US10937414B2 (en) 2018-05-08 2021-03-02 Facebook Technologies, Llc Systems and methods for text input using neuromuscular information
US11481030B2 (en) 2019-03-29 2022-10-25 Meta Platforms Technologies, Llc Methods and apparatus for gesture detection and classification
US11961494B1 (en) 2019-03-29 2024-04-16 Meta Platforms Technologies, Llc Electromagnetic interference reduction in extended reality environments
US10592001B2 (en) 2018-05-08 2020-03-17 Facebook Technologies, Llc Systems and methods for improved speech recognition using neuromuscular information
EP4241661A1 (en) 2018-08-31 2023-09-13 Facebook Technologies, LLC Camera-guided interpretation of neuromuscular signals
CN112789577B (zh) 2018-09-20 2024-04-05 元平台技术有限公司 增强现实系统中的神经肌肉文本输入、书写和绘图
US11797087B2 (en) 2018-11-27 2023-10-24 Meta Platforms Technologies, Llc Methods and apparatus for autocalibration of a wearable electrode sensor system
US11583218B2 (en) * 2019-11-20 2023-02-21 Advancer Technologies, Llc EMG device
US11868531B1 (en) 2021-04-08 2024-01-09 Meta Platforms Technologies, Llc Wearable device providing for thumb-to-finger-based input gestures detected based on neuromuscular signals, and systems and methods of use thereof
CN113222049A (zh) * 2021-05-26 2021-08-06 中国石油天然气股份有限公司 一种特殊螺纹上扣扭矩的分类识别方法及识别设备

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2009064216A (ja) * 2007-09-06 2009-03-26 Univ Of Tsukuba 関数近似装置、強化学習システム、関数近似システムおよび関数近似プログラム

Family Cites Families (32)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5269318A (en) * 1982-08-16 1993-12-14 Neurocom International Inc. Apparatus and method for movement coordination analysis
US5645073A (en) * 1989-07-31 1997-07-08 Synectics Medical Ab Method and an apparatus for use in electromyography to determine risk of muscular disorder
US5505208A (en) * 1993-12-10 1996-04-09 Toomin Research Group Method for determining muscle dysfunction
US6496724B1 (en) * 1998-12-31 2002-12-17 Advanced Brain Monitoring, Inc. Method for the quantification of human alertness
US7499745B2 (en) * 2000-02-28 2009-03-03 Barbara Ann Karmanos Cancer Institute Multidimensional bioelectrical tissue analyzer
US6440067B1 (en) * 2000-02-28 2002-08-27 Altec, Inc. System and method for remotely monitoring functional activities
US7148879B2 (en) * 2000-07-06 2006-12-12 At&T Corp. Bioacoustic control system, method and apparatus
ATE292452T1 (de) * 2001-01-16 2005-04-15 Univ Montreal Myoelektrisch aktivierte respiratorische leckdichtung
JP2002358149A (ja) * 2001-06-01 2002-12-13 Sony Corp ユーザ入力装置
US6643541B2 (en) * 2001-12-07 2003-11-04 Motorola, Inc Wireless electromyography sensor and system
CA2396202A1 (en) * 2002-07-31 2004-01-31 Anthony Michael Galea Sincerity index system and program therefor
US7359750B2 (en) * 2003-02-22 2008-04-15 Chul Gyu Song Apparatus and method for analyzing motions using bio-impedance
GB0306629D0 (en) * 2003-03-22 2003-04-30 Qinetiq Ltd Monitoring electrical muscular activity
JP4078419B2 (ja) 2003-06-18 2008-04-23 独立行政法人産業技術総合研究所 ポインティングデバイス
AU2004290588A1 (en) * 2003-11-18 2005-06-02 Vivometrics, Inc. Method and system for processing data from ambulatory physiological monitoring
WO2005077260A1 (en) * 2004-02-12 2005-08-25 Biopeak Corporation Non-invasive method and apparatus for determining a physiological parameter
US7554549B2 (en) * 2004-10-01 2009-06-30 Sony Corporation System and method for tracking facial muscle and eye motion for computer graphics animation
TWI266217B (en) * 2004-12-28 2006-11-11 Ind Tech Res Inst Methods and devices of a multi-functional operating interface for a nursing machine
WO2006119015A1 (en) * 2005-04-30 2006-11-09 Medtronic, Inc. Impedance-based stimulation adjustment
US20070060830A1 (en) * 2005-09-12 2007-03-15 Le Tan Thi T Method and system for detecting and classifying facial muscle movements
US8280169B2 (en) * 2005-12-21 2012-10-02 Michael Linderman Recordation of handwriting and hand movement using electromyography
US8437844B2 (en) * 2006-08-21 2013-05-07 Holland Bloorview Kids Rehabilitation Hospital Method, system and apparatus for real-time classification of muscle signals from self-selected intentional movements
US8255045B2 (en) * 2007-04-03 2012-08-28 Nuvasive, Inc. Neurophysiologic monitoring system
JP4367663B2 (ja) * 2007-04-10 2009-11-18 ソニー株式会社 画像処理装置、画像処理方法、プログラム
EP2144560B1 (en) * 2007-04-30 2016-09-28 Ictalcare A/S A device and method for monitoring muscular activity
JP5154558B2 (ja) * 2007-08-31 2013-02-27 公益財団法人東京都医学総合研究所 定量的運動機能評価システム及び運動機能評価用プログラム
JP4958128B2 (ja) * 2008-06-20 2012-06-20 国立大学法人大阪大学 筋トーヌス計測装置
US8447704B2 (en) * 2008-06-26 2013-05-21 Microsoft Corporation Recognizing gestures from forearm EMG signals
US8170656B2 (en) * 2008-06-26 2012-05-01 Microsoft Corporation Wearable electromyography-based controllers for human-computer interface
WO2010065067A1 (en) * 2008-11-20 2010-06-10 Bodymedia, Inc. Method and apparatus for determining critical care parameters
US8126542B2 (en) * 2008-12-10 2012-02-28 Somaxis, Inc. Methods for performing physiological stress tests
WO2011020446A1 (en) * 2009-08-21 2011-02-24 The Chinese University Of Hong Kong Systems and methods for reproducing body motions via networks

