Analysis method of surface electromyography of upper limb fusing depth and width learning

Abstract

This invention discloses a method for analyzing upper limb surface electromyography (sEMG) signals by integrating deep and wide learning, belonging to the field of upper limb surface EMG signal analysis technology. The method includes the following steps: acquiring multi-channel EMG signals to establish a complete sEMG dataset; designing the same number of convolutional neural network branches with residual structures for the acquired multi-channel sEMG signals; inputting the output features of the residual convolutional neural network structure of each channel into its respective wide learning sub-model; calculating weights through pseudo-inverse regression; obtaining a contribution index by normalizing the absolute value; stabilizing the contribution index through cross-validation; analyzing the muscle activation weights corresponding to upper limb movements; and validating the representation of muscle activation degree. This achieves the collaborative extraction of multi-channel muscle features and the quantification of muscle activation contribution, thereby solving the problem of insufficient representation ability of traditional muscle co-activation indices for the real neuromuscular control process, and realizing high-precision, interpretable quantification of muscle co-activation relationships.

Description

Technical Field

[0001] This invention relates to the field of upper limb surface electromyography signal analysis technology, and in particular to an upper limb surface electromyography signal analysis method that integrates depth and width learning. Background Technology

[0002] With the rising incidence of neurodegenerative diseases, patients are experiencing upper limb motor dysfunction (such as muscle weakness and decreased motor coordination). The upper limb is the most flexible organ for human interaction with the outside world. Current traditional assessment methods rely on experience-based assessment scales, lacking quantitative assessment methods for upper limb muscle activation based directly on the analysis of patients' physiological signals. Surface electromyography (sEMG) signals are electrophysiological signals generated by skeletal muscle fibers during contraction, reflecting the neuromuscular activity state. They are widely used in rehabilitation training, prosthetic control, and human-computer interaction. Therefore, by analyzing surface EMG signals and extracting their features through deep learning representation methods, we can quantify the activation state of muscle synergy during human movement. This provides technical support for personalized rehabilitation and functional reconstruction for patients with upper limb dysfunction, offering quantifiable examination tools and visually demonstrating the coordination state of each muscle.

[0003] For example, Chinese invention patent CN114041808B discloses a method for transmission entropy coupling analysis based on multi-channel surface electromyography (EMG) signal decomposition. First, a convolution kernel compensation method is used to decompose the multi-channel EMG signal into motor units, separating the EMG signal that was originally composed of multiple superimposed motor units. After motor unit decomposition, a correlation matrix is ​​established using transmission entropy. Weak connections are removed using a thresholding method or a fixed-weighted edge method, constructing an intermuscular network model related to motor function. Undirected graphs of each EMG frequency band are plotted, and intermuscular network indices such as connectivity and small-world properties are calculated to establish a complex network. Through motor unit decomposition, the transmission entropy is calculated, and a complex network is constructed to fundamentally analyze the coupling characteristics of EMG signals.

[0004] For example, Chinese Invention Patent CN116269450B discloses a patient limb rehabilitation status assessment system and method based on electromyography (EMG) signals: a central processing unit acquires the patient's limb movements; a multi-channel signal acquisition unit acquires EMG signals corresponding to each EMG electrode within a set time period at a set acquisition frequency; a multi-channel preamplifier amplifies these EMG signals; a multi-channel filter filters these amplified EMG signals; a multi-channel analog-to-digital converter performs analog-to-digital conversion on these filtered EMG signals; the central processing unit analyzes and corrects these analog-to-digital converted EMG signals, and inputs the corrected EMG signals from all EMG electrodes into a pre-trained convolutional neural network model corresponding to the patient's limb movements for model training and prediction to assess the patient's limb rehabilitation status; and a display outputs the assessed limb rehabilitation status.

[0005] In existing technologies, human movement is initiated by neural commands issued by the brain, which activate specific muscle groups through neural pathways. Multiple electrodes simultaneously collect electrical signals from the surface of muscles, with each channel corresponding to a specific muscle. In multi-channel sEMG signals, since different channels represent different activation states of different muscles in the same movement, directly extracting and fusing the signals from each channel will result in the loss of unique information of individual muscle groups, reducing the ability to identify the activation level of specific muscle groups. This can lead to some muscles or channels being over-represented or under-represented, thereby affecting the accuracy of the overall assessment.

[0006] Furthermore, deep learning representation methods tend to overlook the essential differences between channels when extracting the correlation and coupling features between channels. These differences actually affect the analysis of muscle activation patterns and may suppress certain important personalized features, thus affecting the accurate representation of muscle activation state. Therefore, there is a technical problem that the accuracy of the quantitative indicators of muscle co-activation is insufficient and cannot effectively reflect the real neuromuscular control process in the human body.

