Paralyzed lower limb motor intention decoding method based on spatiotemporal decoupling of residual electromyographic signals

Abstract

This invention discloses a method for decoding motor intent in paralyzed lower limbs based on spatiotemporal decoupling of residual electromyography (EMG) signals, belonging to the fields of biosignal processing and rehabilitation engineering. In the temporal dimension, this invention achieves automatic optimization of the number of modes and penalty factors through dual adaptive constraint variational mode decomposition. Combined with motion template correlation constraints, it accurately separates weak motion signals from background noise, improving the signal-to-noise ratio compared to traditional fixed-parameter VMD. In the spatial dimension, it effectively eliminates abnormal muscle coordination components caused by compensation through sparse non-negative matrix decomposition with health priors, reconstructing effective motor coordination patterns and improving feature discriminability. The two dimensions form a progressive decoupling of signal purification and feature correction, far superior to single-dimensional improvements, systematically solving the dual problems of weak residual EMG and coordination disorder in polio.

Description

Technical Field

[0001] This invention relates to the field of biosignal processing and rehabilitation engineering technology, specifically a method for decoding the motor intention of paralyzed lower limbs based on spatiotemporal decoupling of residual electromyographic signals. Background Technology

[0002] Sequelae of poliomyelitis (infantile paralysis) can lead to irreversible paralysis of the lower limb muscles in children, with weak residual muscle contraction and surface electromyography (EMG) signal amplitude typically only 10%–30% of that in healthy individuals. Simultaneously, long-term compensatory movements cause abnormal and disordered muscle coordination patterns, and a large number of compensatory EMG signals can interfere with the extraction of true motor intent. The core of lower limb rehabilitation training equipment is accurate decoding of motor intent, and surface EMG signals, as a direct electrophysiological representation of neuromuscular activity, are the primary input source for decoding motor intent.

[0003] Most existing electromyographic motor intention decoding technologies are designed for healthy individuals or stroke patients, and their direct application to children with polio has the following core limitations:

[0004] Insufficient denoising capability for weak signals: Traditional time-domain filtering, wavelet thresholding and other methods are difficult to effectively separate motion-related signals with extremely low amplitude in a strong noise background; Conventional variational mode decomposition (VMD) requires manual preset of the number of modes and penalty factors, which cannot adapt to the signal differences of different patients and different rehabilitation stages, and is prone to under-decomposition or over-decomposition, resulting in loss of effective information or noise residue.

[0005] Abnormal coordination patterns were not corrected: Conventional muscle coordination analysis directly extracts the coordination matrix from the raw signal without distinguishing between normal motor coordination and compensatory abnormal coordination. The features contain a large amount of redundant compensatory information, causing feature confusion for different motor intentions. Furthermore, the existing non-negative matrix factorization model does not combine physiological priors and sparse constraints, and the extracted coordination patterns do not conform to the laws of neuromuscular control.

[0006] Poor adaptability of fixed thresholds: Traditional decoding models often use fixed classification thresholds. However, during the rehabilitation process of children, neural remodeling occurs, residual muscle function gradually improves, and electromyographic characteristics continue to drift. Fixed thresholds are prone to false triggering or missed triggering, and the accuracy continues to decline with long-term use, requiring repeated manual calibration.

[0007] Current improvements to electromyography decoding for paralyzed lower limbs are mostly focused on single-dimensional noise reduction or feature optimization. They fail to systematically address the dual problems of weak signals and abnormal coordination from the perspective of spatiotemporal two-dimensional joint decoupling, and also lack an adaptive update mechanism to adapt to the neural remodeling process. As a result, they cannot meet the precise decoding needs of children with polio for long-term rehabilitation training.

[0008] Therefore, a new solution is needed to address the above problems. Summary of the Invention

[0009] The purpose of this invention is to provide a method for decoding the motor intent of paralyzed lower limbs based on spatiotemporal decoupling of residual electromyographic signals, so as to solve the technical problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution: a method for decoding the motor intent of a paralyzed lower limb based on spatiotemporal decoupling of residual electromyographic signals, comprising the following steps:

[0011] S1. Collect multi-channel surface electromyography (EMG) signals of the target muscle group of the paralyzed lower limb and obtain the raw EMG matrix after preprocessing;

[0012] S2, Adaptive decoupling in time dimension: Constrained variational mode decomposition with dual adaptive modality number and penalty factor is adopted to separate motion-related signals and background noise from the original electromyography matrix channel by channel, and the highly correlated modal components are selected to reconstruct the motion-related electromyography sequence;

[0013] S3. Spatial Dimension Collaborative Reconstruction and Decoupling: Using the standard muscle synergy basis matrix of healthy individuals as a prior template, a non-negative matrix decomposition model with prior constraints and L1 sparse regularization is constructed. The muscle synergy basis and activation coefficient matrix are solved, and abnormal compensatory synergy components are removed to obtain the corrected muscle synergy activation matrix.

