Lower limb motion intention prediction method based on multi-source signals

Abstract

A lower limb motion intention prediction method based on multi-source signals comprises the following steps: firstly, analyzing motion characteristics of lower limbs of a human body, screening to-be-detected muscles, collecting electromyographic signals of a plurality of periods, and preprocessing the electromyographic signals; then linear features and nonlinear features of the electromyographic signals are extracted; the linear features and the nonlinear features of the electromyographic signals and motion capture information are fused through a series splicing method; a random forest method (RF-Bagging) is adopted to carry out feature selection on the feature matrix with the higher dimension; and finally, inputting the selected feature matrix by using a long short-term memory network (LSTM), performing joint continuous motion amount regression prediction according to the amputation type, and judging the motion intention recognition performance through various regression indexes. According to the method, joint continuous quantity prediction with a good effect under different amputation body types is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of human-computer interaction signal processing, and in particular to a method for predicting lower limb movement intention based on multi-source signals. Background Art

[0002] The development of human-computer interaction technology has promoted the shift of the focus of lower limb prosthesis research from passive prostheses to intelligent prostheses. The characteristics of intelligent prostheses are that they can recognize the user's movement intention. The accuracy and real-time nature of the recognition results directly affect the movement performance of intelligent prostheses and the smoothness and safety of lower limb movements. However, most of the intelligent powered lower limb prostheses currently on the market still have problems such as high development costs and low intention recognition rates in terms of how to automatically and accurately recognize the human lower limb movement intention, which reduces the coordination between the lower limb prosthesis and the wearer and the human-computer interaction experience.

[0003] The recognition of human lower limb movement intention usually involves the following two aspects: recognition of discrete limb movement modes, such as walking on flat ground, ramps and stairs; and estimation of continuous joint movement using sEMG, such as joint torque, joint angle and other continuous quantities. Movement classification can only predict a few discrete limb movements, and the robot cannot complete human-like continuous smooth movement by using the prediction results. Ensuring continuous matching of human-machine movement is the premise for realizing the safe control of various service robots. Continuous movement intention recognition is more valuable because it can predict the specific value of movement in real time.

[0004] In actual research, too many and repeated electromyographic signals will lead to signal redundancy, which will affect subsequent data processing and application. However, too few or random selection of lower limb muscles to collect electromyographic signals will not accurately represent the lower limb movement of the human body (Wang Junyao. Research on human lower limb movement recognition technology based on multi-source information fusion [D]. Xi'an: University of Electronic Science and Technology, 2023). In existing research, the selection of muscles is mostly based on previous experience or theoretical knowledge of anatomy, and there is no data support. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the purpose of the present invention is to provide a method for predicting lower limb movement intention based on multi-source signals, expand the dimension of surface electromyographic signals based on the VMD algorithm, perform linear and nonlinear feature layer fusion, perform multi-source information fusion from the data source, introduce feature selection algorithm to optimize the regression model, and realize good joint continuous quantity prediction under different amputation types.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is:

[0007] A method for predicting lower limb movement intention based on multi-source signals comprises the following steps:

[0008] 1) Analyze the movement characteristics of the human lower limbs, select the muscles to be tested, collect several cycles of electromyographic signals, and pre-process the electromyographic signals;

[0009] 2) Extract the linear and nonlinear features of the electromyographic signal. The linear features use the time domain and frequency domain features, and the nonlinear features use the VMD intrinsic mode function IMF;

[0010] 3) The linear and nonlinear features of the electromyographic signal and the motion capture information are integrated through the serial splicing method; the serial splicing method is to directly connect different types of feature values ​​into a higher-dimensional feature matrix;

[0011] 4) Using the Random Forest with Bagging (RF-Bagging) method to perform feature selection on the higher-dimensional feature matrix obtained in step 3);

[0012] 5) Using a long short-term memory network (LSTM) to input the feature matrix selected in step 4), perform regression prediction of joint continuous motion according to the amputation type, and use multiple regression indicators to judge the motion intention recognition performance.

