Adaptive control method and system of upper limb rehabilitation robot based on game theory and sEMG

Abstract

The present invention discloses an upper limb rehabilitation robot adaptive control method and system based on game theory and sEMG. First, the motion trajectory of the robot during the operation of the subject is designed; the nonlinear dynamic relationship between the surface electromyographic signal and the terminal force is established by constructing a three-layer neural network based on the back propagation neural network muscle force estimation model of sEMG; the human-computer interaction system is analyzed by using the game theory principle, the role of the robot is derived, the control rate between the two is updated by using the Nash equilibrium, and the adaptive weight factor of the robot and the subject is determined, and finally the robot is made to adaptively adjust its own training mode according to the weight coefficient obtained according to the game theory principle according to the subject's motion intention during the operation, so as to realize that the training mode of the rehabilitation robot can adapt to the human upper limb motion intention. The method fully considers the changes in the training process of the subject and provides corresponding training tasks in real time, thereby improving the training efficiency of the rehabilitation robot.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning technology, and in particular to the field of machine learning technology for rehabilitation robots and rehabilitation training, and mainly relates to an upper limb rehabilitation robot adaptive control method and system based on game theory and sEMG. Background Art

[0002] According to the World Health Organization (WHO) and existing official population statistics, the world's aging problem is becoming increasingly serious, leading to a continuous increase in the incidence of age-related diseases such as stroke and Parkinson's disease. Stroke is also known as cerebral infarction. In China, there are more than 2 million new stroke patients each year, and more than 1.5 million people die from stroke each year. In addition, with the influence of external factors such as lifestyle, the incidence and mortality of stroke have been on the rise in recent years. Since brain damage is irreversible, the lost function cannot be fully restored, and motor dysfunction is its most common sequelae. As an important limb in body movement, posture maintenance, balance, etc., the decline or loss of the upper limbs not only seriously affects the quality of life of patients, but also brings pain and heavy economic burden to patients and their families. Many stroke patients will experience long-term motor disability and need to undergo labor-intensive exercise therapy as early as possible and for a long time, which brings a heavy burden to the medical system.

[0003] With the rapid development of artificial intelligence, robotics has gradually been applied to the field of rehabilitation medicine. Rehabilitation robots that integrate virtual reality, intelligent force feedback, intelligent evaluation and other technologies have been developed at home and abroad. Such equipment can increase sensory input through computer screens, interactive virtual reality devices, and tactile sensing systems. The computer system displays virtual scenes and directional tasks to patients, and patients perform resistance and non-resistance exercise training through mechanical parts attached to their limbs to restore their motor functions. The main advantage of this type of technology is that it can provide patients with high-dose, high-intensity, high-repeatability, and targeted training, and the training mode is more interesting than traditional rehabilitation training, which can mobilize the enthusiasm of patients for training; sensors can record the patient's movement trajectory, exercise time and other information and conduct intelligent evaluation and analysis. Therapists or doctors can adjust the treatment plan in time based on this information, which is conducive to accelerating the patient's rehabilitation process.

[0004] Rehabilitation training methods are one of the important research contents of rehabilitation robot systems. They are divided into passive and active methods, which are suitable for different rehabilitation stages of patients. Passive rehabilitation training refers to the way in which the rehabilitation robot drives the human arm for rehabilitation training; active training refers to the training method in which the human arm exerts force and the rehabilitation robot provides assistance or resistance. Studies have shown that even if the patient does not have the ability to complete the movement, the willingness to actively exercise is necessary for rehabilitation training. Therefore, human-computer interaction control is very important in rehabilitation training. The desired control scheme is to identify the human body's movement intention by estimating the human joint torque to adaptively select the robot training mode. The surface electromyography signal (sEMG) of human skin is a signal generated by biological organisms that can accurately and quickly reflect the user's movement intention. The electromyography signal generated by muscle contraction is 20 to 150 ms ahead of the joint torque, and has a smaller delay and higher signal-to-noise ratio than the force sensor. The surface electromyography signal may be an ideal source for estimating the active joint torque. However, in existing robot-assisted therapies, most of them are based on force sensors for force measurement, and surface electromyographic signals are rarely used to estimate force during movement. On the other hand, as the patient's rehabilitation progresses, resistance training is crucial for strengthening muscles. Therefore, during the patient's rehabilitation training, adaptive mode switching and flexible control of terminal force output are very important. Most of the current human-computer interaction control methods are not flexible enough for collaborative tasks. Generally, they can only perform single auxiliary training or resistance training, or hard switching between multiple controllers, which will cause the overall control input to be discontinuous, and then cause the robot to shake violently during mode switching, and even endanger human safety. Summary of the invention

