JP2015047404A - Control signal generation device, and power assist apparatus - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a power assist apparatus capable of performing a robust control for the instability of a signal acquirement from a sensor, even in the case of detecting a non-invasive biological signal.SOLUTION: For each combination, in which a predetermined number (e.g., one) of sensors (or measurement channel) becomes defective, estimation models of a joint torque are prepared in advance by the machine learning. Next, a defective sensor is detected from a correlation of a plurality of myoelectric signals so that an estimation model is suitably switched into an estimation model containing no defective channel detected, thereby to achieve a robust torque estimation.

Description

  The present invention relates to a control signal generation device for generating a control signal from a biological signal of a user (user) and a power assist device for supporting exercise performed by the user.

  In Japan and many other countries, an aging society with a declining birthrate has become a problem, and there is an increasing demand for assist devices that apply robotics technology. On the other hand, robots that can balance and walk have been developed. For example, there is a robot that can optimally distribute an action force necessary for movement to a plurality of contact points in space and generate torque of each joint in the same way as a human (see Patent Document 1).

  In recent years, there has been an increasing demand for the development of robots that support rehabilitation such as exoskeleton-type robots that aim to support lower limb and trunk movements. For example, an exoskeleton type robot is used for a patient with spinal cord injury in rehabilitation that promotes a patient's independent life (see Patent Document 2).

  Thus, in robot control using biological signals for compensation and recovery of motor function, from the viewpoint of patient burden and system simplicity, surface electromyogram (sEMG), electroencephalogram (Electroencephalogram) : Non-patent document 1, non-patent document 2, non-patent document 3, and non-patent document 4 (see Non-patent Document 1, Non-patent Document 2, Non-patent Document 4).

  In a system that non-invasively detects a biological signal, measurement using a plurality of electrodes is often performed. However, since the non-invasive measurement method has a small signal-to-noise ratio, it is susceptible to noise and electrode impedance changes.

  Thus, conventionally, an attempt has been made to extract useful information from a redundant information by a statistical / machine learning technique by increasing the number of electrodes. In the case of using EEG, a method of extracting brain activity information by a multi-channel fusion model (see Non-Patent Document 5) or a mesh-type sEMG (non-patent document) for acquiring a large number of muscle activity information for determining human motion intentions. (See Patent Document 6). As a specific example of the impedance change, it has been pointed out that the signal obtained by sEMG changes due to muscle fatigue due to a task, sweating, displacement of the position of a sensor to which the sensor is attached, and a slight difference in muscle activity (Non-Patent Document 7). reference). Panagoitis et al. Proposes to extract and feed back a feature value considering temporal changes (see Non-Patent Document 8).

WO2007 / 139135 Publication JP 2012-045194 A

T. Kagawa and Y. Uno.Gait pattern generation for a power-assist device of paraplegic gait.The 18th IEEE International Symposium on Robot and Human Interactive Communication, pp. 633-638, 2009. Tomoyuki Noda, Norikazu Sugimoto, Junichiro Furukawa, Masa aki Sato, Sang-Ho Hyon, and Jun Morimoto.Brain-controlled exoskeleton robot for bmirehabilitaion.IEEE/RAS International Conference on Humanoid Robotics (Humanoids2012), Osaka, pp.21-27, 2012. Tomoyuki Noda, Jun ichiro Furukawa, Tatsuya Teramae, and Jun Morimoto.An electromyogram based force control coordinated in assistive interaction.IEEE International Conference on Robotics and Automation (ICRA2013), Karlsruhe, Germany, p. WeC6.4, May 2013. Tomohiro Hayashi, Hiroaki Kawamoto, and Yoshiyuki Sankai.Control method of robot suit hal working as operator's muscle using biological and dynamical information.IEEE/RSJ International Conference on Inteligent Robots and Systems, pp. 3063-3068, 2005. Lalit Gupta, Beomsu Chung, Mandyam D Srinath, Dennis L Molfese, and Hyunseok Kook.Multichannel fusion models for the parametric classifcation of differential brain activity.IEEE TRANSACTION ON BIOMEDICAL ENGINEERING, Vol. 52, No. 11, pp. 1869-1880, 2005. Kentaro Nagata and Kazushige Magatani.Basic study on combined motion estimation using multi-channel surface emg signals.33rd Annual International Conference of the IEEE EMBS Boston, Massachusetts USA, August 30-September 3, pp. 7865-7868, 2011. N.A.Dimitrova and G.V.Dimitrov.Interpretation of emg changes with fatigue: facts, pitfalls, and fallacies.Journal of Electromyography and Kinesiology, Vol. 13, pp. 13-16, 2003. Panagiotis K artemiads and Kostas J Kyriakopoulos.An emg-based robot control scheme robust to time-varying emg signal features.IEEE TRANSACTION ON INFORMATION TECHNOLOGY IN BIOMEDICAL, Vol. 14, No. 3, pp. 582-588, 2010.

  However, robustness (robustness) against a failure such as a sensor failure or disconnection is required at the time of actual operation of such a system that performs non-invasive detection of a biological signal.

  For example, the electrodes and wiring placed on the body surface are always exposed to physical contact, such as failure of electrodes and amplifiers, disconnection and disconnection, and the risk of loss of some signals of sensors using many There is. However, conventionally, when a sensor failure or disconnection problem occurs, it is not assumed that a target estimation amount is appropriately derived from myoelectricity in real time.

  The present invention has been made to solve the above-described problems, and its purpose is to prevent instability of signal acquisition from a sensor even when a biological signal is detected non-invasively. A control signal generation device capable of performing robust control and a power assist device using the control signal generation device are provided.

  According to one aspect of the present invention, there is provided a control signal generation device that estimates a user's intended musculoskeletal motion from a biological signal of a subject and generates a control signal for driving an external device. A plurality of sensors are arranged for a plurality of parts on the body surface of a person, and the intended musculoskeletal movement is estimated by receiving a sensor for measuring a biological signal and a biological signal from the plurality of sensors for each part. An estimation unit, and the estimation unit detects a defective measurement channel in which an abnormality has occurred among measurement channels from a plurality of sensors for each part based on a correlation between the plurality of measurement channels, and a predetermined number The estimation model prepared by excluding the detected defective measurement channel is selectively used from the estimation models prepared in advance for each combination in which the measurement channel is defective. And a processing unit, based on the estimated musculoskeletal movement, further comprising a control signal calculating means for calculating a control signal for driving the external device.

  Preferably, the sensors are arranged corresponding to the joints to be estimated by the subject, and the estimation means receives the signals from the plurality of sensors for each part and generates joint driving torque generated in the corresponding joints. The estimation processing means selectively selects an estimation model prepared by excluding the detected defective measurement channel from the joint torque estimation models prepared in advance for each combination in which a predetermined number of measurement channels are defective. To estimate the joint drive torque.

  Preferably, the detection means includes a plurality of discriminators that discriminate whether or not an aspect prepared in advance by machine learning has occurred for each defective aspect of each measurement channel.

  Preferably, each discriminator uses logistic regression so as to discriminate the presence or absence of a defect based on the feature vector when the non-diagonal independent component of the covariance matrix for a set of signals from a plurality of measurement channels is used as the feature vector. It is a discriminator learned by analysis.

  Preferably, the detection means detects a failure of each measurement channel by a graphical Gaussian model for a plurality of measurement channels.

  Preferably, in the graphical Gaussian model, the presence / absence of a defect is determined based only on the distance of the current graph to the graph corresponding to the normal time.

  According to another aspect of the present invention, a power assist device for assisting a target human musculoskeletal motion, which is provided for each joint to be assisted and generates a force to assist the motion of the joint A plurality of actuators for each of the parts of the body surface corresponding to the assist target joint, electrodes for measuring surface myoelectric potential, and a plurality of electrodes for each part Torque estimation means for estimating the joint drive torque generated in the corresponding joint, and the torque estimation means is a defective measurement channel in which an abnormality has occurred among measurement channels from a plurality of electrodes for each region. And a joint torque estimation model prepared in advance for each combination in which a predetermined number of measurement channels are defective. And an estimation processing means for estimating joint drive torque by selectively using an estimated model prepared excluding the detected defective measurement channel, and is generated by the actuator based on the estimated joint drive torque. Target torque calculation means for calculating the power target torque, and drive means for driving the actuator based on the target torque.

