Upper limb muscle force prediction method, system and device and storage medium

Abstract

The invention provides an upper limb muscle force prediction method, system and device and a storage medium, and the method comprises the steps: collecting original electromyographic signals of a plurality of upper limb muscles, and preprocessing the original electromyographic signals into electromyographic signals; inputting the electromyographic signals into a trained muscle force prediction model for muscle force prediction; according to the muscle force prediction model, a potential muscle cooperation mode and an activation degree are extracted from an electromyographic signal, a biomechanical model and a deep neural network model are constructed, the biomechanical model is utilized to establish a relationship between the activation degree and muscle force, and guidance information is provided for training of the deep neural network model. And the dependence of the deep neural network model on high-quality muscle force truth value labeling data is reduced. Biomechanical information is introduced into a multi-target joint loss function, and during model training, it is ensured that the prediction result of the deep neural network model follows the biomechanical law by minimizing multi-target joint loss.

Description

Technical Field

[0001] This invention relates to a muscle strength prediction method, specifically, to an upper limb muscle strength prediction method, system, device, and storage medium. Background Technology

[0002] With an aging population and rising stroke rates, there is a significant clinical need for rehabilitation treatment of upper limb motor dysfunction. Rehabilitation assistive devices, such as exoskeleton robots and functional electrical stimulation (fEP), have seen tremendous development. However, precise and intelligent assisted rehabilitation relies on the design of algorithms for understanding upper limb rehabilitation intentions, particularly the prediction of high-precision muscle strength. This indicator directly determines the safety of human-computer interaction and the reliability of quantitative assessment of rehabilitation effects.

[0003] Physics-based biomechanical models require establishing a system of differential equations relating muscle activation and force transmission. Solving these equations numerically involves high-dimensional nonlinear optimization problems, with single muscle force predictions taking 200-500 ms, which cannot meet the real-time control requirements of assistive rehabilitation equipment. While simplified biomechanical models can speed up predictions, they sacrifice accuracy and reduce generalization ability. Traditional deep neural network models are primarily data-driven in their learning process, relying heavily on large amounts of high-quality labeled datasets for training. However, obtaining ground truth muscle force values ​​requires multi-axis force measurement platforms costing millions of yuan, and labeling single-patient data is extremely time-consuming. The scarcity of high-quality data makes it difficult for these models to obtain accurate and reliable predictions, severely hindering their clinical adoption. More importantly, these black-box models lack prior biomechanical knowledge constraints, often producing results that contradict physiological common sense. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by providing a method, system, device, and storage medium for predicting upper limb muscle strength by incorporating prior biomechanical knowledge into a deep neural network model.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The first aspect of this invention provides a method for predicting upper limb muscle strength, comprising: Raw electromyographic (EMG) signals were collected from several upper limb muscles, and the raw EMG signals were preprocessed into EMG signals. Electromyographic signals are input into a trained muscle strength prediction model to predict muscle strength. The muscle strength prediction model was constructed and trained using the following methods: Potential muscle synergistic patterns and activation levels are extracted from electromyographic signals based on nonnegative matrix factorization. A biomechanical model was constructed, and the theoretical muscle force value of the electromyographic signal was calculated based on the activation level and biomechanical parameters. A deep neural network model is constructed, with the electromyographic signal as input, the theoretical muscle strength value and the muscle synergy pattern as labels, so that the deep neural network model learns the relationship between the electromyographic signal, the muscle strength value and the muscle synergy pattern, and predicts the muscle strength value and the muscle synergy pattern. A multi-objective joint loss function is constructed, and the deep neural network model is trained using the multi-objective joint loss function, which includes a biomechanical information loss function and a deep neural network loss function. The biomechanical information loss function is the electromyography reconstruction loss function, and the deep neural network loss function includes the muscle strength prediction loss function and the muscle synergy prediction loss function. The electromyographic remodeling loss function L1 = ||EMG - W × H|| 2 ; The muscle coordination prediction loss function L2=‖WW pred || 2 ; The muscle strength prediction loss function L3=‖F Hill -F NN || 2 ; The multi-objective joint loss function is L = λ1L1 + λ2L2 + λ3L3; In the above formula, EMG represents the preprocessed electromyographic signal, W represents the muscle synergy pattern extracted from the electromyographic signal, and H represents the activation level extracted from the electromyographic signal; F Hill W represents the theoretical muscle force value calculated by the biomechanical model. pred F represents the muscle coordination pattern predicted by the deep neural network model. NN λ1 represents the muscle strength value predicted by the deep neural network model, λ2 represents the weight coefficient of the electromyography reconstruction loss function, λ3 represents the weight coefficient of the muscle synergy prediction loss function, and λ4 represents the weight coefficient of the muscle strength prediction loss function. The muscle strength prediction model is pre-trained by minimizing the multi-objective joint loss function.

[0006] The pre-training of the muscle strength prediction model utilizes a biomechanical model to establish a relationship between activation level and muscle strength, providing guidance for the training of the deep neural network model and reducing the deep neural network model's dependence on high-quality ground truth muscle strength data. By incorporating biomechanical information into the multi-objective joint loss function, during model training, minimizing the multi-objective joint loss not only minimizes the deep neural network loss but also minimizes the loss of biomechanical information, ensuring that the prediction results of the deep neural network model follow biomechanical laws.

