Joint torque prediction method and device based on electromyography and musculoskeletal physiological information

Abstract

A method and device for predicting the joint torque of a lower limb exoskeleton rehabilitation robot based on electromyographic signals, the method comprising: first collecting the user's electromyographic, joint angle, angular velocity and joint torque data, using a neuromusculoskeletal model to model the four muscles related to the knee joint, and using a sparrow optimization algorithm to identify the physiological parameters used in the model, outputting an optimal solution, and obtaining an estimated torque. Then, the real torque and the estimated torque are assigned weights of respectively to obtain a fusion torque. The collected data and the fusion torque are used to train a neural network model, and at the same time, the neuromusculoskeletal model is incorporated into the loss function to allow the neuromusculoskeletal model to participate in the training process of the neural network model and increase the constraints of the model. Finally, the trained model is saved for subsequent call. The present invention has the advantages of high precision and strong generalization, and at the same time, the model training can converge faster under the condition of less data volume.

Description

Technical Field

[0001] The present invention relates to a joint torque prediction method and device based on myoelectric and musculoskeletal physiological information, which is applied to lower limb rehabilitation exoskeleton robots and suitable for lower limb joint torque prediction in active control strategies. It utilizes the advanced characteristics of electromyographic signals to achieve reliable movement intention prediction. Background Art

[0002] Rehabilitation strategies vary with the stage of rehabilitation. To make users feel more comfortable during training, it is necessary to detect and identify the user's movement intention, such as predicting joint angles or joint torques. This can then be applied to the control strategy to make the training more smooth and comfortable. Currently, the characteristic of electromyographic signals that precede muscle movement by 30-200ms is often used to identify movement intention. There are two main methods for predicting joint torque. The first method is to use a neuromusculoskeletal model to model a single muscle, explain the conversion relationship between electromyographic signals and muscle force, and then obtain the predicted joint torque value of the corresponding joint. The other method is to use a neural network model to extract features from the data, such as using a convolutional neural network to extract spatial features, or using a long short-term memory neural network to extract temporal features, and then predict the joint torque.

[0003] However, neuromusculoskeletal models require individual modeling of each relevant muscle, resulting in numerous physiological parameters, poor individual adaptability, and relatively low accuracy. Training neural network models requires extensive data support, and obtaining high-quality data is also challenging. Furthermore, due to their black-box nature, their mechanisms of action are difficult to interpret, making it difficult to ensure that the neural network model's output meets both demand and safety requirements. Therefore, a lightweight model with fewer parameters is needed that can simultaneously meet the real-time, accuracy, universality, and safety requirements for predicting joint torques during movement. Summary of the Invention

[0004] The present invention aims to overcome the aforementioned shortcomings of the prior art and provides a joint torque prediction method and device based on electromyography and musculoskeletal physiological information. This invention designs a neural network model that integrates muscle physiological information and incorporates the neuromusculoskeletal model into the neural network model's training process. This introduces muscle physiological information into the neural network model, combining the advantages of the neuromusculoskeletal model and the neural network model to construct a new lower limb joint torque prediction model constrained by muscle physiological information. This method achieves improved accuracy compared to methods that use only the neural network model.

[0005] To achieve the above effects, the first aspect of the present invention relates to a joint torque prediction method based on electromyography and musculoskeletal physiological information, comprising the following steps:

[0006] S1: Build a data acquisition platform. Use a Raspberry Pi, MATLAB, myoelectric sensors, and an exoskeleton robot to build the entire data acquisition system for offline data collection. Based on muscle synergy, collect EMG signals from the rectus femoris, vastus medialis, medial head of the gastrocnemius, and lateral head of the gastrocnemius, as well as joint angle signals and angular velocity signals from the hip, knee, and ankle joints.

[0007] S2: Neuromusculoskeletal model establishment. Using the PyCharm platform, we modeled the four muscles using the collected electromyographic and joint angle signals. We derived the torque of each muscle and calculated the knee joint torque as it varied with movement. Since the number of muscles used was limited, we added an output coefficient to the output to more accurately approximate the actual torque, thus obtaining the knee joint torque.

