A human knee joint torque test system method based on surface electromyography signal and its application

Through a system based on surface electromyography signal, combined with three-dimensional motion capture and force measurement treadmill, and using LSTM neural network, the knee torque testing problem that the existing technology cannot be applied to movement at any speed is solved, and diversified scenarios such as military Accurate torque measurement and evaluation in

CN110801226BActive Publication Date: 2024-11-08XI AN JIAOTONG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN201911060681.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2019-11-01
Publication Date
2024-11-08
Estimated Expiration
2039-11-01

AI Technical Summary

Technical Problem

The existing technology cannot be applied to the knee torque test of the human body when it moves at any speed, and the traditional constant speed torque test method is difficult to generalize to different scenarios, especially in applications such as rapid marching and combat rescue in the military field, which lacks Applicable technology.

Method used

The human knee torque test system based on surface electromyography signals is adopted, combined with a three-dimensional motion capture system and a three-dimensional force measurement treadmill, and the inverse dynamic model is constructed by synchronously collecting electromyography signals, kinematic data and sole pressure, and using LSTM The neural network establishes a knee torque test model to achieve accurate measurement and prediction of knee torque.

Benefits of technology

Accurate knee torque measurement under any speed movement, improves the applicability and accuracy of the test, does not rely on the laboratory environment, is suitable for diverse scenarios such as military, and improves the torque control and efficiency evaluation of lower limbs and assists exoskeletons. .

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN110801226B_ABST
    Figure CN110801226B_ABST
Patent Text Reader

Abstract

A human knee joint torque test system, method and application based on surface electromyographic signal, comprising the following steps: step 1: electromyographic signal collection and feature extraction of four muscles of the user's lower limb, namely, the vastus lateralis, semitendinosus, biceps femoris long head and medial gastrocnemius; step 2: synchronously collect the user's leg kinematic data during walking using a three-dimensional motion capture system, solve the user's knee joint angle, and synchronously collect the user's foot plantar pressure during walking using a force measuring treadmill, construct a simplified dynamic model of the user's leg, perform inverse dynamics solution, and obtain the user's knee joint calculation torque; step 3: construct a training data set and a test data set of a long short-term memory (LSTM) neural network for knee joint torque test, and complete the training of the neural network model for knee joint torque test using the data obtained in step 1 and step 2. The present invention has the characteristics of wide application range and high test accuracy.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of lower limb power-assisted exoskeleton and biomechanical-electrical integration, and in particular to a human knee joint torque testing system, method and application based on surface electromyography signals. Background Art

[0002] At present, in the military field, due to the rapid march of soldiers, the demand for different application scenarios such as combat rescue and ammunition transportation is becoming more and more intense. The lower limb active power-assisted exoskeleton has become an important equipment to enhance the strength and endurance of soldiers, improve the speed and range of movement, reduce the physical energy consumption of soldiers themselves, and reduce the damage of weight-bearing transportation, thereby increasing combat capabilities. As a key basic technology for the torque control of the lower limb active power-assisted exoskeleton and the evaluation of the power-assisted efficiency of the lower limb active power-assisted exoskeleton, the human joint torque test technology can not only provide data support for the realization of flexible torque control, but also directly measure the human body's movement state and effort. The isokinetic torque test method is a traditional human joint torque test method. It mainly tests the constant angular velocity movement of a single joint or a combination of multiple joints with the assistance of an isokinetic muscle force test system. It is only applicable to the case of the human body in isokinetic movement, and cannot be applied to the case of the human body in arbitrary speed movement. The inverse dynamics solution of the human joint torque requires the collection of the three-dimensional pressure of the sole of the foot during the human body movement. The collection of this data needs to be collected in a three-dimensional force measurement treadmill in a laboratory environment. This method is difficult to generalize to any scenario. Summary of the invention

[0003] In order to overcome the above-mentioned deficiencies of the prior art, the purpose of the present invention is to provide a human knee joint torque testing system, method and application based on surface electromyography signals, which has the characteristics of wide application range and high testing accuracy.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] A human knee joint torque test system based on surface electromyography signals comprises an 8-channel surface electromyography signal collector, a three-dimensional motion capture system, a three-dimensional force measuring treadmill and a time synchronizer. The 8-channel surface electromyography signal collector can collect surface electromyography signals of the tester's lower limbs, the three-dimensional motion capture system can capture the motion data of the tester's lower limb knee joints, the three-dimensional force measuring treadmill can collect the three-dimensional force between the tester's sole and the treadmill during walking, and the time synchronizer can ensure the synchronization of the three devices, namely, the 8-channel surface electromyography signal collector, the three-dimensional motion capture system and the three-dimensional force measuring treadmill, in data collection.