[0007] As a preferred technical solution of the present invention, when training the muscle strength prediction model, the muscle strength prediction loss function is first used to enable the model to initially learn the direct mapping relationship between electromyographic signals and muscle strength. Then, the multi-objective joint loss function is used to simultaneously optimize the muscle coordination and muscle strength predicted by the model, so that the prediction output results follow biomechanical constraints.

[0008] As a preferred technical solution of the present invention, the non-negative matrix decomposition method is a time-varying non-negative matrix decomposition method based on the sliding window mechanism. The time-varying non-negative matrix decomposition method performs local segmental modeling on the continuous electromyographic signal sequence, performs standard non-negative matrix decomposition in each window, and then extracts the local muscle synergy pattern and its activation degree. The expression for the time-varying nonnegative matrix decomposition method is: EMG i =W i ×H i , i=1,2,...,N; In the above formula, EMG i ∈R m×l W represents the electromyographic signal of length l within the i-th sliding window. i ∈R m×k H represents the local muscle coordination matrix within this time window. i ∈R k×l Let represent the corresponding local activation coefficient matrix, and k represent the number of collaborations, which empirically ranges from 4 to 6.

[0009] The time-varying nonnegative matrix factorization method based on the sliding window mechanism described above exhibits good temporal dynamic adaptability. By reconstructing local coordination patterns within each time window, it adapts to the dynamic changes in muscle coordination over time during the acquisition of movements. The obtained muscle coordination matrix is ​​related to the muscle coordination effect, and the obtained activation coefficient matrix is ​​highly correlated with muscle strength. Using the activation coefficient matrix as input for muscle strength prediction can significantly improve prediction accuracy.

[0010] As a preferred technical solution of the present invention, the biomechanical model adopts the Hill model, and the model parameters are initialized based on the upper limb muscle biomechanical parameters and anatomical criterion values ​​provided by the OpenSim platform, so that the model is more in line with the actual situation of the electromyography signal acquisition experimental subjects. The theoretical expression for muscle strength is: F Hill =F max ×a(t)×f l (l)×f v (v); The total length of the muscle and tendon is: l(t) = l m +l T ; The normalized activation level expression is: a(t) = Sigmoid(H); In the above formula, F max The maximum isometric contractile force of the muscle, a(t) is the normalized activation level, and f l (l) represents the relationship between force and muscle length, f v (v) represents the relationship between force and velocity, l m For muscle fiber length, l T H represents the tendon length, and H represents the degree of activation.

[0011] As a preferred embodiment of the present invention, the deep neural network model adopts an LSTM neural network model to convert electromyographic signals into time window sequence data, and the time window sequence data is used as the input data of the LSTM neural network model. The LSTM neural network model includes a three-layer structure, each containing a 64-unit hidden layer with the activation function of tanh. The time window sequence data passes through the three-layer LSTM network, is processed by a fully connected layer, and then input into the softplus activation function to predict muscle strength and muscle coordination patterns.

[0012] The activation level of the input electromyographic signal is extracted through non-negative matrix factorization. The theoretical muscle force value is calculated by establishing a relationship between the Hill model and muscle force. The muscle synergy pattern extracted through non-negative matrix factorization and the theoretical muscle force value calculated by the Hill model are used as labels for the input data of the LSTM neural network model. Training with a multi-objective joint loss function ensures that the prediction results of the LSTM neural network model conform to biomechanical principles, thereby improving the generalization ability of the LSTM neural network model.

[0013] As a preferred technical solution of the present invention, the acquisition of electromyographic signals of several upper limb muscles includes using the Delsys Trigno Wireless sEMG system to attach sensors to the surface of the anterior deltoid, middle deltoid, posterior deltoid, pectoralis major, biceps brachii, triceps brachii, brachialis, and brachioradialis muscles of the experimental subject to acquire raw electromyographic signals.

[0014] As a preferred technical solution of the present invention, the preprocessing includes passing the original electromyography signal sequentially through a 50Hz power frequency notch filter, a bandpass filter, and a full-wave rectification, and then calculating the root mean square value based on the sliding window method.

[0015] The acquired surface electromyography (EMG) signal is a weak, low-frequency signal with an amplitude ranging from a few microvolts to a few millivolts, making it susceptible to noise contamination. By employing 50Hz power frequency notch filtering, bandpass filtering, and full-wave rectification, interference from power frequency noise is reduced, motion artifacts concentrated in the low-frequency band and useless signals exceeding the frequency band of the surface EMG signal are removed. After further processing using the root mean square calculation based on the sliding window method, the EMG signal is compressed into a low-frequency characteristic time series, which can better reflect the firing rate of motor neurons per unit time.