  Preferably, the detection means includes a plurality of discriminators that discriminate whether or not an aspect prepared in advance by machine learning has occurred for each defective aspect of each measurement channel.

  Preferably, each discriminator uses logistic regression so as to discriminate the presence or absence of a defect based on the feature vector when the non-diagonal components of the covariance matrix for the set of signals from a plurality of measurement channels are used as the feature vector. It is a discriminator learned by analysis.

  Preferably, the detection means detects a failure of each measurement channel by a graphical Gaussian model for a plurality of measurement channels.

  Preferably, in the graphical Gaussian model, the presence / absence of a defect is determined based only on the distance of the current graph to the graph corresponding to the normal time.

  Preferably, the power assist device is an exoskeleton robot.

  According to the control signal generation device and the power assist device of the present invention, it is possible to generate a control signal or generate an assist force with robust control against instability of signal acquisition from a sensor.

It is a figure which shows the structural example of the exoskeleton type robot 1 in this Embodiment. It is a figure which shows the structure of the freedom degree of the exoskeleton type robot. 2 is an example of a block diagram of the exoskeleton robot 1. FIG. It is a conceptual diagram for demonstrating the mechanism which detects a myoelectric potential among the detection mechanisms. It is a figure which shows the concept of the process which estimates the torque provided to the joint used as assistance object by the muscle in the power assist control in this Embodiment. In this Embodiment, it is a figure which shows the control block for estimating the joint drive torque from a myoelectric potential and controlling the actuator of an active joint. It is a figure which shows the external appearance which extracted the part of the aerodynamic hybrid actuator which performs operation | movement for 1 degree of freedom in the structure of an exoskeleton type robot. It is a functional block diagram for demonstrating the structure which controls the system of the pneumatic hybrid actuator which performs operation | movement for 1 degree of freedom. It is a conceptual diagram for demonstrating the aspect which each actuator moves a joint. It is a conceptual diagram for demonstrating the structure of the experiment which verifies the usefulness of the estimation process of the joint drive torque of this Embodiment. It is a figure which shows the time discrimination | determination result of the sensor channel by the time change of sEMG signal and logistic regression. It is a figure which shows the condition where the error channel generate | occur | produced in the middle of the torque estimation process. It is a figure which shows the abnormality score in the case of the disconnection of an electrode element (sensor), or the contact failure to the body surface of an electrode element. It is a figure which shows the abnormal score at the time of the insertion mistake of the channel position of the connector of an electrode element (sensor), or when the short circuit between electrodes by sweat etc. generate | occur | produced. It is a figure which shows the time change of the abnormal score when a channel is disconnected.

  Hereinafter, the configuration of the exoskeleton robot according to the embodiment of the present invention will be described with reference to the drawings. In the following embodiments, components and processing steps given the same reference numerals are the same or equivalent, and the description thereof will not be repeated unless necessary.

  As an example of an actuator for driving the joint of an exoskeleton robot, an “aeroelectric hybrid actuator” described below will be described as an example.

  Therefore, in the present embodiment, an exoskeleton type robot using an electrostatic hybrid actuator for walking / posture rehabilitation will be described below.

  However, the aerodynamic hybrid exoskeleton robot of the present invention can be used not only as an exoskeleton robot for assisting the movement of the lower limbs but also as an exoskeleton robot for assisting the movement of the upper limbs. Is possible.

  Further, in the following description, an exoskeleton type robot that assists exercise as a pair of lower limbs will be described. However, an exoskeleton type that assists exercise of either one of lower limbs or upper limbs. It can also be used as a robot.

  Furthermore, if the aerodynamic hybrid exoskeleton robot of the present invention assists the movement of the target human musculoskeletal system, the above-described “at least one of the lower limbs or the upper limbs” It is not limited to “at least one of these exercises”. For example, it may assist only the exercise of the subject's lower back, or it may be linked with the exercise of the lower limbs during walking or running. And assisting the exercise of the lower back. In the present specification, such assist for human motion as a target is collectively referred to as “support for assisting musculoskeletal motion of a target human (assist)”.

  The exoskeleton type robot of the present embodiment has an exoskeleton. “Exoskeleton” refers to a skeleton structure that a robot has corresponding to a human skeleton structure. More specifically, “exoskeleton” refers to a frame (framework) structure that supports a part of a human body to which an exoskeleton-type robot is mounted from the outside.

  The frame structure is further provided with a joint for moving each part of the frame structure in accordance with the movement based on the human skeleton structure.

In particular, an exoskeleton-type robot that assists the movement of the lower limbs is a robot having a base and a lower half body, and having joints with 6 degrees of freedom in the left and right positions of the ankle, knee, and waist. Further, the six joints are aerodynamic hybrid drive joints. Hereinafter, in such an exoskeleton robot, a joint driven by an actuator to give support force to a user's joint is referred to as an “active joint”.
[Embodiment 1]
FIG. 1 is a diagram showing a configuration example of an exoskeleton robot 1 in the present embodiment. The exoskeleton robot 1 has 10 degrees of freedom.

  Note that a similar configuration is disclosed in Patent Document 2 described above for the configuration example of such an exoskeleton type robot.

  In FIG. 1, FIG. 1 (a) is a diagram showing the appearance of an exoskeleton robot, and FIG. 1 (b) is a diagram showing the main part of the exoskeleton robot 1 extracted from the appearance of FIG. 1 (a). FIG.

  In FIG. 1B, the exoskeleton robot 1 has a frame structure corresponding to both legs, a backpack 101, a flexible sheet 102, a HAA antagonist muscle 103, an HFE extensor muscle 104, an HFE motor 111, a KFE extensor muscle 105, and a KFE motor. 106, an AFE extensor / AAA antagonist muscle 107, an AFE flexor muscle 108, a universal joint 109, and a rotary joint 110 with a pulley provided in a frame structure.

  In FIG. 1B, the backpack 101 is directly attached to the structure for supporting exercise. However, as shown in FIG. 1A, the backpack 101 may be removed from this structure. .

  In addition, for example, an optical encoder is attached to the rotation shaft of the universal joint 109, and the joint angle is measured. An optical encoder is similarly attached to the rotary joint 110 with pulley. The optical encoder may not be attached to the shaft, but may be configured to read the moving direction and the moving amount of the belt wound around the shaft. Note that in the HFE and KFE joints which are hybrid joints, the joint angle may be measured using an encoder attached to the motor.

  FIG. 2 is a diagram showing a configuration of the degree of freedom of the exoskeleton robot 1.

  In FIG. 2, in each joint, the display “R_” indicates the right joint, and the display “L_” indicates the left joint.

  Referring to FIGS. 1 and 2, the HFE joint and the KFE joint out of all 10 degrees of freedom are hybrid drives. In FIG. 2, out of all 10 degrees of freedom, the left and right AFE joints employ antagonistic drive using extensors and flexors. The joints other than the hybrid drive and the antagonist drive are passive drives. However, more joints, for example, all joints may be hybrid driven.

  In FIG. 1, a posture sensor is mounted on the body portion to which both legs are connected to detect the posture of the base portion. In addition, wire type encoders (or encoders attached to motors) are attached to all joints so that joint angles can be measured. By detecting the joint angle and the myoelectric potential of the lower limb as described later, the target torque to be generated in each joint can be calculated.

  In addition, a floor reaction force sensor is mounted on the sole, and it is used as an auxiliary to determine whether the sole that is supposed to touch is actually touching or to correct the model error included in the Jacobian matrix It is good also as composition to do.

  In addition to the controller, the backpack 101 incorporates an air muscle valve and an electric motor driver.

The backpack 101 is also equipped with a battery, a compressed air cylinder (or a CO ₂ gas cylinder), and a regulator, and enables a short time autonomous driving in case the power supply line and the air supply are disconnected. It may be.

  FIG. 3 is an example of a block diagram of the exoskeleton robot 1.