[0006] A human knee joint torque testing method based on surface electromyography signals comprises the following steps:

[0007] Step 1: Collect and extract the electromyographic signals of the four muscles of the user's lower limbs: vastus lateralis, semitendinosus, long head of biceps femoris, and medial head of gastrocnemius;

[0008] Step 2: During step 1, a three-dimensional motion capture system is used to synchronously collect the user's leg kinematic data during walking, and the user's knee joint angle is calculated. At the same time, a force measuring treadmill is used to synchronously collect the plantar pressure of the user's foot during walking, and a simplified dynamic model of the user's leg is constructed. Inverse dynamics is solved to obtain the calculated torque of the user's knee joint;

[0009] Step 3: Construct the training data set and test data set of the long short-term memory (LSTM) neural network for knee joint torque test, and use the data obtained in steps 1 and 2 to complete the training of the neural network model for knee joint torque test.

[0010] The step 1 specifically includes:

[0011] (1) Collection and preprocessing of electromyographic signals:

[0012] The right knee joint of the subject was selected as the test object. Before the test, the subject was asked to do appropriate relaxation activities, and the skin at the vastus lateralis, semitendinosus, long head of biceps femoris, and medial head of gastrocnemius of the right leg was wiped with alcohol as the collection points of electromyographic signals. The electromyographic signals of the four muscles of the right leg of the subject were synchronously collected at a speed of 1500HZ when the subject walked at a speed of 2m / s for 1min, and the original electromyographic signals were high-pass filtered to eliminate motion artifacts and filter out 50Hz power frequency interference using a zero-lag 4th-order Butterworth filter (30Hz) and a 50Hz notch filter, and then full-wave rectified;

[0013] (2) EMG signal feature extraction

[0014] The electromyographic signal is a discrete signal. The preprocessed electromyographic signal is used as a window of 30 sampling points to extract the average electromyographic value (AEMG) and the root mean square value (RMS). The calculation formulas of AEMG and RMS are as follows:

[0015]

[0016]

[0017] Where Data[i] is the electromyographic signal; N is the window size;

[0018] Then the mean power frequency (MPF) and median frequency (MF) features are extracted from the preprocessed electromyographic signal. The calculation formulas of MPF and MF are as follows:

[0019]

[0020]

[0021] where f s is the sampling frequency, S(f) is the power spectral density function of the electromyographic signal, and finally the nearest neighbor method is used to downsample the characteristic signal to 100 Hz.

[0022] The step 2 comprises:

[0023] The 3D motion capture system in step 1 is used to synchronously collect the kinematic data of the user's right leg during the 1-minute walking process, and the user's knee joint angle is calculated. At the same time, the force measuring treadmill is used to synchronously collect the plantar pressure of the user's right foot during the walking process, and a simplified dynamic model of the user's right leg is constructed. The inverse dynamics solution is performed to obtain the calculated torque of the user's knee joint;

[0024] (1) Design a user's right leg kinematic data collection experiment. Paste four markers on the user's right leg's hip joint, knee joint, ankle joint, and toe, respectively. Use a 3D motion capture system to obtain the coordinate changes of the four markers in 3D space.

[0025] Since the user's right leg is parallel to the XOY plane of the 3D motion capture system, only the X and Y coordinates of the marker points are needed to solve the knee joint rotation angle. The coordinates of the marker points of the hip joint, knee joint, ankle joint and the outer center of the foot can be expressed as (x hip ,y hip ), (x knee ,y knee ), (x ankle ,y ankle ) and (x toe ,y toe );

[0026] The centroid coordinates of each limb segment are solved by interpolation method as follows:

[0027]

[0028]

[0029]

[0030] Where (x 1 ,y 1 ), (x 2 ,y 2 ) and (x 3 ,y 3 ) are the coordinates of the center of mass of the thigh, calf and foot, respectively, l 1 , l 2 and l 3The lengths of the thigh, calf and foot are respectively, i.e. the length from the hip joint to the knee joint, from the knee joint to the ankle joint and from the ankle joint to the toe. 1a , l 2a and l 3a The distances from the hip joint to the center of mass of the thigh, the knee joint to the center of mass of the calf, and the ankle joint to the center of mass of the foot;

[0031] In order to obtain the velocity and acceleration of the center of mass motion, the first-order and second-order differentials of the motion trajectory are performed:

[0032]

[0033]

[0034] The angle θn between the centroid-joint line of each limb segment and the horizontal ground is selected as the characteristic angle. The slope of the straight line can be used to calculate θn, and it is necessary to judge θ n Is it greater than 90°?