[0008] S3: Identify the parameters of the neuromusculoskeletal model. The physiological parameters required by the neuromusculoskeletal model, such as the optimal muscle force F m0 , maximum pennation angle φ0, optimal muscle fiber length l m0 , tendon length l t As well as the muscle contraction velocity f(v), which has individual differences. All parameters are taken as parameters to be optimized. A reasonable range of each parameter is given and optimized using the sparrow optimization algorithm. The parameters are then transferred back to the neuromusculoskeletal model to improve the estimation accuracy of the neuromusculoskeletal model.

[0009] S4: Build a neural network model. The collected electromyographic signals, joint angle signals, and joint angular velocity signals are used as input features. An attention mechanism is used to assign weight matrices to different features to indicate their importance. Two convolutional layers and a pooling layer are used to extract spatial relationships between features and perform dimensionality reduction on the data. A bidirectional long short-term memory layer is added to extract long-range temporal dependencies. A dropout layer is added between layers to reduce overfitting. Finally, a fully connected layer is added to output the predicted joint torques.

[0010] S5: Construct a loss function. During training, based on the aforementioned neuromusculoskeletal model, joint torques during movement are estimated using electromyographic signals and joint angles. Combined with the true torque, different weights are assigned to the estimated and true torques, resulting in a fused torque estimate that serves as a label for offline neural network training. The mean absolute error between the fused torque and the predicted torque output by the neural network model is calculated as the loss function for training, ensuring that the model continuously fits and approaches the true torque.

[0011] S6: Design an evaluation metric. The mean absolute error between the predicted moment and the true moment of the neural network model is used as the evaluation metric of the model.

[0012] S7: Model training and saving. The model is trained using data and saved for later use.

[0013] The second aspect of the present invention relates to a joint torque prediction device based on electromyography and musculoskeletal physiological information, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the joint torque prediction method based on electromyography and musculoskeletal physiological information of the present invention.

[0014] The third aspect of the present invention relates to a computer-readable storage medium on which a program is stored, including data acquisition and processing, model structure parameters, parameters of the trained model, and a model training method. When the program is executed by a processor, the joint torque prediction method based on electromyography and musculoskeletal physiological information of the present invention is implemented.

[0015] Compared with the existing solutions, the present invention has the following beneficial effects: the neuromusculoskeletal model only requires 4 muscles in the estimation process, which greatly reduces the number of parameters in the model application process. Introducing the muscle physiological information of the neuromusculoskeletal model into the training process of the neural network model can enhance the interpretability of the neural network model, add a physical constraint to the neural network model, and ensure that the output of the model meets the expected effect, thereby ensuring the safety of the exoskeleton. This method also makes up for the problem of instability and large fluctuations in the collected torque signal. By utilizing the fused torque signal of the two, smoother torque label data can be obtained, which enhances the generalization performance of the neural network model and has higher prediction accuracy under the conditions of shorter training time and less training data. At the same time, this method avoids the problem of numerous parameters and low accuracy when the neuromusculoskeletal model is used directly, making the neural network model more suitable for prediction tasks with high real-time requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a display diagram of the joint torque prediction model experimental platform of the present invention.

[0017] Figure 2 This is a diagram of the positions of the myoelectric signal acquisition patches of the present invention.

[0018] Figure 3 This is the effect diagram of the myoelectric signal wavelet packet decomposition of the present invention.

[0019] Figure 4 This is a diagram showing the joint angle and joint angular velocity processing results of the present invention.

[0020] Figure 5 This is a diagram of the joint torque processing results of the present invention.

[0021] Figure 6 Schematic diagram of the neuromusculoskeletal model of the present invention.

[0022] Figure 7 It is a flow chart of the sparrow optimization algorithm of the present invention.

[0023] Figure 8 It is a structural diagram of the neural network model of the present invention.

[0024] Figure 9 It is a structural diagram of the technical solution of the present invention.

[0025] Figure 10 It is the loss function curve of the neural network model training effect of the present invention.

[0026] Figure 11 It is a loss function curve of the neural network model training effect of the fusion neuromusculoskeletal model of the present invention.

[0027] Figure 12 This is a diagram showing the final knee joint torque prediction effect of the neural network model of the fusion neuromusculoskeletal model of the present invention.