[0016] A second aspect of the present invention provides an upper limb muscle strength prediction system, comprising: The acquisition and preprocessing module acquires raw electromyographic signals from several upper limb muscles and preprocesses the raw electromyographic signals into electromyographic signals. The prediction module inputs electromyographic signals into the trained muscle strength prediction model to predict muscle strength. The muscle strength prediction model was constructed and trained using the following methods: Potential muscle synergistic patterns and activation levels are extracted from electromyographic signals based on nonnegative matrix factorization. A biomechanical model was constructed, and the theoretical muscle force value of the electromyographic signal was calculated based on the activation level and biomechanical parameters. A deep neural network model is constructed, with the electromyographic signal as input, the theoretical muscle strength value and the muscle synergy pattern as labels, so that the deep neural network model learns the relationship between the electromyographic signal, the muscle strength value and the muscle synergy pattern, and predicts the muscle strength value and the muscle synergy pattern. A multi-objective joint loss function is constructed, and the deep neural network model is trained using the multi-objective joint loss function, which includes a biomechanical information loss function and a deep neural network loss function. The biomechanical information loss function is the electromyography reconstruction loss function, and the deep neural network loss function includes the muscle strength prediction loss function and the muscle synergy prediction loss function. The electromyographic remodeling loss function L1 = ||EMG - W × H|| 2 ; The muscle coordination prediction loss function L2=‖WW pred || 2 ; The muscle strength prediction loss function L3=‖F Hill -F NN || 2 ; The multi-objective joint loss function is L = λ1L1 + λ2L2 + λ3L3; In the above formula, EMG represents the preprocessed electromyographic signal, W represents the muscle synergy pattern extracted from the electromyographic signal, and H represents the activation level extracted from the electromyographic signal; FHill W represents the theoretical muscle force value calculated by the biomechanical model. pred F represents the muscle coordination pattern predicted by the deep neural network model. NN λ1 represents the muscle strength value predicted by the deep neural network model, λ2 represents the weight coefficient of the electromyography reconstruction loss function, λ3 represents the weight coefficient of the muscle synergy prediction loss function, and λ4 represents the weight coefficient of the muscle strength prediction loss function. The muscle strength prediction model is pre-trained by minimizing the multi-objective joint loss function.

[0017] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the upper limb muscle strength prediction method described in the first aspect.

[0018] A fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the upper limb muscle strength prediction method described in the first aspect.

[0019] This invention has outstanding substantive features and significant progress compared to the prior art. Specifically, this invention provides an upper limb muscle strength prediction method, system, device and storage medium. Through a biomechanical model, it provides guidance information for the training of deep neural network models, overcomes the limitation of insufficient high-quality muscle strength ground truth labeled data, and by introducing biomechanical information into a multi-objective joint loss function, it enables deep neural network models to follow biomechanical laws and have better generalization ability. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a schematic diagram of the structure of the present invention; Figure 3 This is a schematic diagram of the structure of the LSTM neural network model described in this invention; Figure 4 This is a schematic diagram of the Hill muscle strength three-element model described in this invention. Detailed Implementation

[0021] The technical solution of the present invention will be further described in detail below through specific embodiments.

[0022] Example 1 In this embodiment of the application, a method for predicting upper limb muscle strength is proposed: Raw electromyographic (EMG) signals were collected from several upper limb muscles, and the raw EMG signals were preprocessed into EMG signals. Electromyographic signals are input into a trained muscle strength prediction model to predict muscle strength. The muscle strength prediction model was constructed and trained using the following methods: Potential muscle synergistic patterns and activation levels are extracted from electromyographic signals based on nonnegative matrix factorization. A biomechanical model was constructed, and the theoretical muscle force value of the electromyographic signal was calculated based on the activation level and biomechanical parameters. A deep neural network model is constructed, with the electromyographic signal as input, the theoretical muscle strength value and the muscle synergy pattern as labels, so that the deep neural network model learns the relationship between the electromyographic signal, the muscle strength value and the muscle synergy pattern, and predicts the muscle strength value and the muscle synergy pattern. A multi-objective joint loss function is constructed, and the deep neural network model is trained using the multi-objective joint loss function, which includes a biomechanical information loss function and a deep neural network loss function. The biomechanical information loss function is the electromyography reconstruction loss function, and the deep neural network loss function includes the muscle strength prediction loss function and the muscle synergy prediction loss function. The electromyographic remodeling loss function L1 = ||EMG - W × H|| 2 ; The muscle coordination prediction loss function L2=‖WW pred || 2 ; The muscle strength prediction loss function L3=‖F Hill -F NN || 2 ; The multi-objective joint loss function is L = λ1L1 + λ2L2 + λ3L3; In the above formula, EMG represents the preprocessed electromyographic signal, W represents the muscle synergy pattern extracted from the electromyographic signal, and H represents the activation level extracted from the electromyographic signal; F Hill W represents the theoretical muscle force value calculated by the biomechanical model. pred F represents the muscle coordination pattern predicted by the deep neural network model. NN λ1 represents the muscle strength value predicted by the deep neural network model, λ2 represents the weight coefficient of the electromyography reconstruction loss function, λ3 represents the weight coefficient of the muscle synergy prediction loss function, and λ4 represents the weight coefficient of the muscle strength prediction loss function. The muscle strength prediction model is pre-trained by minimizing the multi-objective joint loss function.

