FMG signal enhancement preprocessing method for upper limb composite action recognition

Abstract

The invention discloses an FMG signal enhancement preprocessing method for upper limb composite action recognition, and relates to the technical field of artificial intelligence and biomechanics crossing. In order to solve the problems that in the prior art, FMG signal feature extraction is insufficient, weak action signals are prone to being covered, and complex composite action classification precision is low, the following innovative scheme is provided: firstly, a muscle tension-deformation decoupling feature extraction algorithm is designed, and the muscle tension-deformation decoupling feature extraction algorithm is obtained by modeling a biomechanical relationship between muscle axial tension and radial deformation; optimizing global feature distribution; secondly, a non-linear transformation gesture fine feature amplification algorithm is adopted, large-amplitude signal interference is compressed, and the identification capacity of weak tension changes of far-end muscles is remarkably enhanced; and in combination with a lightweight deep neural network (DNN) classification model, efficient and accurate composite action classification is realized. According to the method, the classification precision and robustness of upper limb composite action recognition are remarkably improved, and powerful support is provided for intelligent application of the muscle pressure sensing technology in multiple fields.

Description

Technical Field

[0001] The present invention relates to the cross - technical field of artificial intelligence and biomechanics, and particularly to a method for enhancing and pre - processing FMG signals for upper - limb compound action recognition. Background Art

[0002] In the wave of the rapid development of artificial intelligence and human - computer interaction technologies, the upper - limb action recognition technology based on muscle pressure signals (FMG, Force Myography) shows extremely broad application prospects in many frontier fields such as intelligent exoskeleton devices, rehabilitation medicine, and virtual reality. As a non - invasive muscle activity measurement method, FMG signals have gradually become an indispensable data source in the research of upper - limb action recognition due to their significant advantages such as simple and easy operation and excellent real - time performance.

[0003] However, in the process of upper - limb action recognition research based on FMG, how to efficiently extract and amplify key features from signals has become the core problem for accurate action classification. The inherent non - linear characteristics and limited dynamic range of FMG signals make traditional feature extraction methods ineffective in capturing features that reflect subtle changes in actions in the signals. At the same time, most current research focuses on the classification of single gestures or posture actions, and very few studies deeply explore methods for classifying combined actions by integrating upper - arm postures and gesture actions. This limitation in research not only severely restricts the application potential of FMG signals in complex action scenarios but also highlights the obvious shortcomings of traditional methods in feature expression ability and model generalization performance. Particularly crucial is that current research mainly focuses on classifying direct signals of driving muscle groups and pays little attention to the subtle tension features caused by distal muscle activation, especially in the field of joint recognition of gesture and posture combined actions, where relevant research and experiments are almost blank. Summary of the Invention

[0004] The purpose of the present invention is to propose a method for enhancing and pre - processing FMG signals for upper - limb compound action recognition. Through two algorithms, namely muscle - pull - deformation decoupling feature extraction and gesture subtle feature amplification, the problem of feature extraction in complex compound action scenarios is effectively solved, and the classification accuracy and robustness of upper - limb compound action recognition are significantly improved.

[0005] To achieve the above - mentioned purpose, the present invention proposes a method for enhancing and pre - processing FMG signals for upper - limb compound action recognition, and the specific steps are as follows:

[0006] Step S1: Use an FMG sensing armband to collect 8 - channel FMG signals containing upper - limb compound actions and construct a standardized data set;

[0007] Step S2: Extract the muscle force-deformation decoupling features from the collected 8-dimensional raw data. By deriving the relationship between muscle force and radial strain, determine the optimal Young's modulus and Poisson's ratio parameters, and calculate the new 8-dimensional features;

[0008] Step S3: Amplify the weak muscle pressure fluctuation signals caused by gesture changes, and at the same time compress the large fluctuation signals to amplify the subtle features of the gestures;

[0009] Step S4: Perform Z-score normalization on the processed data, introduce random scaling and Gaussian noise, use a lightweight deep neural network DNN for classification, and verify the effectiveness of the algorithm through gesture classification experiments and ablation experiments to determine the optimal data processing strategy.