[0007] For example, when we perform an action like "raising our hand," it requires the coordination of multiple muscles, such as muscle A contracting with great force, muscle B contracting with less force, and muscle C relaxing with even less force. Some patients with limb dysfunction may be unable to perform the action because muscle A is weak; while others may be unable to perform the action because muscle C is overworked. Without accurately quantifying the specific muscles involved, it is difficult to pinpoint the problematic muscle and develop a personalized rehabilitation training plan (such as "strengthening muscle A or relaxing muscle C"). Summary of the Invention

[0008] To address the technical problem that traditional muscle co-activation indices in existing technologies are insufficient in representing real neuromuscular control processes, this invention provides a method for analyzing upper limb surface electromyographic signals by integrating depth and width learning. The technical solution is as follows:

[0009] Multi-channel electromyography (sEMG) signals were collected from local areas of the upper limb to establish a complete sEMG dataset. Data preprocessing was performed, and the sEMG dataset was used to comprehensively and objectively describe the relationship between muscles and movement in the entire upper limb. For the collected multi-channel sEMG signals, a convolutional neural network branch with the same number of residual structures was designed to extract effective upper limb movement features from different channels. These effective upper limb movement features reflect the activation patterns of various muscle groups during upper limb movements. The output features of the residual convolutional neural network structure for each channel were input into their respective width learning sub-models. Weights were calculated using pseudo-inverse regression, and the absolute value was normalized to obtain a contribution index. The mean contribution index was stabilized through cross-validation. The muscle activation weights corresponding to upper limb movements were analyzed, and the representation of muscle activation degree was validated.

[0010] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0011] 1. The network structure combining deep and wide learning improves the accuracy and efficiency of upper limb movement analysis. Deep learning extracts local features from sEMG signals through convolutional neural networks, while wide learning optimizes the contribution of channels to accurately analyze the activation weights of muscle groups. Suppose a patient with limb dysfunction is unable to perform a certain movement, this fusion network can comprehensively extract relevant features from multi-channel signals, accurately analyze insufficient activation of muscle A, and provide corresponding rehabilitation suggestions (such as strengthening muscle A training).

[0012] 2. Since the acquisition of electromyography (EMG) signals needs to cover as many key upper limb areas as possible in order to extract the most representative complete upper limb movements, the noise ratio and the coupling of movement also increase. Existing sEMG data preprocessing mainly focuses on denoising and signal filtering, such as commonly used filtering methods including low-pass filtering or band-pass filtering. These methods can usually only remove noise in specific frequency bands, and the processing is relatively simple, which may affect some key information of the signal. Butterworth filtering and sliding window segmentation processing were adopted. During the Butterworth band-pass filtering denoising process, phase compensation was also performed by time reversal to ensure that the phase information of the signal was not disturbed and to avoid the phase distortion that may be introduced by traditional filtering methods.

[0013] 3. Existing convolutional neural networks typically use different convolutional kernels for feature extraction in each channel, or use different network structures for each channel. While this can extract features from each channel, it may lead to inter-channel interference or excessive parameters. By constructing a residual convolutional neural network, the same convolutional block structure is used in feature extraction for each channel, and the same type of features are extracted from the signals of each channel by sharing the convolutional kernel weights. This design avoids signal interference between channels and ensures decoupling between channels. For example, in the analysis of fist clenching movements, muscles A and B may be activated simultaneously. Through the residual convolutional network and weight sharing mechanism, the model can ensure that the activation features of each muscle group are extracted independently and accurately, without being affected by signals from other channels.

[0014] 4. By using a width-based learning sub-model and combining pseudo-inverse regression to calculate the regression coefficient weights of each channel, and then obtaining contribution indicators through absolute value normalization, these indicators accurately reflect the contribution of each channel in different movements, thereby effectively quantifying the weight of muscle activation. This process enables the network to quantify the activation level of each muscle group, helping the system to better analyze the role of each muscle in upper limb movements. For example, when analyzing a patient's hand-raising movement, the model can quantify the activation weights of muscles A and B through contribution indicators, determining whether muscle A is overly fatigued (under-activated) or whether muscle B is over-activated, thus helping to develop personalized rehabilitation plans. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart of an upper limb surface electromyography signal analysis method that integrates depth and width learning, provided in an embodiment of the present invention;

[0017] Figure 2 A schematic diagram illustrating six upper limb movements shown in the sEMG dataset provided in this embodiment of the invention;

[0018] Figure 3 This is a schematic diagram of the overall structure of CDWS-ResCNN and the local structure of the residual block provided in an embodiment of the present invention;

[0019] Figure 4 This is a structural diagram of the BLS sub-model corresponding to the c-th channel provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0021] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0022] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0023] like Figure 1 The diagram shows a flowchart of an upper limb surface electromyography signal analysis method that integrates depth and width learning, provided by an embodiment of the present invention. The method includes the following steps:

[0024] S1. Multi-channel electromyography (EMG) signals are collected from local areas of the upper limb to establish a complete sEMG dataset. Data preprocessing is then performed. The sEMG dataset is used to comprehensively and objectively describe the relationship between muscles and movement in the entire upper limb.

