A method and system for decoding state-sensing electromyographic signals

Abstract

This application discloses a state-aware electromyography (EMG) signal decoding method and system in the field of signal processing technology, comprising: acquiring the surface EMG signal to be decoded and preprocessing it; inputting the preprocessed surface EMG signal into a pre-trained EMG signal decoding model; wherein the EMG signal decoding model includes a state-aware encoding module, a state-aware decoding module, and a discriminator module; and generating an EMG signal motion decoding result based on the preprocessed surface EMG signal, the state-aware encoding module, the state-aware decoding module, and the discriminator module. This application solves the problems of insufficient versatility and difficulty in extending to diverse time-series data scenarios in existing technologies, effectively ensuring the practicality of the decoded surface EMG signal in human-computer interaction, and enhancing the generalization ability of the EMG signal decoding model in diverse scenarios.

Description

Technical Field

[0001] This application relates to the field of signal processing technology, and in particular to a state-sensing electromyographic signal decoding method and system. Background Technology

[0002] Surface electromyography (sEMG) is a bioelectrical signal collected from the skin surface during skeletal muscle contraction. It contains time-series information about muscle activity and can be used to reflect the body's intention and dynamic state during movement. sEMG has broad application prospects in gesture recognition, rehabilitation medicine, prosthetic control, and human-computer interaction. Traditional sEMG-based algorithms often require manually designed features or complex models to capture the relationship between signals and movements. However, these methods generally suffer from problems such as strong dependence on large amounts of training data and poor adaptability to individual differences in samples.

[0003] Existing research has improved model performance through deep learning methods and attention mechanisms, such as using attention mechanisms to construct encoder-decoder networks to estimate finger joint angles. While these models perform well on specific sub-tasks, they lack generality and struggle to generalize to diverse time-series data scenarios, posing significant challenges in handling individual differences. In sEMG signal decoding, sample size and individual variability are two key issues. Sufficient samples enhance model generalization ability, but acquiring large-scale data is often difficult in practical applications. Deep learning models used for surface electromyography signal decoding not only suffer from decreased reliability but also lack excellent multi-scenario generalization capabilities. Summary of the Invention

[0004] The purpose of this application is to provide a state-aware electromyography signal decoding method and system, which solves the technical problems of existing technologies lacking universality, being difficult to generalize to diverse time series data scenarios, and having difficulty in handling individual differences.

[0005] To achieve the above objectives, this application employs the following technical solution:

[0006] In a first aspect, this application provides a method for decoding state-aware electromyographic signals, including:

[0007] Acquire the surface electromyography signal to be decoded and perform preprocessing;

[0008] The preprocessed surface electromyography (EMG) signal is input into a pre-trained EMG signal decoding model; wherein, the EMG signal decoding model includes a state-aware encoding module, a state-aware decoding module, and a discriminator module;

[0009] The preprocessed surface electromyography signal is encoded using the state-aware encoding module to generate a continuous neuromuscular state sequence.

[0010] The state-aware decoding module processes the continuous neuromuscular state sequence to output a predicted electromyography (EMG) sequence within a time window in parallel. The predicted EMG sequence includes a joint angle sequence and a muscle force sequence. The state-aware encoding module includes a dynamic collaborative Mamba decoder.

[0011] The continuous neuromuscular state sequence is sampled using the dynamic collaborative Mamba decoder based on an event-adaptive sampling mechanism.

[0012] Temporal modeling of the sampled continuous neuromuscular state sequence is performed using dynamic collaborative channel state updates to generate electromyographic state features; wherein, the expression for the dynamic collaborative channel state update is:

[0013]

[0014] In the formula, Indicates the first Time, Number The real-time state vector of the channel type. Indicates the first Time, Number Historical state vector of the channel type; This represents the internal state transition matrix of the c-th channel. This represents the input mapping matrix for the c-th channel; This represents the feature vector input to the c-th channel at time t. This represents the summation over all channel categories except for channel c. This represents the dynamic collaborative weight matrix that maps the historical state of the c'-th channel to the c-th channel at time t. Indicates the first Time, Number Historical state vector of the channel type;

[0015] The temporal dependencies of the electromyographic state features are captured using a state propagation mechanism to generate a cooperative state representation.

[0016] The cooperative state representation is input into a time-windowed parallel prediction unit, which outputs the predicted electromyographic sequence within the time window in parallel.

[0017] The discriminator module constrains and optimizes the predicted electromyography sequence to generate motion decoding results of the electromyography signal.

[0018] Furthermore, the preprocessing includes:

[0019] Noise in the surface electromyography (EMG) signal is removed by filtering the signal.