[0023] To ensure that the upper limb muscle strength prediction results conform to biomechanical principles, several suitable upper limb muscles were selected as samples for data collection based on the temporal characteristics of electromyography (EMG) signals and the biomechanical characteristics of muscle strength. EMG signals were used as input data, and theoretical muscle strength values ​​and muscle synergy patterns were used as labels. This allowed a deep neural network model to learn the relationship between EMG signals, muscle strength values, and muscle synergy patterns, achieving a nonlinear mapping from EMG signals to muscle strength. A multi-objective joint loss function incorporating biomechanical knowledge was used for training, enabling the deep neural network model to fit the data. This effectively reflects that the muscle strength prediction results conform to prior biomechanical knowledge and also gives the muscle strength prediction model strong generalization ability.

[0024] Specifically, such as Figure 1 As shown, the upper limb muscle strength prediction method includes the following steps: Step S1: The subject performs upper limb rehabilitation movements, and electromyographic signals are collected and preprocessed using an sEMG sensor.

[0025] In one optional embodiment, the acquisition of electromyographic signals from several upper limb muscles includes using a DelsysTrigno Wireless sEMG system to attach sensors to the surface of the anterior deltoid, middle deltoid, posterior deltoid, pectoralis major, biceps brachii, triceps brachii, brachialis, and brachioradialis muscles of the experimental subject to acquire raw electromyographic signals.

[0026] Specifically, in this embodiment, surface electromyography (EMG) signals were collected from eight upper limb muscles, including the anterior deltoid, middle deltoid, posterior deltoid, pectoralis major, biceps brachii, triceps brachii, brachialis, and brachioradialis. Each muscle corresponds to one surface EMG channel, therefore the original collected EMG signal is an 8-channel data stream. A Delsys Trigno Wireless sEMG system was used for continuous sampling at a sampling rate of 2000Hz, with a data format of 2000 × 8 sample points per second.

[0027] It should be noted that, due to the excessive complexity of setting parameters for a complete upper limb muscle biomechanical model, this invention selects a subset of the major force-generating muscles during upper limb movement as the original electromyographic (EMG) signal sampling objects to calculate the theoretical muscle force of these upper limb muscles. Surface EMG is an electrophysiological signal collected by electrodes placed on the skin surface, characterized by its ease of acquisition, safety, and non-invasiveness, and it contains rich information about muscle state. Surface EMG serves as the primary source of muscle activation information and is used to establish a model between it and muscle force signals.

[0028] In one optional embodiment, the preprocessing includes sequentially passing the raw electromyographic signal through a 50Hz power frequency notch filter, a bandpass filter, and a full-wave rectification, and calculating the root mean square value based on the sliding window method.

[0029] Specifically, the original 8-channel electromyography (EMG) signals were preprocessed. The EMG signal data dimension was 8×T, where T represents the time length, i.e., the number of sampling points. First, a 50Hz power frequency notch filter with a stopband width of ±2Hz was used to eliminate the influence of 50Hz power frequency noise. Then, a fourth-order Butterworth bandpass filter was used, with a frequency range of 20-500Hz, to effectively filter out signals exceeding the range. Next, the preprocessed signal was full-wave rectified. Finally, the root mean square value of a 250ms sliding window was calculated, with a window step size of 50ms, compressing the original EMG signal into a low-frequency characteristic time series. The low-frequency characteristic time series data dimension was 8×N, where N is the number of sliding window output frames.

[0030] It should be noted that surface electromyography (EMG) signals are low-frequency signals with amplitudes ranging from a few microvolts to a few millivolts, with the main energy concentrated between 0-500 Hz. The temporal characteristics of EMG signals change significantly during muscle contraction, and it is generally believed that the temporal characteristics of EMG signals are correlated with muscle contraction force. EMG signal acquisition is susceptible to noise contamination. Preprocessing steps involve noise removal and signal enhancement of the raw EMG signals, including power frequency notch filtering, bandpass filtering, and full-wave rectification. This aims to provide high-quality input data that meets the modeling prerequisites for subsequent non-negative matrix factorization, biomechanical models, and deep neural network models. The root mean square (RMS) value can effectively reflect the firing rate of motor neurons per unit time.

[0031] Step S2 involves decomposing the electromyographic signal using a time-varying nonnegative matrix to dynamically extract the muscle coordination matrix and activation coefficient matrix.

[0032] In an optional embodiment, the non-negative matrix decomposition method is a time-varying non-negative matrix decomposition method based on a sliding window mechanism. The time-varying non-negative matrix decomposition method performs local segmental modeling on a continuous electromyography signal sequence, performs standard non-negative matrix decomposition within each window, and then extracts local muscle synergy patterns and their activation levels. The expression for the time-varying nonnegative matrix decomposition method is: EMG i =W i ×H i , i=1,2,...,N; In the above formula, EMG i ∈R m×l W represents the electromyographic signal of length l within the i-th sliding window. i ∈R m×k H represents the local muscle coordination matrix within this time window. i ∈R k×l Let represent the corresponding local activation coefficient matrix, and k represent the number of collaborations, which empirically ranges from 4 to 6.