  A command for controlling the exoskeleton robot 1 is given from the external control device 20 to the exoskeleton robot via a communication path. Although not particularly limited, the external control device 20 can use a general-purpose personal computer, and an Ethernet (registered trademark) cable can be used as a communication path. Of course, as a communication path, a wireless communication path such as a wireless LAN (Local Area Network) or another communication standard may be used in addition to a wired communication path of another standard.

  The external control device 20 includes an input unit 208 that receives an instruction input from a user, a non-volatile storage device in which data required for control, such as a program for generating commands and various control parameters, is recorded 206, a memory 204 including a ROM (Read Only Memory) in which firmware for starting the external control device 20 is stored, a RAM (Random Access Memory) that operates as a working memory, and the like. Under the control of the arithmetic device 210, an arithmetic device 210 that executes processing to be generated, an interface (I / F) unit 202 for transmitting a command to the exoskeleton robot via a communication path, and the arithmetic device 210. And a display device 212 for displaying information on the state of control of the robot 1.

  As described above, when the external control device 20 is a general-purpose personal computer, the arithmetic device 210 is configured by a CPU (Central Processing Unit), and the nonvolatile storage device 206 is a hard disk drive, a solid state drive, or the like. Can be used. However, some or all of the functional blocks of the external control device 20 may be configured by dedicated hardware.

  Further, the external control device 20 performs, for example, a process of configuring an estimation model for estimating joint torque from myoelectric potential during calibration for a user wearing an exoskeleton robot. Information about the configured estimation model is stored in a storage device such as the storage device 132, for example.

  The exoskeleton robot 1 includes an exoskeleton part 12 and an internal control device 10.

  The exoskeleton 12 includes a base 121, a lower body 122, an active joint 123, and a detection mechanism 124. Further, the active joint 123 includes an air muscle 1231 (not shown) and an electric motor 1232 (not shown).

  The internal control device 10 also includes an I / F unit 11, a recording device 131, a storage device 132, a measuring device 133, a control unit 134, and an output device 135.

  The I / F unit 11 can receive a torque command or a position command commanded from the external control storage 20.

  The base 121 may be considered to include the skeleton at the waist position and the active joint 123 at the waist position, or may be considered to be only the skeleton at the waist position.

  The lower body 122 may be considered to include the skeleton at the position of the thigh or foot, the active joint 123 at the position of the thigh or foot, or may be considered to be only the skeleton at the position of the thigh or foot.

  The active joints 123 are active joints arranged at left and right ankles, left and right knees, and left and right positions of the waist. Here, the active joint 123 is a joint that can be actively operated by an actuator. That is, the active joint 123 includes an actuator.

  Further, the one or more active joints 123 here are of a hybrid type. That is, at least a part of the active joint 123 is a hybrid type including the air muscle 1231 and the electric motor 1232. The actuator has a function of receiving a torque value as a control target value as a drive signal and controlling based on the received torque value.

  When a servo motor is used as an actuator, the actuator has, for example, a drive circuit capable of current control, and a servo motor that generates a torque proportional to the current has a gear value that is input as a control target value. Torque control for generating the input torque is realized by instructing the drive circuit by multiplying the torque constant determined by the ratio. In particular, by providing a torque sensor at the active joint 123 and feeding back a value detected by the torque sensor to the drive circuit, highly accurate torque control is possible.

  The detection mechanism 124 detects the state of the robot. The detection mechanism 124 includes, for example, an encoder disposed at each joint, a floor reaction force sensor disposed at the foot, a gyro sensor for posture detection disposed at the pelvis, and a load cell that detects the driving force of each air muscle. Etc. The detection mechanism 124 may be an angle sensor that detects the angle of the joint, a posture sensor that acquires the posture of the robot, an external force sensor, or the like.

  In addition, the detection mechanism 124 detects the myoelectric potential of the user using a myoelectric sensor attached to a part of a muscle that moves the joint to be assisted by the user, for example, a leg.

  The internal control device 10 operates the active joint 123. The internal control device 10 operates the active joint 123 in response to the target torque or position command received by the I / F unit 11.

  The measuring device 133 receives various signals (data) indicating detection results from the detection mechanism 124 such as a sensor.

  The control unit 134 performs various calculations such as calculation of a control target value. The calculation performed by the control unit 134 will be described later.

  The output device 135 outputs a control signal to the active joint 123. The output device 135 outputs, for example, a target air muscle pressure value and a motor control value to the active joint 123. More specifically, such a control value is output to a motor driver, a pressure control valve, or the like, and physical energy (compressed air or current) is supplied to the motor or air muscle.

(Control of active joint by myoelectric potential feedback: one degree of freedom system)
In the above description, the configuration of the exoskeleton robot that assists the movement of the lower limbs has been described. In the following, in explaining the power assist operation for the exercise of the wearer of the exoskeleton type robot, for the sake of simplicity of explanation, a system with one degree of freedom (movement of one joint in one direction) will be described as an example. . However, it is possible to apply such control to a one-degree-of-freedom system to each joint of the exoskeleton robot of the lower limb described above. Further, as described above, the present invention can also be applied to a more general exoskeleton type robot.

  Therefore, in the following, the operation principle of the present embodiment will be described as a one-degree-of-freedom system that is assumed to assist upper limb movement.

  FIG. 4 is a conceptual diagram for explaining a mechanism for detecting a myoelectric potential in the detection mechanism 124.

  As described above, the detection mechanism 124 receives, on the exoskeleton robot side, a state detection unit 1242 that receives a signal from a sensor that detects the state of the robot such as a joint angle and outputs the signal to the control unit 134, and the myoelectric potential on the wearer side. And a myoelectric potential detection unit 1244 to detect.

  In FIG. 4, in order to detect the surface potential of the upper arm, for example, a ground electrode ELG is disposed in the tendon portion of the upper arm, and electrodes EL1 to EL4 are disposed in the biceps brachii, for example. For example, electrode elements EL5 to EL6 are arranged on the triceps surae.

  Note that the number and arrangement of the electrodes are merely examples, and can be appropriately changed according to the site where the myoelectric potential is measured. In the following, the “part for measuring myoelectric potential” refers to a part of a muscle that causes a target exercise for generating assist force to a joint to be subjected to power assist. .

  Although not shown, each of the electrode elements EL5 to EL6 may be provided with a preamplifier for amplifying the acquired surface potential, an analog filter for performing filter processing, and the like.

  The myoelectric potential detection unit 1244 includes an amplifier 1250 that further amplifies signals from the electrode elements EL <b> 1 to EL <b> 6 and an A / D converter 1254 that performs analog-digital conversion on the signal from the amplifier 1252.

  Therefore, hereinafter, a “measurement channel” for measuring myoelectric potential is a path through which an analog signal from a body surface via a single electrode is individually converted into a digital signal for measuring the potential of a target muscle. It shall be said. Therefore, the “error channel” means a measurement channel for a sensor mixed with a missing or defective signal due to some physical failure or the like, and includes channels in the following states.

      i) The state of poor contact between the electrode element and the body surface. For example, the electrode element is floating from the body surface, or non-uniformity of gel applied to the electrode surface in order to improve adhesion.

      ii) In addition, disconnection in the path from the electrode elements EL5 to EL6 to the amplifier 1250, connection failure or disconnection of the connector.

      iii) A state in which the connection connector is erroneously connected to an input terminal different from the input terminal of the amplifier 1250 to be originally coupled.

      iv) An energized state generated in the electrode element due to sweat or gel on the body surface (for example, when the electrode element has two electrodes, an energized state between the electrodes).

  FIG. 5 is a diagram showing a concept of processing for estimating the torque applied to the joint to be assisted by the muscle in the power assist control in the present embodiment.

  As shown in FIG. 5, in the present embodiment, when a plurality of sensors are used, a target estimated amount is appropriately derived from the biological signal even if there is a sensor or wiring defect.

  Here, when estimating joint torque from myoelectricity, robust estimation is realized for a plurality of combinations having different failure measurement channels.

  More specifically, it is as follows.