[0013] In step 1), the muscle screening method is based on the musculoskeletal model and correlation analysis to screen out the combination of muscles that best reflect joint movement; first, the musculoskeletal analysis software OpenSim is used to build a model to visualize the changes in lower limb muscle length and the distribution of their change rates during walking, and then the PEARSON parameter is used as the correlation coefficient to calculate the correlation between muscles during exercise, and the test electromyographic signal of the muscle to be tested is determined based on the changes in muscle length, its change rate and correlation analysis, as well as the physiological structure layout of human muscles; superficial muscles with large changes in muscle length and its change rate and weak correlation are selected, and the screened muscles are: rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior.

[0014] In step 2), the adaptive parameter optimization VMD (APOVMD) method adaptively determines the most appropriate decomposition parameters, and the decomposition process is:

[0015] 2.1) Input signal data and set decomposition parameters;

[0016] 2.2) Set the decomposition parameters, decomposition mode number K: the number of eigenmodes decomposed; penalty factor α, the penalty parameter that controls the degree of smoothness of each mode; noise width τ: the subgradient method parameter that speeds up the convergence speed; convergence tolerance ε: the stop condition that controls the algorithm iteration;

[0017] 2.3) Initialize variables Initialize the frequency ω of each mode k , modal component u k and the Lagrange multiplier λ;

[0018] 2.4) Iteratively update the modal and frequency Enter the iterative process and gradually update the following variables until convergence: Update the modal component u k :Use the Lagrangian function to optimize and update the current signal after subtracting the remaining modal components; update the center frequency ω k : Adjust the center position of the spectrum of the calculated mode; Update the Lagrange multiplier λ: Update the overall constraint error of the signal;

[0019] 2.5) Check the convergence condition and calculate the change of the modal component before and after the update. If it is less than the tolerance ε, stop the iteration; otherwise, continue to update the modal component and frequency;

[0020] 2.6) Decompose to obtain the mode and center frequency;

[0021] 2.7) Optimize the L and α parameters through KL divergence to determine the best decomposition effect;

[0022] 2.8) Signal reconstruction: Add up all the decomposed modes to verify whether the original signal can be reconstructed;

[0023] 2.9) Output the optimal modal components, center frequencies and related decomposition parameters;

[0024] Decompose the electromyographic signals of the six muscles to obtain their respective intrinsic mode function IMF components (modal components u k ), constructed as a 5-dimensional matrix, the VMD feature matrix is ​​marked as U i , i = 1-6, representing the selected rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior muscles, respectively. The IMF feature matrix of the rectus femoris is expressed as formula (1), the IMF feature represented by VMD;

[0025]

[0026] u k ——Modal component, k=1-5.

[0027] In step 3), the feature layer fusion is performed by serial concatenation. TD+FD+VMD represents the serial concatenation of the time domain features, frequency domain features and VMD features of the electromyographic signal, as shown in formula (3):

[0028] The time domain of the 6 electromyographic signals is constructed as a 17-dimensional matrix, and the frequency domain is constructed as an 8-dimensional matrix. The feature matrix is ​​marked as F i (i = 1-6, representing rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior, respectively), the feature matrix is ​​represented as a 150-dimensional column vector;

[0029] The corresponding hip joint angle, knee joint angle, and ankle joint angle are expressed as an angle matrix θ:

[0030] θ=[θ h θ k θ a ] (4)

[0031] θ h ——Hip joint angle, θ k ——knee joint angle, θ a ——Ankle joint angle.

[0032] Step 4) Feature selection uses the ensemble learning ability of random forests to evaluate the importance of features, while bagging enhances the robustness of the model; the predictor importance scores generated by the random forest model are sorted in descending order to identify and prioritize the most influential features; by focusing on the top-ranked features, we ensure that key features that contribute significantly to the model performance are selected.