[0020] The filtered surface electromyography (EMG) signal is segmented using a sliding window segmentation strategy to generate an EMG signal matrix consisting of multiple time segments, and the EMG signal matrix is ​​then normalized.

[0021] Further, the step of encoding the preprocessed surface electromyography signal using the state-aware encoding module to generate a continuous neuromuscular state sequence includes:

[0022] The preprocessed surface electromyography signal is encoded by multi-scale temporal convolution operations to obtain multi-scale feature vectors;

[0023] The multi-scale feature vectors are mapped to a low-dimensional latent space by linear state projection, and an adaptive state preservation mechanism is introduced to output the continuous neuromuscular state sequence.

[0024] Furthermore, the expression for the multi-scale temporal convolution operation is:

[0025]

[0026] In the formula, Indicates the first Multi-scale feature vectors at different times; Indicates the first The dilation rate of each convolutional kernel; Indicates the first The surface electromyography signals input at any time; This represents a one-dimensional convolution operation; This represents the total number of convolution kernels.

[0027] Furthermore, the discriminator module includes a local convolution submodule, a global convolution submodule, and a multi-scale first-level convolution submodule;

[0028] The step of constraining and optimizing the predicted electromyographic sequence through the discriminator module to generate electromyographic signal motion decoding results includes:

[0029] The predicted electromyography sequence is input into the local convolutional submodule and the global convolutional submodule, respectively;

[0030] The local feature vector is obtained by extracting local change features of the predicted electromyography sequence through the local convolution submodule.

[0031] The global convolutional submodule extracts the global continuity features of the predicted electromyography sequence to obtain a global feature vector;

[0032] The local feature vector and the global feature vector are concatenated and fused through the multi-scale first-level convolution submodule to generate a discriminative feature vector for evaluating the physiological rationality and smoothness of the predicted electromyography sequence.

[0033] The output of the state-aware decoding module is constrained and optimized based on the discriminative feature vector to generate the final electromyographic signal motion decoding result.

[0034] Furthermore, the training process of the electromyographic signal decoding model includes:

[0035] The state-aware decoding module includes a dynamic cooperative Mamba decoder, which uses the weights of the pre-trained Whisper model as the initial weights of the dynamic cooperative Mamba decoder, and introduces a low-rank residual update mechanism to perform task specialization on the weights of the dynamic cooperative Mamba decoder. The expression for the low-rank residual update mechanism is as follows:

[0036]

[0037] In the formula, This represents the decoder weight matrix obtained after low-rank residual update. This represents the weights of the pre-trained Whisper model. Denotes the first low-rank matrix. Let represent the second low-rank matrix, where is the first low-rank matrix. The dimension is The second low-rank matrix The latitude is , Indicates the input dimension. Indicates the output dimension. Let represent a low-rank parameter that satisfies . ;

[0038] The discriminator module and the state-aware decoding module are jointly optimized through adversarial training, which includes:

[0039] The predicted electromyography (EMG) sequence generated by the state-aware decoding module, and the actual EMG sequence corresponding to the predicted EMG sequence, are input to the discriminator module.

[0040] The discriminator module performs joint modeling of the input predicted electromyography (EMG) sequence and the actual EMG sequence based on a continuous state space model to generate a discrimination result, which is used to constrain the physiological rationality and smoothness of the predicted EMG sequence.

[0041] Based on the low-rank residual update mechanism, the weights of the dynamic collaborative Mamba decoder are task-specificated and subjected to adversarial training to obtain an optimized electromyography signal decoding model.

[0042] Furthermore, the expression for the continuous state-space model is:

[0043]

[0044] In the formula, The discriminator module indicates that in the first... The hidden state vector at time step 1. Indicates the first The hidden state vector at time step; This represents the time window output by the discriminator module. The discrimination results of predicting electromyographic sequences within the body; Indicates the first The predicted or actual electromyographic sequence is input at any given time. Represents a non-linear activation function. This represents the Sigmoid activation function. This represents the weight matrix of the input layer. This represents the weight matrix of the hidden layer. This represents the weight matrix of the output layer; This indicates that the discriminator module is within the time window. The hidden state vector at time step; Indicates the length of the time window.

[0045] Furthermore, the joint optimization includes a generator loss function. and discriminator loss function ;

[0046] The generator loss function The expression is:

[0047]

[0048] In the formula, The weighting coefficients of the reconstruction loss term are represented. Represents the weighting coefficient of the adversarial loss term; Indicates the first The actual target value at any given time; This represents the predicted electromyographic sequence output by the state-aware decoding module. This represents a continuous sequence of input neuromuscular states. This indicates that the state-aware decoding module is in the first... Predicted electromyographic sequences output at specific times; Indicates the total length of the time series; This indicates that the discriminator module interprets the predicted electromyographic sequences generated by the state-aware decoding module. The discrimination probability;

[0049] The discriminator loss function The expression is:

[0050]

[0051] In the formula, Represents the mathematical expectation. Indicates sample Follows the distribution of true electromyography sequence , This represents the actual electromyographic sequence; Indicates sample Follows the distribution of continuous neuromuscular state sequences ; This indicates that the discriminator module recognizes the real electromyography sequence. The discrimination probability.