[0010] Preferably, in step S1, use the FMG sensing armband to collect data, construct a dataset containing 15 upper limb compound actions, generate 8-channel time series signal data and attach action labels, and store it in the form of a standardized matrix, denoted as X ∈ R n ×8 , where X is the standardized matrix form of the dataset, n is the number of samples, and each sample contains 8-dimensional FMG raw data; the i-th sample in the matrix is x i =[x i1 ,x i2 ,…,x i8 , where x ij is the muscle pressure value of the j-th channel of the i-th sample, and the value range of j is 1-8.

[0011] Preferably, the upper limb compound actions include 3 arm postures: horizontal elbow flexion, vertical elbow flexion, shoulder flexion and 5 gestures: relaxation, wrist flexion, wrist extension, fist clenching, and grasping permutations.

[0012] Preferably, in step S2, the relationship between muscle force and radial strain is as follows:

[0013] The radial strain generated by muscle work is defined as:

[0014]

[0015] where R is the average cross-sectional radius in the muscle working state, R 0 is the average cross-sectional radius in the muscle relaxation state, and ε radial is the radial strain generated by muscle work;

[0016] According to the Poisson effect, the relationship between muscle radial strain and axial strain is obtained:

[0017] ε radial =vε axial ;

[0018] Among them, ε axial is the axial strain, and v is the Poisson's ratio;

[0019] Under the muscle working state, the relationship between muscle stress σ and axial strain is:

[0020] σ = Eε axial ;

[0021] Among them, E is the muscle equivalent Young's modulus;

[0022] Combining the above three formulas, the relationship between muscle stress and radial strain is obtained:

[0023]

[0024] The relationship between the tensile force F generated under the muscle working state and the muscle cross-sectional area A is expressed as:

[0025] F = σ·A;

[0026] Among them, the cross-sectional area A during muscle work is expressed as:

[0027]

[0028] Combining the above relationship between muscle stress and radial strain, the relationship between tensile force F and muscle cross-sectional area A, and the cross-sectional area A during muscle work, the expression of muscle tensile force is obtained:

[0029]

[0030] Preferably, based on the relationship between muscle tensile force and radial strain, substituting 8-dimensional original data, new 8-dimensional features are obtained, which form a new feature set with the original features. The formula is as follows:

[0031] Y i = [x i1 , x i2 ,..., x i(8) , F i1 , F i2 ,..., F i(8) = [y i1 , y i2 ,..., y i(16) , Y i ∈R n×16 ;

[0032] Among them, Y i is the 16-dimensional feature of the i-th sample; F ij is the muscle tensile force of the j-th channel of the i-th sample; y im is the feature of the m-th channel of the i-th sample in the new feature set, and the value range of m is 1 - 16.

[0033] Preferably, in step S3, the weak myoelectric pressure fluctuation signal caused by the gesture change is amplified, and the formula is as follows:

[0034]

[0035] Y i ' = [y i1 , y i2 ,..., y i(16) , p i1 , p i2 ,..., p i(8) = [y′ i1 , y′ i2 , …, y′ i(24) , Y i ' ∈ R n×24 ;

[0036] where γ is an adjustable parameter, y ij is the feature of the j-th channel of the i-th sample of the new feature set, is the mean value of the features of the j-th channel, p ij is the feature obtained by performing gesture fine feature amplification processing on y ij ; Y i ' is the i-th sample of Y', y′ ik is the feature of the k-th channel of the i-th sample of Y', and the value range of k is 1-24. Y' is a new feature set obtained by adding the features after gesture fine feature amplification processing to Y.

[0037] Preferably, in step S4, the lightweight deep neural network DNN includes two hidden layers, uses the ReLU activation function to enhance the non-linear feature extraction ability, and randomly masks some neurons through the Dropout layer; the model output is converted into a probability distribution through the Softmax layer, and the cross-entropy loss function is used to optimize the model performance.