[0033] In step 5), the amputation types include hip amputation, thigh amputation, knee amputation, calf amputation, and ankle amputation; the regression prediction of joint continuous motion is divided into the following situations for regression analysis: when the amputation side information is used as the regression model input, if the knee joint exists, 6 muscles are selected for electromyographic signal analysis, that is, the input muscles are rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior; if it does not exist, only thigh muscles can be selected for analysis, that is, the input muscles are rectus femoris, gluteus medius, biceps femoris, and vastus medialis; when the healthy side information is used as the regression model input, all muscles are selected for analysis.

[0034] In step 5), there are multiple regression indicators including mean absolute error (MAE), root mean square error (RMSE), R square score (R 2 ), mean square error (MSE), mean absolute percentage error (MAPE) and relative prediction deviation (RPD), see formula (5);

[0035]

[0036] y i is the observed value, is the predicted value, is the mean of the observed values, is the average value of the predicted value, and n is the number of samples; for MAE, RMSE, MSE, and MAPE, the smaller the value, the higher the prediction accuracy and the smaller the prediction error; R 2 The larger the value, the better the model performance. 2The closer it is to 1, the better the fit; RPD reflects the relative consistency between the predicted value and the actual value. When the RPD value is close to or exceeds 1, the predicted value is considered to be very close to the true value.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] Aiming at the problem that irregular selection of lower limb electromyographic signals leads to incomplete feedback of motion information and low correlation between characteristic values ​​of electromyographic signals, which in turn affects the recognition of motion intention, the present invention proposes a muscle screening method based on musculoskeletal simulation model and correlation analysis, and selects the combination of muscles that can best reflect joint movement from numerous lower limb muscles; based on the VMD algorithm, the dimension of the surface electromyographic signal is expanded, the linear feature and nonlinear feature layer of the signal are fused, and the joint angle and torque information are added to perform multi-source information data layer fusion; characteristic value serial splicing and bagging-based random forest feature selection methods are proposed; the LSTM-based regression prediction model selects input information in a targeted manner according to the type of lower limb amputation, realizes the effective prediction of continuous joint movement under different conditions, and is conducive to promoting the intelligent development of lower limb prostheses and related rehabilitation equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flow chart of motion intention prediction according to an embodiment of the present invention.

[0040] Figure 2 This is a diagram of the Opensim musculoskeletal model of an embodiment of the present invention.

[0041] Figure 3 This is a diagram of the changes in lower limb muscle length during a complete gait cycle of the Opensim musculoskeletal model of an embodiment of the present invention.

[0042] Figure 4 This is a ranking diagram of the rate of change of lower limb muscle length during a complete gait cycle of the Opensim musculoskeletal model according to an embodiment of the present invention.

[0043] Figure 5 It is a variational mode decomposition flow chart of adaptive parameter optimization according to an embodiment of the present invention.

[0044] Figure 6 It is a schematic diagram of feature fusion of an embodiment of the present invention.

[0045] Figure 7 It is a schematic diagram of lower limb amputation types and selected muscle distribution according to an embodiment of the present invention.

[0046] Figure 8 This is the knee joint angle regression prediction evaluation index result of the embodiment of the present invention.

[0047] Fig. 9It is the ankle joint angle regression prediction evaluation index result of the present invention. DETAILED DESCRIPTION

[0048] The technical scheme of the present invention is further described and illustrated in detail below in conjunction with the embodiments and drawings. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0049] Reference Figure 1 , a method for predicting lower limb movement intention based on multi-source signals, comprising the following steps:

[0050] 1) Analyze the movement characteristics of the human lower limbs, select the muscles to be tested, collect several cycles of electromyographic signals, and pre-process the electromyographic signals;

[0051] 2) Extract the linear features (time domain and frequency domain features) and nonlinear features (VMD intrinsic mode function IMF) of the electromyographic signal;

[0052] 3) The linear and nonlinear features of the electromyographic signal and the inverse dynamics and inverse kinematics features obtained from the motion capture data are fused through the serial splicing method; the serial splicing method is to directly connect different types of eigenvalues ​​into a higher-dimensional feature matrix;

[0053] 4) Using the Random Forest with Bagging (RF-Bagging) method to perform feature selection on the higher-dimensional feature matrix obtained in step 3) to reduce feature redundancy and improve prediction results;

[0054] 5) Using a long short-term memory network (LSTM) to input the feature matrix selected in step 4), perform regression prediction of joint continuous motion according to the amputation type, and use multiple regression indicators to judge the motion intention recognition performance.