[0033] Specifically, the low-frequency feature time series output from the preprocessing step undergoes non-negative matrix decomposition. The data input to the time-varying non-negative matrix decomposition method is an 8×N dimensional non-negative matrix. This non-negative matrix is ​​further divided into several segments using a sliding window approach, with the number of windows being K = [((Tl) / s) + 1], where T is the overall time series length, l is the length of each window, s is the sliding window step size, and K is the total number of windows. Standard non-negative matrix decomposition is performed within each time window to construct the time-varying muscle coordination pattern sequence {W}. i} and activation sequence {H i Extract the muscle synergy matrix and activation coefficient matrix.

[0034] It should be noted that not every channel of electromyography (EMG) signals correlates well with muscle strength signals. A time-varying nonnegative matrix factorization method based on a sliding window mechanism is used to decompose the signal into a matrix of muscle activation coefficients that are positively correlated with muscle strength. By removing channels inconsistent with muscle strength changes, the prediction accuracy of muscle strength can be further improved, the signal can be simplified, and computational and storage resources can be saved. For common upper limb functional movements, 4-6 coordinating signals are sufficient to explain most of the variability in EMG signals.

[0035] Step S3: Construct the Hill model based on muscle biomechanics and calculate theoretical muscle force.

[0036] In an optional embodiment, the biomechanical model adopts the Hill model, and the model parameters are initialized based on the upper limb muscle biomechanical parameters and anatomical criterion values ​​provided by the OpenSim platform, so that the model is more in line with the actual situation of the electromyography signal acquisition experimental subjects. The theoretical expression for muscle strength is: F Hill =F max ×a(t)×f l (l)×f v (v); The total length of the muscle and tendon is: l(t) = l m +l T ; The normalized activation level expression is: a(t) = Sigmoid(H); In the above formula, F max The maximum isometric contractile force of the muscle, a(t) is the normalized activation level, and f l (l) represents the relationship between force and muscle length, f v (v) represents the relationship between force and velocity, l m For muscle fiber length, l T H represents the tendon length, and H represents the degree of activation.

[0037] Specifically, such as Figure 4As shown, Hill's three-element muscle force model abstracts muscle into a muscle structural mechanics model composed of three elements: active contractile element (CE), series elastic element (SE), and parallel passive elastic element (PE). The series elastic element (SE) typically represents the tendon structure. Figure 4 Marked with "T", this muscle structure mechanics model reflects the mechanical behavior of the muscle-tendon complex and the contractile characteristics of muscles. The human muscle model can be approximated as a spring-damped system; the muscle force of each muscle unit can be described and calculated using the Hill model. The force-velocity relationship uses a hyperbolic tangent function to describe eccentric and concentric contraction characteristics. The force-length relationship uses a Gaussian curve, with the peak value located at the optimal length l_opt = 0.15m.

[0038] It should be noted that the OpenSim platform provides upper limb muscle biomechanical parameters that have been compiled based on extensive experimental and anatomical studies. These parameters are based on the anatomical structure and biomechanical characteristics of the real human body and are used as the initial settings for the model. However, since the muscle and skeletal structures of each individual are slightly different, directly using standard parameters may lead to simulation results that do not quite match reality. Therefore, anatomical calibration values ​​are obtained through the Subject01_arm26.osim upper limb model on the OpenSim platform, and these calibration values ​​are used to adjust the Hill model, making the parameters in the Hill model more closely reflect the actual situation of the specific experimental subjects.

[0039] Step S4: Construct an LSTM model to predict muscle strength and muscle synergy patterns based on input electromyography.

[0040] In one alternative embodiment, such as Figure 3 As shown, the deep neural network model uses an LSTM neural network model to convert electromyographic signals into time window sequence data, which is used as the input data of the LSTM neural network model. The LSTM neural network model includes a three-layer structure, each containing a 64-unit hidden layer with the activation function of tanh. The time window sequence data passes through the three-layer LSTM network, is processed by a fully connected layer, and then input into the softplus activation function to predict muscle strength and muscle coordination patterns.

[0041] Specifically, a 10-frame low-frequency feature time series is converted into a 500ms time window sequence, forming 8-channel × 10-frame two-dimensional surface electromyography (EMG) signal data. This data serves as input to the LSTM neural network model, enabling it to capture the dynamic information of EMG signals over time. The input data is sequentially processed through three LSTM layers and the tanh activation function, with a dropout rate set to 0.2. This allows the model to learn the mapping relationship between EMG and co-activation coefficients and muscle strength, preventing overfitting. A fully connected regression layer is added at the end of the LSTM network, outputting predicted muscle strength and muscle co-activation mode values, both of which are continuous real numbers. To ensure the non-negativity of the output, the output layer uses the softplus activation function, avoiding the vanishing gradient problem and ensuring that the prediction result is greater than zero, consistent with the physical meaning of muscle strength and activation coefficients.

[0042] Step S5: Define a loss function that includes electromyographic reconstruction error, muscle synergy prediction error, and muscle strength prediction error for joint optimization.

[0043] like Figure 2 As shown, the biomechanical information loss function is the electromyography reconstruction loss function, and the deep neural network loss function includes the muscle strength prediction loss function and the muscle synergy prediction loss function; The expression for the multi-objective joint loss function is as follows: L=λ1‖EMG-W·H‖ 2 +λ2‖WW pred || 2 +λ3‖F Hill -F NN || 2 ; The electromyography reconstruction loss, muscle synergy prediction loss, and muscle strength prediction loss are quantified by a multi-objective joint loss function. The Adam optimizer is then used to train the LSTM neural network model. By minimizing the multi-objective joint loss, the LSTM neural network model can be better fitted to the training data.