[0014] S4. Dynamic Threshold Motion Intent Decoding: Based on the corrected muscle coactivation matrix, spatiotemporal fusion features are extracted, a probability discrimination model is constructed through kernel density estimation, and the discrimination threshold is dynamically adjusted in combination with real-time signal fluctuations to output the lower limb motion intent decoding results.

[0015] S5. Adaptive Update of Neural Remodeling: Periodically assess the drift of electromyographic features and decoding confidence, and automatically update the temporal decoupling parameters, spatial reconstruction parameters and dynamic threshold model parameters to adapt to the neural remodeling process of the child.

[0016] Further, S2 includes the following steps:

[0017] S21. Parameter Initialization: Calculate the initial signal-to-noise ratio (SNR) 0 of the original single-channel signal according to the formula. The initial value of the secondary penalty factor is adaptively set. The lower the signal-to-noise ratio, the larger the penalty factor, thereby improving the frequency resolution capability.

[0018] Initialize the number of modes Set energy threshold Frequency band overlap threshold ;

[0019] S22. Construct a variational model with correlation constraints, with the objective function as follows:

[0020]

[0021] Constraints:

[0022]

[0023] In the formula, This is a single-channel raw electromyography signal. For the first One intrinsic mode component For the first The center frequency of each mode; As a pre-calibrated reference template for movement-evoked electromyography, Let Pearson's cross-correlation coefficient be used. These are the modal weighting coefficients. The threshold for overall correlation constraint;

[0024] S23. The constrained problem is transformed into an unconstrained point problem by using the Alternating Direction Multiplier Method (ADMM), and the modal components and center frequencies are iteratively updated until convergence is obtained. All modal components under the value;

[0025] S24. Adaptive determination of the optimal number of modes: Calculation of the remaining signal The ratio of the energy of the original signal to the total energy of the signal. and the relative interval between the center frequencies of adjacent modes. .like or ,but Self-incrementing Return to step S22 to continue decomposition; if both conditions are met simultaneously, stop the iteration and determine the optimal number of modes. ;

[0026] S25. Modal screening and reconstruction: Calculate the cross-correlation coefficient with the reference template component by component, and retain coefficients greater than the correlation threshold. The modal components are superimposed and reconstructed to form a single-channel motion-related electromyography (EMG) signal; after processing all channels, they are combined to form a motion-related EMG matrix. .

[0027] Further, S3 includes the following steps:

[0028] S31. Construction of a Health Synergy Prior Template: Multi-channel electromyography data of the lower limbs of healthy individuals are collected in advance, and a standard muscle synergy basis matrix is ​​extracted through standard nonnegative matrix decomposition. As a priori template, in The muscle synergy dimension is determined by a cumulative variance contribution rate ≥ 90%.

[0029] S32. Construct a nonnegative matrix factorization objective function with prior constraints and sparse regularization:

[0030]

[0031] In the formula, the first term is the reconstruction error term, which measures the fitting accuracy of matrix decomposition; the second term is the prior constraint term, which guides the coordinating basis to move closer to the healthy and normal movement pattern and suppresses the abnormal compensatory mode; the third term is the L1 sparse regularization term, which conforms to the physiological characteristic that "only a few coordinating modules are activated in a single movement" and at the same time improves the distinguishability of the coordinating mode. These are the prior constraint weighting coefficients. These are the sparse regularization coefficients;

[0032] S33. Solve the above objective function using the multiplicative iteration rule. The iteration formula is as follows:

[0033]

[0034] After iterating until the objective function converges, the optimized collaborative basis matrix is ​​obtained. With activation coefficient matrix ;

[0035] S34. Abnormal Cooperative Component Removal: Calculate the cosine similarity between the cooperative basis vector and the standard cooperative basis vector column by column, and filter out components with a similarity greater than a set threshold. By identifying the effective co-activation components and removing the abnormal co-activation components corresponding to the compensatory patterns, the corrected muscle co-activation matrix is ​​finally obtained. .