1) For each combination in which a predetermined number (for example, one) of sensors (measurement channels) is defective, a plurality of joint torque estimation models are prepared in advance by machine learning. Thus, the torque estimated by each of the models 1 to n prepared in advance is estimated torque τ ⁽¹⁾ (hat) to τ ⁽ⁿ⁾ (hat). (The following characters with “^” at the beginning of the character are represented by adding a (hat) after the character. “^” Means an estimated value.)
2) Next, a defective sensor is detected from the correlation of a plurality of myoelectric signals, and the contribution rate of each estimated value to the final estimated value τ (hat) is changed according to the detected defective sensor. That is, the contribution of the model not based on the detected signal from the defective sensor is increased. More specifically, for example, robust torque estimation is achieved by appropriately switching the estimation model to an estimation model that does not include the detected defective channel.
(Robust joint torque estimation from surface electromyogram (sEMG))
Below, the conversion process of myoelectric potential-torque estimation of this Embodiment is demonstrated.

  In the present embodiment, after obtaining myoelectricity from a plurality of electrodes, in the conversion of myoelectric potential-torque estimation, a plurality of models are built in advance, and when a failure occurs, an estimation model that does not include a sensor is used.

  As a result, a model including an error channel is switched in real time to a model composed of only normal sensor signals, and joint torque estimation is realized without failure.

  Hereinafter, as an example, a mechanism for dealing with a failure that occurs in one channel of the sensors of the system will be considered. This is because it is important to improve that the entire system does not function due to a single failure (SPOF: Single Point of Failure) in the network and database fields, and this is one of the indicators of robustness. To do. However, in the conversion process of the myoelectric potential-torque estimation according to the present embodiment, it is assumed that up to two error channels are generated in advance, and a normal sensor signal is configured from a model including the error channel. It is good also as a structure switched to a different model. The assumed maximum number of error channels can be expanded when there are more than this.

  However, in the following, for the sake of simplicity, it is assumed that the failure / failure that occurs in the sEMG measurement system is only one of the channels on the sensor signal path.

  FIG. 6 is a diagram showing a control block for controlling the actuator of the active joint (here, as an example, an electrostatic hybrid actuator) by estimating the joint drive torque from the myoelectric potential in the present embodiment. .

  Such a control block is achieved as a function of the control unit 134, for example. Alternatively, such a function may be achieved by cooperation between the control unit 134 and the arithmetic device 210.

  Referring to FIG. 6, joint torque conversion unit 2100 sets the myoelectric potential (sEMG) signal measured from assist target (in this example, the user's upper limb) 2140 in advance by the procedure described below. The joint torque is calculated according to the model.

More specifically, the joint torque conversion unit 2100 includes an EMG signal filter unit 2102 that performs a predetermined digital filtering process on the sEMG signal from the A / D converter 1254, and an output u (t) obtained by filtering. A joint torque calculation unit 2106 for estimating the joint drive torque τ _emg .

Further, the target torque calculator 2110 calculates a target torque τ _r ^* to be generated by the actuator from the estimated τ _emg .

The target torque calculator 2110 includes an inverse dynamics action model calculator 2112. The reverse dynamics action model calculator 2112 receives the target torque τ by feedback 2114 from the target torque τ _r ^* , as will be described later. Calculate _r ^* .

The tendon spring model calculation unit 2120 calculates the pressure p ^* supplied to the air muscle and the control value for the motor. The control target 2130 is an aerodynamic hybrid actuator provided at an active joint as described later.

  The joint torque calculation unit 2106 estimates the human joint drive torque in this case using a plurality of myoelectric-torque conversion models as shown in Equation (1).

  Here, the notation is as follows.

In other words, the expression (1) means that when the error channel i is detected, the estimated torque value is switched using a model that does not include the channel.
(Error channel discrimination)
The error channel detection process performed by the joint torque calculation unit 2106 will be described below.

  In this embodiment, attention is paid to signal correlation as an error channel detection method for sEMG.

  That is, using the correlation of neighboring channels, the state of an error channel and a normal channel is discriminated by a supervised learning discriminator by logistic regression. The case where the sEMG signal is normal is defined as y = 1 (class S1), and the case where a failure is included is defined as y = 0 (class S2). If the sEMG interval input is normal, a classifier is set in advance by machine learning in S1 if it is normal and if it is defective, in S2.

  In the following specific example, for example, the feature amount is extracted based on a signal from a rectified sEMG low-pass filter (cut-off 18 Hz, window width 1.0 s), and logistic binary discrimination (logistic regression analysis). The above discriminator is configured as follows.

  The linear discriminant functions that are classified into the two classes S1 and S2 are represented by the sum total weighted to the feature amount as in Expression (2).

  f (X; θ) = 0 is a function that determines the boundary between two classes.

  The “D-dimensional feature vector” in this case is an off-diagonal independent component of the correlation matrix for the signal from the measurement channel. For example, when the number of channels is 6, D = (6 × 6- 6) / 2 = 15 elements.

  Here, if the probability of belonging to one of the classes is expressed as P, the logistic regression obtains a weight such that the following equation (3) is accurately established.

  When this equation (3) is solved for P, the following equation (4) is obtained.

P takes a value from 0 to 1, and when f (X; θ) = 0, P = 0.5. Here, a binary variable (label) y (S ₁ when y = ₁ , S ₂ when y = 0) is introduced, and N input-output data samples (X ₁ , y ₁ ),... , (X _N , y _N ), the likelihood function is expressed as the following equation (5).

Here, N input-output data samples (X ₁ , y ₁ ),..., (X _N , y _N ) are assumed as follows, for example. Assuming the case of one error channel, “input feature vector” Xi = (xi1,..., XiD) means that the number of channels is 6, for example, electrodes EL1 to EL6 at a certain time interval (ti to ti + Δt). Is a non-diagonal independent component of the correlation matrix for the signal from the measurement channel. Of these N input features, N / m when the specific channel j is an error channel and N (m−1) / m when the specific channel j is normal. Prepare and use this as a learning sample.

  In logistic regression, a parameter vector θ that maximizes the likelihood function is calculated. The following classification is performed by the above method:

  By classifying in this way, the state of the sensor channel is discriminated, and torque estimation is performed using equation (1).

  That is, for each channel error mode, a discriminator for determining whether such a mode has occurred is prepared in advance by machine learning. For example, if one channel error is assumed, if there are q channels, q discriminators are prepared.

  In addition, when assuming up to one channel error, if abnormality from two channels is detected at the same time, an abnormality has occurred in the channel with the higher likelihood of being abnormal among the two. It is good also as a structure discriminate | determined as a thing.

  Also, if up to 2 channel errors are assumed, {q + q (q-1) / 2} discriminators are prepared in advance as the sum of the cases where the error channel is 1 channel and 2 channels. Is done. The same applies when the number of assumed error channels is further increased.

In the above example, the feature amount is described as being extracted based on a signal from the rectified sEMG low-pass filter. However, the feature amount may be configured to use raw data from a sensor.
(Joint torque estimation from sEMG)
In this example, the joint drive torque of the elbow is estimated from the sEMG around the elbow.

  In order to estimate the driving torque from sEMG, the dynamic system of the human arm is approximated to a robot model, and the driving torque of the joint is calculated from the acquired joint angle data by inverse dynamics calculation of Equation (6). The calculation of the joint driving torque is executed by the inverse dynamics interaction model calculation unit 2112.

Next, a process in which the joint torque calculation unit 2106 estimates the joint drive torque τ _emg from the sEMG signal will be described.

  In this embodiment, the sEMG signal applied with a second-order low-pass filter expresses the muscle tension well, and the effect of nonlinearity of muscle contraction speed-muscle tension and muscle length-muscle tension is slow. Since the operation with a narrow joint movable range is an experimental task, the torque is estimated by a linear model from the value of the sEMG signal subjected to the low-pass filter as follows.

  Here, there is the following relationship, and M is the number of channels.

Further, the filtered signal u ^m of each channel obtained from the measured sEMG signal is expressed by the following equation.

  For example, the sEMG signal is measured at 250 MHz, rectified, and then subjected to a secondary Butterworth filter with a cutoff of 6.0 Hz. Further, a moving integral value in the 0.04 sec section is calculated.