[0028] Figure 13 It is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0029] The following examples of the invention, combined with the accompanying drawings, provide detailed descriptions of the experimental scheme, data processing, model construction and training, etc., to further clarify the application areas, design ideas and technical solutions of the invention.

[0030] Example 1

[0031] This embodiment relates to a joint torque prediction method based on electromyography and musculoskeletal physiological information, comprising the following steps:

[0032] S1: According to Figure 1 As shown in the figure, an experimental plan is designed on the experimental platform of the built joint torque prediction model to collect offline data. A data acquisition program is built on the PC using Simulink to achieve simultaneous acquisition of joint angles, joint torques, myoelectric signals, and IMU signals. The PC and exoskeleton communicate through the Raspberry Pi, and the myoelectric sensors and IMU sensors are directly connected to the PC through the USB interface. When collecting data, you need to use Figure 1 The device in the exoskeleton is worn, and the exoskeleton's built-in torque sensor and angle sensor are used to collect the joint angle and joint torque signals of the knee and hip joints during movement, and save them in Simulink. The electromyographic signals are collected using a six-channel electromyographic sensor, such as Figure 2Each muscle corresponds to a channel. Yellow and green electrodes are attached to the corresponding muscles, and the red patch is attached to the knee as a reference electrode. In Simulink, the data is collected synchronously with signals such as joint angles. For the angle data of the ankle joint, two IMU sensors are used to collect the pitch angles respectively, and the data is obtained after processing. A total of 7 male subjects were included in the experiment, with a height of 178±7cm and a weight of 78±14kg. The acquisition frequency of the electromyographic signals, joint angles, and joint torque signals was 1000Hz. Data were collected at walking speeds of 0.5km / h, 1.0km / h, and 1.5km / h respectively. Two groups were collected at the same speed, and each group of data was collected for 30s.

[0033] S2: Pre-process the collected data. Use wavelet packet decomposition to analyze the energy distribution of the electromyographic signal. According to the decomposition results, filter the electromyographic signal to remove noise, such as 50Hz power frequency interference, high-frequency noise, etc. In the 10-level wavelet packet decomposition, the high-frequency decomposition results of the first level show that the electromyographic signal has higher energy before each movement begins. At this time, the frequency range is about 0-500Hz. The low-frequency decomposition results of the seventh level better retain the overall trend of the electromyographic signal and have a certain filtering effect. The results are as follows: Figure 3 As shown. After experiments, it was finally determined that the bandpass filter cutoff frequency is 300-450Hz, and the low-pass filter cutoff frequency is 2Hz. After filtering, the original electromyographic signal, high-frequency signal and low-frequency signal are retained respectively. The joint angle and joint torque data are processed by sliding average window to make the data smoother, and the window length is set to 250 and 400 time steps respectively. The angular velocity signal of each joint is calculated by the central difference method, and the processed joint angle, joint angular velocity and joint torque signals are saved. The processing effect is shown as follows Figure 4 、 5 As shown in the figure, the original EMG signal, high-frequency EMG signal, low-frequency EMG signal, joint angle signal, and joint angular velocity signal are used as input features, and the joint torque is used as the label. Finally, the sliding window method is used to process the dataset. All feature data in a window corresponds to the label data of its last time step, which is used as a sample. The window length is 50 time steps (should not exceed the time length that the EMG signal leads the muscle movement), and the sliding step size is 1. This is used to create a dataset for neural network model training, which will be used for subsequent model training.

[0034] S3: According to Figure 6The neuromusculoskeletal model in [1] models the four muscles: the rectus femoris, vastus medialis, medial gastrocnemius, and lateral gastrocnemius. Processed low-frequency electromyographic signals and joint angle data are input into the neuromusculoskeletal model. The electromyographic signals are converted from neural signals to muscle activation through muscle activation dynamics. Muscle contraction dynamics converts muscle activation into muscle tension. Joint dynamics converts muscle tension into joint torque, resulting in the estimated torque output by the neuromusculoskeletal model. The modeling method is as follows:

[0035] For the obtained electromyographic signal e, calculate its neural activation:

[0036] u(t)=ae(tT)-bu(t-1)-cu(t-2) (1)

[0037] Where u(t-1) and u(t-2) represent the neural activation at moments t-1 and t-2, respectively, and e(tT) is the surface electromyographic signal with a time delay of T. The constraints are:

[0038] b=d1+d2,c=d1d2,abc=1

[0039] |d1|<1, |d2|<1 (2)

[0040] Using the u(t) obtained above, calculate the muscle activation a(t) as follows:

[0041]

[0042] Where A is the nonlinear shape factor, ranging from A∈[-3,0].