[0044] Specifically, the expression for the electromyographic remodeling loss function is L1 = ||EMG - W × H|| 2 Under the condition of nonnegative matrix factorization, the difference between the original electromyographic signal matrix (EMG) and the approximate EMG signal matrix (W·H) is measured by calculating the sum of the squared differences of the elements between the product of the original EMG signal matrix and the activation coefficient matrix and the cooperation matrix. The difference is then minimized through an iterative optimization process. When the EMG reconstruction loss is minimized, it indicates that the approximate EMG signal matrix (W·H) is the optimal solution.

[0045] Specifically, the expression for the muscle coordination prediction loss function is L2 = ||WW|| pred || 2Calculate the W matrix and W pred The sum of the squared differences of corresponding elements of the matrix. When the loss between the muscle synergy pattern predicted by the deep neural network model and the muscle synergy pattern obtained by non-negative matrix factorization converges, it indicates that the muscle synergy pattern predicted by the deep neural network model has biomechanical consistency.

[0046] Specifically, the expression for the muscle strength prediction loss function is L3 = ||F|| Hill -F NN || 2 When the deep neural network model predicts the muscle strength value (F) NN ) and muscle strength values ​​(F) of biomechanical muscle strength prediction models Hill The convergence of the loss between the two indicates that the muscle strength values ​​predicted by the deep neural network model conform to prior biomechanical knowledge.

[0047] It should be noted that λ1, λ2, and λ3 are the weight coefficients of the electromyography reconstruction loss function, the muscle synergy prediction loss function, and the muscle strength prediction loss function, respectively. They are used to balance the importance of the three parts of the multi-objective joint loss function L1, L2, and L3, and to integrate the three optimization objectives into a single overall optimization index.

[0048] Step S6: Implement parameter training of the LSTM model guided by the Hill model, and output continuous muscle force prediction values ​​that conform to the muscle contraction law.

[0049] It should be noted that the muscle strength prediction model can be trained using either an end-to-end joint training strategy, utilizing a multi-objective joint loss function to simultaneously optimize the model's predicted muscle synergy and muscle strength; or a staged training strategy, first learning the direct mapping relationship between electromyographic signals and muscle strength, and then utilizing a multi-objective joint loss function to simultaneously optimize the model's predicted muscle synergy and muscle strength.

[0050] In some exemplary embodiments, the training process for the muscle strength prediction model is divided into two stages: In the first stage, the muscle strength prediction loss function is used to enable the model to initially learn the direct mapping relationship between electromyographic signals and muscle strength.

[0051] The muscle synergy matrix is ​​set to an initial random value and kept unchanged. An LSTM neural network model is trained using the muscle strength prediction loss function, with the Adam optimizer's learning rate set to 0.001. The training run is set to 50 epochs. The first stage does not introduce constraints on the muscle synergy pattern, which helps the model initially learn the direct mapping relationship between electromyographic signals and muscle strength. This avoids gradient interference or optimization getting stuck in local optima before the muscle synergy matrix and activation coefficient matrix are fully learned in the initial stage, thus preventing reduced network training efficiency.

[0052] In the second stage, a multi-objective joint loss function is used to simultaneously optimize the model's predictions of muscle synergy and muscle strength, so that the prediction output follows biomechanical constraints.

[0053] The muscle coordination matrix and activation coefficient matrix were solved using nonnegative matrix factorization. An LSTM network was trained by minimizing a multi-objective joint loss function, including EMG reconstruction error, muscle coordination prediction error, and muscle strength prediction error, and the optimal solutions for the muscle coordination matrix and activation coefficient matrix were found. The training dataset was set to 100 epochs with a batch size of 32. The generalization ability of the model was enhanced by adding 20dB Gaussian noise and ±50ms time-domain shift to the trained EMG signal data.

[0054] The phased training strategy, which proceeds from coarse to fine, gradually introduces structured constraints, which helps the network converge from initial learning to refined modeling, thereby improving the stability and robustness of training, as well as the interpretability and prediction accuracy of the model.

[0055] It is important to note that the trained LSTM neural network model needs to be validated to evaluate its generalization ability. A separate test set is used to validate the model's performance. This test set contains electromyographic signal data from subjects different from those in the training set, ensuring the independence of the data distribution. The test set is then input into the trained Hill model and LSTM model to obtain the theoretical muscle strength of Hill and the predicted muscle strength of the LSTM, respectively. The consistency between the model predictions and the physiological modeling is evaluated using the root mean square error of both.

[0056] The expression for the root mean square error is: .