[0052] Secondly, this application provides a state-sensing electromyographic signal decoding device, the device comprising:

[0053] The acquisition module is used to acquire the surface electromyography signal to be decoded and perform preprocessing.

[0054] An input module is used to input the preprocessed surface electromyography (EMG) signal into a pre-trained EMG signal decoding model; wherein, the EMG signal decoding model includes a state-aware encoding module, a state-aware decoding module, and a discriminator module;

[0055] The output module is used to encode the preprocessed surface electromyography signal using the state-aware encoding module to generate a continuous neuromuscular state sequence.

[0056] The state-aware decoding module processes the continuous neuromuscular state sequence to output a predicted electromyography (EMG) sequence within a time window in parallel. The predicted EMG sequence includes a joint angle sequence and a muscle force sequence. The state-aware encoding module includes a dynamic collaborative Mamba decoder.

[0057] The continuous neuromuscular state sequence is sampled using the dynamic collaborative Mamba decoder based on an event-adaptive sampling mechanism.

[0058] Temporal modeling of the sampled continuous neuromuscular state sequence is performed using dynamic collaborative channel state updates to generate electromyographic state features; wherein, the expression for the dynamic collaborative channel state update is:

[0059]

[0060] In the formula, Indicates the first Time, Number The real-time state vector of the channel type. Indicates the first Time, Number Historical state vector of the channel type; This represents the internal state transition matrix of the c-th channel. This represents the input mapping matrix for the c-th channel; This represents the feature vector input to the c-th channel at time t. This represents the summation over all channel categories except for channel c. This represents the dynamic collaborative weight matrix that maps the historical state of the c'-th channel to the c-th channel at time t. Indicates the first Time, Number Historical state vector of the channel type;

[0061] The temporal dependencies of the electromyographic state features are captured using a state propagation mechanism to generate a cooperative state representation.

[0062] The cooperative state representation is input into a time-windowed parallel prediction unit, which outputs the predicted electromyographic sequence within the time window in parallel.

[0063] The discriminator module constrains and optimizes the predicted electromyography sequence to generate motion decoding results of the electromyography signal.

[0064] Compared with the prior art, the beneficial effects achieved by this application are as follows:

[0065] The state-aware electromyography (EMG) signal decoding method provided in this application outputs joint angle and muscle force information by leveraging an EMG signal decoding model. This effectively solves the technical challenges of existing technologies, such as insufficient versatility, difficulty in adapting to diverse time-series data scenarios, and difficulty in handling individual differences. This method utilizes a state-aware encoding module, a state-aware decoding module, and a discriminator module to perform in-depth processing and optimization of surface EMG signals, resulting in more accurate decoding results. Furthermore, the method of this application can output predicted EMG sequences within a time window in parallel, significantly improving decoding efficiency. In the field of human-computer interaction, this allows devices to more sensitively perceive human movement intentions, achieving a more natural and efficient interactive experience. Attached Figure Description

[0066] Figure 1 A schematic flowchart illustrating the state-sensing electromyographic signal decoding method provided in this application embodiment;

[0067] Figure 2 This is a schematic diagram of the electromyography signal decoding model provided in the embodiments of this application;

[0068] Figure 3This is a schematic diagram of the decoding process of the dynamic collaborative Mamba decoder in the embodiments of this application;

[0069] Figure 4 This is a schematic diagram illustrating the normalization of physiological constraint transformation of joint angle sequences provided in an embodiment of this application. Detailed Implementation

[0070] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit this application or its application or use. Example 1

[0071] like Figure 1 As shown in the figure, this application provides a state-sensing electromyography signal decoding method, which includes the following steps:

[0072] Step 1: Acquire the surface electromyography (EMG) signal to be decoded and perform preprocessing;

[0073] In this embodiment, the preprocessing step for the surface electromyography signal includes:

[0074] Noise in the surface electromyography (EMG) signal is removed by filtering the signal.

[0075] Specifically, the surface electromyography signal is filtered by a bandpass filter to remove low-frequency motion artifacts and high-frequency random noise, while retaining effective muscle activity information.

[0076] The filtered surface electromyography (EMG) signal is segmented using a sliding window segmentation strategy to generate an EMG signal matrix consisting of multiple time segments, and the EMG signal matrix is ​​then normalized.

