Finger angle estimation system, finger angle estimation method, and finger angle estimation program

Abstract

Finger angle information is estimated from electromyographic data with less computation. [Solution] The finger angle estimation system 1 comprises an electromyographic signal acquisition unit 11 attached to the surface of the skin to acquire electromyographic signals, a finger angle information acquisition unit 12 that acquires finger joint angle information, and an estimation unit 13 that performs machine learning using the electromyographic signals acquired by the electromyographic signal acquisition unit 11 and the finger joint angle information acquired by the finger angle information acquisition unit 12 as training data to generate a model that estimates the finger joint angle when new electromyographic information is provided. The estimation unit 13 performs machine learning using RC (reservoir computing).

Description

Technical Field

[0001] The present invention relates to a finger angle estimation system, a finger angle estimation method, and a finger angle estimation program.

Background Art

[0002] There is a certain number of finger amputees due to accidents, diseases, etc. both at home and abroad. Finger amputees use prosthetic hands to compensate for the lost part.

[0003] Prosthetic hands include decorative prosthetic hands used as ornaments, active prosthetic hands that are moved by using the force of the remaining part or another part, myoelectric prosthetic hands that are moved by driving an actuator based on the activity information of the remaining muscles, etc. Among these, myoelectric prosthetic hands are being actively researched and developed because they enable multi-degree-of-freedom and intuitive operation.

[0004] Regarding myoelectric prosthetic hands, those using a plurality of discriminators (for example, see Non-Patent Document 1) and those using myoelectric potential subjected to FFT as a feature amount (for example, see Non-Patent Document 2) are known.

[0005] When cutting a finger, if the remaining muscle and soft tissue are sutured to process the stump, the remaining muscle is not given sufficient load and the muscle will atrophy. When such atrophy of the remaining muscle occurs, secondary disorders may occur due to blood circulation disorders around the remaining muscle. Therefore, in recent years, a surgical method of fixing the stump of the remaining muscle to the bone and applying tension to prevent muscle atrophy has become common.

[0006] In myoelectric prosthetic hand systems, when estimating finger joint torque and angles, it is desirable to be able to decode isotonic, isometric, and isokinetic contraction movements from information provided by muscles performing isometric contractions (e.g., surface electromyography signals). For this reason, recent prosthetic hand systems utilize machine learning for control. However, improving the control and prediction performance of machine learning systems generally increases the computational resources required for the system. On the other hand, equipping a prosthetic hand with abundant computational resources leads to an increase in the size, weight, and cost of the myoelectric prosthetic hand. Therefore, it is desirable for the system to be mounted on the prosthetic hand to be simple and provide sufficient freedom of movement for daily life.

[0007] One machine learning technique that can reduce the learning cost is RC (reservoir computing). RC is a type of RNN (Recurrent Neural Network), and by fixing the connection weights of the hidden layers in the RNN, it is possible to learn with less computational cost.

[0008] As a technique for applying RC to myoelectric prosthetic hands, a method has been proposed to read voluntary movements of the hand and arm from surface electromyographic signals (for example, Non-Patent Document 3). [Prior art documents] [Non-patent literature]

[0009] [Non-Patent Document 1] Ayuko Ibe, Manabu Gouko, and Koji Ito: “Discrimination of Combined Motions for Prosthetic Hands Using Surface EMG Signals”. Transactions of the Society of Instrument and Control Engineers 45.12, pp. 717.723, (Jan. 2011). [Non-Patent Document 2] Akira HIRAIWA et al.: “EMG Recognition with a Neural Network Model for Cyber ​​Finger Control”. Transactions of the Society of Instrument and Control Engineers 30.2, pp. 216.224, (1994). doi: 10.9746 / sicetr1965.30.216. [Non-Patent Document 3] Satoshi Yoshida, Seiya Kasai, "Analysis of Surface Electromyographic Signals and Readout of Voluntary Motion Using a Reservoir Calculation Method," Proceedings of the 83rd Autumn Meeting of the Japan Society of Applied Physics. [Non-Patent Document 4] Atsushi Katayama, Deku Shin, and Yasuharu Koike, "Estimation of Finger Joint Angle Using Electromyographic Signals," IEICE Technical Report, MVE2006-79, (2007-3) [Non-Patent Document 5] Herbert Jaeger “The “echo state” approach to analyzing and training recurrent neural networks -with an Erratum note” German National Research Center for Information Technology GMD Technical Report, 148(34):13, (2001) [Overview of the Initiative] [Problems that the invention aims to solve]