[0036] Further, S4 includes the following steps:

[0037] S41. Feature Extraction: Set a sliding time window and step size, and extract features within each time window. Extract time-domain features (mean, root mean square, waveform length) and frequency-domain features (average power frequency, median frequency), and concatenate them to form a spatiotemporal fusion feature vector. ;

[0038] S42. Probability Density Fitting: For each type of motion intention training sample, a Gaussian kernel density estimation is used to fit the probability density function of the corresponding feature vector.

[0039]

[0040] In the formula, The number of training samples for this type. For core bandwidth, The Gaussian kernel function;

[0041] S43. Dynamic Threshold Determination: Calculating the current feature vector based on Bayes' theorem. Belongs to the Posterior probability of motion intention Set adaptive discrimination threshold:

[0042]

[0043] In the formula, As the initial baseline threshold, The variance of the feature vectors within the current sliding window. This is the threshold adjustment coefficient. The greater the signal fluctuation and the higher the variance, the higher the threshold should be to suppress false triggering due to noise; when the signal is stable, the threshold should be lowered to reduce missed transmissions.

[0044] When the maximum posterior probability is greater than the dynamic threshold of the corresponding category, the corresponding motion intention is output; otherwise, it is determined to be a static state.

[0045] Further, S5 includes the following steps:

[0046] S51, Index Calculation: Based on length... The sliding time window is used as the update period to calculate the average electromyographic amplitude within the current window. Average amplitude of the reference period Relative drift:

[0047]

[0048] Simultaneously calculate the average decoding confidence within the current window. The confidence level is defined as the mean of the maximum posterior probability corresponding to the decoded output.

[0049] S52. Update trigger mechanism: Set drift threshold With confidence threshold ,when and When this occurs, a model update is triggered: the correlation screening threshold for time decoupling is re-optimized. With penalty factor The prior constraint weights and sparse coefficients of spatial reconstruction are determined, and the kernel density estimation model and dynamic threshold parameters are refitted based on the newly added effective samples.

[0050] S53, Benchmark Refresh: After the update is completed, the electromyographic features and decoding results of the current cycle are stored in the benchmark library, and the benchmark amplitude is updated. This serves as a reference benchmark for the next update, enabling continuous iterative optimization of the model as the neural remodeling process of the child progresses.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] 1. In the time dimension, this invention achieves automatic optimization of the number of modes and penalty factors through dual adaptive constraint variational mode decomposition. Combined with motion template correlation constraints, it accurately separates weak motion signals from background noise, improving the signal-to-noise ratio compared to traditional fixed-parameter VMD. In the spatial dimension, it effectively eliminates abnormal muscle coordination components caused by compensation through sparse non-negative matrix decomposition with health prior, reconstructs effective motion coordination patterns, and improves feature discrimination. The two form a progressive decoupling of signal purification and feature correction, which is far superior to single-dimensional improvement and systematically solves the dual problems of weak residual electromyography and coordination disorder in polio patients.

[0053] 2. This invention introduces standard muscle coordination of healthy individuals as a priori into the NMF model, and adds L1 sparse regularization constraints, which not only corrects the abnormal compensation patterns of paralyzed patients, but also ensures that the extracted coordination patterns conform to the physiological sparsity of neuromuscular control. This overcomes the defects of strong randomness and susceptibility to noise interference in the extraction results of traditional unconstrained NMF, and significantly improves the interpretability and stability of the coordination patterns.

[0054] 3. This invention uses a dual-dimensional approach of feature drift and decoding confidence to automatically perceive the child's rehabilitation progress and update the parameters of the entire model, eliminating the need for repeated manual calibration. Attached Figure Description

[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the 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.

[0056] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0057] Figure 2 This is a schematic diagram illustrating the principle of muscle collaborative reconstruction with prior constraints in the spatial dimension of the present invention.

[0058] Figure 3 This is a schematic diagram illustrating the principle of muscle collaborative reconstruction with prior constraints in the spatial dimension of the present invention.

[0059] Figure 4 This is a bar chart comparing the decoding performance of the present invention. Detailed Implementation

[0060] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0061] Please see Figures 1-3A method for decoding motor intent of paralyzed lower limbs based on spatiotemporal decoupling of residual electromyographic signals includes the following steps:

[0062] S1. Acquisition and preprocessing of multi-channel residual electromyography signals:

[0063] Multichannel surface electromyography (EMG) signals of the target muscle groups in the paralyzed lower limb were acquired using a surface EMG electrode array. These signals were then sequentially subjected to 50Hz power frequency notch filtering, 20–500Hz bandpass filtering, and baseline drift correction to obtain a discretized raw EMG matrix. ,in This represents the number of electromyographic channels. This represents the total number of sampling points.