[0055] Reference Figure 2 , Figure 3 , Figure 4 In step 1), the muscle screening method is based on the musculoskeletal model and correlation analysis to select the combination of muscles that best reflect joint movement from numerous lower limb muscles; first, the musculoskeletal analysis software OpenSim is used to establish a model to visualize the changes in lower limb muscle length and the distribution of their change rates during walking, and then the PEARSON parameter is used as the correlation coefficient to calculate the correlation between muscles during exercise, and the electromyographic signal of the muscle to be tested is determined based on the changes in muscle length, its change rate and correlation analysis, as well as the physiological structure layout of human muscles, and the electromyographic signal is accurately selected from the data source to further improve the prediction accuracy of the algorithm; shallow muscles with large changes in muscle length and its change rate and weak correlation are selected, and the screened muscles are: rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior.

[0056] Figure 2 This embodiment uses the musculoskeletal analysis software OpenSim to simulate the changes in lower limb muscle length. The human body model 2392_Simbody.osim is selected, and the normal gait motion normal.mot is loaded. The model parameters are: height 1.8m, weight 75kg, and 19 bones and 92 muscles in the lower limbs.

[0057] Figure 3 The length of the muscles in the present embodiment varies during walking. The lengths of the muscles of the lower limbs are different, with the sartorius being the longest and the gluteus medius being the shortest. Figure 3 In (a) and (d), the sartorius muscle is distributed along the thigh in a spiral shape and is responsible for knee flexion and cross-leg movements. The gluteus medius muscle is located deep in the buttocks and is used to abduct the thigh. The vastus lateralis, vastus intermedius, vastus medialis, and rectus femoris are collectively called the quadriceps femoris, which plays a key role in knee extension, flexion, abduction, and internal rotation. ( Figure 3 In (a) and (c)), when the gait cycle is 70% (the period when the toes touch the ground), the four parameters have a peak value, indicating that the muscles are stretched and the knee joint is in a flexed state. Figure 3 In (c) and (d)), the long head of the biceps femoris, the semitendinosus, and the semimembranosus are collectively called the hamstrings. Their changing trends are basically the same. This is because all three originate from the ischial tuberosity and end at the tibia. The muscles also have the same function of extending the hip and flexing the knee. At 70% of the cycle, the three reach their trough values. During walking, the tibialis anterior muscle changes little in length, and only reaches its maximum value at around 65% of the cycle. At this time, the tibialis anterior muscle is stretched and the ankle joint is in a plantar flexion state. The gracilis is a muscle group on the inner side of the thigh. Its function is to maintain hip adduction and external rotation and to help the knee flex and internally rotate. Therefore, its trough and peak values ​​appear alternately, which is beneficial for the human body to perform activities such as standing and walking. ( Figure 3 (d) The medial gastrocnemius can flex the knee or plantar flex the ankle, so the trough occurs at 70% of the cycle.

[0058] Figure 4 The length change rate of muscles in this embodiment is sorted. There are differences in the length change rate of gait muscles in a complete gait cycle. This is mainly because the absolute elongation of different muscles from resting state to moving state is different. Taking the vastus medialis as an example, the vastus medialis muscle is used to extend the knee, and its resting length is small, and its length change rate is the largest. The length change rate of different muscles provides a basis for muscle selection.