[0038] Therefore, the present invention proposes a method for enhancing and preprocessing FMG signals for upper limb compound action recognition, and its beneficial effects are as follows:

[0039] (1) Improving classification accuracy and robustness: The method for enhancing and preprocessing FMG signals for upper limb compound action recognition proposed by the present invention effectively solves the problem of feature extraction in complex compound action scenarios through two algorithms of muscle tension-deformation decoupling feature extraction and gesture fine feature amplification, improves the ability to capture fine features, and significantly improves the classification accuracy and robustness of upper limb compound action recognition.

[0040] (2) Strong hardware adaptability: A method for enhancing and preprocessing FMG signals for upper limb compound motion recognition proposed in the present invention combines a lightweight deep neural network model, achieving high computational performance and excellent classification performance while ensuring low hardware configuration requirements, and demonstrating excellent adaptability in complex compound motion scenarios.

[0041] (3) High scalability: A method for enhancing and preprocessing FMG signals for upper limb compound motion recognition proposed in the present invention enables flexible adaptation to the requirements of diverse motion recognition tasks through modular design of feature processing algorithms, providing strong theoretical and technical support for the application expansion of muscle pressure sensing technology and the intelligent innovation in fields such as upper limb exoskeletons and rehabilitation medicine.

[0042] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Description of the Drawings

[0043] Figure 1 It is a flowchart of the specific solution of the present invention;

[0044] Figure 2 It is a schematic diagram of the relationship between upper arm motion, muscle tension, and deformation in the present invention;

[0045] Figure 3 It is a schematic diagram of the relationship between Young's modulus and Poisson's ratio and muscle tension - radial strain in the present invention;

[0046] Figure 4 It is a schematic diagram of the experimental results of gesture classification in the embodiment of the present invention; among them, Figure 4 in (a) is the classification effect diagram of the arm posture of horizontal elbow flexion based on the original data; Figure 4 in (b) is the classification effect diagram of the arm posture of horizontal elbow flexion based on the amplified features; Figure 4 in (c) is the classification effect diagram of the arm posture of vertical elbow flexion based on the original data; Figure 4 in (d) is the classification effect diagram of the arm posture of vertical elbow flexion based on the amplified features; Figure 4 in (e) is the classification effect diagram of the arm posture of shoulder flexion based on the original data; Figure 4 in (f) is the classification effect diagram of the arm posture of shoulder flexion based on the amplified features;

[0047] Figure 5 It is a schematic diagram of the experimental results of ablation in the embodiment of the present invention;

[0048] Figure 6 It is a schematic diagram of the confusion matrix of upper limb compound motion state classification in the embodiment of the present invention. Detailed Embodiment

[0049] To make the technical solutions, advantages and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be described clearly and completely below. The described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present invention without creative efforts fall within the protection scope of this application.

[0050] Unless otherwise defined, the technical terms or scientific terms used in the present invention should have the ordinary meaning understood by those of ordinary skill in the field to which the present invention belongs.

[0051] As Figure 1 shown, the present invention provides an FMG signal enhancement preprocessing method for upper limb compound action recognition, and innovatively designs two data processing algorithms of muscle tension-deformation decoupling feature extraction and gesture subtle feature amplification to significantly improve the classification accuracy and robustness of upper limb compound action recognition, including the following steps:

[0052] Step S1: Use the FMG sensing armband to collect 8-channel FMG signals containing upper limb compound actions and construct a standardized data set;

[0053] Step S2: Perform muscle tension-deformation decoupling feature extraction on the collected 8-dimensional raw data, and determine the optimal Young's modulus and Poisson's ratio parameters by deriving the relationship between muscle tension and radial strain, and calculate the new 8-dimensional features.