[0010] The method disclosed in Non-Patent Document 3 estimates joint torque from surface electromyography (EMG) signals by applying machine learning using RC. However, this method cannot estimate finger joint angles from surface EMG signals. In Non-Patent Document 3, when attempting to estimate information about joint movement from detected electromyographic signals, integration processing is performed on the signals, following the principles of neural circuit signal processing. Surface EMG signals are a collection of electromyographic signals emitted by multiple muscle fibers, and processing them requires the integration of complex time-series signals. Estimating finger angles from electromyographic signals requires integration of time-series signals over a longer period than estimating finger joint torque. For these reasons, it has been difficult to estimate finger joint angles from electromyographic signals even with machine learning using RC.

[0011] There is a technique for estimating finger joint angles from electromyographic signals, as disclosed in Non-Patent Document 4. However, this technique estimates joint angles based on offline data processing and does not use RC for training, so it is not sufficient in terms of performance such as real-time computation and accuracy.

[0012] This invention was made in view of these circumstances, and its purpose is to estimate finger angle information from electromyographic information with less computation. [Means for solving the problem]

[0013] To solve the above problems, a finger angle estimation system according to one aspect of the present invention comprises: an electromyographic signal acquisition unit attached to the surface of the skin to acquire electromyographic signals; a finger angle information acquisition unit to acquire finger joint angle information; and an estimation unit that performs machine learning using the electromyographic signals acquired by the electromyographic signal acquisition unit and the finger joint angle information acquired by the finger angle information acquisition unit as training data to estimate the finger joint angle from the electromyographic information. The estimation unit performs machine learning using RC (reservoir computing).

[0014] Another aspect of the present invention is a finger angle estimation method. This method includes the steps of: acquiring electromyographic signals using an electromyographic signal acquisition unit attached to the surface of the skin; acquiring finger joint angle information using a finger angle information acquisition unit; and estimating the finger joint angle from the electromyographic information by performing machine learning using the acquired electromyographic signals and the acquired finger joint angle information as training data. The estimation step involves machine learning using RC (reservoir computing).

[0015] A further aspect of the present invention is a finger angle estimation program. This program causes a computer to perform the following steps: acquiring electromyographic signals using an electromyographic signal acquisition unit attached to the surface of the skin; acquiring finger joint angle information using a finger angle information acquisition unit; and estimating the finger joint angle from the electromyographic information by performing machine learning using the acquired electromyographic signals and the acquired finger joint angle information as training data. The estimation step uses machine learning with RC (reservoir computing).

[0016] Furthermore, any combination of the above components, or any substitution of the components or expressions of the present invention between methods, apparatus, programs, temporary or non-temporary storage media recording programs, systems, etc., are also valid embodiments of the present invention. [Effects of the Invention]

[0017] According to the present invention, finger angle information can be estimated from electromyographic information with less computational effort. [Brief explanation of the drawing]

[0018] [Figure 1] This is a functional block diagram of the finger angle estimation system according to the first embodiment. [Figure 2] This is a functional block diagram of the finger angle estimation system, as a modified example of the finger angle estimation system shown in Figure 1. [Figure 3]It is a flowchart showing the processing procedure of the finger angle estimation method according to the second embodiment. [Figure 4] It is a schematic diagram showing the outline of the experimental system of this verification experiment. [Figure 5] It is a schematic diagram showing the state of the electrodes attached to the left arm of the subject. [Figure 6] It is a diagram showing the result of learning by ENS. [Figure 7] It is a diagram showing the result of learning by RNN.

Mode for Carrying Out the Invention

[0019] Hereinafter, the present invention will be described with reference to each drawing based on preferred embodiments. In the embodiments and modification examples, the same or equivalent components and parts are denoted by the same reference numerals, and redundant explanations will be omitted as appropriate. In addition, the dimensions of the parts in each drawing are appropriately enlarged or reduced for easy understanding. Also, in each drawing, some elements that are not important in explaining the embodiment are omitted. In addition, terms including ordinal numbers such as first and second are used to describe various components, but this term is used only for the purpose of distinguishing one component from another, and the components are not limited by this term.