[0005] The present invention is aimed at the problems existing in the multi-mode switching process of robots in the prior art, and provides an upper limb rehabilitation robot adaptive control method based on game theory and sEMG. First, a trajectory following task is designed, and the motion trajectory of the robot is controlled within the training time during the operation of the subject. Then, a muscle force estimation model based on the sEMG back propagation neural network (BPNN) is designed, and a nonlinear dynamic relationship between the surface electromyographic signal and the terminal force is established by constructing a three-layer neural network; the human-computer interaction system is analyzed using the principles of game theory, the role of the robot is derived, the Nash equilibrium is used to update the control rate between the two, and the adaptive weight factors of the robot and the subject are determined. When the weight factor Q of the robot is positive, the robot provides auxiliary force; when the weight factor Q of the robot is negative, the robot provides impedance force. Finally, the robot adaptively adjusts its own training mode according to the weight coefficient obtained according to the game theory principle according to the subject's movement intention during the operation, so as to achieve the training mode of the rehabilitation robot being able to adapt to the human upper limb movement intention. The control method in this case can adaptively change the working mode of the rehabilitation robot according to the human body's movement intention, and can fully consider the changes in the subject's training progress to provide corresponding training tasks in real time, thereby ensuring the subject's involvement in the training process, and effectively stimulate the nerves to induce neural function reorganization, thereby improving the training efficiency of the rehabilitation robot.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is: an adaptive control method of an upper limb rehabilitation robot based on game theory and sEMG, comprising the following steps:

[0007] S1, designing a trajectory following task: the trajectory following task is to control the robot to move within the training time during the subject's operation, and the task at least includes the reference trajectory of the subject's movement and the training time;

[0008] S2, recognition of the subject's movement intention: a muscle force estimation model based on sEMG back-propagation neural network is designed, and a nonlinear dynamic relationship between surface electromyographic signals and terminal force is established by constructing a three-layer neural network; the muscle force estimation model based on back-propagation neural network is divided into three layers, namely input layer, hidden layer and output layer; wherein the input data is the surface electromyographic signal smoothed after filtering and normalization; the output is the estimated force F;

[0009] S3, adaptive switching of training mode: Analyze the human-computer interaction system using game theory principles, deduce the role of the robot, use Nash equilibrium to update the control rate between the two, and determine the adaptive weight factors of the robot and the subject.

[0010] Q+Q h ≡C

[0011] Among them, Q is the weight factor of the robot; Qh is the weight factor of the subject; C is a constant;

[0012] When the robot's weight factor Q is positive, the robot provides auxiliary force; when the robot's weight factor Q is negative, the robot provides impedance force;

[0013] S4, adaptive control strategy: During the training process, the robot adaptively adjusts its own training mode according to the weight coefficients obtained according to the game theory principles based on the subject's movement intention during the operation.

[0014] In order to achieve the above purpose, the present invention also adopts a technical solution: an upper limb rehabilitation robot adaptive control system based on game theory and sEMG, including a trajectory design module, a motion intention recognition module and an adaptive control module.

[0015] The trajectory design module is used to design the trajectory following task, that is, the motion trajectory design of the robot during the subject's operation;

[0016] The movement intention recognition module is used to design a muscle force estimation model based on the sEMG back propagation neural network, and a nonlinear dynamic relationship between the surface electromyography signal and the end force is established by constructing a three-layer neural network;

[0017] The adaptive control module analyzes the human-computer interaction system based on the information obtained from the trajectory design module and the motion intention recognition module using the principles of game theory, derives the role of the robot, adopts Nash equilibrium to update the control rate between the robot and the subject, and determines the adaptive weight factor between the two. When the robot's weight factor is positive, the robot provides auxiliary force; when the robot's weight factor is negative, the robot provides impedance force; when the robot's weight factor is zero, the robot does not provide force and is in free training mode.

