Upper limb joint angle prediction method based on long short-term memory network and sliding mode control

Abstract

The invention provides an upper limb joint angle prediction method based on a long short-term memory network and sliding mode control, which comprises the steps of data acquisition, data processing, upper limb dynamic model construction, design of the upper limb joint angle prediction method based on the long short-term memory network and the sliding mode control and the like. According to the method, the prediction accuracy is improved by fusing the advantages of the long and short-term memory network in sequence data processing and the characteristics of the sliding mode control in the aspects of processing uncertainty and robustness, the uncertainty and external interference of a system model are processed by using the sliding mode control, good control performance and prediction precision are kept, and the prediction accuracy is improved. The robustness and adaptability of the system are enhanced, the stable prediction effect is rapidly adjusted and maintained, accurate prediction of the upper limb joint angle is achieved, and the generalization ability and accuracy of a prediction model and the stability of system control are improved.

Description

Technical Field

[0001] The present invention belongs to the field of biomechanics, machine learning and control technology, and relates to a joint angle prediction method, and in particular to an upper limb joint angle prediction method based on a long short-term memory network and sliding mode control. Background Art

[0002] In the prior art, the prediction of upper limb joint angles is mainly achieved by constructing complex physiological models or directly using regression models. However, the physiological model is complex to construct and difficult to apply directly. Although the direct regression model is widely used, it has a strong dependence on the structure selection and network parameter determination, and the prediction results are easily affected by the training data, resulting in insufficient generalization ability. In addition, when dealing with complex dynamic systems, traditional control methods often face the challenges of uncertain factors such as parameter changes and disturbances, which affect the prediction accuracy and stability. The present invention hopes to provide an upper limb joint angle prediction method based on long short-term memory network and sliding mode control, which realizes accurate prediction of upper limb joint angles by integrating the advantages of long short-term memory network in processing sequence data and the characteristics of sliding mode control in processing uncertainty and robustness, thereby improving the generalization ability of the prediction model and the stability of system control. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides an upper limb joint angle prediction method based on long short-term memory network and sliding mode control, comprising the following steps:

[0004] Step 1: Data Collection

[0005] Collect the position, velocity, acceleration and other motion parameter data of each joint during the movement of the upper limbs in the sagittal plane, and collect the surface electromyographic signals of each joint;

[0006] Furthermore, the joints include shoulder joints, elbow joints and wrist joints; the surface electromyography signals include surface electromyography signals of anterior deltoid muscle, posterior deltoid muscle, biceps brachii, triceps brachii, radial extensor carpi and ulnar extensor carpi.

[0007] Step 2: Data Processing

[0008] Perform data preprocessing on surface electromyographic signals, including bandpass filtering, notch filtering, full-wave rectification, subsampling, smoothing envelope and normalization;

[0009] Furthermore, the bandpass filtering process is as follows: a low-pass filter is used to allow signals with a frequency lower than 500 Hz to pass through, and a high-pass filter is used to allow signals with a frequency higher than 20 Hz to pass through;

[0010] The notch filtering process is as follows: using a 60Hz second-order notch filter to eliminate power line noise;

[0011] The mathematical expression of the full-wave rectification is:

[0012] sEMG r (n)=|sEMG(n)|

[0013] Where sEMG(n) is the nth amplitude sample of the surface electromyographic signal after bandpass filtering and notch filtering, sEMG r (n) is the nth amplitude sample of the surface electromyographic signal after full-wave rectification;

[0014] The sub-sampling process is: to make the position, velocity, acceleration and sampling frequency of the surface electromyography signal of the joints consistent during the upper limb movement, and to sub-sample the surface electromyography signal. The mathematical expression is:

[0015]

[0016] Where N is the number of times the surface electromyography signal needs to be sub-sampled, sEMG s (n) is the amplitude of the nth surface electromyographic signal after subsampling;

[0017] The smoothing envelope process is: using a Buttonworth filter to perform smoothing envelope processing on the sub-sampled surface electromyography signal;

[0018] The normalization mathematical expression is:

[0019]

[0020] in, represents the value after data normalization, f i represents sample data, f min and f max Indicates the minimum and maximum values ​​in the sample data.