[0025] S2. For the acquired multi-channel sEMG signals, a convolutional neural network branch with the same number of residual structures is designed to extract effective upper limb movement features from different channels. These effective upper limb movement features can reflect the activation patterns of various muscle groups in the upper limb movement. Since different channels represent different activation states of different muscles in the same movement, when mining the activation weight patterns of muscles at different locations, a multi-channel residual convolutional neural network structure (Channel-Decoupled Weight-Shared Residual Convolutional Neural Network, CDWS-ResCNN) is constructed to ensure that each channel extracts decoupled features with high representational power and independent channels, while also ensuring that features with the same attributes are extracted for all channels, thereby ensuring the fairness of evaluating muscle activation states.

[0026] In each branch of CDWS-ResCNN, each branch consists of 6 residual blocks. Input is passed across layers via shortcut connections and added to the output of the convolutional layers (i.e., identity mapping). Specifically, each residual block consists of two layers of one-dimensional convolutions and ReLU activation functions. A 1×1 convolution is introduced to achieve channel dimension scaling, ensuring the effectiveness of feature fusion and mitigating gradient vanishing and model degradation issues in deep networks through cross-layer propagation. Furthermore, to enhance the network's efficient extraction of key features from sEMG signals, a SoftPool is introduced after each residual block for temporal (1D) downsampling, extracting highly discriminative deep features block by block.

[0027] S3, the residual convolutional neural network structure of each channel outputs features and inputs them into its respective width learning sub-model. The weights are calculated through pseudo-inverse regression, and the absolute value is normalized to obtain the contribution index. The contribution index is stabilized by cross-validation of the mean. The muscle activation weights corresponding to upper limb movements are analyzed, and the representation of the degree of muscle activation is verified.

[0028] In this embodiment, for the task of assessing human upper limb motor function, by deeply exploring the patterns of activation weights of upper limb muscle groups in different movements, a set of interpretable algorithms for measuring the contribution of human upper limb muscle activation weights is constructed, providing important quantitative indicators and theoretical support for upper limb motor function assessment.

[0029] An architecture integrating deep learning and breadth learning is constructed to deeply mine and quantify the muscle activation weights corresponding to upper limb movements, obtaining interpretable muscle activation contributions. Based on CDWS-ResCNN deep learning, deep features of sEMG signals are extracted from a "vertical" perspective through multi-layer nonlinear stacking, backpropagation, and repeated parameter tuning. Building upon this, a breadth learning system (BLS) is combined to further mine and quantify muscle activation weights from a "horizontal" perspective, making model parameters more transparent, the model more interpretable, and easier to tune and deploy.

[0030] like Figure 2 The diagram shown is a schematic of six upper limb movements displayed in the sEMG dataset provided in this embodiment of the invention. The main content is: six upper limb movements were designed based on human daily behavior, in which the movement involves grasping a thermos cup filled with water (weighing about 550g).

[0031] Furthermore, multi-channel electromyography (sEMG) signals were collected from local areas of the upper limb to establish a complete sEMG dataset. The specific process was as follows: based on the analysis results of the upper limb movement process, the location of the electrode pads of the electromyography device for collecting sEMG signals was determined; through upper limb movement tasks, sEMG signals of different muscle groups under different contraction states were collected for each channel through the electrode pads, constructing a multi-channel sEMG dataset including several types of upper limb movements, and analyzing the muscle contraction corresponding to each movement; the relationship between muscle contraction patterns and movement commands was extracted, and the changes in electromyography signals under different movement states were analyzed.

[0032] In this embodiment, upper limb movement is a highly coordinated process jointly completed by the central nervous system (CNS) and the peripheral nervous system (PNS), involving the cerebral cortex, spinal cord, peripheral nerves and the skeletal muscles they innervate. It is a comprehensive effect produced by the synergistic action of multiple muscles. The activation characteristics of a single muscle are not consistent when the upper limb performs different movement postures. After conducting an in-depth analysis of human upper limb movements from the perspective of exercise physiology, this invention comprehensively considers the muscle contraction state at the sEMG acquisition location and establishes an electromyographic signal dataset of human upper limb movements based on the entire upper limb.

[0033] In order to collect sEMG signals of key muscles or muscle groups under different contraction states during upper limb movement, six upper limb movements were designed based on human daily behavior. The specific muscle contraction under different movements is shown in Table 1. The six types of movements evenly mobilize the corresponding muscle groups, so that most muscles can show different contraction states in different movements. In this way, the collected sEMG signals can more comprehensively and objectively describe the relationship between muscles and movement.

[0034] Table 1. Muscle contraction patterns of the lower limbs during different movements.

[0035]

[0036] Among them, CA represents concentric contraction, which occurs when a muscle generates active force and shortens in length, thus shortening the distance between the proximal and distal attachment points of the muscle. EA represents eccentric contraction, which occurs when a muscle generates active force (attempting to contract) but is simultaneously pulled to a longer length by a more dominant external force. IA represents isometric contraction, which occurs when a muscle generates active force while maintaining a constant length.