[0035] If θ<90°, then:

[0036]

[0037] If θ>90°, then:

[0038]

[0039] In order to obtain the angular velocity and acceleration of the characteristic angle of each limb segment, the first-order and second-order differentials of the angle are performed:

[0040]

[0041]

[0042] The knee joint angle is the angle between the calf and the thigh, which can be calculated using the following formula:

[0043] θ K =θ 1 -θ 2

[0044] (2) Construct a simplified dynamic model of the human lower limbs, and calculate the knee joint torque of the human body during walking based on the kinematic information of the human lower limbs and the plantar pressure. The human lower limbs can be simplified into a rigid three-link model. During walking, the gait is divided into a support phase and a swing phase according to whether the foot is in contact with the ground. The simplified model of the lower limbs in the swing phase can be regarded as a three-link model with s fixed, and the simplified model of the lower limbs in the support phase can be regarded as a three-link model with the plantar fixed.

[0045] First, the foot is studied separately and regarded as a rigid body. According to the Newton-Euler equation, the dynamic equation of the foot can be expressed as:

[0046]

[0047]

[0048] Among them, F 2,3 is the force exerted by the calf on the foot; F y F is the vertical component of the ground force on the foot; x G is the parallel ground force on the foot; 3 is the weight of the foot; M 2,3 M is the moment of force exerted by the calf on the foot; 3 is the foot driving torque; I 3 I is the moment arm of the calf to the center of mass of the foot; y F y The moment arm about the center of mass of the foot; I x F x The moment arm about the center of mass of the foot;

[0049] The ankle joint torque can be calculated according to the foot dynamics equation, which includes the plantar pressure and is also the ankle joint torque in the support phase:

[0050]

[0051] Removing the plantar pressure term from the above equation gives the ankle joint torque in the swing phase. The above equation is applicable to the entire gait cycle. When calculating the swing phase, only the plantar pressure needs to be set to 0.

[0052]

[0053] The calf is studied and regarded as a rigid body. According to the Newton-Euler equation, the dynamic equation of the calf can be expressed as:

[0054]

[0055] Among them, F 1,2 G is the force exerted by the thigh on the calf; 2 is the weight of the calf; M 1,2 M is the moment of force from the thigh to the calf; 2 I is the driving torque of the calf; 2 I is the moment arm of the calf; 2a The moment arm of the calf's center of mass where the thigh acts on the calf;

[0056] According to the calf dynamics equation, the knee joint torque can be solved:

[0057]

[0058] The step 3 constructs a training data set and a test data set for the knee joint torque test LSTM neural network, and trains the neural network based on the data obtained in steps 1 and 2.

[0059] The four eigenvalues ​​of the four muscles, the knee angle and the knee torque in 1 minute obtained in step 1 and step 2 are divided into the first 40s data set and the last 20s data set according to time. The first 40s data set is used as the training set of the neural network, and the last 20s data set is used as the test set of the neural network. The four eigenvalues ​​of the four muscles and the knee angle data in the training set are captured respectively. 30 data points x starting from the tth data point <t> As the input of the LSTM neural network, the 30 data points y of the knee joint torque data in the training set starting from the tth data point are captured <t> As the output of the LSTM neural network;

[0060] The input to the hidden layer of the LSTM neural network includes the current input x of the network <t> and the hidden layer output vector a at the previous moment <t-1> , and the hidden layer state C <t-1> , the hidden layer adds a forget gate f <t> , input gate i <t> And the output gate o <t> . Forget Gate f <t> Decide to discard state C <t> What information in x <t> and a <t-1> Can it be used in state C? <t> The update of the input gate i <t> Decision, after the forget gate and input gate, state C <t> The update is completed, the output gate o <t> To determine the state C <t> How does it work on a <t> In the calculation of the forget gate, input gate, output gate and hidden layer output a <t> and status update C <t> The calculation expression is as follows:

[0061] f <t> =σ(W f ·[a <t-1> ,x <t> ]+b f )

[0062] i <t> =σ(W i ·[a <t-1> ,x <t> ]+b i )

[0063] o <t> =σ(W o ·[a <t-1> ,x <t> ]+b o )

[0064] c <t> =f <t> ·a <t-1> +x <t> tanh(W c ·[a <t-1> ,x <t> ]+b c )

[0065] a <t> =o <t> ·tanh(c <t> )

[0066] Among them, σ is the sigmoid function, tanh is the hyperbolic tangent function, W f , W i , W c , W o and b f , b i , b c , b o They are the weight parameters and bias parameters that need to be trained in the forget gate, input gate, and output gate respectively;

[0067] The output layer loss function uses a square loss function to evaluate the error between the actual output and the expected output, and uses the error as the basis for back propagation optimization to minimize the error. The loss function is as follows:

[0068]

[0069] Among them, Y is the calculated torque value of the knee joint; f(x) is the estimated torque value of the knee joint; L(Y,f(x)) is the loss function.