[0020] [First Embodiment] FIG. 1 is a functional block diagram of a finger angle estimation system 1 according to the first embodiment. The finger angle estimation system 1 includes an electromyogram signal acquisition unit 11, a finger angle information acquisition unit 12, and an estimation unit 13. As will be described later, this embodiment is a composite of a plurality of hardware such as an electrode (electromyogram signal acquisition unit 11) attached to the skin, a bending sensor (finger angle information acquisition unit 12) attached to the finger, and a computer (estimation unit 13) that receives and estimates signals from each. In this sense, it is referred to as the "finger angle estimation system 1". However, since this can also be regarded as an integrated hardware, in that case, it may be referred to as the "finger angle estimation device 1" or the like.

[0021] The electromyography signal acquisition unit 11 is attached to the surface of the subject's skin and acquires the subject's preferred electromyography signal. For example, the electromyography signal acquisition unit 11 may acquire time-series data of the electromyography signal. By performing integral calculations on such time-series data, the angle of the finger joint can be estimated from the electromyography signal.

[0022] The number and mounting positions of the electromyographic signal acquisition units 11 can be any suitable configuration. For example, the electromyographic signal acquisition units 11 are electrodes attached to the surface of the subject's skin. Preferably, the electromyographic signal acquisition units 11 include multiple electrodes, which are arranged randomly or regularly. Because the electromyographic signal acquisition units 11 include multiple electrodes, and these electrodes are arranged randomly or regularly, electromyographic information can be acquired from multiple locations on the skin. By using these signals for analysis, the finger joint angle can be estimated with higher accuracy.

[0023] Preferably, the electromyographic signal acquisition unit 11 acquires electromyographic signals from multiple fingers. By acquiring electromyographic signals from multiple fingers, the finger joint angles can be estimated for multiple fingers.

[0024] Preferably, the electromyography signal acquisition unit 11 acquires electromyography signals for all five fingers. By acquiring electromyography signals for all five fingers, the finger joint angles can be estimated for all five fingers.

[0025] The finger angle information acquisition unit 12 acquires angle information of the finger joints. The finger angle information acquisition unit 12 can be anything that can accurately measure the angle of the finger joints. For example, the finger angle information acquisition unit 12 may include a bending sensor, a pressure sensor, a potentiometer, or a camera. By including a bending sensor, a pressure sensor, a potentiometer, or a camera in the finger angle information acquisition unit 12, the bending angle of the finger can be measured directly, so finger angle information can be acquired more accurately.

[0026] Preferably, the electromyographic signal acquisition unit 11 and the finger angle information acquisition unit 12 are attached to the skin surface of the arms or fingers on opposite sides of the body. The electromyographic signal acquisition unit 11 acquires electromyographic signals from the arm corresponding to the muscles on the side where fingers are missing. The finger angle information acquisition unit 12 acquires finger angle information from the healthy side. For example, the electromyographic signal acquisition unit 11 is attached to the skin surface of the left arm, and the finger angle information acquisition unit 12 acquires angle information from the fingers of the right hand. In this case, a correlation can be obtained between the electromyographic signals originating from the muscles of the left arm (corresponding to the muscles on the side where fingers are missing) and the angle of the fingers of the right hand (corresponding to the fingers on the healthy side). By arranging the electromyographic signal acquisition unit 11 and the finger angle information acquisition unit 12 on opposite sides of the body, data closer to that of an actual myoelectric prosthetic hand can be obtained.

[0027] Preferably, the electromyographic signal acquisition unit 11 acquires electromyographic signals under isometric contraction conditions. Isometric contraction refers to the movement of contracting a muscle while keeping its length constant. For example, it is a form of static contraction in which force is applied to a muscle without moving a joint, and the length of the muscle can be kept constant by fixing both ends of the muscle. As mentioned above, in recent years, when amputating fingers, it has become common to fix the stumped ends of the remaining muscles to the bone and apply tension to prevent muscle atrophy, so myoelectric prosthetic hand systems are generally based on electromyographic signals under isometric contraction conditions. Therefore, by acquiring electromyographic signals under isometric contraction conditions, the electromyographic signal acquisition unit 11 can decode isotonic, isometric, and isokinetic contraction movements from the information of the muscle performing isometric contraction, thus obtaining more useful electromyographic signals for the actual myoelectric prosthetic hand system.