[0037] Furthermore, determining the location of the electromyography (EMG) electrode pads for sEMG acquisition also includes full-arm coverage acquisition. The specific process is as follows: several acquisition points covering key muscle groups of the entire arm are collected. Based on the muscle contraction pattern, upper limb movement quantification indicators, and the distribution size of muscles on the human upper limb skin surface, representative muscles affecting the shoulder, upper arm, and forearm are selected.

[0038] In this embodiment, existing sEMG datasets are mostly focused on gesture recognition, with few describing sEMG signals across the entire upper limb. A complete upper limb movement is the result of the coordinated action of four parts: the shoulder, upper arm, forearm, and hand (the result of the coordinated action of multiple muscle groups). Therefore, electromyography signal acquisition should cover as many key upper limb areas as possible in order to extract the most representative complete upper limb movement.

[0039] It should be noted that, considering that sEMG is caused by the contraction of muscle fibers during the contraction of superficial muscles under the skin, and since the hand muscles are not easy to collect stable electromyographic signals with ordinary electrode pads, the electromyographic signals of the forearm muscles that control hand movements are collected to reflect the hand posture in upper limb movements.

[0040] Further, data preprocessing includes Butterworth filtering for noise reduction. The specific process is as follows: acquire the raw electromyographic signal, and perform time reversal and filtering to improve the signal processing accuracy. The raw electromyographic signal represents the raw time-domain electrical signal sequence directly recorded by the sensor without processing. Set the filter frequency range to the allowable range of the signal, and filter out frequency components that exceed the allowable range of the signal.

[0041] In this embodiment, a Butterworth bandpass filter is applied to the original signal to obtain a forward-filtered intermediate signal; the sequence of the forward-filtered intermediate signal is reversed; the same Butterworth bandpass filter is applied again to the time-reversed intermediate signal for reverse filtering; the sequence of the reverse-filtered signal is restored to obtain the final filtering result, thereby achieving complete phase compensation, essentially canceling the delay while achieving zero phase distortion.

[0042] Electromyography (EMG) signal acquisition should cover key upper limb areas as much as possible in order to extract the most representative complete upper limb movements. Forward-backward Infinite Impulse Response (IIR) Butterworth Band-pass Filter (Butterworth BPF) is used for time-domain filtering.

[0043] A bandpass filter can retain a portion of a signal within a specific frequency range, while removing frequency components outside the allowed range. When the original signal is x[n], the normalized cutoff frequencies are as follows:

[0044]

[0045] Where, ω low ω is the low-frequency cutoff angular frequency. high f is the high-frequency cutoff angular frequency. s f is the sampling frequency. N =f s / 2 is the Nyquist frequency, f low For low cutoff frequency, f high It is a high cutoff frequency.

[0046] The IIR Butterworth filter achieves a faster roll-off with a lower order, reducing computational complexity. Its difference equation can be written as:

[0047] ,

[0048] Where N is the order of the filter, representing the maximum historical time point involved in the IIR filter processing; a k With b k The gain coefficient of the filter is determined by the Butterworth design method. y[nk] is the sample of the filter output sequence y[n] k times ago, while x[nk] is the sample of the filter input sequence x[n] k times ago. n is the index of the current time, indicating the time being calculated, and k is the index of the time delay, used to represent the values ​​of y[n] and x[n] at different times.

[0049] Conventional IIR filters introduce group delay, which requires compensation to align with the original signal. To address this issue, this invention employs a forward-backward filtering strategy. First, the signal is forward-filtered; then, the result is time-reversed and inversely filtered; finally, the result is reversed back to the original sequence. This achieves complete phase compensation, essentially canceling out the delay while achieving zero phase distortion. This process can be represented as:

[0050] ,

[0051] Among them, F -1 It is the inverse Fourier transform, used to convert frequency domain signals back to the time domain. Let x[n] be the spectrum of the original signal. Here is the frequency response function of the Butterworth bandpass filter. This represents the corresponding frequency response after time reversal. The canceled phase response ensures that the filtered signal retains the original phase structure.

[0052] Furthermore, data preprocessing also includes sliding window segmentation. The specific process is as follows: based on the window length required for feature extraction and calculation, the denoised sEMG signal is divided into a set number of sliding windows to generate signal segments for subsequent feature extraction and analysis. The signal after the sliding window can expand the training sample set without significantly increasing the computational complexity of the model, thereby improving the model's ability to express features and enabling the extraction of more useful features from non-stationary sEMG signals.

[0053] Once a window has finished processing, the sliding window continues to slide forward, adding new data points to the end of the window while discarding older data points, until the entire signal has been processed.