[0018] Compared with the prior art, the present invention has the following beneficial effects:

[0019] (1) This method uses the back propagation neural network (BPNN) method based on surface electromyography (sEMG) to establish a force estimation model and predict the human body's movement intention. It can adaptively provide the best training mode according to the human body's movement intention during rehabilitation training, and provide the best training intensity according to the degree of human recovery during training, thereby maintaining the subject's active training state and maximizing the training efficiency.

[0020] (2) Adopting the game theory control principle for adaptive mode switching, this method has stable switching without severe jitter, and can provide a training method that adapts to the human body's movement intention, stimulates the subject's neural perception of force and movement control, always maintains the best training state, and avoids useless training process as much as possible. This control strategy ensures the stability and real-time performance of adaptive switching, and provides a solution for intelligent and efficient neurorehabilitation. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 It is a schematic diagram of the overall adaptive closed-loop control framework of the control method of the present invention;

[0022] Figure 2 Schematic diagram of the trajectory following task designed in step S1 of the control method of the present invention;

[0023] Figure 3 The flow chart of generating a smoothed surface electromyographic signal after filtering and normalization in step S2 of the control method of the present invention;

[0024] Figure 4 Schematic diagram of a force estimation model based on sEMG in the control method of the present invention;

[0025] Figure 5 This is a brief example diagram of the adaptive control process of the test example of the present invention. DETAILED DESCRIPTION

[0026] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0027] Example 1

[0028] Adaptive control method of upper limb rehabilitation robot based on game theory and sEMG, such as Figure 1 As shown, the following steps are included:

[0029] S1, designing a trajectory following task: the trajectory following task is to control the robot to move within the training time during the subject's operation, and the task at least includes the reference trajectory of the subject's movement and the training time;

[0030] like Figure 2 As shown, in this embodiment, the target trajectory is designed as a 1 / 4 curve of a standard circle, and three reference points A, B, and C are evenly displayed on the trajectory as guide points. During the subject's operation, the robot end is controlled to follow the trajectory within a specified time based on the reference points. The subject moves from the horizontal starting point to the vertical ending point and then returns to the horizontal starting point in one experiment, which is a cycle, that is, the motion trajectory is ABCBA, and each experiment is performed for 5-8 cycles.

[0031] S2, design of sEMG-based back propagation neural network (BPNN) muscle force estimation model:

[0032] The nonlinear dynamic relationship between the surface electromyographic signal and the end force is established by constructing a three-layer neural network. The BPNN model is divided into three layers, namely the input layer, the hidden layer and the output layer. The input data is the smoothed surface electromyographic signal (RSEMG) after filtering and normalization. For the convenience of representation, RSEMG is used in this paper to represent the smoothed surface electromyographic signal after filtering and normalization; the estimated force F is the output of BPNN.

[0033] In rehabilitation training, the movement state is generally monitored and characterized by the surface electromyographic signal (sEMG) of the human skin. Because the original electromyographic signal is subject to a lot of interference, the smoothed surface electromyographic signal (RSEMG) after filtering and normalization is used, such as Figure 3 As shown, that is, the original surface electromyographic signal collected is first subjected to high-pass filtering, full-wave rectification and low-pass filtering in sequence, and then normalized to obtain a smooth surface electromyographic signal (RSEMG) after filtering and normalization. The original electromyographic signal in this embodiment can select the biceps brachii, the long head of the triceps brachii, the short head of the triceps brachii and the brachioradialis muscle, which are mainly responsible for the motor function of the upper limbs, as the muscle groups to be analyzed, because the above muscle groups are mainly responsible for the motor function of the upper limbs, which is more accurate and efficient.

[0034] In addition, in view of the differences between the subjects and the purpose of providing each subject with a personalized training method in the subsequent process, each subject was given a three-minute pre-experiment before the experiment to collect electromyographic data and establish a force estimation model. The electromyographic sensor and force sensor data collected in a short period of time can be used to train the model, and this modeling process can avoid fatigue in the pre-experiment of the subjects.