Here, emg _k is the value of the sEMG signal at time k.

The parameter w in Expression (7) is determined so as to minimize the square error between τ determined by Expression (6) and τ _all (hat) or τ _−i (hat) in Expression (7).

The estimated value thus obtained (in this example, τ _all (hat) or τ _-i (hat)) is output to the inverse dynamics interaction model calculating unit 2112 as the joint drive torque τ _emg .

  In addition, in order to estimate a torque from a sEMG signal, it is not necessarily limited to the linear model as mentioned above, For example, you may use the muscle tension of the Hill model disclosed by the following literature.

Known Document 1: A. Hill, “The heat of shortening and the dynamic constants of muscle,” Proceedings of the Royal Society of London. Series B, Biological Sciences, vol. 126, no. 843, pp. 136-195, 1938 .
Known Document 2: S. Stroeve, “Learning combined feedback and feedforward control of a musculoskeletal system,” Biological Cybernetics, vol. 75, no. 1, pp. 73-83, 1996.
Alternatively, a non-linearity of the sEMG signal and force expressed by an exponential function of the Napier number may be used.

Known Document 3: David G Lloyd and Thor F Besier. An emg-driven musculoskeletal model to estimate muscle forces and knee joint moments in vivo. Journal of Biomechenics, Vol. 36, pp. 765 {776, 2003.
However, the linear model of the equation (7) as described above has an advantage that it is not necessary to identify a parameter for converting to muscle tension for each muscle, and analysis is easy.
[Experiments with a single-axis assist device]
(Torque controller for pneumatic hybrid actuator)
In the following, a single-axis assist device capable of force control by a torque controller for an aeroelectric hybrid actuator will be described. Since the air muscle PAM generates a very large force, even if it is a high-strength and high-elasticity synthetic fiber, elongation causes a model error. Dense joint torque control is possible by force control assuming a “tendon spring model” and torque generated by a small motor.

  Such an air / electric hybrid actuator and a torque controller for the air / hybrid hybrid actuator are disclosed in Non-Patent Document 2 described above, and will be outlined.

  As will be described below, such a torque controller (function executed by the tendon spring model calculation unit 2120) can be configured to be executed by causing the control unit 134 to execute machine learning. It is.

  The pneumatic air muscle is lightweight, and can generate a large force by converting the energy of compressed air (or compressed gas) into a contracting force by a rubber tube.

  The principle that an air muscle generates force is that a spiral fiber embedded with a pneumatic air bag shrinks in the longitudinal direction (longitudinal direction) when compressed air is fed and the air bag expands. is there.

  More specifically, the rubber tube with plugs at both ends is covered with a spirally wound fiber on the surface so as to restrain the radial direction. When air is fed into the rubber tube, the rubber tube expands due to the pressure of the air. However, since the radial direction is constrained by the fiber, it cannot expand, and is pulled by the radial expansion and contracts in the vertical direction. The place that resembles an animal's muscles that swell and contract is called an artificial muscle.

  The actuator itself is light and soft. Furthermore, since the entire inner surface of the rubber tube contributes to the contraction of the actuator, it is easier to obtain a larger power / weight ratio than a general air cylinder or the like having a structure that receives pressure only by a cross-sectional area. On the other hand, as described above, in general, device control by air pressure has a large control delay due to air contraction / expansion and is not good at quick operation.

  As long as the “air bag” is inflated or contracted by a fluid, what flows into the bag is not limited to air, and is more generally expressed as a “fluid bag”.

  FIG. 7 is a diagram showing an external appearance of a part of an aerodynamic hybrid actuator that performs an operation for one degree of freedom in the configuration of an exoskeleton type robot.

  In FIG. 7, the upper air muscle 302 of the pneumatic air muscle corresponds to the flexor (PAM1), and the lower air muscle 304 corresponds to the extensor (PAM2).

  The extensor PAM2 generates a force opposite to the flexor PAM1.

  Driving force from the electric motor is also applied to the rotary joint 310 with pulley, and the driving force from the pneumatic air muscle and the driving force from the electric motor are combined.

  The driving force from the air muscle 302 is transmitted to the rotary joint 310 with pulley by the wire 306, and the driving force from the air muscle 304 is transmitted to the rotary joint 310 with pulley through the wire 308. With such a configuration, it is possible to reduce the weight of the driving force transmission mechanism itself.

  The leg (or arm) 350 is driven by the driving torque applied to the rotary joint 310 with pulley.

  FIG. 8 is a functional block diagram for explaining a configuration for controlling the system of the aerodynamic hybrid actuator that performs the operation for one degree of freedom shown in FIG.

  In FIG. 8, the internal control device 10 is configured as a multi function board.

  The multifunction board 10 connected to the external control device 20 controls the actuator in accordance with a command from the external control device 20. Specifically, the multifunction board 10 controls the valve 301 and the valve 302 for controlling the contraction of the pneumatic air muscles 302 and 304, and the motor driver 311 for controlling the electric motor 312. Furthermore, the multifunction board 10 reads the measurement data from the angle encoder 324 that detects the joint angle θ, the load cell that detects the driving force from the air muscle, and the torque sensor that detects the torque applied to the active joint. Based on this, control as described below is executed.

  As the angle encoder 324, for example, a quadrature encoder can be used.

  The driving forces from the air muscles 302 and 304 and the electric motor 312 are combined at the rotary joint 310 with a pulley and give a torque τ to the leg 350.

  FIG. 9 is a conceptual diagram for explaining a mode in which each actuator moves a joint.

  In the following, first, let us consider a case where the extension of the tendon wire can be ignored as shown in FIG.

  The driving force fPAMi (i = 1, 2,...) From the pneumatic air muscle force is transferred / converted as torque to the arm (or leg) by the wire and the pulley. The driving force of the electric motor is also transmitted through this pulley. The driving force fPAMi from the pneumatic air muscle force is detected by a load cell provided at the junction between the pneumatic air muscle and the wire that transmits the driving force to the pulley.

  The description of the following model is not limited to the combination of the antagonistic muscle and the electric motor, but can be widely applied to the combination of the extensor and the electric motor, for example. However, in the following, for simplicity of explanation, a simple two antagonistic muscle model is considered.

The torque τ _PAM is _expressed by the following formula _using the driving force f _PAMi .

τ _PAM is the torque due to the driving force of the pneumatic air muscle, while r ₀ is the pulley radius, and in this simple model, r ₀ is a constant.

  Torque due to the driving force of the motor is transmitted in parallel with the driving force of the pneumatic air muscle. For example, a small torque is transmitted via a pulley by a mechanical belt.

The pneumatic air muscle is excellent in generating direct current or low frequency torque. In addition, since the driving force by the electric motor is a quick and high frequency torque and covers the error of the torque τ _PAM due to the driving force of the pneumatic air muscle, the value of the electric torque τmotor is the torque by the pneumatic air muscle. A smaller value is acceptable.

  The total torque is expressed by the following formula.

τ = τ _PAM + τmotor
Therefore, although not particularly limited, for example, the internal control device 10 causes the electric motor 312 to additionally operate with respect to torque that is at a higher frequency than the first threshold (assuming that “more” includes “more”). Is preferred. The additional operation is to operate the electric motor 312 in addition to the air muscles 302 and 304. Here, the first threshold is, for example, 3 Hz.

  From the pneumatic air muscle, the wire transmits the driving force to the pulley and drives the joint in one designated direction.

  As the wire, a liquid crystal polymer fiber can be adopted because it is lighter, stronger and more flexible than a metal wire, chain, or mechanical belt. As the liquid crystal polymer fiber, for example, Vectran (registered trademark) manufactured by Kuraray Co., Ltd. can be used, which is known as a material having high strength and high elastic modulus even though it is a polymer fiber.

(Mechanical model of pneumatic air muscle)
Pneumatic actuators, including pneumatic air muscles, wires and pulleys, have much in common with human muscles.

  When an exoskeleton robot supports the weight of a human, the wire resembles a human tendon in its characteristics.