[0064] S2. Adaptive Decoupling in Time Dimension: The separation of motion signal and noise adopts constrained variational mode decomposition (AC-VMD) with dual adaptive mode number and penalty factor to decompose the electromyographic signal of each channel in the time domain. Based on the traditional VMD, motion template correlation constraints are introduced and the parameters are adaptive to ensure that the decomposed modes effectively retain the motion-induced features. Highly correlated mode components are screened by cross-correlation coefficient to reconstruct the denoised motion-related electromyographic sequence.

[0065] S3. Spatial Dimension Coordination Reconstruction and Decoupling: Abnormal Compensation Mode Elimination. For abnormal coordination modes caused by muscle compensation in children, non-negative matrix factorization with prior constraints and L1 sparse regularization (PC-SNMF) is used to spatially decouple the motor-related electromyography matrix. Standard muscle coordination of healthy individuals is introduced as a prior guide. Combined with the physiological sparsity constraints of muscle activation, the normal motor coordination component and the abnormal compensation coordination component are separated, and the corrected muscle coordination activation matrix is ​​reconstructed.

[0066] S4. Dynamic threshold motion intent decoding extracts multi-dimensional spatiotemporal fusion features based on the corrected coactivation matrix, constructs a dynamic threshold discrimination model based on kernel density estimation, adaptively adjusts the discrimination threshold according to real-time signal fluctuations, and outputs the lower limb motion intent decoding results.

[0067] S5. Adaptive update of the neural remodeling model: Periodically collect electromyographic signals and decoding feedback results within a set time window, calculate the electromyographic feature drift and decoding confidence, automatically sense the progress of neural remodeling in the child, trigger model parameter updates, and achieve long-term adaptive optimization of the decoding model.

[0068] The specific process of the dual adaptive constraint variational mode decomposition in S2 is as follows:

[0069] S21. Initialize the number of modes K=2, and adaptively set the initial value of the secondary penalty factor based on the initial signal-to-noise ratio of the original electromyographic signal.

[0070] S22. Construct a variational objective function with motion template correlation constraints, and solve it using the alternating direction multiplier method to obtain K intrinsic mode components;

[0071] S23. Calculate the ratio of the remaining signal energy to the total energy of the original signal, and the relative interval of the center frequencies of adjacent modes; when the ratio of the remaining energy is greater than the energy threshold, or the overlap of the frequency bands of adjacent modes is greater than the overlap threshold, increment K by 1 and return to step S22; when both conditions are met, stop the iteration and determine the optimal number of modes K*.

[0072] S24. Calculate the cross-correlation coefficient between each modal component and the motion-evoked electromyography reference template, filter and retain modal components with correlation coefficients greater than the correlation threshold, and superimpose and reconstruct to obtain single-channel motion-related electromyography signals. All channels are combined to form a motion-related electromyography matrix.

[0073] The variational objective function with correlation constraints is:

[0074]

[0075] st ,

[0076] in, For the first One modal component, For the first The center frequency of each mode This is a single-channel raw electromyography signal. The unit impulse function, This is a convolution operation; As a reference template for exercise-induced electromyography, Let Pearson's cross-correlation number be denoted as . These are the modal weighting coefficients. The threshold is the correlation constraint threshold.

[0077] Secondary penalty factor The adaptive calculation formula is:

[0078]

[0079] in, As the benchmark penalty factor, The initial signal-to-noise ratio of the original signal. This is the signal-to-noise ratio adjustment coefficient.

[0080] The objective function for nonnegative matrix factorization with prior constraints and sparse regularization in S3 is:

[0081]

[0082] in, For motor-related electromyography (EMG) matrix, This represents the number of electromyographic channels. Number of sampling points; For muscle coordination basis matrix, For the collaborative dimension; This is the co-activation coefficient matrix; A standard muscle synergistic prior template for healthy individuals; These are the prior constraint weighting coefficients. These are the sparse regularization coefficients; It is the Frobenius norm. It is an L1 norm.

[0083] The objective function is solved using a multiplicative iteration rule, and the iteration formula is as follows:

[0084]

[0085]

[0086] After iterating until the objective function converges, the cosine similarity between the cooperating basis vector and the standard cooperating basis vector is calculated column by column. Valid cooperating components with similarity greater than a set threshold are selected, and abnormal compensatory cooperating components are removed to obtain the corrected muscle cooperating activation matrix.