[0059] Reference Figure 5In step 2), the nonlinear feature extraction of the electromyographic signal is carried out by using an adaptive parameter optimized VMD (APOVMD) to expand the dimension of the surface electromyographic signal, and the linear feature and nonlinear feature layer are fused by feature extraction means, and multi-source information data layer is fused based on joint angle and torque; the adaptive parameter optimized VMD (APOVMD) method can adaptively determine the most appropriate decomposition parameters and improve the performance of VMD. The decomposition process is as follows:

[0060] 2.1) Input signal data and set decomposition parameters;

[0061] 2.2) Set the decomposition parameters, decomposition mode number K: the number of eigenmodes decomposed (need to be adjusted); penalty factor α, penalty parameter that controls the degree of smoothness of each mode (balance accuracy and overfitting); noise width τ: subgradient method parameter that speeds up convergence; convergence tolerance ε: stop condition that controls algorithm iteration;

[0062] 2.3) Initialize variables Initialize the frequency ω of each mode k , modal component u k and the Lagrange multiplier λ;

[0063] 2.4) Iteratively update the modal and frequency Enter the main iteration process and gradually update the following variables until convergence: Update the modal component u k :Use the Lagrangian function to optimize and update the current signal after subtracting the remaining modal components; update the center frequency ω k : Adjust the center position of the spectrum of the calculated mode; Update the Lagrange multiplier λ: Update the overall constraint error of the signal;

[0064] 2.5) Check the convergence condition and calculate the change of the modal component before and after the update. If it is less than the tolerance ε, stop the iteration; otherwise, continue to update the modal component and frequency;

[0065] 2.6) Decompose to obtain the mode and center frequency;

[0066] 2.7) Optimize K and α parameters through KL divergence (relative entropy) to determine the best decomposition effect;

[0067] 2.8) Signal reconstruction: Add up all the decomposed modes to verify whether the original signal can be reconstructed;

[0068] 2.9) Output the optimal modal components, center frequencies and related decomposition parameters;

[0069] The IMF (intrinsic mode function) components obtained by decomposing the electromyographic signals of the six muscles, i.e., the modal components u k Constructed as a 5-dimensional matrix, the characteristic matrix is ​​marked as Ui (i = 1-6, representing the selected rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior muscles, respectively). The IMF feature matrix of the rectus femoris is expressed as formula (1), which is a 30-dimensional column vector, and the IMF feature represented by VMD;

[0070]

[0071] u k ——Modal component, k=1-5.

[0072] Reference Figure 6 In step 3), the feature layer fusion is performed by the serial splicing method. In this embodiment, the time domain features of the six electromyographic signals (integrated electromyographic value iEMG, mean square error RMS, variance VAR, number of zero crossing points ZC, waveform length WL, standard deviation SD, kurtosis Kurtosis, mean absolute value MAV, mean amplitude change SSC, Willison amplitude, logarithmic detection LD, simple square integral SSI, absolute value integral AVI, waveform factor WFF, pulse factor IPF, kurtosis factor PF, margin factor MF) are constructed as a 17-dimensional matrix, and the frequency domain features (fourth-order autoregressive coefficient, average power frequency MPF, median frequency MDF, power spectrum ratio PSR, energy spectrum density ESD) are constructed as an 8-dimensional matrix. The feature matrix is ​​marked as F i (i = 1-6, representing the selected rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior muscles, respectively), the feature matrix is ​​expressed as formula (2), which is a 150-dimensional column vector, TD and FD represent the time domain characteristics and frequency domain characteristics of the electromyographic signal, respectively;

[0073]

[0074] Based on the serial concatenation method, feature layer fusion is performed. TD+FD+VMD represents the serial concatenation of the time domain features, frequency domain features and VMD features of the electromyographic signal, which is a 180-dimensional feature, as shown in formula (3): TD+FD+VMD=[F1 F2…F6 U1 U2…U6] (3)

[0076] The time domain of the 6 electromyographic signals is constructed as a 17-dimensional matrix, and the frequency domain is constructed as an 8-dimensional matrix. The feature matrix is ​​marked as F i (i = 1-6, representing rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior, respectively), the feature matrix is ​​represented as a 150-dimensional column vector;