[0054] Step S3: Amplify the weak muscle pressure fluctuation signals caused by gesture changes, and at the same time compress the large fluctuation signals to achieve gesture subtle feature amplification.

[0055] Step S4: Perform Z-score standardization processing on the processed data, introduce random scaling and Gaussian noise, use a lightweight deep neural network DNN for classification, and verify the effectiveness of the algorithm through gesture classification experiments and ablation experiments to determine the optimal data processing strategy.

[0056] Embodiment 1

[0057] 1. Use the FMG sensing armband to collect 8-channel FMG signals of upper limb compound actions and construct a standardized data set.

[0058] In this study, 15 kinds of upper limb compound actions were designed, covering all permutations of 3 common arm postures (horizontal elbow flexion, vertical elbow flexion, shoulder flexion) and 5 common gestures (relaxation, wrist flexion, wrist extension, fist clenching, grasping). This combination strategy fully considers the diversity of upper limb movements and their high-frequency usage scenarios in daily life, and has strong practical significance.

[0059] Data acquisition is completed through the FMG sensing armband, generating time series signal data containing 8 channels. Each channel records the muscle pressure changes in the corresponding part, accurately capturing the muscle dynamic characteristics caused by upper limb movements. An action label is appended to the end of each frame of data to identify a specific combination of arm postures and gestures, achieving an accurate correspondence between the signal and the action type. The collected data is based on the signal records of a single action as the basic unit and stored in the form of a standardized matrix, denoted as X ∈ R n ×8 , where n is the number of samples, and each sample contains 8-dimensional FMG data. The i-th sample in the matrix is x i = [x i1 , x i2 ,..., x i8 , corresponding to the muscle pressure values of 8 channels. With this efficient data organization method, the data can not only comprehensively reflect the action characteristics but also has highly structured features, providing high-quality input for feature extraction and classification model training, and fully supporting the accurate recognition of upper limb compound actions.

[0060] 2. Extract muscle tension-deformation decoupled features, determine the optimal Young's modulus and Poisson's ratio parameters, and calculate new 8-dimensional features.

[0061] The design inspiration for muscle tension-deformation decoupled feature extraction comes from the close connection between muscle mechanical behavior and deformation in biomechanics. Different action states of the upper arm correspond to different joint maintenance torques, and the essence of joint torque comes from the axial tension of muscle work. As Figure 2 shown, during muscle contraction, radial deformation (i.e., the change in muscle cross-section) occurs concomitantly and is captured by the FMG sensor. The amount of deformation is positively correlated with the data obtained by the sensor; the relationship between muscle tension and radial strain is as follows:

[0062] The radial strain generated by muscle work can be defined as:

[0063]

[0064] where R is the average cross-sectional radius in the muscle working state, and R 0 is the average cross-sectional radius in the muscle relaxed state. According to Poisson's effect, the relationship between muscle radial strain and axial strain is approximately obtained:

[0065] ε radial = vε axial ;

[0066] where υ is Poisson's ratio, and the Poisson's ratio of human soft tissues or muscles is generally between 0.4 - 0.5. In the muscle working state, the relationship between muscle stress σ and axial strain is:

[0067] σ = Eε axial ;

[0068] Among them, E is the muscle equivalent Young's modulus, a physical quantity used to describe the ability of a material to resist deformation within the elastic range. Combining the above three equations, the relationship between muscle stress and radial strain can be obtained:

[0069]

[0070] The relationship between the tensile force F generated during muscle work and the cross-sectional area A of the muscle can be approximately expressed as:

[0071] F = σ·A;

[0072] Among them, the cross-sectional area A during muscle work can be approximately expressed as:

[0073]

[0074] Combining the above three equations, the expression for muscle tensile force can be obtained:

[0075]

[0076] Such as Figure 3 shown, the selection of Young's modulus and Poisson's ratio has a significant impact on muscle tensile force - radial strain. By performing fitting analysis on the data collected by the sensor, the optimal Young's modulus and Poisson's ratio parameters are determined and applied to the model to improve the accuracy and applicability of the model.