[0087] S4 specifically includes:

[0088] The temporal and frequency domain features are extracted from the corrected muscle coactivation matrix by sliding time windows and concatenated to form a spatiotemporal fusion feature vector. The probability density function of the feature corresponding to various motion intentions is fitted by Gaussian kernel density estimation, and the posterior probability of the current feature corresponding to various motions is calculated based on Bayes' theorem.

[0089] The formula for calculating the dynamic discrimination threshold is:

[0090]

[0091] in, As the initial baseline threshold, The variance of the feature vectors within the current sliding window. This is the threshold adjustment coefficient. For the first similar movements The dynamic discrimination threshold at any given time; when the maximum posterior probability exceeds the dynamic threshold of the corresponding category, the corresponding motion intention is output.

[0092] S5 specifically includes: using a sliding time window of a set length as the update period, calculating the relative drift of the average electromyographic amplitude relative to the baseline period within the current period, as well as the average decoding confidence; when the relative drift is greater than or equal to the drift threshold and the average decoding confidence is less than or equal to the confidence threshold, triggering a model update: re-optimizing the correlation threshold and penalty factor for time decoupling, the prior constraint weights and sparsity coefficients for spatial reconstruction, and the kernel density estimation model and dynamic threshold parameters; after the update is completed, refreshing the baseline amplitude and entering the next period.

[0093] The method for obtaining the exercise-evoked electromyography reference template is as follows:

[0094] During the initial calibration phase, guide the child to complete the maximum voluntary contraction of each target movement, with each contraction lasting 3-5 seconds, and repeat 3 times.

[0095] The middle stable segment of each contraction is extracted, and after the same preprocessing, it is synchronously averaged in the time domain to generate a single-channel reference template for the corresponding action.

[0096] After taking the union of all motion templates and performing normalization, a universal motion-evoked electromyography reference template is obtained. .

[0097] The method for obtaining the standard muscle synergetic basis prior template for healthy individuals is as follows: Healthy subjects matched to the target child's age group are recruited. Inclusion criteria include the absence of neuromuscular diseases, lower limb muscle strength grade 5, and no motor dysfunction. Multichannel electromyography data from the same site and under the same movement paradigm are collected, and individual muscle synergetic basis is extracted using standard nonnegative matrix decomposition. Synergetic basis matching and alignment of all individuals is performed using spectral clustering, and the within-group mean is taken to obtain the standardized healthy synergetic basis matrix. The collaborative dimension is determined by the cumulative variance contribution rate being ≥90%.

[0098] See Figure 4 Based on the above embodiments, the following experimental verification is proposed:

[0099] 1.1 Experimental Platform and Hardware System Setup

[0100] This embodiment establishes a complete rehabilitation verification system for paralyzed lower limbs, consisting of four main units, with variables controlled throughout to ensure experimental reproducibility:

[0101] Signal acquisition unit: Employs a medical-grade 16-channel surface electromyography (EMG) acquisition system. The analog front-end uses an ADS1299 chip with a sampling rate of 1000Hz, an analog-to-digital conversion resolution of 16 bits, and an input impedance ≥100MΩ. It is paired with disposable Ag / AgCl disc electrodes, with electrode placement strictly adhering to SENIAM international standards and an electrode spacing of 20mm. Simultaneously, an optical motion capture system (sampling frequency 200Hz) is used to acquire ankle and knee joint angle data via reflective balls attached to bony landmarks of the lower limbs, serving as the gold standard for motion intent annotation.

[0102] Main control processing unit: Equipped with an ARM Cortex-A72 embedded processor and an FPGA coprocessor, running a real-time operating system, and deploying the spatiotemporal dual decoupling decoding algorithm of this invention; the actual measured decoding latency of a single frame of 200ms window data is ≤42ms, which meets the latency requirements of real-time rehabilitation equipment.

[0103] Rehabilitation execution unit: It adopts a unilateral lower limb rehabilitation exoskeleton, covering the hip, knee and ankle joints, and is equipped with a DC servo motor and torque sensor, with a maximum output torque of 50 N·m; it is equipped with an 8-channel functional electrical stimulation (FES) module, which outputs constant current pulses with a current range of 0~100mA and a pulse width of 200μs, and can target and stimulate the target muscle groups according to the decoding results.