[0070] The root mean square error (RMSE) and the cross correlation coefficient (CORR) are used to evaluate the quality of the human knee joint torque prediction model. RMSE is used to measure the error between the model estimate and the true value, and CORR is used to describe the correlation between two variables. The closer the RMSE value is to 0, the more accurate the corresponding neural network model estimate is. On the contrary, the closer the CORR value is to 1, the better the model fitting effect is. The calculation formulas of RMSE and CORR are as follows:

[0071]

[0072]

[0073] Among them, Y act is the knee joint torque calculated by inverse dynamics, is the average value of the calculated torque, Y pre is the estimated moment of the neural network model, is the average value of the estimated moment of the model, and N is the length of the test sample sequence data.

[0074] The parameters obtained through the testing method are used to control the lower limb power-assisted exoskeleton robot.

[0075] Beneficial effects of the present invention:

[0076] The present invention is easy to operate and can maintain test accuracy. The three-dimensional motion capture system can accurately measure the user's kinematic data, which is used to calculate the user's knee joint angle, and the user's knee joint angle data obtained thereby has high accuracy. The force measuring treadmill can accurately collect the user's plantar pressure, which is used to solve the user's knee joint torque by inverse dynamics, thereby obtaining the accurate user's knee joint torque. The surface electromyograph has a high sampling rate and can monitor the user's muscle activity in real time to reflect the size of the muscle strength. All three devices are equipped with a time synchronization interface, which can perform synchronous data acquisition.

[0077] The established human knee joint torque test neural network model has higher practicality than the traditional human body torque test method. This method is not only independent of the laboratory environment, but also applicable to the human body moving at any speed. The present invention utilizes the nonlinear fitting and long-term memory capabilities of the LSTM neural network, introduces the surface electromyographic signal reflecting the size of human muscle force, and establishes a mapping relationship between the surface electromyographic signal and the human knee joint torque. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] Figure 1 It is a block diagram of an overall method for estimating the human knee joint torque using the human lower limb electromyographic signals and knee joint angle signals.

[0079] Figure 2 This is a schematic diagram of the positions of the vastus lateralis, vastus lateralis, biceps femoris long head and medial head of gastrocnemius in the human lower limbs.

[0080] Figure 3 It is a schematic diagram of the kinematic model of a single lower limb of the human body.

[0081] Figure 4 This is a simplified diagram of the dynamic model of the human body's single lower limb support phase (right) and swing phase (left).

[0082] Figure 5 This is a schematic diagram of the hidden layer nodes of the LSTM neural network for knee joint torque test.

[0083] Figure 6Schematic diagram of an embodiment of knee joint torque test.

[0084] Figure 7 This is a comparison chart of the torque estimated by the LSTM neural network and the torque calculated by inverse dynamics. DETAILED DESCRIPTION

[0085] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0086] See also Figure 1 , the present invention constructs a LSTM neural network model for knee joint torque test, and realizes the knee joint torque estimated by inputting surface electromyographic signals and outputting knee joint angle. On the one hand, the surface electromyographic signals at the positions of the vastus lateralis, semitendinosus, long head of biceps femoris, and medial head of gastrocnemius of the lower limbs of the user during walking, as well as the knee joint angle signals of the user during walking, calculate the time domain characteristics and frequency domain characteristics of the electromyographic signals, as the input of the neural network model, and output the estimated torque of the knee joint torque based on the initial weights of the neural network model. On the other hand, the user's lower limb kinematic information and plantar pressure information are collected, and the Newton-Euler method is used to perform inverse dynamics solution to obtain the calculated torque of the knee joint. The square of the error between the calculated torque and the estimated torque is used as the objective function of LSTM neural network training, and the Adam algorithm is used to perform offline iterative training of the LSTM neural network to achieve the weight update of the neural network, and finally the input lower limb surface electromyographic signals and knee joint angles are obtained, and the corresponding stable knee joint torque test model is output. The specific implementation plan is as follows:

[0087] Step 1: Collection and feature extraction of muscle electrical signals of the vastus lateralis, semitendinosus, long head of biceps femoris, and medial head of gastrocnemius in the human lower limbs. The specific methods are as follows:

[0088] (1) Collection and preprocessing of electromyographic signals:

[0089] See also Figure 2 The right knee joint of the subject was selected as the test object. Before the test, the subject was asked to do appropriate relaxation activities, and the skin at the vastus lateralis, semitendinosus, long head of biceps femoris and medial head of gastrocnemius of the right leg was wiped with alcohol as the collection points of electromyographic signals. A four-channel surface electromyograph was used to synchronously collect the electromyographic signals of the four muscles of the right leg of the subject when he walked at a speed of 2m / s for 1min at a speed of 1500HZ. A zero-lag 4th-order Butterworth filter (30Hz) and a 50Hz notch filter were used to high-pass filter the original electromyographic signals to eliminate motion artifacts and filter out 50Hz power frequency interference, and then full-wave rectified.