  Therefore, as described below, a torque controller for a pneumatic air muscle is derived with the assumption of equilibrium.

  For controlling the pressure of the pneumatic air muscle, for example, a valve that adjusts the pressure proportionally is used, and the pressure p is controlled by a closed feedback loop and is sufficiently stable.

  In the transient state, there is an aerodynamic movement between the valve pressure and the pressure of the pneumatic air muscle, but after a certain time constant, there is an external movement such as the weight of the robot and human. The pneumatic air muscle contracts until the driving force is balanced with the above constraints.

  Therefore, the influence of air circuit dynamics is small and can be ignored in quasi-static operation.

  At this equilibrium point, the driving force generation depends on the internal pressure and contraction rate, and the driving model of the pneumatic air muscle is given as follows:

Here, ε is a contraction ratio, and D ₀ and ψ ₀ are the diameter of the pneumatic air muscle at normal pressure, and the spiral fiber in which a pneumatic air bag is embedded in the pneumatic air muscle. It is the angle of inclination of the winding direction with respect to the direction orthogonal to the shrinking direction.

  Such a “pneumatic air muscle driving force model” is disclosed in the following document, for example.

Known Document 4: K. Inoue. Rubbertuators and applications for robots. In Proceedings of the 4th international symposium on Robotics Research, pp. 57-63. MIT Press, 1988.
Known Document 5: DG Caldwell, A. Razak, and MJ Goodwin. Braided pneumatic muscle actuators. In Proceedings of the IFAC Conference on Intelligent Autonomous Vehicles, pp. 507-512, 1993.
Unlike air cylinders, torque changes nonlinearly as joint angles change.
Under the assumption that the motion constraints are always unchanged, the pressure of the pneumatic air muscle will always give the contraction rate ε (p) at the same equilibrium point. That is, the shrinkage ratio ε (p) is a function of the pressure p.

  At this time, an inverse model of the “pneumatic air muscle driving force model” can be learned by machine learning. That is, it corresponds to obtaining the pressure p for obtaining a desired driving force f. The internal control device 10 outputs a desired driving force by controlling the valve so as to apply such a pressure p to the air muscle.

  For example, Hartman et al. Have proposed dynamic learning in the following literature.

Known Document 6: Christoph Hartmann, Joschka Boedecker, Oliver Obst, Shuhei Ikemoto, and Minoru Asada. Real-time inverse dynamics learning for musculoskeletal robots based on echo state gaussian process regression. In Accepted at RSS 2012, 2012.
However, in general, since the constraints on motion are dynamically changed and balanced with different external forces F, this assumption does not hold strictly in an exoskeleton robot as described above.

  Therefore, the contraction ratio ε at the equilibrium point in this case is expressed as ε (p, F) as a function of pressure and external force.

(Tendon spring equilibrium model)
Since the wire made of the liquid crystal polymer fiber as described above has a high strength and a high elastic modulus, the approximation that the length does not change with respect to the tension usually holds well. Examples of the high-strength / high-modulus fiber include ultra high molecular weight polyethylene fiber and PBO fiber in addition to polyarylate fiber such as Vectran (registered trademark).

  However, to hold a human (eg, weigh 60 kg), a pneumatic air muscle typically produces 3000 N as a tension (for example, a maximum of 5000 N is required). .

  When such a large force is applied, even if it is a wire (tendon wire) made of a liquid crystal polymer fiber, the extension amount cannot be ignored and the equilibrium point is changed.

  Below, as shown in FIG.9 (b), the model which considered the influence of the expansion | extension of a tendon wire is examined.

  Therefore, a linear tendon spring model as described below is introduced as a mechanical model of a tendon made of liquid crystal polymer fiber.

f = kΔε (S2)
That is, in this model, the tendon wire (wire made of liquid crystal polymer fiber such as polyarylate fiber) is approximated to a spring, Δε is the extra air muscle contraction caused by the force, and k is the spring constant. is there. In other words, the pneumatic air muscle needs to be contracted excessively as much as the tendon extends by force. Such a tendon wire approximated to a spring will be referred to as a “tendon spring”.

  If the wire corresponding to such a tendon is not approximated as a tendon spring, the pneumatic air muscle driving force model g (...) Has three parameters (two of which are dependent on each other). Represented by the following formula (S3):

here,
The following relationship holds for the coefficients of the quadratic expression.

At the equilibrium point, the driving force f ^* decreases as follows due to the extension of the tendon.

Here, g ′ (...) Is a driving force model under the assumption of a tendon spring, ε ^* is an apparent contraction rate, ε is an actual contraction rate of a pneumatic air muscle, and Δε Is an additional term to the actual contraction rate of the pneumatic air muscle.

From the equation (S2), Δε = f ^* / k.

It is difficult to directly measure the actual contraction ratio ε with the detection mechanism 124 of the exoskeleton robot as described above. Therefore, the estimated shrinkage ratio εest (= ε ^* −Δε) is used instead. The contraction ratio ε ^* is calculated from the joint angle as in the following equation.

Here, L ₀ is the initial length of the pneumatic air muscle when the joint angle θ = 0. In other words, the “apparent contraction ratio” is a contraction ratio derived from a change in the angle of the joint driven by the actuator, and is based on an amount that can be detected by the angle sensor, whereas “actual contraction ratio”. The “ratio” is the rate at which the air muscle is actually contracted to generate the driving force f ^* .

When the driving force f ^* is required at the contraction ratio ε ^* , the necessary pressure p ^* is derived from the inverse model g ⁻¹ () as follows.

  Pneumatic air muscle parameters a, b and c can be estimated from known pneumatic air muscle data using a least squares algorithm.

  Furthermore, for example, when the equilibrium point driving force is measured by a one-degree-of-freedom system in a calibration experiment, an optimum value can be experimentally obtained by manually adjusting the spring constant k.

The main torque that assists the motion (torque of a component below a predetermined frequency, for example, DC component torque) is supplied by a pneumatic air muscle, and the actual torque τ _PAMs can be measured by a load cell. As described above, high-frequency torque (component torque exceeding a predetermined frequency: torque of AC component) is generated by the electric motor.

  Therefore, the torque of the motor is expressed as the following formula (S7).

As described above, by applying the tendon spring model to the tendon wire, in the configuration for detecting the joint angle and the driving force applied to the tendon wire, the inverse model that correctly reflects the driving force of the pneumatic air muscle is For example, it can be constructed by machine learning. As a result, it is possible to improve the frequency response when controlling the aeroelectric actuator.
(Experiment with 1-axis assist system: Robust joint torque estimation experiment)
(Experimental settings)
FIG. 10 is a conceptual diagram for explaining the configuration of an experiment for verifying the usefulness of the joint drive torque estimation process according to the present embodiment.

  In this experiment, a robot arm with one degree of freedom was used, and the task of moving up and down fixed on a human upper limb as shown in FIG. Human joint angle data is obtained from the encoder of the robot arm, and sEMG signals are measured from a total of four locations, two from the biceps brachialis and two from the triceps surae, and the joint drive torque is estimated by the method described above. .

  In this case, consider a case where an error occurs in one channel of the biceps, which is the main for torque estimation. In this experiment, it is assumed that the sEMG signal sensor code is poorly connected to the amplifier or the like as an error channel, and the connection between the sensor code and the amplifier is turned on / off during the experiment so that such a signal enters. Switched intermittently.

  FIG. 11 is a diagram illustrating the sensor channel state determination result based on the temporal change of the sEMG signal and logistic regression in such an experiment.

  FIG. 11A shows the state of sEMG during operation. A section in which the magnitude of the voltage is extremely small is an error.

  FIG. 11B shows the result of categorizing the sEMG signal so that the label is 1 when it is determined as a normal channel and 0 when it is determined as an error channel. The correct answer rate was 91%.

(Result of joint torque estimation)
FIG. 12 is a diagram illustrating a situation in which an error channel has occurred during the torque estimation process.

  FIG. 12 shows a result of robust joint torque estimation corresponding to a case where an error occurs in a channel that is the main for estimation when torque is estimated from the sEMG signal.