[0028] The estimation unit 13 uses the electromyographic signal acquired by the electromyographic signal acquisition unit 11 and the finger joint angle information acquired by the finger angle information acquisition unit 12 as training data to perform machine learning and generate a model that estimates the finger joint angle when new electromyographic information is provided. At this time, the estimation unit 13 performs machine learning using RC (reservoir computing).

[0029] Since there is a valid correlation between human electromyography (EMG) signals and finger angles, machine learning can be performed using actually acquired EMG signals and finger joint angle information as training data. This allows for the estimation of finger joint angles when new EMG information is provided.

[0030] RC is a type of Recurrent Neural Network (RNN) with recurrent connections. The most distinctive feature of RC is that the network weights, called the input layer and reservoir layer, are set randomly and are not trained. RC trains only the output layer using linear regression such as the least squares method, and this process is extremely fast compared to conventional RNNs. In particular, RC is extremely useful for training using time-series signals, such as electromyography signals, as described in this disclosure.

[0031] Preferably, the RNN used for machine learning is an ESN (Echo State Network). An ESN is a type of RNN that learns only some parameters to improve learning efficiency and reduce computational complexity. By adjusting some of the connection weights of an RNN using a simple learning algorithm, an ESN enables high-speed machine learning, especially in time series forecasting. As shown in the experimental example described later, by applying an ESN to the machine learning of the estimation unit 13, the computation time for estimating finger angle information from electromyographic information can be significantly reduced compared to a conventional RNN.

[0032] Figure 2 is a functional block diagram of the finger angle estimation system 2, which is a modified version of the finger angle estimation system 1 in Figure 1. The finger angle estimation system 2 comprises an electromyographic signal acquisition unit 11, a finger angle information acquisition unit 12, an estimation unit 13, and a preprocessing unit 14. In other words, the finger angle estimation system 2 adds the preprocessing unit 14 to the configuration of the finger angle estimation system 1. The other components of the finger angle estimation system 2 are the same as those of the finger angle estimation system 1.

[0033] The preprocessing unit 14 squares the electromyographic signal acquired by the electromyographic signal acquisition unit 11 and performs preprocessing to normalize the entire squared value so that it is between 0 and 1.

[0034] By performing preprocessing by the preprocessing unit 14 on the electromyographic signal acquisition unit 11, the finger angle can be estimated more accurately and with less computation.

[0035] According to the embodiments described above, a system can be provided that estimates finger angle information from electromyographic information with less computational effort.

[0036] [Second Embodiment] Figure 3 is a flowchart showing the processing procedure for a finger angle estimation method according to the second embodiment. This finger angle estimation method includes a step S10 to acquire electromyographic signals, a step S20 to acquire finger joint angle information, and a step S30 to estimate the finger joint angle from the electromyographic information.

[0037] In step S10, this method acquires electromyographic signals using an electromyographic signal acquisition unit attached to the surface of the skin.

[0038] In step S20, this method uses a finger angle information acquisition unit to acquire angle information of the finger joints.

[0039] In step S30, this method performs machine learning using the acquired electromyographic signals and acquired finger joint angle information as training data to generate a model that estimates the finger joint angle when new electromyographic information is provided.

[0040] The estimation step involves machine learning using RC (reservoir computing).

[0041] According to this embodiment, a method for estimating finger angle information from electromyographic information with less computational effort can be provided.

[0042] [Third Embodiment] The third embodiment is a finger angle estimation program. The processing procedure of this program will be explained using Figure 3. This program causes the computer to perform the following steps: step S10 to acquire electromyographic signals, step S20 to acquire finger joint angle information, and step S30 to estimate the finger joint angle from the electromyographic information.

[0043] In step S10, this program acquires electromyographic signals using an electromyographic signal acquisition unit attached to the surface of the skin. Here, the electromyographic signal acquisition unit is a device that acquires electromyographic information, such as electrodes.