[0090] (2) EMG signal feature extraction

[0091] The electromyographic signal is a discrete signal. The preprocessed electromyographic signal is used as a window of 30 sampling points to extract the average electromyographic value (AEMG) and the root mean square value (RMS). The calculation formulas of AEMG and RMS are as follows:

[0092]

[0093]

[0094] Where Data[i] is the electromyographic signal; N is the window size.

[0095] Then, the mean power frequency (MPF) and median frequency (MF) features are extracted from the preprocessed electromyographic signal. The calculation formulas of MPF and MF are as follows:

[0096]

[0097]

[0098] where f s is the sampling frequency, S(f) is the power spectral density function of the electromyographic signal, and finally the nearest neighbor method is used to downsample the characteristic signal to 100 Hz.

[0099] Step 2: During step 1, use the 3D motion capture system to synchronously collect the kinematic data of the user's right leg during 1 minute of walking, calculate the user's knee joint angle, and use the force measuring treadmill to synchronously collect the plantar pressure of the user's right foot during walking, build a simplified dynamic model of the user's right leg, perform inverse dynamics solution, and obtain the calculated torque of the user's knee joint.

[0100] (1) Design a user's right leg kinematic data collection experiment. Figure 3 , four marker points are respectively attached to the hip joint, knee joint, ankle joint and toe of the user's right leg, and the coordinate changes of the four marker points in the three-dimensional space are obtained by using the three-dimensional motion capture system.

[0101] Since the user's right leg is parallel to the XOY plane of the 3D motion capture system, only the X and Y coordinates of the marker points are needed to solve the knee joint rotation angle. The coordinates of the marker points of the hip joint, knee joint, ankle joint and the outer center of the foot can be expressed as (x hip ,y hip ), (x knee ,y knee ), (x ankle ,y ankle ) and (x toe ,y toe ).

[0102] The centroid coordinates of each limb segment are solved by interpolation method as follows:

[0103]

[0104]

[0105]

[0106] Where (x 1 ,y 1 ), (x 2 ,y 2 ) and (x 3 ,y 3 ) are the coordinates of the center of mass of the thigh, calf and foot, respectively, l 1 , l 2 and l 3 The lengths of the thigh, calf and foot are respectively, i.e. the length from the hip joint to the knee joint, from the knee joint to the ankle joint and from the ankle joint to the toe. 1a , l 2a and l 3a The distances from the hip joint to the center of mass of the thigh, from the knee joint to the center of mass of the calf, and from the ankle joint to the center of mass of the foot.

[0107] In order to obtain the velocity and acceleration of the center of mass motion, the first-order and second-order differentials of the motion trajectory are performed:

[0108]

[0109]

[0110] In order to facilitate the subsequent dynamics calculation, the angle θn between the center of mass of each limb segment and the joint line and the horizontal ground is selected as the characteristic angle. The slope of the straight line can be used to calculate θn, and it is necessary to judge θ n Is it greater than 90°?

[0111] If θ<90°, then:

[0112]

[0113] If θ>90°, then:

[0114]

[0115] In order to obtain the angular velocity and acceleration of the characteristic angle of each limb segment, the first-order and second-order differentials of the angle are performed:

[0116]

[0117]

[0118] The knee joint angle is the angle between the calf and the thigh, which can be calculated using the following formula:

[0119] θ K =θ 1 -θ 2

[0120] (2) Construct a simplified dynamic model of the human lower limbs and calculate the knee joint torque of the human body during walking based on the kinematic information of the human lower limbs and the plantar pressure; see Figure 4 The human lower limbs can be simplified into a rigid three-link model. During walking, the gait is divided into a support phase and a swing phase according to whether the foot is in contact with the ground. The simplified model of the lower limbs in the swing phase can be regarded as a three-link model with s fixed, and the simplified model of the lower limbs in the support phase can be regarded as a three-link model with the sole fixed.

[0121] First, the foot is studied separately and regarded as a rigid body. According to the Newton-Euler equation, the dynamic equation of the foot can be expressed as:

[0122]

[0123]

[0124] Among them, F 2,3 is the force exerted by the calf on the foot; F y F is the vertical component of the ground force on the foot; x G is the parallel ground force on the foot; 3 is the weight of the foot; M 2,3 M is the moment of force exerted by the calf on the foot; 3 is the foot driving torque; I 3 I is the moment arm of the calf to the center of mass of the foot; y F y The moment arm about the center of mass of the foot; I x F x The moment arm about the center of mass of the foot.

[0125] The ankle joint torque can be calculated based on the foot dynamics equation, which includes the plantar pressure and is therefore also the ankle joint torque in the support phase:

[0126]

[0127] Removing the plantar pressure term from the above equation yields the ankle joint torque during the swing phase. The above equation is applicable to the entire gait cycle. When calculating the swing phase, only the plantar pressure needs to be set to 0.