[0077] The corresponding hip joint angle, knee joint angle, and ankle joint angle are expressed as an angle matrix θ:

[0078] θ=[θ h θk θ a ] (4)

[0079] θ h ——Hip joint angle, θ k ——knee joint angle, θ a - Ankle joint angle;

[0080] This embodiment aims to predict joint continuous quantities (joint angles). It uses the electromyographic features (time domain features, frequency domain features, VMD decomposition features) of the 6-channel EMG signals of the unilateral lower limb and the inverse dynamics and inverse kinematics features (joint angles, etc.) of the motion capture data as the input of the regression model. It proposes a multi-source signal fusion method that combines electromyographic features and joint angles and uses it for continuous prediction of joint angles.

[0081] Step 4) Feature selection uses the Random Forest with Bagging (RF-Bagging) method, which uses the ensemble learning ability of random forests to evaluate the importance of features, while bagging enhances the robustness of the model; the predictor importance scores generated by the random forest model are arranged in descending order to identify and prioritize the most influential features; by focusing on these top-ranked features, this method ensures that key features that contribute significantly to model performance are selected, thereby facilitating efficient and accurate analysis.

[0082] Reference Figure 7 In step 5), the amputation types include hip amputation, thigh amputation, knee amputation, calf amputation, and ankle amputation; the regression prediction of joint continuous motion is divided into the following situations for regression analysis. When the information of the affected side (amputation side) is used as the regression model input, if the knee joint exists, the electromyographic signal can select the 6 muscles for analysis, that is, the input muscles are the rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior; if it does not exist, only the thigh muscles can be selected for analysis, that is, the input muscles are the rectus femoris, gluteus medius, biceps femoris, and vastus medialis; when the information of the healthy side is used as the regression model input, all muscles can be selected for analysis.

[0083] Reference Figure 8-Figure 9 In step 5), multiple regression indicators are used to quantitatively evaluate the accuracy of continuous prediction of joint angles. The multiple regression indicators include mean absolute error (MAE), root mean square error (RMSE), R square score (R 2 ), mean square error (MSE), mean absolute percentage error (MAPE) and relative prediction deviation (RPD), see formula (5);

[0084]

[0085] yi is the observed value, is the predicted value, is the mean of the observed values, is the average value of the predicted value, and n is the number of samples. For MAE, RMSE, MSE, and MAPE, the smaller the value, the higher the prediction accuracy and the smaller the prediction error. 2 The larger the value, the better the model performance. 2 The closer it is to 1, the better the fit; RPD reflects the relative consistency between the predicted value and the actual value. When the RPD value is close to or exceeds 1, the predicted value is considered to be very close to the true value.

[0086] In this embodiment, the data of each subject are trained 10 times using the LSTM model, and then the best result among the 10 times is selected for each subject, that is, their optimal prediction model; based on the optimal prediction models of the 10 subjects, the regression prediction results of selecting 4 muscles and 6 muscles (that is, the muscles retained by different amputation types) as feature inputs are obtained.

[0087] Reference Figure 8 In the prediction of knee joint angle, the input features include electromyographic signal features (EMG features), intrinsic mode function (IMFs) components, and hip joint angle. The input features are divided into two categories: 1) including EMG features and IMFs components of 4 muscles, as well as hip joint angle; 2) corresponding features extended to 6 muscles. The evaluation indicators of the knee joint angle prediction results were obtained based on the optimal prediction model of 10 subjects. The results showed that there was no significant difference between the two, and both achieved satisfactory prediction results. However, there were differences in the knee joint angle prediction performance between different subjects. Subject 7 performed best, while subject 10 performed worst. For subject 7, the six evaluation indicators (MAE, RMSE, R 2 , MSE and RPD) are 1.731, 2.582, 0.977, 6.668 and 0.989 respectively, while the corresponding indicators of the prediction effect based on the four muscle features are 1.590, 2.353, 0.981, 5.537 and 0.984 respectively. The inclusion of calf muscle features has improved the prediction performance.