  FIG. 12 (a) shows a torque estimation result when an error occurs in a main channel in the conventional method, and FIG. 12 (b) shows a robust torque estimation result for the same situation. FIG. 12 shows both the torque calculated from the inverse dynamics and the torque estimated from the sEMG signal.

  From FIG. 12, when an error channel is detected, the conventional method cannot cope with it, and the error of the estimated joint torque becomes large, so that the estimation is not properly performed.

  On the other hand, in the estimation method of the present embodiment, the estimation model is switched as soon as an error channel is detected, so that it can be seen that there is little error and appropriate estimation is performed.

  In FIG. 12A, the overall mean square error is 17.9, whereas in FIG. 12B, it is 8.2.

As described above, in this embodiment, by using a plurality of sensor channels for detecting biological signals and combining a plurality of models for estimating joint drive torque from the sEMG signal, a channel such as a failure may be generated in a channel from some sensors. Even if an error occurs, the torque can be estimated robustly.
[Embodiment 2]
In the first embodiment, in order to detect an error channel, a discriminator that determines whether such an aspect has occurred is prepared in advance for each channel error aspect by machine learning. .

In the second embodiment, a graphical Gaussian model as described below is used for error channel detection. Except for the configuration of error channel detection, it is the same as in the first embodiment.
(Error channel detection)
Below, the process which performs the abnormality detection of the sEMG channel based on the method of Embodiment 2 is demonstrated.

  The method according to the second embodiment calculates an abnormal score for each channel using a graphical Gaussian model.

  Such techniques are disclosed in the following documents, for example.

Known Document 7: T. Ide, AC Lozano, N. Abe, and Y. Liu, “Proximity-based anomaly detection using sparse structure learning,” SIAM International Conference on Data Mining, pp. 97-108, 2009.
That is, given two data sets D _A (N _A sets) and D _B (N _B sets) as follows from a sensor, first, sparse based on a given variance matrix Learn the structure.

Here, ^RM indicates an M-dimensional real vector.

  The two sparse graphs are then compared and an abnormal score is calculated for each variable.

  This will be described in detail below.

1) Graphical Gaussian model (GGM)
GGM is equivalent to comparing two multivariate data and calculating the contribution of each variable to the difference.

For an M-dimensional random variable x (∈R ^M ), GGM assumes an M-dimensional Gaussian distribution as shown in the following equation (T2).

Here, det represents a determinant, and Λ (∈RM ^{× M} ) represents an accuracy matrix. Further, a Gaussian distribution (normal distribution) of the mean vector μ and the dispersion matrix Σ is represented by N (... | Μ, Σ).

  In GGM, a Gaussian distribution is associated with a graph (V, E), where V is a set of nodes that contain all M variables, and E is a set of edges.

In the graph, given all other variables, if and only if they are conditionally independent, there will be no edge between variable x _i and variable x _j .

  Under the assumption of Gaussian distribution, it is expressed by the following equation.

Here, the symbol between the variable x _i and the variable x _j represents statistical independence.

  Further, formally, the definition of “neighborhood” is given as follows.

That is, “when Λij ≠ 0 holds, and only in that case, the variable node xi is in the vicinity of the variable node xj.”
The neighborhood graph of the variable node x _{i is} a graph in which the node x _i and its neighboring nodes are connected by an edge and includes the node x _i and its neighboring nodes.

  Furthermore, “selecting neighboring nodes” means listing all neighboring nodes of each node.

2) Sparse Structure Learning Next, a method for learning a sparse structure from data will be described.

  “Structural learning” means finding an essential dependency between variables (corresponding to electrode elements (sensors)) from data. That is, it corresponds to “obtaining a graph with a variable node as a vertex (weighted)”. Edges are stretched between variables that have dependencies, otherwise edges are not stretched. Therefore, in the case of “sparse structure learning”, the goal is to obtain a graph having edges only in essential dependency relationships.

  Since this process is common to both data A and B, the subscript A or B is omitted for the time being, and if both data are represented by D as in the following equation, the sample covariance matrix S is given by equation (T4).

3) Normalized maximum likelihood estimation:
In GGM, from a different point of view, structural learning means finding a multivariable Gaussian distribution accuracy matrix Λ. Here, the accuracy matrix Λ corresponds to an inverse matrix of the sample covariance matrix S.

  However, in general, the sample covariance matrix is often insufficient in rank and there is no inverse matrix.

  Therefore, it is necessary to solve the maximum likelihood equation with the L1 normalization term expressed by the following equation (T5).

Thanks to the normalization term (penalty term), many of the elements in the precision matrix Λ ^* are exactly zero. The weight ρ of the normalization term is an input parameter and can be interpreted as an index of “up to which value of the correlation coefficient is regarded as a significant correlation”.

4) The graph's LASSO algorithm:
The LASSO (least absolute shrinkage and selection operator) algorithm for a graph reduces the problem to be solved into a quadratic programming problem with a series of L1 regularization terms by using a gradient method in which the matrix is blocked.

  Such an algorithm is disclosed in the following known document 8, for example.

Known Document 8: O. Banerjee, LE Ghaoui, and G. Natsoulis, “Convex optimization techniques for fitting sparse gaussian graphical models,” In proc. Intl. Conf. Machine Learning, pp. 89-96, 2006.
In order to use the gradient method in which the inside of the matrix is blocked, attention is paid to one specific variable x _i as shown in the following equation, and the accuracy matrix Λ and its inverse matrix are divided.

Here, it is assumed that the rows and columns of the matrix are arranged so that the elements associated with the variable x _i are located in the last row and last column.

  Therefore, “reducing into a series of quadratic programming problems with a L1 regularization term” means that an optimal problem is derived for a portion of l in the accuracy matrix Λ, with the submatrix L being a constant, and the accuracy matrix is 1 column This is equivalent to optimizing each line. As a result, a sparse accuracy matrix can be obtained without explicit inverse matrix calculation.

Further, the sample covariance matrix Σ (corresponding to the matrix S) is divided by the same method corresponding to the division of the matrix based on the variable x _i .

5) Abnormal score By sparse structure learning, a sparse structure corresponding to a set of variables (corresponding to electrode elements (sensors)) at each time point is obtained at two different time points on the time axis. These are distinguished by subscripts A and B.

In the context of GGM, this is synonymous with obtaining two probability models p _A (x) and p _B (x) that generate data. As will be described below, by comparing the two, a score (abnormal score) indicating which variable contributes to how much the two have changed is calculated.

In order to quantify how much each variable contributes to the difference between two sets of data in a probabilistic model given data sets D _A and D _B , the most natural measure is Kullback- Leibler (KL) distance.

When focusing on one variable _{x i, p A (x i} | z i) and p _B | an expected value of KL distance between (x _i z _i), is calculated by the distribution p _A (z _i), The following equation (T6) is obtained.

Furthermore, the KL distance d ^AB _i can be rewritten as follows.

Here, the accuracy matrix Λ _A and its inverse matrix Σ _A are divided as follows.

The KL distance d ^AB _i is a measure of the change in the neighborhood graph of the i-th variable node. The larger this amount, the greater the change with respect to the variable x _i .

  Therefore, assuming that neighboring nodes are preserved, it is reasonable to define the abnormal score of the i-th variable as the following equation (T8).

  However, the calculation cost for the variance matrix and the accuracy matrix is very high. In order to control a real-time robot, it is desirable to minimize such calculation costs.

  Therefore, the abnormal score of the i-th variable is simply defined as the following equation (T9).

  Experiments have confirmed that the definition of equation (T9) is sufficient to discriminate channel errors.

  FIG. 13 is a diagram illustrating an abnormal score in the case of disconnection of an electrode element (sensor) or poor contact of the electrode element with the body surface.

  FIG. 13A shows an abnormal score value of each channel when the second channel is disconnected or missing, and FIG. 13B shows one of the two electrodes of the first channel electrode element. The value of the abnormal score when it is in a floating state is shown.

  In FIG. 13A, the abnormality in the second channel appears in the abnormality score, and in FIG. 13B, the abnormality in the first channel appears in the abnormality score.