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2009064216A (ja) * 2007-09-06 2009-03-26 Univ Of Tsukuba 関数近似装置、強化学習システム、関数近似システムおよび関数近似プログラム

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
MASAHIKO MORITA: "Efficiency improvement of reinforce learning using a selective desensitization neural network", IEICE TECHNICAL REPORT, vol. 107, no. 542, 5 March 2008 (2008-03-05), pages 355 - 359 *
TOMOYUKI SHINBO: "Reinforcement Learning Using Selective Desensitization Neural Networks in the State Space with Redundant Dimensions", IEICE TECHNICAL REPORT, vol. 108, no. 383, 12 January 2009 (2009-01-12), pages 7 - 12 *

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2016030096A (ja) * 2014-07-29 2016-03-07 株式会社きもちラボ 生体電気ノイズ同定システム及び生体電気ノイズ除去システム
JP2018054765A (ja) * 2016-09-27 2018-04-05 日本電気株式会社 データ処理装置、データ処理方法、およびプログラム
JP2018196603A (ja) * 2017-05-24 2018-12-13 日本電信電話株式会社 行動推定装置、及び行動推定方法
WO2022195657A1 (ja) * 2021-03-15 2022-09-22 日本電信電話株式会社 筋音抽出方法、筋音抽出装置及びプログラム

Also Published As

Publication number Publication date
US9392954B2 (en) 2016-07-19
JP5545574B2 (ja) 2014-07-09
JPWO2011007569A1 (ja) 2012-12-20
US20120209134A1 (en) 2012-08-16

Similar Documents

Publication Publication Date Title
JP5545574B2 (ja) 分類推定システムおよび分類推定プログラム
CN110337269B (zh) 基于神经肌肉信号推断用户意图的方法和装置
Khan et al. Accelerometer signal-based human activity recognition using augmented autoregressive model coefficients and artificial neural nets
Allard et al. A convolutional neural network for robotic arm guidance using sEMG based frequency-features
CN107378944B (zh) 一种基于主成分分析法的多维表面肌电信号假手控制方法
Long et al. Single-accelerometer-based daily physical activity classification
EP3836836B1 (en) Real-time spike detection and identification
CN112074225A (zh) 用于减轻神经肌肉信号伪像的方法和装置
Mekruksavanich et al. Smartwatch-based sitting detection with human activity recognition for office workers syndrome
Muscillo et al. Early recognition of upper limb motor tasks through accelerometers: real-time implementation of a DTW-based algorithm
CN110333783B (zh) 一种用于鲁棒肌电控制的无关手势处理方法及系统
CN110826625B (zh) 一种基于表面肌电信号的手指手势分类方法
AlOmari et al. Analysis of extracted forearm sEMG signal using LDA, QDA, K-NN classification algorithms
US11772259B1 (en) Enhanced activated exoskeleton system
Shin et al. Korean sign language recognition using EMG and IMU sensors based on group-dependent NN models
Chowdhury et al. Physical activity recognition using posterior-adapted class-based fusion of multi-accelerometers data
Chang et al. A hierarchical hand motions recognition method based on IMU and sEMG sensors
Andronache et al. Towards extending real-time EMG-based gesture recognition system
Arteaga et al. EMG-based adaptive trajectory generation for an exoskeleton model during hand rehabilitation exercises
Kumar et al. A critical review on hand gesture recognition using semg: Challenges, application, process and techniques
Koch et al. Inhomogeneously stacked rnn for recognizing hand gestures from magnetometer data
Kurzynski et al. Multiple classifier system applied to the control of bioprosthetic hand based on recognition of multimodal biosignals
Dev et al. Performance analysis of classifiers for EMG signal of different hand movements
Nieuwoudt et al. Investigation of real-time control of finger movements utilising surface EMG signals
Ren et al. PDCHAR: Human activity recognition via multi-sensor wearable networks using two-channel convolutional neural networks

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 10799633

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2011522736

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 10799633

Country of ref document: EP

Kind code of ref document: A1