[0043] Then the muscle contraction force F is obtained according to the following calculation formula m (t):

[0044] F m (t)=f(v)f(l)a(t)F m0 (4)

[0045] Where, F m (t) is the time-varying muscle fiber contraction force, f(v) is the normalized muscle contraction velocity-force relationship, f(l) is the normalized muscle fiber length-force relationship, F m0 is the maximum isometric muscle fiber contraction force. Since the muscle fiber contraction force is less affected by the muscle contraction speed under slow walking conditions, f(v) can be regarded as a constant in the modeling process.

[0046] The muscle contraction force is obtained and then converted into muscle tendon force. The relationship between muscle force and muscle tendon force is as follows:

[0047] F mt=F m cosφ (5)

[0048] Here φ represents the feathering angle.

[0049] Finally, the joint torque of the knee joint is calculated according to the following formula:

[0050]

[0051] Among them, T J (θ, t) is the joint torque of joint J, r i (θ) is the moment arm of the i-th muscle of joint J when the joint angle is θ, F mt(i) (θ,t) refers to the muscle tendon force of the i-th muscle at angle θ and time t.

[0052] In addition, the time-varying muscle-tendon length l required in the calculation process mt The estimation can be done according to the following method:

[0053]

[0054]

[0055] Where θ1, θ2, θ3, θ4 are the joint angles of muscle-related joints, a1, a2, a3, ..., a n is the parameter to be identified.

[0056] S4: There are four muscles involved in the neuromusculoskeletal model, each of which contains the maximum isometric muscle force F m0 , optimal feather angle φ0, optimal muscle fiber length l m0 , tendon length l t , normalized muscle force-velocity relationship f(v), and individual-specific physiological parameters d1, d2, A, a total of 23 parameters need to be identified. The sparrow optimization algorithm is used for parameter identification. Table 1 lists the reasonable range of each parameter. The sparrow optimization algorithm first randomly generates multiple populations as possible solutions based on the reasonable range of each parameter in Table 1, and calculates the fitness of each solution separately. Each population is sorted according to the fitness, and all populations are updated to the population with the highest fitness. It is continuously iterated until the stopping condition is met to generate the optimal solution, and then the optimal solution is passed back to the neuromusculoskeletal model to obtain the estimated torque of the knee joint that is closest to the actual torque. The flowchart of the sparrow optimization algorithm is as follows Figure 7 As shown. Among them, the fitness function uses the mean absolute error between the estimated torque and the true torque of the neuromusculoskeletal model, and the fitness function is:

[0057]

[0058] Where y represents the true moment, represents the estimated torque output by the neuromusculoskeletal model, and n represents the number of data points.

[0059] Table 1 Parameter range setting list

[0060]

[0061]

[0062] The conditions for determining whether optimization should stop are:

[0063] (1) If the fitness improvement is less than 0.01 for seven consecutive rounds, or the difference between the maximum and minimum fitness in the same round is less than 0.01, and the fitness function value is less than 1.45, the optimization is stopped. Otherwise, the population is regenerated for parameter optimization.

[0064] (2) The maximum number of iterations is 30 rounds. If the condition of less than 1.45 is not met after 30 rounds, the population with the smallest fitness function value in all rounds is taken.

[0065] S5: Build Figure 8 The neural network model shown in the figure begins with a feature selection attention mechanism, assigning different weights to different features. Two 2D convolutional layers with different numbers of convolutional kernels, both with a kernel size of 3×3, are then used to extract spatial relationships between different features. The first convolutional layer has 32 kernels, and the second has 16 kernels. Both layers have a sliding stride of (1,1), and use a Reluctant Unit (ReLU) activation function. An average pooling layer is added between the two convolutional layers to reduce the data dimensionality. The pooling window size is 3×3 with a stride of (1,1). A dropout layer with a dropout rate of 0.25 is also added to prevent overfitting. A layer normalization operation is added before the activation function of the second convolutional layer to stabilize the training process and prevent vanishing or exploding gradients. A bidirectional long short-term memory neural network is then used to extract long-range temporal dependencies between features. The number of hidden units in the layer is 32. Finally, a fully connected layer with 64 neurons is used to integrate features and improve overall model performance. The output layer uses a linear activation function. The model uses the Adam optimizer with a learning rate of 0.001.