[0128]

[0129] The calf is studied and regarded as a rigid body. According to the Newton-Euler equation, the dynamic equation of the calf can be expressed as:

[0130]

[0131] Among them, F 1,2 G is the force exerted by the thigh on the calf; 2 is the weight of the calf; M 1,2 M is the moment of force from the thigh to the calf; 2 I is the driving torque of the calf; 2 I is the moment arm of the calf; 2a It is the moment arm of the calf's center of mass where the thigh acts on the calf.

[0132] According to the calf dynamics equation, the knee joint torque can be solved:

[0133]

[0134] Step 3: Construct the training data set and test data set of the LSTM neural network for knee torque test, and train the neural network based on the data obtained in steps 1 and 2.

[0135] The four eigenvalues ​​of the four muscles, the knee angle and the knee torque in 1 minute obtained in step 1 and step 2 are divided into the first 40s data set and the last 20s data set according to time. The first 40s data set is used as the training set of the neural network, and the last 20s data set is used as the test set of the neural network. The four eigenvalues ​​of the four muscles and the knee angle data in the training set are captured respectively. 30 data points x starting from the tth data point are <t> As the input of the LSTM neural network, the 30 data points y of the knee joint torque data in the training set starting from the tth data point are captured <t> As the output of the LSTM neural network.

[0136] The hidden layer node design of LSTM neural network can be found in Figure 5 , the input of the hidden layer of the LSTM neural network includes the current input x of the network <t> and the hidden layer output vector a at the previous moment <t-1> , and the hidden layer state C <t-1> The hidden layer adds a forget gate f <t> , input gate i <t> And the output gate o <t> . Forget Gate f <t> Decide to discard state C <t> What information in x <t> and a <t-1> Can it be used in state C? <t> The update of the input gate i <t> After the forget gate and input gate, state C<t> The update of is completed. Output gate o <t> To determine the state C <t> How does it work on a <t> The forget gate, input gate, output gate, and hidden layer output a <t> and status update C <t> The calculation expression is as follows:

[0137] f <t> =σ(W f ·[a <t-1> ,x <t> ]+b f )

[0138] i <t> =σ(W i ·[a <t-1> ,x <t> ]+b i )

[0139] o <t> =σ(W o ·[a <t-1> ,x <t> ]+b o )

[0140] c <t> =f <t> ·a <t-1> +x <t> tanh(W c ·[a <t-1> ,x <t> ]+b c )

[0141] a <t> =o <t> ·tanh(c <t> )

[0142] Where σ is the sigmoid function and tanh is the hyperbolic tangent function. f , W i , W c , W o and b f , b i , b c , b o They are the weight parameters and bias parameters that need to be trained in the forget gate, input gate, and output gate respectively.

[0143] The output layer loss function uses a square loss function to evaluate the error between the actual output and the expected output, and uses the error as the basis for back propagation optimization to minimize the error. The loss function is as follows:

[0144]

[0145] Among them, Y is the calculated torque value of the knee joint; f(x) is the estimated torque value of the knee joint; L(Y,f(x)) is the loss function.

[0146] The root mean square error (RMSE) and the cross correlation coefficient (CORR) are used to evaluate the quality of the human knee joint torque prediction model. RMSE is used to measure the degree of error between the model estimate and the true value, and CORR is used to describe the correlation between two variables. The closer the RMSE value is to 0, the more accurate the corresponding neural network model estimate is. On the contrary, the closer the CORR value is to 1, the better the model fitting effect is. The calculation formulas of RMSE and CORR are as follows:

[0147]

[0148]

[0149] Among them, Y act is the knee joint torque calculated by inverse dynamics, is the average value of the calculated torque, Y pre is the estimated moment of the neural network model, is the average value of the estimated moment of the model, and N is the length of the test sample sequence data.

[0150] Example: A healthy male was selected as the subject, with a height of 171 cm, a weight of 55.2 kg, and no major diseases. The subject was asked to perform a walking experiment. For specific scenarios, see Figure 6 The subject walked on the three-dimensional force measuring treadmill at a speed of 1.2 m / s for 1 min, and the surface electromyographic signals of the vastus lateralis, semitendinosus, long head of biceps femoris, and medial head of gastrocnemius of the subject's right leg were synchronously collected. At the same time, the three-dimensional coordinate information of the marking points of the hip joint, knee joint, ankle joint and the center of the outer side of the foot of the subject's right leg, as well as the three-dimensional interaction force between the subject's right foot and the force measuring treadmill were collected. These data were used to verify the knee joint torque test method of the present invention. The test results are shown in Figure 7 .