[0077] In one specific implementation, the electromyography (EMG) signal sequence matrix is ​​normalized. Specifically, a Z-score normalization method based on statistical characteristics is used to process the data of each signal channel separately, eliminating inherent amplitude differences and baseline shifts between different individuals and between different signal channels. This maps the signal data of each channel to a standard numerical distribution range with zero mean and unit variance, thereby obtaining normalized multi-channel surface EMG signals.

[0078] Step 2: Input the preprocessed surface electromyography (EMG) signal into the pre-trained EMG signal decoding model; wherein the EMG signal decoding model includes a state-aware encoding module, a state-aware decoding module, and a discriminator module;

[0079] In this embodiment, the structural schematic diagram of the electromyography signal decoding model is as follows: Figure 2 As shown, the training process of the electromyography signal decoding model uses the AdamW optimizer, with an initial learning rate of... The specific training process for the electromyography signal decoding model includes:

[0080] The state-aware decoding module includes a dynamic cooperative Mamba decoder. It uses the weights of a pre-trained Whisper model as the initial weights for the dynamic cooperative Mamba decoder and introduces a low-rank residual update mechanism to specialize the weights of the dynamic cooperative Mamba decoder for the task. The expression for the low-rank residual update mechanism is as follows:

[0081]

[0082] In the formula, This represents the decoder weight matrix obtained after low-rank residual update. This represents the weights of the pre-trained Whisper model. Denotes the first low-rank matrix. Let represent the second low-rank matrix, where is the first low-rank matrix. Dimensions for The second low-rank matrix Dimensions for , Indicates the input dimension. Indicates the output dimension. Let represent a low-rank parameter that satisfies . ;

[0083] The discriminator module and the state-aware decoding module are jointly optimized through adversarial training, which includes:

[0084] The predicted electromyography (EMG) sequence generated by the state-aware decoding module, and the actual EMG sequence corresponding to the predicted EMG sequence, are input to the discriminator module.

[0085] The discriminator module performs joint modeling of the input predicted electromyography (EMG) sequence and the actual EMG sequence based on a continuous state space model to generate a discrimination result, which is used to constrain the physiological rationality and smoothness of the predicted EMG sequence.

[0086] Based on the low-rank residual update mechanism, the weights of the dynamic collaborative Mamba decoder are task-specificated and subjected to adversarial training to obtain an optimized electromyography signal decoding model.

[0087] In one feasible implementation, the expression for the continuous state-space model is:

[0088]

[0089] In the formula, The discriminator module indicates that in the first... The hidden state vector at time step 1. Indicates the first The hidden state vector at time step; This represents the time window output by the discriminator module. The discrimination results of predicting electromyographic sequences within the body; Indicates the first The predicted or actual electromyographic sequence is input at any given time. Represents a non-linear activation function. This represents the Sigmoid activation function. This represents the weight matrix of the input layer. This represents the weight matrix of the hidden layer. This represents the weight matrix of the output layer; This indicates that the discriminator module is within the time window. The hidden state vector at time step; Indicates the length of the time window.

[0090] The joint optimization includes a generator loss function. and discriminator loss function ;

[0091] The generator loss function The expression is:

[0092]

[0093] In the formula, The weighting coefficients of the reconstruction loss term are represented. Represents the weighting coefficient of the adversarial loss term; Indicates the first The actual target value at any given time; This represents the predicted electromyographic sequence output by the state-aware decoding module. This represents a continuous sequence of input neuromuscular states. This indicates that the state-aware decoding module is in the first... Predicted electromyographic sequences output at specific times; Indicates the total length of the time series; This indicates that the discriminator module interprets the predicted electromyographic sequences generated by the state-aware decoding module. The discrimination probability;

[0094] The discriminator loss function The expression is:

[0095]

[0096] In the formula, Represents the mathematical expectation. Indicates sample Follows the distribution of true electromyography sequence , This represents the actual electromyographic sequence; Indicates sample Follows the distribution of continuous neuromuscular state sequences ; This indicates that the discriminator module recognizes the real electromyography sequence. The discrimination probability.

[0097] Step 3: Encode the preprocessed surface electromyography signal using the state-aware encoding module to generate a continuous neuromuscular state sequence;

[0098] In this embodiment, the specific steps for generating a continuous neuromuscular state sequence by inputting the preprocessed normalized multichannel surface electromyography signal into the state-aware coding module include:

[0099] Normalized multi-channel surface electromyography signals are encoded by multi-scale temporal convolution operations to obtain multi-scale feature vectors;

[0100] The multi-scale feature vectors are mapped to a low-dimensional latent space by linear state projection, and an adaptive state preservation mechanism is introduced to output the continuous neuromuscular state sequence.