[0077] Based on the relational expression between muscle tensile force and radial strain, substituting the 8-dimensional original data, new 8-dimensional features are further calculated, and the formula is as follows:

[0078] Y i = [x i1 , x i2 ,..., x i(8) , F i1 , F i2 , …, F i(8) = [y i1 , y i2 , …, y i(16) , Y i ∈ R n×16 ;

[0079] Among them, Y i is the 16-dimensional feature of the i-th sample; F ij is the muscle tensile force of the j-th channel of the i-th sample; y im is the feature of the m-th channel of the i-th sample in the new feature set, and the value range of m is 1 - 16.

[0080] 3. Amplification of subtle gesture features.

[0081] The changes in gestures are mainly driven by the activation of forearm muscles and generally do not directly cause significant changes in the volume of upper-arm muscles, but can have a subtle indirect impact on upper-arm muscles. When performing wrist or finger joint movements, upper-arm muscles often need to work together to maintain the stability of the arm. For example, in the actions of making a fist or grasping, the biceps brachii slightly tightens to assist in completing the action; during the process of wrist flexion or extension, some muscles in the driving muscle group (such as flexor digitorum, extensor digitorum, extensor carpi radialis longus, etc.) have their origins attached near the upper arm or elbow, and their contractions will cause certain tension changes in the upper-arm muscles through the traction effect.

[0082] However, the muscle tension fluctuations caused by gesture changes are usually relatively weak, and it is difficult to directly capture the impact on the volume of upper-arm muscles. In the acquisition of muscle tension signals, this indirect impact usually manifests as low-amplitude signal fluctuations, which are easily masked by the large-amplitude signals caused by the direct working state changes of upper-arm muscles. Therefore, in the data processing or feature extraction stage, it is necessary to effectively distinguish and extract these weak signals to ensure the accuracy and robustness of the analysis results. In this study, the following formula is used to compress the large-amplitude fluctuation signals and simultaneously amplify the weak signal changes caused by gesture changes to more clearly capture the subtle activity characteristics of upper-arm muscles under different gestures:

[0083]

[0084] Y i ' = [y i1 , y i2 ,..., y i(16) , p i1 , p i2 ,..., p i(8) = [y′ i1 , y′ i2 , …, y′ i(24) , Y i ' ∈ R n×24 ;

[0085] Among them, γ is an adjustable parameter, y ij is the feature of the j-th channel of the i-th sample in the new feature set, is the mean value of the features of the j-th channel, p ij is the feature obtained after magnifying the subtle features of gestures for y ij ; Y i ' is the i-th sample of Y', y′ ik is the feature of the k-th channel of the i-th sample of Y', and the value range of k is 1 - 24, and Y' is the new feature set obtained by adding the features after magnifying the subtle features of gestures to Y.

[0086] Through this feature transformation method, the influence of large-amplitude signals is effectively suppressed, while the weak signals caused by gestures are significantly amplified, making the subtle changes in the upper-arm muscles under different gestures more prominently reflected in the feature space, laying an important foundation for the accurate modeling of the upper-limb composite motion posture recognition task.

[0087] 4. Experimental verification.

[0088] The data is processed by Z-score standardization to have zero mean and unit variance, thereby improving the learning efficiency and convergence performance of the model. Subsequently, random scaling and Gaussian noise are introduced to enrich the diversity and variability of the data and enhance the robustness of the model under different input conditions.

[0089] To ensure application convenience and reduce dependence on hardware configuration, a lightweight deep neural network (DNN) is used as the classification model in the experiment. The DNN takes the processed feature vectors as input and outputs the classification results of the upper-limb composite motions. The model structure contains two hidden layers, uses the ReLU activation function to enhance the non-linear feature extraction ability, and at the same time randomly masks some neurons through the Dropout layer to effectively suppress the overfitting phenomenon. Finally, the model output is converted into a probability distribution through the Softmax layer, and the cross-entropy loss function is used to optimize the model performance.