[0035] In this step, the sEMG-based back propagation neural network (BPNN) muscle force estimation model is as follows: Figure 4 As shown, RSEMG is used as the input of the neural network, and the estimated force (F) is used as the output. A three-layer neural network is constructed to establish a nonlinear dynamic relationship between the surface electromyography signal and the end force.

[0036] S3, adaptive switching of training mode:

[0037] The human-computer interaction system is analyzed using the principles of game theory. A cost function is established based on the input states of the robot and the subject. The Nash equilibrium is used to update the control rate between the two, the weight of the cost function is estimated, the adaptive weight factors of the robot and the subject are determined, and the role of the robot is determined based on the output weight factor.

[0038] Cost function:

[0039]

[0040]

[0041] Nash equilibrium:

[0042] u=-Lξ,L≡B T P

[0043]

[0044] u h =-L h ξ,L h ≡B T P h

[0045]

[0046] Adaptive weight factor: Q+Q h ≡C

[0047] Where: ξ represents the motion state (trajectory error and speed), QQ h represents the weight coefficient of the robot and the subject, C is a constant; uu h represents the motor commands of the robot and the subject, LL h represents the feedback gain of the robot and the subject, AB represents the state and input matrix, PP h represents the Riccati matrix of the robot and the subject; A r A h Represents the state matrix of the subject and the robot; (.) T Indicates transpose.

[0048] S4, adaptive control strategy:

[0049] The control strategy of adaptively adjusting the control training mode according to the human body's movement intention is as follows: During the training process, the subject's movement intention is first estimated through the surface electromyographic signals of the subject's arm during the operation, that is, the arm end force is estimated.

[0050] Furthermore, according to the subject's movement intention and the robot's auxiliary status, the corresponding weight coefficient is obtained by using the game theory principle and fed back to the robot's motor, thereby adaptively adjusting the robot's training mode.

[0051] When the robot weight is positive, the human's neural intention and muscle ability cannot fully guide the robot, and the robot is in auxiliary mode; on the contrary, the subject's movement intention is enhanced, and when the robot weight is zero, the robot does not provide force in free training mode; when the robot weight is negative, it provides resistance in resistance training mode. That is, the human body's movement intention can automatically control the working mode of the rehabilitation robot, and adaptive rehabilitation training can be completed through the subject's own movement intention.

[0052] Test example:

[0053] Assume that some of the surface electromyographic signals collected by the subjects during the training process are as follows Figure 5 As shown in (a), after filtering and normalization shown in step S2 above, a smooth surface electromyographic signal RSEMG will be obtained, and the muscle force estimation model described in step S2 will obtain an estimated muscle force Force, such as Figure 5 (b), and then through the adaptive mode switching method described in step S3, the respective role assignments of the subject and the robot are obtained, and then the corresponding auxiliary or resistance training mode is provided, such as Figure 5 (c) shows that at 0 seconds, the subject’s adaptive weight factor Q h The value is 0, and the robot's adaptive weight factor Q value is also 0, that is, no force is applied; between 0-20s, according to Q h The robot will adaptively update its own Q value when the subject's Q h When the value is larger, that is, the movement ability is enhanced, the robot's Q value will decrease, that is, the auxiliary force will be reduced or the corresponding impedance will be provided; when the subject's Q h When the value decreases, that is, the movement ability decreases, the robot's Q value will increase, that is, reduce the impedance or provide corresponding auxiliary force. h If the value is stable, that is, the movement ability remains basically unchanged, the Q value of the robot will also remain stable, and the human-computer interaction will reach a balanced state.

[0054] The control method in this case uses a desktop force feedback robot as the carrier and research object, and the non-advantageous arm of a healthy subject as the experimental object. The study obtains the end-of-arm force based on sEMG and then estimates the human body's movement intention to achieve adaptive switching of the robot training mode, ensuring that the subject is provided with the best assistance or resistance intensity at any stage of the training process, and always using the best personalized and intelligent training mode, as well as ensuring the subject's involvement during the training process, effectively improving the training efficiency of robot-assisted rehabilitation, and providing technical support for the rehabilitation robot to enter clinical application faster.