[0101] Specifically, the expression for multi-scale temporal convolution operation is:

[0102]

[0103] In the formula, Indicates the first Multi-scale feature vectors at different times; Indicates the first The dilation rate of each convolutional kernel; Indicates the first The surface electromyography signals input at any time; This represents a one-dimensional convolution operation; This represents the total number of convolution kernels.

[0104] In some embodiments, the rules of the adaptive state preservation mechanism include: in response to electromyographic activity being in a stable phase, the potential state vector of the state-aware encoding module remains unchanged at its current value;

[0105] In response to a significant change in muscle activation, the adaptive state preservation mechanism triggers a latent state update process, and the latent state vector of the state-aware encoding module is updated in real time.

[0106] Step 4: The continuous neuromuscular state sequence is processed by the state-aware decoding module to output the predicted electromyography sequence within the time window in parallel, wherein the predicted electromyography sequence includes joint angle sequence and muscle force sequence;

[0107] In this embodiment, the state-aware decoding module includes a dynamic cooperative Mamba decoder. The steps for decoding a continuous neuromuscular state sequence using the dynamic cooperative Mamba decoder are as follows: Figure 3 As shown, the specific steps include:

[0108] The continuous neuromuscular state sequence is sampled using the dynamic collaborative Mamba decoder based on an event-adaptive sampling mechanism.

[0109] Temporal modeling of the sampled continuous neuromuscular state sequence is performed using dynamic collaborative channel state updates to generate electromyographic state features; wherein, the expression for the dynamic collaborative channel state update is:

[0110]

[0111] In the formula, Indicates the first Time, Number The real-time state vector of the channel type. Indicates the first Time, Number Historical state vector of the channel type; This represents the internal state transition matrix of the c-th channel. This represents the input mapping matrix for the c-th channel; This represents the feature vector input to the c-th channel at time t. This represents the summation over all channel categories except for channel c. This represents the dynamic collaborative weight matrix that maps the historical state of the c'-th channel to the c-th channel at time t. Indicates the first Time, Number Historical state vector of the channel type;

[0112] In some embodiments, to achieve refined modeling of the dynamic changes in muscle coordination patterns with the motion content, the dynamic coordination weight matrix... Adaptive computation through attention mechanisms:

[0113]

[0114] In the formula, , The query matrix represents the hidden states of all channels in the previous time step. Trainable matrix Obtained by projection; , The key matrix is ​​represented by the hidden states of all channels at the previous time step. Trainable matrix Obtained by projection; Indicates the feature dimension scaling factor; This represents the matrix transpose operation; This represents the Softmax function.

[0115] The temporal dependencies of the electromyographic state features are captured using a state propagation mechanism to generate a cooperative state representation.

[0116] The cooperative state representation is input into the time-windowed parallel prediction unit, which outputs the predicted electromyographic sequence within the time window in parallel.

[0117] Step 5: Constrain and optimize the predicted electromyography sequence through the discriminator module to generate the electromyography signal motion decoding result.

[0118] In this embodiment, the discriminator module includes a local convolution submodule, a global convolution submodule, and a multi-scale first-level convolution submodule. The discriminator module constrains and optimizes the predicted electromyography (EMG) sequence to generate motion decoding results for the EMG signal. The specific steps include:

[0119] The predicted electromyography sequence is input into the local convolutional submodule and the global convolutional submodule, respectively;

[0120] The local feature vector is obtained by extracting local change features of the predicted electromyography sequence through the local convolution submodule.

[0121] The global convolutional submodule extracts the global continuity features of the predicted electromyography sequence to obtain a global feature vector;

[0122] The local feature vector and the global feature vector are concatenated and fused through the multi-scale first-level convolution submodule to generate a discriminative feature vector for evaluating the physiological rationality and smoothness of the predicted electromyography sequence.

[0123] In some embodiments, the multi-scale first-level convolutional submodule comprises a first-level convolutional layer, a batch normalization layer, a max pooling layer, and a ReLU activation layer in sequence. In practice, the multi-scale first-level convolutional submodule concatenates and fuses the input local and global feature vectors to form a feature representation. The first-level convolutional layer extracts discriminative patterns from the feature representation, and the batch normalization layer standardizes the data to accelerate training convergence. The max pooling layer downsamples the feature representation, enhancing feature robustness and scale invariance. The ReLU activation function introduces nonlinearity, ultimately obtaining the discriminative feature vector.

[0124] The output of the state-aware decoding module is constrained and optimized based on the discriminative feature vector to generate the final electromyographic signal motion decoding result. Example 2

[0125] Based on the same technical concept as in Embodiment 1, this embodiment provides a state-sensing electromyographic signal decoding system, the system comprising:

[0126] The acquisition module is used to acquire the surface electromyography signal to be decoded and perform preprocessing.