[0090] The experimental verification is mainly divided into two parts: gesture classification experiment and ablation experiment:

[0091] (a) Gesture classification experiment;

[0092] Evaluate the improvement effect of the gesture subtle feature amplification algorithm on the gesture classification performance under the same arm posture. The experiment is divided into three groups according to three different arm postures, and five gestures are classified respectively, including classification based on the original data and classification after feature amplification. The classification confusion matrix is as Figure 4 shown, where the left column is the classification effect based on the original data, and the right column is the classification effect after feature amplification. Figure 4 In (a) and (b) of Figure 4 the arm postures are horizontal elbow flexion, Figure 4 in (c) and (d) of

[0093] the arm postures are vertical elbow flexion,

[0094] (b) Ablation experiment;

[0095] By removing specific data processing modules, taking the average performance over multiple rounds of experiments, and comparing with the control group, the individual contributions and their synergistic effects of the muscle force-deformation decoupling feature extraction and gesture subtle feature amplification modules were comprehensively analyzed, thus further verifying the key role of each module in the classification performance.

[0096] As Figure 5 shown, Method 1 uses only the raw data after standardization as the benchmark. Method 2 and Method 3 respectively introduce single data processing strategies of muscle force decoupling and gesture subtle feature amplification on this basis. Method 4 integrates the muscle force decoupling and gesture subtle feature amplification modules, showing the best data processing performance and classification effect.

[0097] By comparing the classification accuracy (Accuracy), precision (Precision), recall (Recall), and F1Score, the following conclusions were drawn from the experiments: The single processing strategies (Method 2, Method 3) improved the classification performance compared to the benchmark (Method 1), verifying the effectiveness of the muscle force decoupling and gesture subtle feature amplification modules in optimizing feature expression. When the two strategies are used in combination (Method 4), all performance indicators reached 96%, showing the best performance. This fully demonstrates the synergistic effect of the dual-module fusion and its advantages in improving the overall performance of the classification model.

[0098] The best confusion matrix of the experiment is as Figure 6 shown, where elbow flexion 1 represents the arm posture of horizontal elbow flexion, and elbow flexion 2 represents vertical elbow flexion; the research finally determined the optimal data processing strategy, providing an efficient technical solution for compound action recognition.

[0099] Therefore, the present invention provides a method for enhancing and preprocessing FMG signals for upper limb compound action recognition. Through two algorithms of muscle force-deformation decoupling feature extraction and gesture subtle feature amplification, it effectively solves the problem of feature extraction in complex compound action scenarios and improves the ability to capture subtle features; combined with a lightweight deep neural network model, while ensuring low hardware configuration requirements, it achieves high computational performance and excellent classification performance, significantly improving the classification accuracy and robustness of upper limb compound action recognition.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A FMG signal enhancement preprocessing method for upper limb complex action recognition, characterized in that: The specific steps are as follows: Step S1, using the FMG sensor armband to collect 8-channel FMG signals including upper limb compound movements and construct a standardized data set; Step S2, extracting muscle tension-deformation decoupling features from the collected 8-dimensional raw data, and determining the optimal Young's modulus and Poisson's ratio parameters by deducing the relationship between muscle tension and radial strain, and calculating new 8-dimensional features; Step S3, amplifying the weak muscle pressure fluctuation signal caused by the gesture change, and compressing the large fluctuation signal, so as to amplify the subtle features of the gesture; Step S4: perform Z-score normalization on the processed data, introduce random scaling and Gaussian noise, use a lightweight deep neural network DNN for classification, and verify the effectiveness of the algorithm through gesture classification experiments and ablation experiments to determine the optimal data processing strategy.