Claims

1. A method for testing human knee joint torque based on surface electromyography signals, characterized in that: The steps include: Step 1: Collect and extract the electromyographic signals of the four muscles of the user's lower limbs: vastus lateralis, semitendinosus, long head of biceps femoris, and medial head of gastrocnemius; Step 2: During step 1, a three-dimensional motion capture system is used to synchronously collect the user's leg kinematic data during walking, and the user's knee joint angle is calculated. At the same time, a force measuring treadmill is used to synchronously collect the plantar pressure of the user's foot during walking, and a simplified dynamic model of the user's leg is constructed. Inverse dynamics is solved to obtain the calculated torque of the user's knee joint; Step 3: Construct a training data set and a test data set for the long short-term memory neural network of the knee joint torque test, and use the data obtained in steps 1 and 2 to complete the training of the neural network model of the knee joint torque test; The step 2 comprises: The 3D motion capture system in step 1 is used to synchronously collect the kinematic data of the user's right leg during the 1-minute walking process, and the user's knee joint angle is calculated. At the same time, the force measuring treadmill is used to synchronously collect the plantar pressure of the user's right foot during the walking process, and a simplified dynamic model of the user's right leg is constructed. The inverse dynamics solution is performed to obtain the calculated torque of the user's knee joint; (1) Design a user's right leg kinematic data collection experiment. Paste four markers on the user's right leg's hip joint, knee joint, ankle joint, and toe, respectively. Use a 3D motion capture system to obtain the coordinate changes of the four markers in 3D space. Since the user's right leg is parallel to the XOY plane of the 3D motion capture system, only the X and Y coordinates of the marker points are needed to solve the knee joint rotation angle. The coordinates of the marker points of the hip joint, knee joint, ankle joint and the outer center of the foot can be expressed as (x hip ,y hip ), (x knee ,y knee )(x ankle ,y ankle ) and (x toe ,y toe ) The centroid coordinates of each limb segment are solved by interpolation method as follows: Where (x1, y1), (x2, y2) and (x3, y3) are the coordinates of the center of mass of the thigh, calf and foot respectively, l1, l2 and l3 are the lengths of the thigh, calf and foot respectively, i.e. the lengths from the hip joint to the knee joint, from the knee joint to the ankle joint and from the ankle joint to the toe, l 1a , l 2a and l 3a The distances from the hip joint to the center of mass of the thigh, the knee joint to the center of mass of the calf, and the ankle joint to the center of mass of the foot; In order to obtain the velocity and acceleration of the center of mass motion, the first-order and second-order differentials of the motion trajectory are performed: The angle θn between the centroid-joint line of each limb segment and the horizontal ground is selected as the characteristic angle. The slope of the straight line can be used to calculate θn, and it is necessary to judge θ n Is it greater than 90°? If θ<90°, then: If θ>90°, then: In order to obtain the angular velocity and acceleration of the characteristic angle of each limb segment, the first-order and second-order differentials of the angle are performed: The knee joint angle is the angle between the calf and the thigh, which can be calculated using the following formula: i k =θ1-θ2 (2) Construct a simplified dynamic model of the human lower limbs, and calculate the knee joint torque of the human body during walking based on the kinematic information of the human lower limbs and the plantar pressure. The human lower limbs can be simplified into a rigid three-link model. During walking, the gait is divided into a support phase and a swing phase according to whether the foot is in contact with the ground. The simplified model of the lower limbs in the swing phase can be regarded as a three-link model with s fixed, and the simplified model of the lower limbs in the support phase can be regarded as a three-link model with the plantar fixed. First, the foot is studied separately and regarded as a rigid body. According to the Newton-Euler equation, the dynamic equation of the foot can be expressed as: Among them, F 2,3 is the force exerted by the calf on the foot; F y F is the vertical component of the ground force on the foot; x is the parallel ground force on the foot; G3 is the weight of the foot; M 2,3 is the moment of force acting on the foot by the calf; I3 is the force arm of the calf to the center of mass of the foot; I y F y The moment arm about the center of mass of the foot; I x F x The moment arm about the center of mass of the foot; The ankle joint torque can be calculated according to the foot dynamics equation, which includes the plantar pressure and is also the ankle joint torque in the support phase: Removing the plantar pressure term from the above equation gives the ankle joint torque in the swing phase. The above equation is applicable to the entire gait cycle. When calculating the swing phase, only the plantar pressure needs to be set to 0. The calf is studied and regarded as a rigid body. According to the Newton-Euler equation, the dynamic equation of the calf can be expressed as: Among them, F 1,2 is the force of the thigh on the calf; G2 is the weight of the calf; M 1,2 is the moment of force from the thigh to the calf; I2 is the force arm of the calf; I 2a The moment arm of the calf's center of mass where the thigh acts on the calf; According to the calf dynamics equation, the knee joint torque can be solved: The three-dimensional motion test system includes an 8-channel surface electromyography signal acquisition instrument, a three-dimensional motion capture system, a three-dimensional force measurement treadmill and a time synchronizer. The 8-channel surface electromyography signal acquisition instrument can collect surface electromyography signals of the tester's lower limbs, the three-dimensional motion capture system can capture the motion data of the tester's lower limb knee joints, and the three-dimensional force measurement treadmill can collect the three-dimensional force between the tester's sole and the treadmill during walking. The time synchronizer can ensure the synchronization of the three devices, the 8-channel surface electromyography signal acquisition instrument, the three-dimensional motion capture system and the three-dimensional force measurement treadmill, in data acquisition. The step 1 specifically includes: (1) Collection and preprocessing of electromyographic signals: The right knee joint of the subject was selected as the test object. Before the test, the subject was asked to do appropriate relaxation activities, and the skin at the vastus lateralis, semitendinosus, long head of biceps femoris, and medial head of gastrocnemius of the right leg was wiped with alcohol as the collection points of electromyographic signals. The electromyographic signals of the four muscles of the right leg of the subject were synchronously collected at a speed of 1500HZ when the subject walked at a speed of 2m / s for 1min, and the original electromyographic signals were high-pass filtered to eliminate motion artifacts and filter out 50Hz power frequency interference using a zero-lag 4th-order Butterworth filter (30Hz) and a 50Hz notch filter, and then full-wave rectified; (2) EMG signal feature extraction The electromyographic signal is a discrete signal. The preprocessed electromyographic signal is used as a window with 30 sampling points to extract the average electromyographic value AEMG and the electromyographic root mean square value RMS. The calculation formulas of the average electromyographic value AEMG and the electromyographic root mean square value RMS are as follows: Where Data[i] is the electromyographic signal; N is the window size; Then, the preprocessed electromyographic signal is subjected to the feature extraction of mean power frequency MPF and median frequency MF. The calculation formulas of mean power frequency MPF and median frequency MF are as follows: where f s is the sampling frequency, S(f) is the power spectral density function of the electromyographic signal, and finally the nearest neighbor method is used to downsample the characteristic signal to 100 Hz.