[0127] An input module is used to input the preprocessed surface electromyography (EMG) signal into a pre-trained EMG signal decoding model; wherein, the EMG signal decoding model includes a state-aware encoding module, a state-aware decoding module, and a discriminator module;

[0128] The output module is used to encode the preprocessed surface electromyography signal using the state-aware encoding module to generate a continuous neuromuscular state sequence.

[0129] The state-aware decoding module processes the continuous neuromuscular state sequence to output a predicted electromyographic sequence within a time window in parallel. The predicted electromyographic sequence includes a joint angle sequence and a muscle force sequence.

[0130] The discriminator module constrains and optimizes the predicted electromyography sequence to generate motion decoding results of the electromyography signal.

[0131] The specific implementation process of each module function in this embodiment can be found in Embodiment 1, which has the same technical effect as the method provided in Embodiment 1, and will not be described in detail here. Example 3

[0132] To evaluate the state-aware electromyography (EMG) signal decoding method of Example 1, this example used the NinaPro DB2 database containing 40 subjects to establish an EMG signal decoding model between joint angles and hand forces. In this database, hand kinematic information was acquired using a CyberGlove data glove equipped with 22 sensors; hand dynamics information was measured by a finger force linear sensor; and muscle activity was recorded using Delsys dual-differential surface EMG electrodes, which were fixed to the forearm using an elastic armband.

[0133] For each subject, the collected signals included surface electromyography (EMG) signals, joint angle signals, and finger force signals. Surface EMG signals underwent spectral transformation and Mel-scale transformation to extract frequency domain features. The joint angle signals were normalized using a physiologically constrained transformation method; the specific transformation formula is as follows:

[0134]

[0135] In the formula, This indicates the joint angle signal.

[0136] like Figure 4 As shown, joint angle signals are mapped to a standardized numerical range to accommodate biomechanical differences among individuals. Normalization of finger force signals is based on a comprehensive analysis of 10 subjects, setting a normalization boundary for the observed data range; outliers exceeding this boundary are truncated to enhance signal robustness and representation stability.

[0137] To evaluate the performance of the electromyography signal decoding model in this application, several comparative experiments were conducted. The baseline models used included: a CNN-based Regression model, a CNN-LSTM model, a CNN-Transformer model, and a Whisper-Decoder model. All models were trained under the same data partitioning, input window length, and training strategy.

[0138] The performance of the electromyography signal decoding model was quantitatively evaluated using the following metrics: mean absolute error (MAE), root mean square error (RMSE), and inference delay, as shown in Tables 1 and 2.

[0139] Table 1. Performance Comparison of Joint Angle Prediction Tasks

[0140]

[0141] Table 2: Performance Comparison of Hand Muscle Strength Prediction Task

[0142]

[0143] As shown in Table 1, in the joint angle prediction task, the electromyography signal decoding model of this application achieved the lowest mean absolute error and root mean square error compared to other baseline models, demonstrating the best joint angle prediction accuracy. Compared to the CNN-based Regression model, this model significantly reduced MAE and RMSE by 26.6% and 22.3%, respectively. Compared with other state-of-the-art comparative models, this model also maintained its accuracy advantage: for example, compared to the CNN-Transformer model, this model further reduced MAE and RMSE by 7.8% and 5.0%, respectively.

[0144] In terms of inference speed, this model also performs exceptionally well. Compared to the Whisper-Decoder model, this model reduces the inference latency to 142ms while maintaining similar prediction errors, achieving a 30.0% speed improvement and a better balance between accuracy and efficiency. Although the CNN-LSTM model and the CNN-based Regression model have lower latency, their prediction errors are significantly higher than this model, indicating that their trade-off between accuracy and efficiency is far less effective than that of this model.

[0145] As shown in Table 2, in the hand muscle strength prediction task, the electromyography signal decoding model proposed in this application achieved the best estimation accuracy among all the comparison models, with MAE and RMSE reaching 1.34 degrees and 1.95 degrees, respectively.

[0146] Compared to the second-best performing Whisper-Decoder model, this model further reduces the mean absolute error by 3.6% and the root mean square error by 3.0%. This clearly demonstrates that this model effectively improves both the accuracy and stability of hand muscle strength prediction. Simultaneously, this model maintains good real-time performance, with an inference latency of 138 milliseconds, significantly lower than the slowest CNN-Transformer model, and also exhibits an advantage in processing speed compared to the Whisper-Decoder model. This indicates that this model significantly improves decoding accuracy while also achieving a good balance between computational efficiency and performance.

[0147] Experimental data show that the model of this invention significantly outperforms the traditional baseline model in all evaluation indicators in the task of predicting joint angles and hand forces, verifying its superiority in accuracy and efficiency.