[0021] Step 3: Construct the upper limb dynamics model

[0022] Step 301: construct an upper limb dynamics model, the mathematical expression is:

[0023]

[0024] Among them, θ, are the position, velocity and acceleration vector of the joint respectively, M(θ) is the inertia matrix, is the centrifugal force and Coriolis force matrix, G(θ) is the upper arm and forearm gravity matrix, τ h is the upper limb joint torque;

[0025] Step 302: Obtain the upper limb joint torque τ by collecting the position, velocity, and acceleration data of the upper limb joints and based on the upper limb dynamics modelh ;

[0026] Step 4: Design an upper limb joint angle prediction method based on long short-term memory network and sliding mode control

[0027] Step 401: Use the long short-term memory network to establish an upper limb joint torque prediction model based on the surface electromyography signal, use the surface electromyography signal after data processing as the input of the long short-term memory network, and the joint torque is the output of the long short-term memory network to obtain the predicted joint torque based on the long short-term memory network.

[0028] Step 402: Design a discrete upper limb joint angle prediction model using predicted joint torque and sliding mode control. The mathematical expression is:

[0029]

[0030] in, is the predicted state variable, is the derivative of the predicted state variable, θ k and are the predicted positions and velocities of the shoulder, elbow, and wrist joints, δ s is the sampling time, y(Θ k ,δ k ) represents the upper limb motion state variable, u(δ k ) is the sliding mode control, δ k is the sampling time of upper limb motion state variables and sliding mode control;

[0031] y(Θ k ,δ k ) is:

[0032]

[0033] u(δ k ) is:

[0034]

[0035] S k =E k

[0036] E k =Θ dk -Θ k

[0037] Among them, γ and ξ are coefficients, S k is the sliding surface, Θ dk is the expected state variable, is the derivative of the desired state variable, E kis the error between the expected state variable and the predicted state variable.

[0038] Beneficial effects of the present invention:

[0039] The present invention discloses an upper limb joint angle prediction method based on long short-term memory network and sliding mode control. The upper limb joint angle prediction method is established by using long short-term memory network and sliding mode control. The advantages of long short-term memory network in processing sequence data and the characteristics of sliding mode control in processing uncertainty and robustness are integrated to improve the accuracy of prediction. The sliding mode control is used to process the uncertainty and external interference of the system model, maintain good control performance and prediction accuracy, enhance the robustness and adaptability of the system, quickly adjust and maintain a stable prediction effect, realize accurate prediction of upper limb joint angles, and improve the generalization ability and accuracy of the prediction model and the stability of system control. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 A method and system flow chart of upper limb joint angle prediction based on long short-term memory network and sliding mode control according to the present invention;

[0041] Figure 2 A comparison diagram of predicted shoulder joint angles of an upper limb joint angle prediction method based on a long short-term memory network and sliding mode control according to the present invention;

[0042] Figure 3 A comparison diagram of predicted elbow joint angles of an upper limb joint angle prediction method based on a long short-term memory network and sliding mode control according to the present invention;

[0043] Figure 4 A comparison diagram of predicted wrist joint angles of an upper limb joint angle prediction method based on a long short-term memory network and sliding mode control according to the present invention;

[0044] Figure 5 A comparison diagram of predicted shoulder joint angle errors of an upper limb joint angle prediction method based on a long short-term memory network and sliding mode control according to the present invention;

[0045] Figure 6 A comparison diagram of predicted elbow joint angle errors of an upper limb joint angle prediction method based on a long short-term memory network and sliding mode control according to the present invention;

[0046] Figure 7 This is a comparison diagram of the predicted wrist joint angle errors of the upper limb joint angle prediction method based on long short-term memory network and sliding mode control described in the present invention. DETAILED DESCRIPTION

[0047] Figure 1As shown, the present invention provides an upper limb joint angle prediction method based on long short-term memory network and sliding mode control, and the specific steps are as follows:

[0048] Step 1: Data Collection

[0049] Step 101: collecting motion parameter data such as position, velocity, acceleration, etc. of the shoulder joint, elbow joint, and wrist joint during the movement of the upper limb on the sagittal plane;

[0050] Step 102: The muscles related to the movement of the shoulder joint in the sagittal plane are the anterior deltoid muscle and the posterior deltoid muscle, and surface electromyographic signals of the anterior deltoid muscle and the posterior deltoid muscle are collected;

[0051] Step 103: The muscles related to the movement of the elbow joint in the sagittal plane are the biceps and triceps, and surface electromyographic signals of the biceps and triceps are collected;