[0044] In step S20, the program uses a finger angle information acquisition unit to acquire finger joint angle information. Here, the finger angle information acquisition unit is a device that acquires finger joint angle information, such as a bending sensor, potentiometer, pressure sensor, or vibration sensor.

[0045] In step S30, the program performs machine learning using the acquired electromyographic signals and finger joint angle information as training data to generate a model that estimates the finger joint angle when new electromyographic information is provided.

[0046] The estimation step involves machine learning using RC (reservoir computing).

[0047] According to this embodiment, a program can be provided that estimates finger angle information from electromyographic information with less computational effort.

[0048] [Verification experiment] The inventors conducted experiments to verify the effects of the present disclosure. Figure 4 is a schematic diagram showing the outline of the experimental system used in this verification experiment.

[0049] The electromyography (EMG) signal acquisition unit consists of multiple electrodes, which are attached to the subject's left arm. The subject's left hand is secured by a pouch while gripping a cylindrical rod with a diameter of 30 mm and a total length of 150 mm. This fixes the end of the muscle in the left arm, maintaining a constant muscle length. When the subject applies force to their left arm to bend and straighten their fingers in this state, an EMG signal is generated under isometric contraction conditions. In this way, the electrodes acquire the EMG signal under isometric contraction conditions.

[0050] A bending sensor is attached to the fingers of the subject's right hand, and the bending angle of the finger joints is measured when the subject exerts force on their left arm.

[0051] The electromyographic signal information acquired by electrodes on the left arm is input to the intermediate layer of the estimation unit. The estimation unit uses machine learning with the electromyographic signal from the left arm and the angle information of the finger joints of the right hand as training data to generate a model that estimates the finger joint angle when new electromyographic information is provided. This estimation is performed using an ESN.

[0052] Figure 5 is a schematic diagram showing the electrodes attached to the subject's left arm. In this experiment, 20 electrodes (4 electrodes x 5 rows) were attached to acquire electromyographic signals related to the movement of the five fingers. The electrodes were spaced 20 mm apart, and the electrode rows were arranged to be equally spaced. This experiment employed a bipolar induction method using a total of three electrodes: one GND electrode attached near the wrist and two active electrodes.

[0053] The electromyography (EMG) and finger angle information necessary for learning were measured using a custom-made measuring instrument. The EMG signal measurement circuit consisted of a differential amplifier and a fourth-order high-pass filter (f c The circuit consisted of a 20Hz (=20Hz) and a non-inverting amplifier. The output from the measuring instrument was sampled at 1kHz. The measurement was performed with both hands facing each other, and data was sampled during a 90-second exercise while imagining both hands moving in the same way.

[0054] As mentioned above, the input to the Estimation System Number (ESN) and the training data consisted of 10 channels of electromyography (EMG) information and finger angle information. Each EMG signal was preprocessed by squaring it and normalizing the result so that the overall value ranged from 0 to 1. The measured data (90,000 steps × 10 channels) was split in half along the time axis, with the first half used as training data and the second half as test data. Training was performed offline, and ridge regression was used to adjust the output weights.

[0055] As a comparative example, we conducted a training experiment using an RNN with a network structure similar to that of the disclosed model. PyTorch 2.1.0 was used to construct this RNN. To reduce the training time for the RNN and facilitate comparison with the ESN, the training epoch count was set to 20.

[0056] In this experiment, the learning system using ESN did not utilize parallel computing on the GPU. On the other hand, the comparative learning using RNN utilized parallel computing on the GPU.

[0057] The parameters of the ESN are shown below. Number of inputs: 10 Number of outputs: 5 Number of nodes: 512 Leak rate: 0-1 Spectrum radius: 0.98

[0058] The parameters for the RNN are shown below. Number of inputs: 10 Number of outputs: 5 Number of nodes: 512 Batch size: 8 Optimization algorithm: SGD Loss function: MSELoss

[0059] Figure 6 shows the results of training using ENS. Figure 7 shows the results of training using RNN. In both figures, the horizontal axis represents time (minutes), and the vertical axis represents the finger bending angle. The graphs, from top to bottom, relate to the first finger (thumb), second finger (index finger), third finger (middle finger), fourth finger (ring finger), and fifth finger (little finger). The left half of each graph is used for training, and the right half is used for testing. The dotted line represents the measured value from the bending sensor, and the solid line represents the estimated value from training. To remove the effects of noise, the output values ​​are averaged over 20 data points.