2. The FMG signal enhancement preprocessing method for upper limb complex action recognition according to claim 1, characterized in that: In step S1, the FMG sensor armband is used to collect data, and a data set containing 15 types of upper limb compound movements is constructed to generate 8-channel time series signal data and attach action labels, which are stored in a standardized matrix form, represented as X∈R n×8 , where X is the standardized matrix form of the data set, n is the number of samples, and each sample contains 8-dimensional FMG raw data; the i-th sample in the matrix is ​​x i =[x i1 ,x i2 ,…,x i8 ], where x ij is the muscle pressure value of the jth channel of the ith sample, and the value range of j is 1-8.

3. The FMG signal enhancement preprocessing method for upper limb complex action recognition according to claim 2, characterized in that: The upper limb compound movements include 3 arm postures: horizontal elbow flexion, vertical elbow flexion, shoulder flexion, and 5 gestures: relaxation, wrist flexion, wrist extension, fist clenching, and full arrangement of grasping.

4. The FMG signal enhancement preprocessing method for upper limb complex action recognition according to claim 1, characterized in that: In step S2, the relationship between muscle tension and radial strain is as follows: The radial strain produced by muscle work is defined as: Among them, R is the average cross-sectional radius when the muscle is in working state, R0 is the average cross-sectional radius when the muscle is in relaxed state, ε radial radial strains produced for muscle work; According to the Poisson effect, the relationship between muscle radial strain and axial strain is obtained: e radial e axial ; Among them, ε axial is the axial strain, v is the Poisson’s ratio; When the muscle is in working state, the relationship between muscle stress σ and axial strain is: σ=Eε axial ; Where, E is the equivalent Young's modulus of muscle; Combining the above three equations, the relationship between muscle stress and radial strain is obtained: The relationship between the tension F generated by the muscle in the working state and the muscle cross-sectional area A is expressed as: F = σ·A; Among them, the cross-sectional area A of the muscle when working is expressed as: Combining the above relationship between muscle stress and radial strain, the relationship between tension F and muscle cross-sectional area A, and the cross-sectional area A when the muscle is working, the expression of muscle tension is obtained:

5. The FMG signal enhancement preprocessing method for upper limb complex action recognition according to claim 4, characterized in that: Based on the relationship between muscle tension and radial strain, the 8-dimensional original data is substituted to obtain a new 8-dimensional feature, which is combined with the original feature to form a new feature set. The formula is as follows: Y i =[x i1 ,x i2 ,…,x i(8) ,F i1 ,F i2 ,…,F i(8) ]=[y i1 ,y i2 ,…,y i(16) ],Y i ∈R n×16 ; Among them, Y i is the 16-dimensional feature of the i-th sample; F ij is the muscle tension of the jth channel of the ith sample; y im is the feature of the mth channel of the i-th sample in the new feature set, and the value range of m is 1-16.

6. The FMG signal enhancement preprocessing method for upper limb complex action recognition according to claim 1, characterized in that: In step S3, the weak muscle pressure fluctuation signal caused by the gesture change is amplified, and the formula is as follows: AND i '=[and i1 ,and i2 ,…,and i(16) ,p i1 ,p i2 ,…,p i(8) ]=[and' i1 ,and' i2 ,…,and' i(24) ],AND i '∈R n×24 ; Among them, γ is an adjustable parameter, y ij is the feature of the jth channel of the i-th sample in the new feature set, is the mean of the jth channel feature, p ij For y ij The features obtained after the subtle features of the gesture are magnified; Y i ' is the i-th sample of Y', y' ik is the kth channel feature of the ith sample of Y', the value range of k is 1-24, and Y' is a new feature set obtained by adding the features of gesture subtle features to Y after amplification processing.

7. The FMG signal enhancement preprocessing method for upper limb complex action recognition according to claim 1, characterized in that: In step S4, the lightweight deep neural network DNN includes two hidden layers, uses the ReLU activation function to enhance the nonlinear feature extraction capability, and randomly shields some neurons through the Dropout layer; the model output is converted into a probability distribution through the Softmax layer, and the cross entropy loss function is used to optimize the model performance.

Images