[0052] Step 104: The muscles related to the movement of the wrist joint in the sagittal plane are the radial extensor carpi and the ulnar extensor carpi, and surface electromyographic signals of the radial extensor carpi and the ulnar extensor carpi are collected;

[0053] Step 2: Data Processing

[0054] The surface electromyographic signals of the anterior deltoid, posterior deltoid, biceps, triceps, radial extensor carpi and ulnar extensor carpi are preprocessed. The specific process is as follows:

[0055] Step 201: performing band-pass filtering, the specific process is: using a low-pass filter to allow the frequency signal below 500 Hz to pass, and using a high-pass filter to allow the frequency signal above 20 Hz to pass;

[0056] Step 202: performing notch filtering, the specific process is: using a 60 Hz second-order notch filter to eliminate power line noise;

[0057] Step 203: Perform full-wave rectification, the mathematical expression is:

[0058] sEMG r (n)=|sEMG(n)|

[0059] Where sEMG(n) is the nth amplitude sample of the surface electromyographic signal after bandpass filtering and notch filtering, sEMG r (n) is the nth amplitude sample of the surface electromyographic signal after full-wave rectification;

[0060] Step 204: sub-sampling is performed. The specific process is: the position, velocity, acceleration and sampling frequency of the surface electromyography signal of the joints during the upper limb movement are kept consistent, and the surface electromyography signal is sub-sampled. The mathematical expression is:

[0061]

[0062] Where N is the number of times the surface electromyography signal needs to be sub-sampled, sEMG s (n) is the amplitude of the nth surface electromyographic signal after subsampling;

[0063] Step 205: performing smoothing envelope, the specific process is: using a Buttonworth filter to perform smoothing envelope processing on the sub-sampled surface electromyography signal;

[0064] Step 206: Perform normalization, the mathematical expression is:

[0065]

[0066] in, represents the value after data normalization, f i represents sample data, f min and f max Indicates the minimum and maximum values ​​in the sample data;

[0067] Step 3: Construct the upper limb dynamics model

[0068] The upper limb dynamics model lays the foundation for designing the upper limb joint angle prediction method. The specific process is as follows:

[0069] Step 301: construct an upper limb dynamics model, the mathematical expression is:

[0070]

[0071] Among them, θ, are the position, velocity and acceleration vector of the joint respectively, M(θ) is the inertia matrix, is the centrifugal force and Coriolis force matrix, G(θ) is the upper arm and forearm gravity matrix, τ h is the upper limb joint torque;

[0072] Step 302: Obtain the upper limb joint torque τ by collecting the position, velocity, and acceleration data of the shoulder joint, elbow joint, and wrist joint and based on the upper limb dynamics model h ;

[0073] Step 4: Design an upper limb joint angle prediction method based on long short-term memory network and sliding mode control

[0074] In this step, in order to improve the generalization ability of the prediction model and the stability of the system control, an upper limb joint angle prediction method based on long short-term memory network and sliding mode control is designed to achieve accurate prediction of the upper limb joint angle. The specific process is as follows:

[0075] Step 401: Using the long short-term memory network, an upper limb joint torque prediction model based on the surface electromyographic signal is established. The surface electromyographic signals of the anterior deltoid, posterior deltoid, biceps, triceps, radial extensor carpi and ulnar extensor carpi are used as the input of the long short-term memory network after data processing. The joint torques of the shoulder joint, elbow joint and wrist joint are the output of the long short-term memory network. The long short-term memory network sets the number of hidden layer neurons to 300, the learning rate to 0.005 and the threshold to 1, and obtains the predicted joint torque based on the long short-term memory network.

[0076] Step 402: Design a discrete upper limb joint angle prediction model using predicted joint torque and sliding mode control. The mathematical expression is:

[0077]

[0078] in, is the predicted state variable, is the derivative of the predicted state variable, θ k and are the predicted positions and velocities of the shoulder, elbow, and wrist joints, δ s is the sampling time, y(Θ k ,δ k ) represents the upper limb motion state variable, u(δ k ) is the sliding mode control, δ k is the sampling time of upper limb motion state variables and sliding mode control;

[0079] y(Θ k ,δ k )The specific mathematical expression is:

[0080]

[0081] u(δ k )The specific mathematical expression is:

[0082]

[0083] S k =E k

[0084] E k =Θ dk -Θ k

[0085] Among them, γ and ξ are coefficients, S k is the sliding surface, Θ dk is the expected state variable, is the derivative of the desired state variable, E k is the error between the expected state variable and the predicted state variable.