[0057] Example 2 Embodiment 2 of the present invention provides an upper limb muscle strength prediction system, comprising: The acquisition and preprocessing module acquires raw electromyographic signals from several upper limb muscles and preprocesses the raw electromyographic signals into electromyographic signals. The prediction module inputs electromyographic signals into the trained muscle strength prediction model to predict muscle strength. The muscle strength prediction model was constructed and trained using the following methods: Potential muscle synergistic patterns and activation levels are extracted from electromyographic signals based on nonnegative matrix factorization. A biomechanical model was constructed, and the theoretical muscle force value of the electromyographic signal was calculated based on the activation level and biomechanical parameters. A deep neural network model is constructed, with the electromyographic signal as input, the theoretical muscle strength value and the muscle synergy pattern as labels, so that the deep neural network model learns the relationship between the electromyographic signal, the muscle strength value and the muscle synergy pattern, and predicts the muscle strength value and the muscle synergy pattern. A multi-objective joint loss function is constructed, and the deep neural network model is trained using the multi-objective joint loss function, which includes a biomechanical information loss function and a deep neural network loss function. The biomechanical information loss function is the electromyography reconstruction loss function, and the deep neural network loss function includes the muscle strength prediction loss function and the muscle synergy prediction loss function. The electromyographic remodeling loss function L1 = ||EMG - W × H|| 2 ; The muscle coordination prediction loss function L2=‖WW pred || 2 ; The muscle strength prediction loss function L3=‖F Hill -F NN || 2 ; The multi-objective joint loss function is L = λ1L1 + λ2L2 + λ3L3; In the above formula, EMG represents the preprocessed electromyographic signal, W represents the muscle synergy pattern extracted from the electromyographic signal, and H represents the activation level extracted from the electromyographic signal; F Hill W represents the theoretical muscle force value calculated by the biomechanical model. pred F represents the muscle coordination pattern predicted by the deep neural network model. NN λ1 represents the muscle strength value predicted by the deep neural network model, λ2 represents the weight coefficient of the electromyography reconstruction loss function, λ3 represents the weight coefficient of the muscle synergy prediction loss function, and λ4 represents the weight coefficient of the muscle strength prediction loss function. The muscle strength prediction model is pre-trained by minimizing the multi-objective joint loss function.

[0058] Example 3 An electronic device provided in Embodiment 3 of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the upper limb muscle strength prediction method provided in Embodiment 1.

[0059] In one alternative embodiment, the electronic device further includes an input device and a display device, which can be connected to a memory and a processor via a bus or other means.

[0060] Memory, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules. The processor executes various functional applications and data processing of electronic devices by running the software programs, instructions, and modules stored in memory. Input devices are used to receive electromyographic signals, and display devices are used to output the predicted results through devices such as displays.

[0061] Example 4 Embodiment 4 of the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the upper limb muscle strength prediction method as provided in Embodiment 1.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for predicting upper limb muscle strength, characterized in that, Includes the following steps: Raw electromyographic (EMG) signals were collected from several upper limb muscles, and the raw EMG signals were preprocessed into EMG signals. Electromyographic signals are input into a trained muscle strength prediction model to predict muscle strength. The muscle strength prediction model was constructed and trained using the following methods: Potential muscle synergistic patterns and activation levels are extracted from electromyographic signals based on nonnegative matrix factorization. A biomechanical model was constructed, and the theoretical muscle force value of the electromyographic signal was calculated based on the activation level and biomechanical parameters. A deep neural network model is constructed, with the electromyographic signal as input, the theoretical muscle strength value and the muscle synergy pattern as labels, so that the deep neural network model learns the relationship between the electromyographic signal, the muscle strength value and the muscle synergy pattern, and predicts the muscle strength value and the muscle synergy pattern. A multi-objective joint loss function is constructed, and the deep neural network model is trained using the multi-objective joint loss function, which includes a biomechanical information loss function and a deep neural network loss function. The biomechanical information loss function is the electromyography reconstruction loss function, and the deep neural network loss function includes the muscle strength prediction loss function and the muscle synergy prediction loss function. The electromyographic remodeling loss function L1 = ||EMG - W × H|| 2 ; The muscle coordination prediction loss function L2=‖WW pred || 2 ; The muscle strength prediction loss function L3=‖F Hill -F NN || 2 ; The multi-objective joint loss function is L = λ1L1 + λ2L2 + λ3L3; In the above formula, EMG represents the preprocessed electromyographic signal, W represents the muscle synergy pattern extracted from the electromyographic signal, and H represents the activation level extracted from the electromyographic signal. F Hill W represents the theoretical muscle force value calculated by the biomechanical model. pred F represents the muscle coordination pattern predicted by the deep neural network model. NN λ1 represents the muscle strength value predicted by the deep neural network model, λ2 represents the weight coefficient of the electromyography reconstruction loss function, λ3 represents the weight coefficient of the muscle synergy prediction loss function, and λ4 represents the weight coefficient of the muscle strength prediction loss function. The muscle strength prediction model is pre-trained by minimizing the multi-objective joint loss function.

2. The upper limb muscle strength prediction method according to claim 1, characterized in that: When training the muscle strength prediction model, the muscle strength prediction loss function is first used to enable the model to learn the direct mapping relationship between electromyographic signals and muscle strength. Then, the multi-objective joint loss function is used to simultaneously optimize the model's prediction of muscle synergy and muscle strength, so that the prediction output follows biomechanical constraints.