[0054] In this embodiment, in order to extract more useful features from the non-stationary sEMG signal, a sliding window is used to filter each segment of the sEMG signal y=[y1,y2,...,y...]. n ] R T Segmentation is performed, and the signal after the sliding window can expand the training sample set without significantly increasing the model's computational complexity, thereby improving the model's ability to express features. Subsequences are continuously extracted from the original time-series sEMG signal, where the signal of the i-th (i≥1) window is:

[0055] ,

[0056] Where w is the set window length and st is the sliding step size. This represents the feature sequence extracted by the i-th sliding window.

[0057] like Figure 3 The diagram shown illustrates the overall structure of CDWS-ResCNN and the local structure of the residual blocks provided in this embodiment of the invention. The main content is as follows: A deep learning model based on the CDWS-ResCNN structure is demonstrated. This model extracts features from sEMG signals using a convolutional neural network (CNN) and residual connections. By processing, fusing, and classifying signals from multiple channels, the model can extract rich muscle activity information from the original signals for subsequent motion recognition and muscle state analysis.

[0058] Furthermore, extracting effective upper limb movement features from different channels also includes constructing a weight sharing mechanism in the network. The specific process is as follows: each branch uses a residual block with the same structure and parameters to extract local features, and all branches share the same set of convolutional kernel weights.

[0059] The features extracted from each branch are merged at the back end of the network, and the difference between the current network output and the target is calculated using a loss function to obtain the loss value.

[0060] The loss value is passed back to the network through the backpropagation algorithm to adjust the weight parameters, thereby continuously optimizing the network and improving the accuracy of feature extraction.

[0061] In this embodiment, the loss value reflects the prediction error of the model. In action classification tasks, the cross-entropy loss function is usually used to calculate the difference between the output class probability distribution and the true class label. In muscle activation quantization tasks, the global optimality of weight solution is ensured by solving a convex optimization problem, thereby measuring the difference between the predicted value and the true value.

[0062] When processing signals from different channels, although each channel has an independent input path, they need to share the same set of convolutional kernel weights and network module structure during the local feature extraction stage. This ensures that the muscle groups in each channel extract the same type of features. In the CDWS-ResCNN network, although each channel corresponds to different muscle group signals (such as muscle A, muscle B, and muscle C), by sharing convolutional kernels and network modules, the activation features of muscles A, B, and C will be extracted as consistent feature types (e.g., mean, variance, etc.). This ensures that the network can fairly evaluate the activation state of each muscle group and will not dominate feature extraction because some muscle groups have stronger activation.

[0063] The shared weight mechanism ensures that the network processes signals from different muscle groups in a consistent manner, effectively extracting features of muscle synergy. For example, muscles A and B both play a role in the same movement, but their activation levels differ. By sharing convolutional kernels and residual blocks, the network can accurately assess the contribution of each muscle group to the entire movement, ensuring the extraction of muscle synergy features.

[0064] Furthermore, the learning sub-model for the width of the input feature of the residual convolutional neural network structure output by each channel also includes feature standardization processing of the input features. The specific process is as follows: extract the feature data to be standardized in different channels, and calculate the mean and standard deviation of the original feature value of each feature channel.

[0065] The standard score algorithm is used for standardization. The standardized feature data is then used as input to the BLS model for learning and prediction, ensuring the matching degree of feature value scale and the balanced contribution of each feature to the training.

[0066] Before standardization, the raw feature data for each channel are the same type of deep features extracted from sEMG signals through the CDWS-ResCNN network. After standardization, the feature values ​​of each channel are transformed into data with zero mean and unit variance. For example, after standardization, muscle signals that were originally in the range of 0 to 1000 may become feature values ​​with a mean of 0 and a standard deviation of 1.

[0067] The purpose of feature standardization is to transform features at different scales into standard normal distribution data with a uniform scale. Through this process, the mean of the feature data becomes 0 and the standard deviation becomes 1, so that all features are at the same scale. This avoids the impact of differences in feature scales on model training, making the feature data suitable for machine learning algorithms such as BLS models, which helps to accelerate model convergence and improve performance.

[0068] In this embodiment, before inputting the features extracted from each channel into the BLS sub-model, it is necessary to ensure the matching degree of feature value scales and the balanced contribution of each feature to the training. To this end, the features are standardized using the z-score algorithm, and the specific formula is as follows:

[0069] , , ,

[0070] Among them, X C Let x represent the eigenvalue of the c-th channel after standardization. c This represents the original feature value of the c-th channel. The original feature mean, is the standard deviation of the original features, M is the total number of features, and j represents the feature value of the j-th sample in the c-th channel when calculating the mean and standard deviation. The calculation is performed by traversing all M samples.

[0071] like Figure 4 The diagram shown illustrates the structure of the BLS sub-model corresponding to the c-th channel in this embodiment of the invention. The main content is the relationship between the feature mapping nodes, enhanced feature nodes, and output in the BLS sub-model. Through layer-by-layer processing by these modules, the model can extract useful features from the input sEMG signal and ultimately make predictions.