[0060] The training time for the ESN and RNN datasets (90,000 steps x 10 channels) was approximately 16 seconds and 900 seconds, respectively.

[0061] The following shows the prediction errors for ESN and RNN. Mean Absolute Error (MAE) was used to calculate the prediction error. ESN MAE = 0.320 RNN MAE=0.491 Thus, when comparing the predicted values ​​of ESN and RNN, it can be seen that the RNN output has more noise superimposed on it.

[0062] As shown in Figure 6, the estimated finger angles in the learning results using ESNs agree well with the measured values. However, for the ring finger and little finger, there are occasional instances where the activity of one finger is estimated to be the activity of both fingers. This is thought to be due to the fact that the muscles that cause movement in the ring finger and little finger are very close together, as well as the fact that these two fingers are anatomically connected by tendons.

[0063] Both ESN and RNN training results show noise in the predicted values ​​overall. However, comparing ESN and RNN, the RNN appears to have more noise. In fact, comparing the MAE values, the prediction results from RNN are higher.

[0064] The reason the RNN had more noise compared to the ESN is likely because the number of epochs (epochs) set in this case was insufficient for learning the difference in input and output timescales. It is thought that increasing the number of epochs in the RNN could improve the learning results. However, the time required for learning increases proportionally to the number of epochs, which would impose a burden on the user in practical terms. Therefore, improving learning accuracy by increasing the number of epochs is not realistic.

[0065] The overall noise observed in both ESN and RNN is thought to be caused by the fact that the frequency of the input electromyographic information is very fast compared to the frequency of the predicted time-series information of finger angle, as well as noise introduced during measurement.

[0066] These noises can be reduced by taking a moving average of the output values ​​over a wide time interval, which is thought to yield a more stable output. Another approach is to introduce a network that further filters the output. However, digitally filtering the output values ​​consumes computational resources and is therefore not ideal. For this reason, it is considered better to improve prediction accuracy by applying analog preprocessing to the input information.

[0067] Regarding training time, one possible reason why RNNs take significantly longer to train than ESNs is that updating the weights in the hidden layers of RNNs is extremely complex.

[0068] Furthermore, while this experiment used a GPU to train the RNN, equipping an actual prosthetic hand with an expensive and heavy GPU is extremely difficult, both from an economic standpoint and in terms of the physical fatigue experienced by the subjects. Therefore, it is likely that training a real prosthetic hand would take an even longer time.

[0069] As demonstrated by the experimental results above, learning using RC, such as ESN, significantly reduces computation time and overall noise compared to conventional RNNs. This is expected to reduce the computational resources required for myoelectric prosthetic hand systems, lower the cost and weight of the prosthetics, and greatly contribute to reducing the economic and physical burden on prosthetic hand users.

[0070] [Each aspect of this disclosure] The embodiments of this disclosure described above are summarized below. One embodiment of the finger angle estimation system comprises: an electromyographic signal acquisition unit attached to the surface of the skin to acquire electromyographic signals; a finger angle information acquisition unit to acquire finger joint angle information; and an estimation unit that performs machine learning using the electromyographic signals acquired by the electromyographic signal acquisition unit and the finger joint angle information acquired by the finger angle information acquisition unit as training data to generate a model that estimates the finger joint angle when new electromyographic information is provided. The estimation unit performs machine learning using RC (reservoir computing).

[0071] According to this embodiment, a system can be provided that estimates finger angle information from electromyographic information with less computational effort.

[0072] In one embodiment of the finger angle estimation system, RC may be an ESN (Echo State Network).

[0073] According to this embodiment, a system can be provided that estimates finger angle information from electromyographic information with even less computational effort.

[0074] In one embodiment of the finger angle estimation system, the electromyography signal acquisition unit acquires time-series data of electromyography signals.

[0075] According to this embodiment, the angle of the finger joint can be estimated based on time-series data of electromyographic signals.