2. A human knee joint torque testing method based on surface electromyography signals according to claim 1, characterized in that: The step 3 constructs a training data set and a test data set for the knee joint torque test LSTM neural network, and trains the neural network based on the data obtained in steps 1 and 2; The four eigenvalues ​​of the four muscles, the knee angle and the knee torque in 1 minute obtained in step 1 and step 2 are divided into the first 40s data set and the last 20s data set according to time. The first 40s data set is used as the training set of the neural network, and the last 20s data set is used as the test set of the neural network. The four eigenvalues ​​of the four muscles and the knee angle data in the training set are captured respectively. 30 data points x starting from the tth data point As the input of the LSTM neural network, the 30 data points y of the knee joint torque data in the training set starting from the tth data point are captured As the output of the LSTM neural network; The input to the hidden layer of the LSTM neural network includes the current input x of the network and the hidden layer output vector a at the previous moment <i-1>< / i-1> , and the hidden layer state C <i-1>< / i-1> , the hidden layer adds a forget gate f , input gate i And the output gate o ; Forget gate f Decide to discard state c What information in x and a <i-1>< / i-1> Can it be used in state C? The update of the input gate i Decision, after the forget gate and input gate, state C The update is completed, the output gate o To determine the state C How does it work on a In the calculation of the forget gate, input gate, output gate and hidden layer output a and status update C The calculation expression is as follows: f <t>< / t> =σ(W f ·[a <i-1 ,x ]+b f ) I <τ> =σ(W i ·[a <i-1>< / i-1> ,x ]+b i ) o =σ(W o ·[a <i-1>< / i-1> ,x ]+b o ) c =f ·a <i-1>< / i-1> +x ·tanh(W c ·[a <i-1>< / i-1> ,x ]+b c ) a =0 ·tanh(c ) Among them, σ is the sigmoid function, tanh is the hyperbolic tangent function, W f , W i , W c , W o and b f , b i , b c , b o They are the weight parameters and bias parameters that need to be trained in the forget gate, input gate, and output gate respectively; The output layer loss function uses a square loss function to evaluate the error between the actual output and the expected output, and uses the error as the basis for back propagation optimization to minimize the error. The loss function is as follows: Where Y is the calculated torque value of the knee joint; f(x) is the estimated torque value of the knee joint; L(Y,f(x)) is the loss function; The root mean square error (RMSE) and the correlation coefficient (CORR) are used to evaluate the quality of the human knee joint torque prediction model. RMSE is used to measure the error between the model estimate and the true value, and CORR is used to describe the correlation between two variables. The closer the RMSE value is to 0, the more accurate the corresponding neural network model estimate is. On the contrary, the closer the CORR value is to 1, the better the model fitting effect is. The calculation formulas of RMSE and CORR are as follows: Among them, y oci is the knee joint torque calculated by inverse dynamics, is the average value of the calculated torque, is the estimated moment of the neural network model, is the average value of the estimated moment of the model, and N is the length of the test sample sequence data.