[0088] Reference Fig. 9In the prediction of ankle joint angle, inspired by the prediction results of knee joint angle, joint angle information is added. It is divided into two groups of input feature combinations, one of which includes the EMG features and IMFs components of 4 muscles, as well as the hip joint angle; the other group is expanded to the features of 6 muscles, and the knee and hip joint angles are added at the same time. The evaluation indicators of the ankle joint angle prediction results show that the ankle joint angle prediction performance based on the features of 6 muscles is better than the ankle joint angle prediction performance based on the features of 4 muscles. Fig. 9 In (a), (b) and (d), the MAE, RMSE and MSE indicators of the ankle joint angle prediction performance based on the features of 6 muscles are lower than the results of the ankle joint angle prediction performance based on the features of 4 muscles for almost all subjects. Fig. 9 In (c), the ankle joint angle prediction performance R based on the characteristics of the six muscles 2 For both feature combinations, most RPD values ​​are greater than 1. The evaluation results show that angle information as input improves the motion intention recognition results.

[0089] In order to evaluate the overall regression performance of joint angle prediction, the average values ​​of the evaluation indicators of the optimal prediction models for knee and ankle angles in 10 subjects were calculated. The results are as follows: Figure 8 (f) and Fig. 9 By comparing the two sets of results for knee joint angle prediction, it can be seen that the models based on different feature fusions (amputation types) all achieved good prediction performance.

[0090] The above implementation modes are only used to illustrate the present invention rather than to limit the present invention. Ordinary technicians in the relevant technical field may make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention. The patent protection scope of the present invention should be defined by the claims.

Claims

1. A method for predicting lower limb movement intention based on multi-source signals, characterized in that: The following steps are involved: 1) Analyze the movement characteristics of the human lower limbs, select the muscles to be tested, collect several cycles of electromyographic signals, and pre-process the electromyographic signals; 2) Extract the linear and nonlinear features of the electromyographic signal. The linear features use the time domain and frequency domain features, and the nonlinear features use the VMD intrinsic mode function IMF; 3) The linear and nonlinear features of the electromyographic signal and the motion capture information are integrated through the serial splicing method; the serial splicing method is to directly connect different types of feature values ​​into a higher-dimensional feature matrix; 4) Using the Random Forest with Bagging (RF-Bagging) method to perform feature selection on the higher-dimensional feature matrix obtained in step 3); 5) Using a long short-term memory network (LSTM) to input the feature matrix selected in step 4), perform regression prediction of joint continuous motion according to the amputation type, and use multiple regression indicators to judge the motion intention recognition performance.

2. The prediction method according to claim 1, characterized in that: In step 1), the muscle screening method is based on the musculoskeletal model and correlation analysis to screen out the combination of muscles that best reflect joint movement; first, the musculoskeletal analysis software OpenSim is used to build a model to visualize the changes in lower limb muscle length and the distribution of their change rates during walking, and then the PEARSON parameter is used as the correlation coefficient to calculate the correlation between muscles during exercise, and the test electromyographic signal of the muscle to be tested is determined based on the changes in muscle length, its change rate and correlation analysis, as well as the physiological structure layout of human muscles; superficial muscles with large changes in muscle length and its change rate and weak correlation are selected, and the screened muscles are: rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior.