[0076] In one embodiment of a finger angle estimation system, the electromyographic signal acquisition unit includes multiple electrodes, which are arranged randomly or regularly.

[0077] According to this embodiment, electromyographic information can be obtained from multiple parts of the skin, allowing for more accurate estimation of finger joint angles.

[0078] In one embodiment of the finger angle estimation system, the electromyographic signal acquisition unit acquires electromyographic signals from multiple fingers.

[0079] According to this embodiment, the finger joint angles can be estimated for multiple fingers.

[0080] In one embodiment of the finger angle estimation system, the electromyography signal acquisition unit acquires electromyography signals for each of the five fingers.

[0081] According to this embodiment, the finger joint angles can be estimated for all five fingers.

[0082] In one embodiment of the finger angle estimation system, the electromyography signal acquisition unit is attached to the surface of the skin on either the left or right arm or finger. The electromyography signal acquisition unit acquires electromyography signals from the arm corresponding to the muscle on the side where the fingers are missing. The finger angle information acquisition unit acquires finger angle information from the healthy side.

[0083] According to this embodiment, electromyographic signals can be acquired from one arm, which is assumed to have missing fingers, while angle information of the fingers on the other side can be acquired, thus enabling the acquisition of data that is closer to that of an actual myoelectric prosthetic hand.

[0084] In one embodiment of the finger angle estimation system, the electromyographic signal acquisition unit may acquire electromyographic signals under isometric contraction conditions.

[0085] According to this embodiment, isotonic, isometric, and isokinetic contraction movements can be decoded from information about muscles performing isometric contraction.

[0086] In one embodiment of the finger angle estimation system, the finger angle information acquisition unit may include a bending sensor, a pressure sensor, a potentiometer, or a camera.

[0087] According to this embodiment, finger angle information can be obtained more accurately.

[0088] In one embodiment of the finger angle estimation system, a preprocessing unit may further include a preprocessing unit that squares the electromyographic signal acquired by the electromyographic signal acquisition unit and normalizes the entire squared value so that it is between 0 and 1.

[0089] According to this embodiment, the finger angle can be estimated more accurately and with less computation.

[0090] A finger angle estimation method in one embodiment includes the steps of: acquiring electromyographic signals using an electromyographic signal acquisition unit attached to the surface of the skin; acquiring finger joint angle information using a finger angle information acquisition unit; and estimation, which involves performing machine learning using the acquired electromyographic signals and the acquired finger joint angle information as training data to generate a model that estimates the finger joint angle when new electromyographic information is provided. The estimation step involves machine learning using RC (reservoir computing).

[0091] According to this embodiment, a method for estimating finger angle information from electromyographic information with less computational effort can be provided.

[0092] A further aspect of the present invention is a finger angle estimation program. This program causes a computer to perform the following steps: acquiring electromyographic signals using an electromyographic signal acquisition unit attached to the surface of the skin; acquiring finger joint angle information using a finger angle information acquisition unit; and generating an estimation step in which machine learning is performed using the acquired electromyographic signals and the acquired finger joint angle information as training data to generate a model that estimates the finger joint angle when new electromyographic information is given. The estimation step uses machine learning with RC (reservoir computing).

[0093] According to this embodiment, a program can be provided that estimates finger angle information from electromyographic information with less computational effort.

[0094] The present invention has been described above based on embodiments. The embodiments are illustrative, and it will be understood by those skilled in the art that various modifications are possible in combinations of their components and processing processes, and that such modifications also fall within the scope of the present invention.

[0095] In the above embodiment, the finger angle information acquisition unit 12 was a bending sensor. However, the finger angle information acquisition unit 12 is not limited to this, and may be any system or device that can appropriately measure the angle of the finger joint. For example, the finger angle information acquisition unit 12 may use a potentiometer, pressure sensor, vibration sensor, etc. Alternatively, the finger angle information acquisition unit 12 may be equipped with a camera and calculate the angle of the finger joint based on the image captured by this camera.

[0096] This modified version allows for greater flexibility in configuration.

[0097] In the verification experiment described above, the electrodes constituting the electromyographic signal acquisition unit were arranged in a vertical line along the subject's arm. However, this is not the only way in which the electrodes can be arranged; for example, they may be arranged randomly.