[0055] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. An adaptive control method for upper limb rehabilitation robot based on game theory and sEMG, characterized in that: The steps include: S1, designing a trajectory following task: the trajectory following task is to control the robot to move within the training time during the subject's operation, and the task at least includes the reference trajectory of the subject's movement and the training time; S2, recognition of the subject's movement intention: a muscle force estimation model based on sEMG back propagation neural network is designed, and a nonlinear dynamic relationship between surface electromyographic signals and terminal force is established by constructing a three-layer neural network; the muscle force estimation model based on sEMG back propagation neural network is divided into three layers, namely, input layer, hidden layer and output layer; wherein the input data is the surface electromyographic signal smoothed after filtering and normalization; the output is the estimated force F; S3, adaptive switching of training mode: Analyze the human-computer interaction system using game theory principles, deduce the role of the robot, use Nash equilibrium to update the control rate between the two, and determine the adaptive weight factors of the robot and the subject. Q+Q h ≡C Among them, Q is the weight factor of the robot; Q h is the weight factor of the subject; C is a constant; When the robot's weight factor Q is positive, the robot provides auxiliary force; when the robot's weight factor Q is negative, the robot provides impedance force; S4, adaptive control strategy: During the training process, the robot adaptively adjusts its own training mode according to the weight coefficients obtained according to the game theory principles based on the subject's movement intention during the operation.

2. The adaptive control method for an upper limb rehabilitation robot based on game theory and sEMG as claimed in claim 1, characterized in that: The input data of the muscle force estimation model based on the back-propagation neural network in step S2 is the surface electromyography signal that is smoothed after filtering and normalization. The surface electromyography signal that is smoothed after filtering and normalization is obtained by sequentially processing the collected original surface electromyography signal with high-pass filtering, full-wave rectification, low-pass filtering, and normalization.

3. The adaptive control method for an upper limb rehabilitation robot based on game theory and sEMG as claimed in claim 2, characterized in that: The original surface electromyographic signal is taken from the electromyographic signal of the biceps brachii, the long head of the triceps brachii, the short head of the triceps brachii or the brachioradialis muscle as the muscle group to be analyzed.

4. The adaptive control method for an upper limb rehabilitation robot based on game theory and sEMG as claimed in claim 1 or 3, characterized in that: In the process of analyzing the human-computer interaction system using the game theory principle in step S3, a cost function is established according to the input states of the robot and the subject. The cost function is: Among them, U and U h denote the cost functions of the robot and the subject, respectively; t denotes time; ξ denotes motion state; u and u h represent the motor commands of the robot and the subject respectively; Q is the weight factor of the robot; Q h is the weight factor of the subject; (.) T represents transpose; t0 represents the start time.

5. The adaptive control method for upper limb rehabilitation robot based on game theory and sEMG as claimed in claim 4, characterized in that: In step S3, Nash equilibrium is used to update the control rate between the robot and the subject. Specifically, the Nash equilibrium is: u=-Lξ,L≡B T P the h =-L h ξ,L h ≡B T P h Where: L and L h represents the feedback gain of the robot and the subject, AB represents the state and input matrix, PP h represents the Riccati matrix of the robot and the subject; A r A h Represents the state matrix of the subject and the robot; (.) T Indicates transpose.

6. The adaptive control method for an upper limb rehabilitation robot based on game theory and sEMG as claimed in claim 5, characterized in that: In step S4, when the robot weight is zero, the robot does not provide force and is in free training mode.

7. An adaptive control system for an upper limb rehabilitation robot based on game theory and sEMG using the method of claim 1, characterized in that: Including trajectory design module, motion intention recognition module and adaptive control module, The trajectory design module is used to design the trajectory following task, that is, the motion trajectory design of the robot during the subject's operation; The movement intention recognition module is used to design a muscle force estimation model based on the sEMG back propagation neural network, and a nonlinear dynamic relationship between the surface electromyography signal and the end force is established by constructing a three-layer neural network; The adaptive control module analyzes the human-computer interaction system based on the information obtained from the trajectory design module and the motion intention recognition module using the principles of game theory, derives the role of the robot, adopts Nash equilibrium to update the control rate between the robot and the subject, and determines the adaptive weight factor between the two. When the robot's weight factor is positive, the robot provides auxiliary force; when the robot's weight factor is negative, the robot provides impedance force; when the robot's weight factor is zero, the robot does not provide force and is in free training mode.

Images