[0072] Furthermore, the residual convolutional neural network structure of each channel outputs features that are input into their respective width learning sub-models. The specific process is as follows: the standardized features are input into the width learning sub-model. In the input layer of the width learning sub-model, the standardized features are divided by channel and nonlinearly mapped using the hyperbolic tangent activation function to obtain feature mapping nodes.

[0073] The feature mapping nodes are re-mapped nonlinearly to generate enhanced feature nodes, which further improves the learning performance of the network and enables it to fit more complex data structures. The feature mapping nodes and enhanced feature nodes are used together as the hidden layer of the corresponding channel width learning sub-model and passed to the output layer to obtain the output value of the width learning sub-model and the connection weights from the hidden layer to the output layer.

[0074] In this embodiment, existing technologies, in order to ensure the interpretability of muscle activation algorithms, often employ structurally transparent "white-box" algorithms, such as traditional machine learning algorithms like SVM, LDA, and KNN. This invention fully leverages the advantages of width learning, which also possesses structural transparency, and combines this with strong constraints on CNN deep learning algorithms. The constructed architecture, integrating deep learning and width learning, can establish interpretable and quantifiable muscle activation level analysis based on extracted sEMG deep features, making it more interpretable and scientific.

[0075] Width learning does not directly input features into the network structure at the input layer; instead, it performs feature mapping on the input data. Feature mapping not only effectively improves the representation ability of features in high-dimensional space but also enhances the model's discriminative performance.

[0076] After standardization, features are divided by channel. A fixed order for electrode / muscle group channels is determined (e.g., Ch1→Ch2→…→ChC). The same type of deep features is extracted for each channel using a CDWS-ResCNN network. This ensures that the timestamps of each row (sample / window) across all channels are consistent within the same batch. The standardized features of each sample are then concatenated in blocks according to the predetermined channel order: first the Ch1 feature column, then the Ch2 feature column, and so on until ChC. This results in a "total feature table," whose columns are arranged in blocks by channel. Each channel's corresponding column block range is recorded. When input from a specific channel is needed, the entire column block for that channel is retrieved from the total feature table according to the mapping.

[0077] The partitioned features are input into the width learning sub-model, and the hyperbolic tangent (Tanh) activation function is used to perform nonlinear feature mapping. This function outputs continuous values ​​symmetric about the origin and in the interval [-1, 1]. During the nonlinear transformation of the input features, it effectively suppresses extreme values, stabilizes the feature distribution, and reduces the risk of numerical overflow. The model input data consists of standardized features for each channel. X C Then the first p Group feature mapping node Z P for:

[0078] ,

[0079] in, and These are random weights and biases, respectively. n 1 The feature mapping nodes are merged to obtain To further improve the network's learning performance and enable it to fit more complex data structures, feature mapping nodes are... Further enhanced feature nodes are generated through hyperbolic tangent nonlinear mapping, then the... q Group Enhanced Feature Nodes H q for:

[0080] ,

[0081] This yields n² enhanced feature nodes, which, after merging, result in... Then, the feature mapping nodes and enhanced feature nodes are used together as the hidden layer of the corresponding channel's BLS sub-model and passed to the output layer. The output of the c-th channel's BLS sub-model and the connection weights from the hidden layer to the output layer are then... W c The interval can be represented as: .

[0082] Furthermore, the weights are calculated through pseudo-inverse regression, and the absolute value is normalized to obtain the contribution index. The specific process is as follows: perform pseudo-inverse operation on the output value of the width learning sub-model to solve the regression coefficient weights of each channel width learning branch.

[0083] The absolute values ​​of the regression coefficient weights on all mapping nodes are summed, and the sum of all channels in the same type of upper limb movement is normalized to obtain the contribution index of each channel for each type of movement. The contribution index is used to reflect the weight contribution of channel features in the model discrimination process, and provides an effective quantitative basis for channel importance ranking and muscle function analysis.

[0084] The relative contribution index is calculated using a cross-validation algorithm to obtain the normalized contribution of each channel to each category at each fold. The average value of the results after a set number of folds is taken as the final contribution index. The final contribution index is used to reduce the randomness of sample partitioning and improve the reliability of activation weights.

[0085] In this embodiment, BLS is pseudo-inverseed, that is, the formula is... The transformation yields:

[0086] ,

[0087] in, It is a pseudo-inverse matrix obtained through ridge regression approximation, and the parameters... , , and All of these are random variables that remain unchanged during network training, so the calculation process is a convex optimization problem. Through optimization... The value of determines the network output fitted value. With known upper limb movement target labels Closer. The optimization problem is modeled as follows:

[0088] ,

[0089] in, It minimizes the loss function, thereby optimizing the weight matrix. , Let ||·|| be the regularization coefficient. F The Frobenius norm is used to control the minimization of training error; due to the... Adding a positive number to the diagonal of the original inverse pseudo-equation gives it exactly one solution.