[0066] S6: Construct a loss function. Incorporating the neuromusculoskeletal model into the training process of the neural network model, through the neuromusculoskeletal model, the physiological information of the muscle and the transformation process from electromyographic signals to joint torque are introduced into the training process of the neural network model, which can add constraints to the system and enhance its interpretability. By constructing a suitable loss function, the fusion of the neuromusculoskeletal model and the neural network model can be achieved. The method adopted by the present invention is to convert the output estimated torque (TNMS ) and true torque (T true ) is assigned a weight and added together to obtain a fusion torque (T label ), in specific implementation, the estimated moments are given and true torque The weight of the neural network model is defined as the prediction moment (T Pred ) and true torque (T true The mean absolute error between the two is the loss function, and the neural network model is trained by calculating and minimizing the loss function. That is,

[0067] Fusion torque:

[0068] T label =αT NMS +βT true (7)

[0069] Loss function L:

[0070]

[0071] Where y represents the true moment, Represents the fusion torque, and n represents the number of data points in the training set.

[0072] S7: Define the model evaluation index. To evaluate the final effect of the model, define the absolute average error between the model's predicted output and the true torque as the model evaluation index:

[0073]

[0074] Where y represents the true moment, Represents the prediction moment of the model output, and n represents the number of data points in the test set.

[0075] S8: Train the model and save it. Divide the data into training set and test set in the ratio of 8:2 for training the fusion model. After all the data are trained, save the model for subsequent model calls. The overall design idea is as follows Figure 9 Framework diagram of the technical solution presented.

[0076] Finally, using the sparrow optimization algorithm for parameter identification, the average error between the estimated torque output by the neuromusculoskeletal model and the actual torque was 1.604NM. Compared with the simulated annealing algorithm, the average error was 2.805NM, and the sparrow optimization algorithm showed a significant improvement. When only the neural network model was trained, the training error of the predicted joint torque output was approximately 1.457NM, and the model gradually converged after more than 80 rounds of training. Figure 10As shown in the figure. The neural network model that integrates the neuromusculoskeletal model has an output prediction joint torque training error of approximately 0.617 N·M. In the model evaluation index, it is only 0.939 N·M. In addition, the model gradually converges after only 40 rounds of training, as shown in the figure. Figure 11 The final knee joint torque prediction effect is shown as Figure 12 shown.

[0077] Example 2

[0078] This embodiment relates to a joint torque prediction device based on electromyography and musculoskeletal physiological information, including a memory and one or more processors. The memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the joint torque prediction method based on electromyography and musculoskeletal physiological information of Example 1.

[0079] Example 3

[0080] The third aspect of the present invention relates to a computer-readable storage medium on which a program is stored, including data collection and processing, modeling of a neuromusculoskeletal model, neural network model structure and parameters, parameters of the trained model, and a model training method. When the program is executed by a processor, the joint torque prediction method based on electromyography and musculoskeletal physiological information of the present invention is implemented.

[0081] In summary, the present invention uses electromyography to predict joint torque during movement. Combined with physiological prior knowledge of the neuromusculoskeletal model, it can increase the constraints on the output of the neural network model, more reliably predict joint torque during movement while wearing an exoskeleton, and better ensure user safety. At the same time, the reduced number of muscles makes it more convenient to use and more suitable for tasks with strong real-time requirements. It provides reliable prediction values ​​for motion control, which is crucial for improving the compliance and comfort of the exoskeleton rehabilitation robot's motion control process.

[0082] The above embodiment describes one embodiment of the present invention. Those skilled in the art may make various adjustments, changes, and improvements to the present invention. Any adjustments, changes, and improvements made within the scope of the core principles of the present invention are considered to be within the scope of the present invention.