[0098] This modified version allows for greater flexibility in how the electrodes are arranged.

[0099] Any combination of the embodiments and modifications described above is also useful as an embodiment of this disclosure. The new embodiments resulting from such combinations possess the combined effects of both the combined embodiments and the modifications.

[0100] In understanding the technical concept abstracted from the embodiments, that technical concept should not be interpreted restrictively to the content of the embodiments. The embodiments and modifications described above are merely examples, and many design changes, such as changes, additions, and deletions of components, are possible. In the embodiments, the content in which such design changes are possible is emphasized with the notation "embodiment." However, design changes are also permitted in content without such notation. [Explanation of Symbols]

[0101] 1…Finger angle estimation system, 2…Finger angle estimation system, 11...Electromyography signal acquisition unit, 12...Finger angle information acquisition unit, 13...estimation section, 14…Pre-treatment section, S10...Step to acquire electromyographic signals. S20... A step to acquire finger joint angle information. S30... A step to estimate the angle of the finger joint from electromyographic information.

Claims

1. A unit that attaches to the surface of the skin to acquire electromyographic signals, A finger angle information acquisition unit that acquires finger joint angle information, An estimation unit generates a model that estimates the angle of a finger joint when new electromyographic signal information is provided, by performing machine learning using the electromyographic signal acquired by the electromyographic signal acquisition unit and the finger joint angle information acquired by the finger angle information acquisition unit as training data. Equipped with, The aforementioned estimation unit is a finger angle estimation system characterized by performing machine learning using RC (reservoir computing).

2. The finger angle estimation system according to claim 1, characterized in that the RC is ESN (Echo State Network).

3. The finger angle estimation system according to claim 1, characterized in that the electromyographic signal acquisition unit acquires time-series data of electromyographic signals.

4. The finger angle estimation system according to claim 1, characterized in that the electromyographic signal acquisition unit includes a plurality of electrodes, and the plurality of electrodes are arranged randomly or regularly.

5. The finger angle estimation system according to claim 3, characterized in that the electromyographic signal acquisition unit acquires electromyographic signals for multiple fingers.

6. The finger angle estimation system according to claim 3, characterized in that the electromyographic signal acquisition unit acquires electromyographic signals relating to five fingers.

7. The electromyographic signal acquisition unit and the finger angle information acquisition unit are attached to the surface of the skin on opposite arms or fingers. The electromyographic signal acquisition unit acquires electromyographic signals from the arm corresponding to the muscles on the side where the fingers are missing. The finger angle estimation system according to claim 1, characterized in that the finger angle information acquisition unit acquires finger angle information from the healthy side.

8. The finger angle estimation system according to claim 1, characterized in that the electromyographic signal acquisition unit acquires electromyographic signals under isometric contraction conditions.

9. The finger angle estimation system according to claim 1, characterized in that the finger angle information acquisition unit includes a bending sensor, a pressure sensor, a potentiometer, or a camera.

10. The finger angle estimation system according to claim 1, further comprising a preprocessing unit that squares the electromyographic signal acquired by the electromyographic signal acquisition unit and performs a preprocessing operation to normalize the entire squared value so that it is between 0 and 1.

11. A step of acquiring electromyographic signals using an electromyographic signal acquisition unit attached to the surface of the skin, The steps include: acquiring finger joint angle information using a finger angle information acquisition unit; The estimation step involves performing machine learning using acquired electromyographic signals and acquired finger joint angle information as training data to generate a model that estimates the finger joint angle when new electromyographic information is provided, and Includes, The aforementioned estimation step is a finger angle estimation method characterized by performing machine learning using RC (reservoir computing).

12. A step of acquiring electromyographic signals using an electromyographic signal acquisition unit attached to the surface of the skin, The steps include: acquiring finger joint angle information using a finger angle information acquisition unit; The estimation step involves performing machine learning using acquired electromyographic signals and acquired finger joint angle information as training data to generate a model that estimates the finger joint angle when new electromyographic information is provided, and Have the computer run it, The aforementioned estimation step is a finger angle estimation program characterized by performing machine learning using RC (reservoir computing).

Images