[0090] ,

[0091] Where I is the identity matrix, Then the analytical solution to this optimization problem is:

[0092] ;

[0093] The contribution index of channel C to the normalized processing of the k-th type of upper limb movement is defined as:

[0094]

[0095] Where 'a' is the index of the sample, representing the feature contribution of the 'a'th sample, and 'b' is the index of the feature dimension, representing the contribution of the 'b'th feature dimension to the final result.

[0096] This contribution index quantifies the relative activation weight of each channel in different upper limb movement categories. The larger the contribution value for a certain category, the higher the contribution of that channel in representing and judging that category, and the stronger the activation degree.

[0097] To reduce the randomness of sample partitioning, this invention employs... Folded cross-validation improves the reliability of activation weights. Specifically, it calculates the normalized contribution of each channel to each category at each fold, and finally selects the optimal value. The average of the calculated results is used as the final contribution index of that channel to each category. This strategy effectively improves the stability and repeatability of the contribution evaluation, as shown in the following formula:

[0098] ,

[0099] Where f represents the index of the f-th fold, which is the number used to mark the current fold in F-fold cross-validation.

[0100] The muscle activation weight quantification method based on the BLS sub-model ensures the rigor of measuring the activation degree of muscles or muscle groups in upper limb movements and has good interpretability. It is an effective proof of the differential role of each sEMG channel in multi-classification tasks.

[0101] Furthermore, the muscle activation weights corresponding to upper limb movements were analyzed, and the representation of muscle activation degree was verified. The specific process is as follows: the output values ​​of the corresponding channels of the learning sub-model of a set width are concatenated column by column to obtain the prediction matrix to be classified; the prediction matrix to be classified is input into a multilayer perceptron (MLP) to perform category classification prediction and output the action classification probability. The action classification probability is used to reflect the degree of activation of each action category, thereby indirectly representing the activation degree of the corresponding muscle group.

[0102] The quality of representation information in the output of each channel width learning sub-model is judged based on the action classification probability:

[0103] If the movement classification probability is greater than the set movement classification probability, then the muscle activation weight of this group is determined to successfully represent the degree of muscle activation of the movement, and the greater the contribution value of the muscle group when performing the movement, the stronger the degree of muscle activation.

[0104] If the movement classification probability is not greater than the set movement classification probability, then the muscle activation weight of that group is determined to be unable to represent the degree of muscle activation for that movement, and retraining is required until the movement classification probability is greater than the set movement classification probability.

[0105] In this embodiment, to verify the action representation capability of the BLS sub-model weight matrix and thus demonstrate the effectiveness of the muscle activation weight contribution model, this invention establishes an MLP-based classification model for the output of each BLS sub-module. The model's classification performance is used to illustrate its representation capability. Specifically, the output matrices of the corresponding channels of each BLS sub-model are sorted column-wise. The matrix S to be classified is obtained by concatenating the matrix.

[0106]

[0107] in, Indicates will Each BLS submodel corresponds to a channel output matrix The columns are concatenated to obtain the prediction matrix S to be classified. S is then fed into the MLP for classification to output the final classification probability. The MLP makes full use of the complementarity and high-order synergy between channels, so that the channel contribution value analysis has the full-process explanatory power from linear discrimination to deep fusion, and can deeply explore the correlation between key channels and action recognition.

[0108] The above embodiments can be implemented, in whole or in part, by software, hardware (such as circuits), firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0109] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0110] In various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0111] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0112] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices, apparatuses, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0113] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for analyzing upper limb surface electromyographic signals by integrating depth and width learning, characterized in that, Includes the following steps: S1. Collect multi-channel electromyography signals from local areas of the upper limb, establish a complete sEMG dataset, and perform data preprocessing. The sEMG dataset is used to comprehensively and objectively describe the relationship between muscles and movement in the entire upper limb. S2, For the acquired multi-channel sEMG signals, design the same number of convolutional neural network branches with residual structures, and extract effective upper limb movement features from different channels. The effective upper limb movement features can reflect the activation patterns of each muscle group in upper limb movements. S3, the residual convolutional neural network structure of each channel outputs features and inputs them into their respective width learning sub-models. The weights are calculated through pseudo-inverse regression, and the absolute value is normalized to obtain the contribution index. The contribution index is stabilized by cross-validation of the mean. The muscle activation weights corresponding to upper limb movements are analyzed, and the representation of muscle activation degree is verified. The extraction of effective upper limb movement features from different channels also includes constructing a weight-sharing mechanism in the network, the specific process of which is as follows: Each branch uses residual convolutional blocks with the same structure and parameters to extract local features, and all branches share the same set of convolutional kernel weights; The features extracted from each branch are merged at the back end of the network, and the difference between the current network output and the target is calculated using a loss function to obtain the loss value. The loss value is passed back to the network through the backpropagation algorithm to adjust the weight parameters; The residual convolutional neural network structure of each channel outputs features that are input into their respective widths to learn a sub-model. The specific process is as follows: The standardized features are input into the width learning sub-model. In the input layer of the width learning sub-model, the standardized features are divided by channel and non-linearly mapped using the hyperbolic tangent activation function to obtain feature mapping nodes. The feature mapping nodes are re-mapped nonlinearly to generate enhanced feature nodes; The feature mapping nodes and the enhanced feature nodes are used together as the hidden layer of the corresponding channel width learning sub-model and passed to the output layer to obtain the output value of the width learning sub-model and the connection weights from the hidden layer to the output layer. The analysis of upper limb movements corresponds to muscle activation weights, and the verification of muscle activation levels are performed. The specific process is as follows: The output values ​​of the corresponding channels of the learning sub-models with a set width are concatenated column by column to obtain the prediction matrix to be classified; The prediction matrix to be classified is input into a multilayer perceptron to perform category classification prediction and output the action classification probability. The action classification probability is used to reflect the degree of activation of each action category, thereby indirectly representing the degree of activation of the corresponding muscle group. The quality of representation information in the output of each channel width learning sub-model is judged based on the action classification probability: If the action classification probability is greater than the set action classification probability, the muscle activation weight is determined to successfully represent the degree of muscle activation of the action, and the greater the contribution value of the muscle group when performing the action, the stronger the degree of muscle activation. If the movement classification probability is not greater than the set movement classification probability, it is determined that the muscle activation weight cannot represent the degree of muscle activation of the movement, and retraining is required until the movement classification probability is greater than the set movement classification probability.