Claims

1. A joint torque prediction method based on myoelectric and musculoskeletal physiological information includes the following steps: S1: Build a data acquisition platform; use Raspberry Pi, MATLAB, myoelectric sensors, and an exoskeleton robot to build the entire data acquisition system to achieve offline data acquisition; determine the collection of myoelectric signals of the rectus femoris, vastus medialis, medial head of the gastrocnemius, and lateral head of the gastrocnemius, as well as the joint angle signals and joint angular velocity signals of the hip, knee, and ankle joints based on muscle synergy; S2: Neuromusculoskeletal model establishment: Using the PyCharm platform, the four muscles were modeled using the collected electromyographic signals and joint angle signals. The torque of each muscle was obtained and the torque of the knee joint that changes with movement was calculated. Due to the small number of muscles used, an output coefficient was added to the output to make the model output closer to the actual torque, thus obtaining the joint torque of the knee joint. The modeling method is as follows: For the obtained electromyographic signal e, calculate its neural activation: u(t)=ae(tT)-bu(t-1)-cu(t-2) (1) in, u(t-1) and u(t-2) represent the neural activation at moments t-1 and t-2, respectively, e(tT) is the surface electromyographic signal with a time delay of T, and the constraints are: b=d1+d2,c=d1d2,abc=1 |d1|<1, |d2|<1 (2) Using the u(t) obtained above, calculate the muscle activation a(t) as follows: Where A is the nonlinear shape factor, ranging from A∈[-3,0]; Then the muscle contraction force F is obtained according to the following calculation formula m (t): F m (t)=f(v)f(l)a(t)F m0 (4) Where, F m (t) is the time-varying muscle fiber contraction force, f(v) is the normalized muscle contraction velocity-force relationship, f(l) is the normalized muscle fiber length-force relationship, F m0 is the maximum isometric muscle fiber contraction force; since the muscle fiber contraction force is less affected by the muscle contraction speed under slow walking conditions, f(v) is considered a constant in the modeling process; The muscle contraction force is obtained and then converted into muscle tendon force; the relationship between muscle force and muscle tendon force is as follows: F t =F m cosφ (5) Here φ represents the feathering angle; Finally, the joint torque of the knee joint is calculated according to the following formula: Among them, T J (θ, t) is the joint torque of joint J, r i (θ) is the moment arm of the i-th muscle of joint J when the joint angle is θ, F mt (i)(θ,t) refers to the muscle-tendon force of the i-th muscle at angle θ, time t; In addition, the time-varying muscle-tendon length l required in the calculation process mt The estimation of is based on formula (7): Where θ1, θ2, θ3, θ4 are the joint angles of muscle-related joints, a1, a2, a3, ..., a n is the parameter to be identified; S3: Identify the parameters of the neuromusculoskeletal model; the physiological parameters required by the neuromusculoskeletal model, including the optimal muscle force F m0 , maximum pennation angle φ0, optimal muscle fiber length l m0 , tendon length l t As well as the muscle contraction velocity f(v), which has individual differences; all parameters are taken as parameters to be optimized, a reasonable range of each parameter is given, and the optimization is performed using the sparrow optimization algorithm, which is then passed back to the neuromusculoskeletal model to improve the estimation accuracy of the neuromusculoskeletal model; S4: Build a neural network model; use the collected electromyographic signals, joint angle signals, and joint angular velocity signals as input features, and use the attention mechanism to assign weight matrices to different features to indicate their importance. Use two convolutional layers and a pooling layer to extract the spatial relationship between features and perform dimensionality reduction on the data. Add a bidirectional long short-term memory layer to extract long-term temporal dependencies. Add a dropout layer in between to reduce overfitting. Finally, add a fully connected layer to output the predicted joint torque. S5: Construct a loss function. During training, based on the aforementioned neuromusculoskeletal model, use electromyographic signals and joint angles to estimate joint torques during movement. Combined with the true torque, different weights are assigned to the estimated torque and the true torque, respectively, to obtain a fused torque estimate as a label for offline training of the neural network. The mean absolute error between the fused torque and the predicted torque output by the neural network model is calculated as the loss function for training, so that the model continuously fits and approaches the true torque. S6: Design evaluation indicators; use the mean absolute error between the predicted torque and the true torque of the neural network model as the evaluation indicator of the model; S7: Model training and saving: Train the model using data and save it for subsequent use.