[0148] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0149] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0150] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0151] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0152] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the technical principles of this invention, and these forms are all within the protection scope of this application.

Claims

1. A method for decoding state-aware electromyographic signals, characterized in that, include: Acquire the surface electromyography signal to be decoded and perform preprocessing; The preprocessed surface electromyography (EMG) signal is input into a pre-trained EMG signal decoding model; wherein, the EMG signal decoding model includes a state-aware encoding module, a state-aware decoding module, and a discriminator module; The preprocessed surface electromyography signal is encoded using the state-aware encoding module to generate a continuous neuromuscular state sequence. The state-aware decoding module processes the continuous neuromuscular state sequence to output a predicted electromyography (EMG) sequence within a time window in parallel. The predicted EMG sequence includes a joint angle sequence and a muscle force sequence. The state-aware decoding module includes a dynamic collaborative Mamba decoder. The continuous neuromuscular state sequence is sampled using the dynamic collaborative Mamba decoder based on an event-adaptive sampling mechanism. Temporal modeling of the sampled continuous neuromuscular state sequence is performed using dynamic collaborative channel state updates to generate electromyographic state features; wherein, the expression for the dynamic collaborative channel state update is: In the formula, Indicates the first Time, Number The real-time state vector of the channel type. Indicates the first Time, Number Historical state vector of the channel type; This represents the internal state transition matrix of the c-th channel. This represents the input mapping matrix for the c-th channel; This represents the feature vector input to the c-th channel at time t. This represents the summation over all channel categories except for channel c. This represents the dynamic collaborative weight matrix that maps the historical state of the c'-th channel to the c-th channel at time t. Indicates the first Time, Number Historical state vector of the channel type; The temporal dependencies of the electromyographic state features are captured using a state propagation mechanism to generate a cooperative state representation. The cooperative state representation is input into a time-windowed parallel prediction unit, which outputs the predicted electromyographic sequence within the time window in parallel. The discriminator module constrains and optimizes the predicted electromyography sequence to generate motion decoding results of the electromyography signal.

2. The state-aware electromyographic signal decoding method according to claim 1, characterized in that, The preprocessing includes: Noise in the surface electromyography (EMG) signal is removed by filtering the signal. The filtered surface electromyography (EMG) signal is segmented using a sliding window segmentation strategy to generate an EMG signal matrix consisting of multiple time segments, and the EMG signal matrix is ​​then normalized.

3. The state-sensing electromyographic signal decoding method according to claim 1 or 2, characterized in that, The step of encoding the preprocessed surface electromyography signal using the state-aware encoding module to generate a continuous neuromuscular state sequence includes: The preprocessed surface electromyography signal is encoded by multi-scale temporal convolution operations to obtain multi-scale feature vectors; The multi-scale feature vectors are mapped to a low-dimensional latent space by linear state projection, and an adaptive state preservation mechanism is introduced to output the continuous neuromuscular state sequence.

4. The state-aware electromyographic signal decoding method according to claim 3, characterized in that, The expression for the multi-scale temporal convolution operation is: In the formula, Indicates the first Multi-scale feature vectors at different times; Indicates the first The dilation rate of each convolutional kernel; Indicates the first The surface electromyography signals input at any time; This represents a one-dimensional convolution operation; This represents the total number of convolution kernels.

5. The state-aware electromyographic signal decoding method according to claim 1, characterized in that, The discriminator module includes a local convolution submodule, a global convolution submodule, and a multi-scale first-level convolution submodule; The step of constraining and optimizing the predicted electromyographic sequence through the discriminator module to generate electromyographic signal motion decoding results includes: The predicted electromyography sequence is input into the local convolutional submodule and the global convolutional submodule, respectively; The local feature vector is obtained by extracting local change features of the predicted electromyography sequence through the local convolution submodule. The global convolutional submodule extracts the global continuity features of the predicted electromyography sequence to obtain a global feature vector; The local feature vector and the global feature vector are concatenated and fused through the multi-scale first-level convolution submodule to generate a discriminative feature vector for evaluating the physiological rationality and smoothness of the predicted electromyography sequence. The output of the state-aware decoding module is constrained and optimized based on the discriminative feature vector to generate the final electromyographic signal motion decoding result.