  FIG. 14 is a diagram showing an abnormal score when the channel position of the electrode element (sensor) connector is misplaced or when a short circuit between the electrodes due to sweat or the like occurs.

  FIG. 14A shows the abnormal score value of each channel when the first channel and the fourth channel are connected in reverse, and FIG. 14B shows the two electrodes of the electrode element of the second channel. Shows the value of the abnormal score when energization is caused by sweat (or gel).

  In FIG. 14A, the first and fourth channels have an abnormality, and in FIG. 14B, the second channel has an abnormality, respectively.

  FIG. 15 is a diagram illustrating a temporal change in the abnormality score when the channel is disconnected.

  FIG. 15A shows a comparison of an abnormal score when the electrode element (sensor) of the second channel is disconnected with other channels, and FIG. 15B shows an abnormality with the sEMG signal of the second channel. It is a figure which shows the time change of a score.

  During the period in which the sEMG signal is zero due to the disconnection, the abnormality score is a constant value, and it can be seen that the abnormality of the second channel can be detected in real time by the abnormality score.

  Therefore, using the method as described above, the abnormal score of each sEMG channel can be calculated to detect that the sensor channel contains an error. According to the detected channel error, the torque estimation model can be switched in the same manner as in the first embodiment.

  Since the other configuration of the power assist device of the second embodiment is the same as that of the power assist device of the first embodiment, description thereof will not be repeated.

  Even with the configuration of error channel detection as described above, it is possible to achieve the same effect as in the first embodiment.

  In addition, even when the number of channels is increased, only essential relationships between variables can be extracted by sparse structure learning, so that it is possible to discriminate channel errors more suitable for multi-channeling.

  In the first and second embodiments described above, the myoelectric potential is used as the biological signal. However, the present invention is not limited to such a case. For example, signals from a plurality of sensors are used. By integrating and analyzing, the present invention can be applied to a system that estimates the joint driving force (or driving torque) in the motion that the user is going to perform. For example, in addition to or instead of myoelectric potential, for example, a brain activity signal detected by a plurality of sensors from the surface side of the skull (eg, NIRS (Near-Infrared Spectroscopy) by brain waves or near infrared) System using a signal etc.).

  In the first and second embodiments described above, the system for estimating the joint driving force (or driving torque) in the motion that the user is going to perform using the myoelectric potential has been described. Is not limited to such a configuration, for example, by integrating and analyzing biological signals such as myoelectric potentials and brain activity signals as described above, the direction of movement the user is going to perform It can also be used in a system that estimates an intended mode of motion, such as intensity, speed, or the like, and drives an external device, such as an exoskeleton robot, according to a corresponding control signal. For example, in the case of controlling the lower limb exoskeleton robot and applying it to rehabilitation, if the user can determine whether he is trying to stand up or sit down, weight compensation is performed by the limb exoskeleton robot. It is also possible to perform control such as switching whether to perform or not according to the determination result.

  Moreover, in Embodiment 1 and 2 demonstrated above, although demonstrated by the structure of estimating a user's own joint drive force (or joint drive torque) by measurement of a user's (user) 's own myoelectric potential, For example, the measurement of myoelectric potential (more generally, the measurement of vital signs) measures the myoelectric potential of the therapist, and it is possible to use it like a user (rehabilitation subject) who supports joint movement It is. In this case, the configuration of the power assist is not necessarily limited to the configuration of the exoskeleton type robot, and any other configuration can be used as long as the drive system can support the user's movement.

  Embodiment disclosed this time is an illustration of the structure for implementing this invention concretely, Comprising: The technical scope of this invention is not restrict | limited. The technical scope of the present invention is shown not by the description of the embodiment but by the scope of the claims, and includes modifications within the wording and equivalent meanings of the scope of the claims. Is intended.

  DESCRIPTION OF SYMBOLS 1 Exoskeleton type robot, 10 Internal controller, 11 I / F part, 12 Exoskeleton, 20 External controller, 121 Base, 122 Lower body, 123 Active joint, 124 Detection mechanism, 131 Recording processing part, 132 Storage device, 133 Measuring device, 134 control unit, 135 output device, 302, 304 air muscle, 310 rotary joint with pulley, 312 electric motor.

Claims (12)

A control signal generation device that estimates a musculoskeletal motion of the intended user from a biological signal of a subject and generates a control signal for driving an external device,
A plurality of sensors arranged for each of a plurality of sites on the subject's body surface, and sensors for measuring biological signals,
An estimation means for receiving biological signals from a plurality of sensors for each region and estimating the intended musculoskeletal movement;
The estimation means includes
Among the measurement channels from the plurality of sensors for each part, a detection unit that detects a defective measurement channel in which an abnormality has occurred based on the correlation between the plurality of measurement channels;
The musculoskeletal motion is estimated by selectively using an estimated model prepared by excluding a detected defective measurement channel from among estimated models prepared in advance for each combination in which a predetermined number of measurement channels are defective. Estimation processing means to
A control signal generation device further comprising control signal calculation means for calculating a control signal for driving the external device based on the estimated musculoskeletal motion. The sensor is arranged corresponding to a joint to be estimated by the subject,
The estimation means receives signals from a plurality of sensors for each part, estimates joint drive torque generated in the corresponding joint,
The estimation processing means selectively uses an estimation model prepared by excluding a detected defective measurement channel from among joint torque estimation models prepared in advance for each combination in which a predetermined number of measurement channels are defective. The control signal generation device according to claim 1, wherein the joint driving torque is estimated.   The control signal generation apparatus according to claim 2, wherein the detection unit includes a plurality of discriminators that discriminate whether or not the mode prepared in advance by machine learning has occurred for each mode of failure of the measurement channels.   Each of the discriminators uses a logistic so as to discriminate the presence / absence of the defect based on the feature vector when a non-diagonal component of a covariance matrix for a set of signals from the plurality of measurement channels is used as a feature vector. The control signal generation device according to claim 3, wherein the control signal generation device is a discriminator learned by regression analysis.   The control signal generation apparatus according to claim 2, wherein the detection unit detects a failure of each of the measurement channels by a graphical Gaussian model for the plurality of measurement channels.   6. The control signal generation device according to claim 5, wherein in the graphical Gaussian model, presence / absence of a defect is determined based only on a distance of a current graph with respect to a graph corresponding to a normal state. A power assist device for assisting a target human musculoskeletal movement,
An actuator that is provided for each joint to be assisted, and that generates a force to assist the movement of the joint;
A plurality of electrodes are arranged for each of a plurality of sites on the subject's body surface corresponding to the assist target joint, and electrodes for measuring surface myoelectric potential;
Torque estimation means for receiving a signal from a plurality of electrodes for each part and estimating a joint driving torque generated in a corresponding joint;
The torque estimation means includes
Among the measurement channels from the plurality of electrodes for each part, a detection means for detecting a defective measurement channel in which an abnormality has occurred based on a correlation between the plurality of measurement channels;
By selectively using an estimated model prepared by excluding a detected defective measurement channel from among estimated models of joint torque prepared in advance for each combination in which a predetermined number of measurement channels are defective, the joint drive torque Estimation processing means for estimating
A target torque calculating means for calculating a target torque to be generated by the actuator based on the estimated joint driving torque;
A power assist device comprising: drive means for driving the actuator based on the target torque.   The power assist device according to claim 7, wherein the detection unit includes a plurality of discriminators that discriminate whether or not the mode prepared in advance by machine learning has occurred for each mode of failure of the measurement channels.   Each of the discriminators uses a logistic so as to discriminate the presence / absence of the defect based on the feature vector when a non-diagonal component of a covariance matrix for a set of signals from the plurality of measurement channels is used as a feature vector. The power assist device according to claim 8, which is a discriminator learned by regression analysis.   The power assist device according to claim 7, wherein the detection unit detects a defect of each of the measurement channels by a graphical Gaussian model for the plurality of measurement channels.   The power assist device according to claim 10, wherein in the graphical Gaussian model, presence / absence of a defect is determined based only on a distance of a current graph with respect to a graph corresponding to a normal state.   The power assist device according to any one of claims 7 to 11, wherein the power assist device is an exoskeleton robot.