[0086] Figure 2-Figure 7 The upper limb joint angle prediction method based on long short-term memory network and sliding mode control is compared with the upper limb joint angle prediction method based on long short-term memory network. The long short-term memory network and sliding mode control are used to achieve accurate prediction of the upper limb joint angle, which is specifically reflected in that the predicted upper limb joint angle based on long short-term memory network and sliding mode control accurately tracks the expected angle, and the predicted angle error tends to zero.

Claims

1. A method for predicting upper limb joint angles based on long short-term memory network and sliding mode control, Features: The following steps are involved: Step 1: Data Collection Collect the position, velocity, acceleration motion parameter data of each joint during the movement of the upper limbs in the sagittal plane, and collect the surface electromyographic signals of each joint; Step 2: Data Processing Perform data preprocessing on surface electromyographic signals, including bandpass filtering, notch filtering, full-wave rectification, subsampling, smoothing envelope and normalization; Step 3: Construct the upper limb dynamics model Step 301: construct an upper limb dynamics model, the mathematical expression is: Among them, θ, are the position, velocity and acceleration vector of the joint respectively, M(θ) is the inertia matrix, is the centrifugal force and Coriolis force matrix, G(θ) is the upper arm and forearm gravity matrix, τ h is the upper limb joint torque; Step 302: Obtain the upper limb joint torque τ by collecting the position, velocity, and acceleration data of the upper limb joints and based on the upper limb dynamics model h ; Step 4: Design an upper limb joint angle prediction method based on long short-term memory network and sliding mode control Step 401: Use the long short-term memory network to establish an upper limb joint torque prediction model based on the surface electromyography signal, use the surface electromyography signal after data processing as the input of the long short-term memory network, and the joint torque is the output of the long short-term memory network to obtain the predicted joint torque based on the long short-term memory network. Step 402: Design a discrete upper limb joint angle prediction model using predicted joint torque and sliding mode control. The mathematical expression is: in, is the predicted state variable, is the derivative of the predicted state variable, θ k and are the predicted positions and velocities of the shoulder, elbow, and wrist joints, δ s is the sampling time, y(Θ k ,δ k ) represents the upper limb motion state variable, u(δ k ) is the sliding mode control, δ k is the sampling time of upper limb motion state variables and sliding mode control; y(Θ k ,δ k ) is: u(δ k ) is: S k =E k E k =Θ dk -I k Among them, γ and ξ are coefficients, S k is the sliding surface, Θ dk is the expected state variable, is the derivative of the desired state variable, E k is the error between the expected state variable and the predicted state variable.

2. According to claim 1, a method for predicting upper limb joint angles based on long short-term memory network and sliding mode control, Features: The joints described in step 1 include shoulder joints, elbow joints and wrist joints; the surface electromyographic signals include surface electromyographic signals of anterior deltoid muscle, posterior deltoid muscle, biceps brachii, triceps brachii, radial extensor carpi and ulnar extensor carpi.

3. The upper limb joint angle prediction method based on long short-term memory network and sliding mode control according to claim 1, Features: The bandpass filtering process described in step 2 is: using a low-pass filter to allow signals with frequencies below 500 Hz to pass through, and using a high-pass filter to allow signals with frequencies above 20 Hz to pass through; The notch filtering process is as follows: using a 60Hz second-order notch filter to eliminate power line noise; The mathematical expression of the full-wave rectification is: sEMG r (n)=|sEMG(n)| Where sEMG(n) is the nth amplitude sample of the surface electromyographic signal after bandpass filtering and notch filtering, sEMG r (n) is the nth amplitude sample of the surface electromyographic signal after full-wave rectification; The sub-sampling process is: to make the position, velocity, acceleration and sampling frequency of the surface electromyography signal of the joints consistent during the upper limb movement, and to sub-sample the surface electromyography signal. The mathematical expression is: Where N is the number of times the surface electromyography signal needs to be sub-sampled, sEMG s (n) is the amplitude of the nth surface electromyographic signal after subsampling; The smoothing envelope process is: using a Buttonworth filter to perform smoothing envelope processing on the sub-sampled surface electromyography signal; The normalization mathematical expression is: in, represents the value after data normalization, f i represents sample data, f min and f max Indicates the minimum and maximum values ​​in the sample data.

Images