6. The state-aware electromyographic signal decoding method according to claim 1, characterized in that, The training process of the electromyography signal decoding model includes: The state-aware decoding module includes a dynamic cooperative Mamba decoder. It uses the weights of a pre-trained Whisper model as the initial weights for the dynamic cooperative Mamba decoder and introduces a low-rank residual update mechanism to specialize the weights of the dynamic cooperative Mamba decoder for specific tasks. The expression for the low-rank residual update mechanism is as follows: In the formula, This represents the decoder weight matrix obtained after low-rank residual update. This represents the weights of the pre-trained Whisper model. Denotes the first low-rank matrix. Let represent the second low-rank matrix, where is the first low-rank matrix. The dimension is The second low-rank matrix The dimension is , Indicates the input dimension. Indicates the output dimension. Let represent a low-rank parameter that satisfies . ; The discriminator module and the state-aware decoding module are jointly optimized through adversarial training, which includes: The predicted electromyography (EMG) sequence generated by the state-aware decoding module, and the actual EMG sequence corresponding to the predicted EMG sequence, are input to the discriminator module. The discriminator module performs joint modeling of the input predicted electromyography (EMG) sequence and the actual EMG sequence based on a continuous state space model to generate a discrimination result, which is used to constrain the physiological rationality and smoothness of the predicted EMG sequence. Based on the low-rank residual update mechanism, the weights of the dynamic collaborative Mamba decoder are task-specificated and subjected to adversarial training to obtain an optimized electromyography signal decoding model.

7. The state-aware electromyographic signal decoding method according to claim 6, characterized in that, The expression for the continuous state-space model is: In the formula, The discriminator module indicates that in the first... The hidden state vector at time step 1. Indicates the first The hidden state vector at time step; This represents the time window output by the discriminator module. The discrimination results of predicting electromyographic sequences within the body; Indicates the first The predicted or actual electromyographic sequence is input at any given time. Represents a non-linear activation function. This represents the Sigmoid activation function. This represents the weight matrix of the input layer. This represents the weight matrix of the hidden layer. This represents the weight matrix of the output layer; This indicates that the discriminator module is within the time window. The hidden state vector at time step; Indicates the length of the time window.

8. The state-aware electromyographic signal decoding method according to claim 6, characterized in that, The joint optimization includes a generator loss function. and discriminator loss function ; The generator loss function The expression is: In the formula, The weighting coefficients of the reconstruction loss term are represented. Represents the weighting coefficient of the adversarial loss term; Indicates the first The actual target value at any given time; This represents the predicted electromyographic sequence output by the state-aware decoding module. This represents a continuous sequence of input neuromuscular states. This indicates that the state-aware decoding module is in the first... Predicted electromyographic sequences output at specific times; Indicates the total length of the time series; This indicates that the discriminator module interprets the predicted electromyographic sequences generated by the state-aware decoding module. The discrimination probability; The discriminator loss function The expression is: In the formula, Represents the mathematical expectation. Indicates sample Obeying the distribution of true electromyographic sequences , This represents the actual electromyographic sequence; Indicates sample Follows the distribution of continuous neuromuscular state sequences ; This indicates that the discriminator module recognizes the real electromyography sequence. The discrimination probability.

9. A state-aware electromyographic signal decoding system, characterized in that, The system includes: The acquisition module is used to acquire the surface electromyography signal to be decoded and perform preprocessing. An input module is used to input the preprocessed surface electromyography (EMG) signal into a pre-trained EMG signal decoding model; wherein, the EMG signal decoding model includes a state-aware encoding module, a state-aware decoding module, and a discriminator module; The output module is used to encode the preprocessed surface electromyography signal using the state-aware encoding module to generate a continuous neuromuscular state sequence. The state-aware decoding module processes the continuous neuromuscular state sequence to output a predicted electromyography (EMG) sequence within a time window in parallel. The predicted EMG sequence includes a joint angle sequence and a muscle force sequence. The state-aware decoding module includes a dynamic collaborative Mamba decoder. The continuous neuromuscular state sequence is sampled using the dynamic collaborative Mamba decoder based on an event-adaptive sampling mechanism. Temporal modeling of the sampled continuous neuromuscular state sequence is performed using dynamic collaborative channel state updates to generate electromyographic state features; wherein, the expression for the dynamic collaborative channel state update is: In the formula, Indicates the first Time, Number The real-time state vector of the channel type. Indicates the first Time, Number Historical state vector of the channel type; This represents the internal state transition matrix of the c-th channel. This represents the input mapping matrix for the c-th channel; This represents the feature vector input to the c-th channel at time t. This represents the summation over all channel categories except for channel c. This represents the dynamic collaborative weight matrix that maps the historical state of the c'-th channel to the c-th channel at time t. Indicates the first Time, Number Historical state vector of the channel type; The temporal dependencies of the electromyographic state features are captured using a state propagation mechanism to generate a cooperative state representation. The cooperative state representation is input into a time-windowed parallel prediction unit, which outputs the predicted electromyographic sequence within the time window in parallel. The discriminator module constrains and optimizes the predicted electromyography sequence to generate motion decoding results of the electromyography signal.

Images