3. The method for predicting upper limb muscle strength according to claim 1 or 2, characterized in that: The non-negative matrix decomposition method is a time-varying non-negative matrix decomposition method based on a sliding window mechanism. The time-varying non-negative matrix decomposition method performs local segmental modeling on a continuous electromyography signal sequence, performs standard non-negative matrix decomposition within each window, and then extracts local muscle synergy patterns and their activation levels. The expression for the time-varying nonnegative matrix decomposition method is: EMG i =W i ×H i , i=1,2,...,N; In the above formula, EMG i ∈R m×l W represents the electromyographic signal of length l within the i-th sliding window. i ∈R m×k H represents the local muscle coordination matrix within this time window. i ∈R k×l Let represent the corresponding local activation coefficient matrix, and k represent the number of collaborations, which empirically ranges from 4 to 6.

4. The method for predicting upper limb muscle strength according to claim 1 or 2, characterized in that: The biomechanical model adopted is the Hill model. The model parameters are initialized based on the upper limb muscle biomechanical parameters and anatomical criterion values ​​provided by the OpenSim platform, so that the model is more consistent with the actual situation of the electromyography signal acquisition experiment subjects. The theoretical expression for muscle strength is: F Hill =F max ×a(t)×f l (l)×f v (v); The total length of the muscle and tendon is: l(t) = l m +l T ; The normalized activation level expression is: a(t) = Sigmoid(H); In the above formula, F max The maximum isometric contractile force of the muscle, a(t) is the normalized activation level, and f l (l) represents the relationship between force and muscle length, f v (v) represents the relationship between force and velocity, l m For muscle fiber length, l T H represents the tendon length, and H represents the degree of activation.

5. The method for predicting upper limb muscle strength according to claim 1 or 2, characterized in that: The deep neural network model uses an LSTM neural network model to convert electromyographic signals into time window sequence data, which is used as the input data for the LSTM neural network model. The LSTM neural network model includes a three-layer structure, each containing a 64-unit hidden layer with the activation function of tanh. The time window sequence data passes through the three-layer LSTM network, is processed by a fully connected layer, and then input into the softplus activation function to predict muscle strength and muscle coordination patterns.

6. The upper limb muscle strength prediction method according to claim 1 or 2, characterized in that: The acquisition of electromyographic signals from several upper limb muscles included using the Delsys Trigno Wireless sEMG system to attach sensors to the surface of the anterior deltoid, middle deltoid, posterior deltoid, pectoralis major, biceps brachii, triceps brachii, brachialis, and brachioradialis muscles of the experimental subject to acquire raw electromyographic signals.

7. The upper limb muscle strength prediction method according to claim 1 or 2, characterized in that: The preprocessing includes passing the raw electromyographic signal sequentially through a 50Hz power frequency notch filter, a bandpass filter, and a full-wave rectification, and calculating the root mean square value based on the sliding window method.

8. An upper limb muscle strength prediction system, characterized in that, include: The acquisition and preprocessing module acquires raw electromyographic signals from several upper limb muscles and preprocesses the raw electromyographic signals into electromyographic signals. The prediction module inputs electromyographic signals into the trained muscle strength prediction model to predict muscle strength. The muscle strength prediction model was constructed and trained using the following methods: Potential muscle synergistic patterns and activation levels are extracted from electromyographic signals based on nonnegative matrix factorization. A biomechanical model was constructed, and the theoretical muscle force value of the electromyographic signal was calculated based on the activation level and biomechanical parameters. A deep neural network model is constructed, with the electromyographic signal as input, the theoretical muscle strength value and the muscle synergy pattern as labels, so that the deep neural network model learns the relationship between the electromyographic signal, the muscle strength value and the muscle synergy pattern, and predicts the muscle strength value and the muscle synergy pattern. A multi-objective joint loss function is constructed, and the deep neural network model is trained using the multi-objective joint loss function, which includes a biomechanical information loss function and a deep neural network loss function. The biomechanical information loss function is the electromyography reconstruction loss function, and the deep neural network loss function includes the muscle strength prediction loss function and the muscle synergy prediction loss function. The electromyographic remodeling loss function L1 = ||EMG - W × H|| 2 ; The muscle coordination prediction loss function L2=‖WW pred || 2 ; The muscle strength prediction loss function L3=‖F Hill -F NN || 2 ; The multi-objective joint loss function is L = λ1L1 + λ2L2 + λ3L3; In the above formula, EMG represents the preprocessed electromyographic signal, W represents the muscle synergy pattern extracted from the electromyographic signal, and H represents the activation level extracted from the electromyographic signal. F Hill W represents the theoretical muscle force value calculated by the biomechanical model. pred F represents the muscle coordination pattern predicted by the deep neural network model. NN λ1 represents the muscle strength value predicted by the deep neural network model, λ2 represents the weight coefficient of the electromyography reconstruction loss function, λ3 represents the weight coefficient of the muscle synergy prediction loss function, and λ4 represents the weight coefficient of the muscle strength prediction loss function. The muscle strength prediction model is pre-trained by minimizing the multi-objective joint loss function.

9. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the upper limb muscle strength prediction method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the upper limb muscle strength prediction method as described in any one of claims 1 to 7.

Images