3. The prediction method according to claim 1, characterized in that: In step 2), the adaptive parameter optimization VMD (APOVMD) method adaptively determines the most appropriate decomposition parameters, and the decomposition process is: 2.1) Input signal data and set decomposition parameters; 2.2) Set the decomposition parameters, decomposition mode number K: the number of eigenmodes decomposed; penalty factor α, the penalty parameter that controls the degree of smoothness of each mode; noise width τ: the subgradient method parameter that speeds up the convergence speed; convergence tolerance ε: the stop condition that controls the algorithm iteration; 2.3) Initialize variables Initialize the frequency ω of each mode k , modal component u k and the Lagrange multiplier λ; 2.4) Iteratively update the modal and frequency Enter the iterative process and gradually update the following variables until convergence: Update the modal component u k :Use the Lagrangian function to optimize and update the current signal after subtracting the remaining modal components; update the center frequency ω k : Adjust the center position of the spectrum of the calculated mode; Update the Lagrange multiplier λ: Update the overall constraint error of the signal; 2.5) Check the convergence condition and calculate the change of the modal component before and after the update. If it is less than the tolerance ε, stop the iteration; otherwise, continue to update the modal component and frequency; 2.6) Decompose to obtain the mode and center frequency; 2.7) Optimize K and α parameters through KL divergence to determine the best decomposition effect; 2.8) Signal reconstruction: Add up all the decomposed modes to verify whether the original signal can be reconstructed; 2.9) Output the optimal modal components, center frequencies and related decomposition parameters; The EMG signals of the six muscles are decomposed to obtain their respective intrinsic mode function IMF components, which are constructed into a 5-dimensional matrix. The characteristic matrix is ​​marked as U i , i = 1-6, representing the selected rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior muscles, respectively. The IMF feature matrix of the rectus femoris is expressed as formula (1), the IMF feature represented by VMD; u k ——Modal component, k=1-5.

4. The prediction method according to claim 1, characterized in that: In step 3), the feature layer fusion is performed by serial concatenation. TD+FD+VMD represents the serial concatenation of the time domain features, frequency domain features and VMD features of the electromyographic signal, as shown in formula (3): TD+FD+VMD=[F1 F2 … F6 U1 U2 … U6] (3) The time domain of the 6 electromyographic signals is constructed as a 17-dimensional matrix, and the frequency domain is constructed as an 8-dimensional matrix. The feature matrix is ​​marked as F i , i = 1-6, representing the rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior, respectively, then the feature matrix is ​​represented as a 150-dimensional column vector; The corresponding hip joint angle, knee joint angle, and ankle joint angle are expressed as an angle matrix θ: θ=[θ h i k i a ] (4) θ h ——Hip joint angle, θ k ——knee joint angle, θ a ——Ankle joint angle.

5. The prediction method according to claim 1, characterized in that: Step 4) Feature selection uses the ensemble learning ability of random forests to evaluate the importance of features, while bagging enhances the robustness of the model; the predictor importance scores generated by the random forest model are sorted in descending order to identify and prioritize the most influential features; By focusing on the top-ranked features, you ensure that you select key features that contribute significantly to the model performance.

6. The prediction method according to claim 1, characterized in that: In step 5), the amputation types include hip amputation, thigh amputation, knee amputation, calf amputation, and ankle amputation; the regression prediction of joint continuous motion is divided into the following situations for regression analysis: when the amputation side information is used as the regression model input, if the knee joint exists, 6 muscles are selected for electromyographic signal analysis, that is, the input muscles are rectus femoris, gluteus medius, biceps femoris, vastus medialis, medial gastrocnemius, and tibialis anterior; if it does not exist, only thigh muscles can be selected for analysis, that is, the input muscles are rectus femoris, gluteus medius, biceps femoris, and vastus medialis; when the healthy side information is used as the regression model input, all muscles are selected for analysis.

7. The prediction method according to claim 1, characterized in that: In step 5), there are multiple regression indicators including mean absolute error (MAE), root mean square error (RMSE), R square score (R 2 ), mean square error (MSE), mean absolute percentage error (MAPE) and relative prediction deviation (RPD), see formula (5); y i is the observed value, is the predicted value, is the mean of the observed values, is the average value of the predicted value, and n is the number of samples; for MAE, RMSE, MSE, and MAPE, the smaller the value, the higher the prediction accuracy and the smaller the prediction error; R 2 The larger the value, the better the model performance. 2 The closer it is to 1, the better the fit; RPD reflects the relative consistency between the predicted value and the actual value. When the RPD value is close to or exceeds 1, the predicted value is considered to be very close to the true value.

Images