2. The method for analyzing upper limb surface electromyographic signals by fusing depth and width learning as described in claim 1, characterized in that, The process of acquiring multi-channel electromyography (EMG) signals from local areas of the upper limb to establish a complete sEMG dataset is as follows: Based on the analysis results of upper limb movement process, the location of the electrode pads of the electromyography device for collecting sEMG was determined; Through upper limb movement tasks, sEMG signals of different muscle groups under different contraction states were collected for each channel using electrode pads. A multi-channel sEMG dataset including several types of upper limb movements was constructed, and the muscle contraction corresponding to each movement was analyzed. The relationship between muscle contraction patterns and motor commands was extracted, and the changes in electromyographic signals under different movement states were analyzed.

3. The method for analyzing upper limb surface electromyographic signals by fusing depth and width learning as described in claim 2, characterized in that, The determination of the location for sEMG acquisition by the electrode pads of the electromyography device also includes full-arm coverage acquisition, the specific process of which is as follows: Collect data at several collection points covering key muscle groups throughout the arm; Based on muscle contraction patterns, quantitative indicators of upper limb movement, and the distribution and size of muscles on the skin surface of the human upper limb, representative muscles affecting the shoulder, upper arm, and forearm were selected.

4. The method for analyzing upper limb surface electromyographic signals by fusing depth and width learning as described in claim 1, characterized in that, The data preprocessing includes Butterworth filtering for noise reduction, and the specific process is as follows: The raw electromyographic signal is acquired and subjected to time reversal and filtering to improve the signal processing accuracy. The raw electromyographic signal represents the unprocessed raw time-domain electrical signal sequence directly recorded by the sensor. Set the filter frequency range to the allowable range of the signal, and filter out frequency components that exceed the allowable range of the signal.

5. The method for analyzing upper limb surface electromyographic signals by fusing depth and width learning as described in claim 4, characterized in that, The data preprocessing also includes sliding window segmentation, the specific process of which is as follows: Based on the required window length for feature extraction and calculation, the denoised sEMG signal is divided into a set number of sliding windows to generate signal segments for subsequent feature extraction and analysis. Once a window has finished processing, the sliding window continues to slide forward, adding new data points to the end of the window while discarding older data points, until the entire signal has been processed.

6. The method for analyzing upper limb surface electromyographic signals by fusing depth and width learning as described in claim 1, characterized in that, The residual convolutional neural network structure for each channel outputs features, and its respective width learning sub-model also includes feature standardization processing of the input features. The specific process is as follows: Extract the feature data to be standardized from different channels, and calculate the mean and standard deviation of the original feature values ​​for each feature channel. The standard score algorithm is used for standardization. The standardized feature data will be used as input to the BLS model for learning and prediction.

7. The method for analyzing upper limb surface electromyographic signals by fusing depth and width learning as described in claim 1, characterized in that, The contribution index is obtained by calculating weights through pseudo-inverse regression and normalizing the absolute value. The specific process is as follows: The pseudo-inverse operation is performed on the output value of the width learning sub-model to solve the regression coefficient weights of each channel width learning branch; The absolute values ​​of the regression coefficient weights on all mapping nodes are summed, and the sum of all channels in the same type of upper limb movement is normalized to obtain the contribution index of each channel for each type of movement. The contribution index is used to reflect the weight contribution of channel features in the model discrimination process. The relative contribution index is calculated using a cross-validation algorithm to obtain the normalized contribution of each channel to each category at each fold. The average value of the results after a set number of folds is taken as the final contribution index. The final contribution index is used to reduce the randomness of sample partitioning and improve the reliability of activation weights.

Images