2. The joint torque prediction method based on myoelectricity and musculoskeletal physiological information according to claim 1, characterized in that: Step S3 specifically includes: There are four muscles involved in the neuromusculoskeletal model, each of which contains the maximum isometric muscle force F m0 , optimal feather angle φ0, optimal muscle fiber length l m0 , tendon length l t , normalized muscle force-velocity relationship f(v), and individual-specific physiological parameters d1, d2, and A, a total of 23 parameters need to be identified; parameter identification uses the sparrow optimization algorithm. The sparrow optimization algorithm first randomly generates multiple populations as possible solutions based on the reasonable range of each parameter, and calculates the fitness of each solution separately. Each population is sorted according to fitness, and all populations are updated to move closer to the population with the highest fitness. It is continuously iterated until the stopping condition is met to generate the optimal solution. The optimal solution is then passed back to the neuromusculoskeletal model to obtain the estimated torque of the knee joint that is closest to the true torque; the fitness function uses the mean absolute error between the estimated torque and the true torque of the neuromusculoskeletal model. The fitness function is: Where y represents the true moment, represents the estimated torque output by the neuromusculoskeletal model, and n represents the number of data points; The conditions for determining whether optimization should stop are: (1) If the fitness improvement is less than 0.01 for seven consecutive rounds, or the difference between the maximum and minimum fitness in the same round is less than 0.01, and the fitness function value is less than 1.45, the optimization is stopped. Otherwise, the population is regenerated for parameter optimization. (2) The maximum number of iterations is 30 rounds. If the condition of less than 1.45 is not met after 30 rounds, the population with the smallest fitness function value in all rounds is taken.

3. The joint torque prediction method based on myoelectricity and musculoskeletal physiological information according to claim 1, characterized in that: Step S4 specifically includes: First, a feature selection attention mechanism is used to assign different weights to different features. Then, two two-dimensional convolutional layers with different numbers of convolutional kernels and a kernel size of 3×3 are used to extract the spatial relationship between different features. The first convolutional layer has 32 convolutional kernels, and the second convolutional layer has 16 convolutional kernels. The sliding step size is (1,1), and the ReLU activation function is used. An average pooling layer is added between the two convolutional layers to reduce the dimensionality of the data. The pooling window size is 3×3 and the step size is (1,1). At the same time, a dropout layer with a dropout rate of 0.25 is added to prevent overfitting. A layer normalization operation is added before the activation function of the second convolutional layer to stabilize the training process and prevent gradient vanishing or gradient exploding. Then, a bidirectional long short-term memory neural network is used to extract the long-range temporal dependencies of features. The number of hidden layer units is 32. Finally, a fully connected layer is used to achieve feature integration and improve the overall performance of the model. The number of neurons in the fully connected layer is 64, and the final output layer uses a linear activation function. The model uses the Adam optimizer with a learning rate of 0.

001.

4. The joint torque prediction method based on myoelectricity and musculoskeletal physiological information according to claim 1, characterized in that: The specific method of constructing the loss function in step S5 is: the output estimated torque (T NMS ) and true torque (T true ) is assigned a weight and added together to obtain a fusion torque (T label ), in specific implementation, the estimated moments are given and true torque The weight of the neural network model; define the prediction moment (T Pred ) and true torque (T true ) is the loss function, and the neural network model is trained by calculating and minimizing the loss function; that is, Fusion torque: T label =αT NMS +βT true (7) Loss function L: Where y represents the true moment, Represents the fusion torque, and n represents the number of data points in the training set.

5. The joint torque prediction method based on myoelectricity and musculoskeletal physiological information according to claim 1, characterized in that: The model evaluation indicators of step S6 are specifically: Where y represents the true moment, Represents the prediction moment of the model output, and n represents the number of data points in the test set.

6. A joint torque prediction device based on myoelectricity and musculoskeletal physiological information, characterized in that: It includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the joint torque prediction method based on electromyography and musculoskeletal physiological information as described in any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that A program is stored thereon, and when the program is executed by the processor, the joint torque prediction method based on myoelectricity and musculoskeletal physiological information according to any one of claims 1 to 5 is implemented.

Images