[0104] Motion restraint and feedback unit: Equipped with an adjustable lower limb fixation brace to fix the child's trunk and proximal thigh, restricting trunk compensation and hip joint synkinesis; set up screen visual feedback to display decoding results and target movements in real time.

[0105] 1.2 Subjects and Experimental Paradigm

[0106] Inclusion criteria for subjects: Multiple children with unilateral lower limb paralysis due to polio sequelae were selected, with manual muscle strength grade of 2-3 in the affected lower limb, and all of them had varying degrees of muscle atrophy and compensatory movement patterns; all guardians of the children signed informed consent forms, and the experimental protocol was approved by the institution's ethics committee.

[0107] Healthy control group: 15 healthy children of the same age, without neuromuscular diseases and with normal lower limb muscle strength were recruited to construct a healthy collaborative prior template.

[0108] Experimental Paradigm: The experiment was conducted in a temperature-controlled, electromagnetically shielded rehabilitation room. The child was seated with the affected lower limb fixed to a brace. The experiment included five modes: resting, ankle dorsiflexion, ankle plantarflexion, knee flexion, and knee extension. Each movement task was completed autonomously by the child following visual cues on a screen. Each muscle contraction lasted 3 seconds, followed by a 5-second rest period. Each movement was repeated 40 times, and 200 valid movement samples were collected from each subject. The experiment was conducted in four phases, each 7 days apart, to verify the long-term adaptive update effect.

[0109] 1.3 Reference Template and Health Collaboration Template Calibration

[0110] (1) Obtaining reference templates for exercise-induced electromyography

[0111] Calibration was performed before the first experiment: Children were guided to sequentially perform maximal voluntary contractions (MVCs) of five different movements, each lasting 4 seconds, repeated three times with a 2-minute rest interval to avoid fatigue. A 2-second stable segment was extracted from the middle of each contraction. After preprocessing as in S1, the data from the three repetitions were simultaneously averaged in the time domain to obtain a single-channel MVC template for each movement. The five movement templates were then normalized by energy and their union was used to generate a universal motion-evoked electromyography (EMG) reference template. .

[0112] (2) Obtaining standard muscle synergist templates from healthy individuals

[0113] Six-channel electromyography (EMG) data were collected from 15 healthy children, using the same body parts and movement paradigms. Thirty effective movement cycles were extracted for each movement type. Standard normalized matrix factorization (NMF) was used to perform co-factor decomposition on all healthy data. The co-factor dimension was gradually increased starting at M=2, and the optimal co-factor dimension M=4 was determined when the cumulative variance contribution rate reached 90%. A spectral clustering algorithm was used to match and align the co-factor vectors of all individuals, eliminating individual order differences, and the standardized health co-factor matrix was obtained by calculating the within-group mean. , as a priori template.

[0114] 1.4 Complete Execution Flow of Spatiotemporal Dual Decoupling Decoding

[0115] Signal preprocessing: The raw electromyography (EMG) signal is first subjected to bad segment removal, removing large motion artifact segments with amplitudes exceeding ±5mV; then, 50Hz second-order IIR power frequency notch filtering, 20-500Hz fourth-order Butterworth bandpass filtering, and third-order polynomial baseline correction are performed sequentially to obtain a 6-channel raw EMG matrix, which is then slid segmented according to a 200ms time window and a 50ms step size.

[0116] Time-dimensional adaptive decoupling: setting a baseline penalty factor Signal-to-noise ratio adjustment coefficient Energy threshold Frequency band overlap threshold Relevant screening thresholds For each channel signal, a dual adaptive constraint VMD decomposition is performed window by window to automatically determine the optimal number of modes (the optimal K for measured children's data is concentrated in 4~6), and the highly correlated modes are screened to reconstruct the motion-related electromyography matrix.

[0117] Spatial Dimension Collaborative Reconstruction: Setting Prior Constraint Weights sparse regularization coefficient Initial Collaboration Dimension PC-SNMF decomposition was performed using a multiplicative iterative rule, iterating 150 times until convergence. The cosine similarity between the cooperative basis and the standard basis was calculated, and a similarity threshold of 0.6 was set to select 3-4 effective normal cooperative components and remove abnormal compensatory components to obtain the corrected cooperative activation matrix.

[0118] Dynamic threshold decoding: Nine-dimensional spatiotemporal fusion features (3 co-components × 3 types of temporal features) are extracted for each time window. Gaussian kernel density estimation is used to fit the probability distribution of the five modes, and the kernel bandwidth is optimized to 0.15 using five-fold cross-validation. An initial baseline threshold is set. Threshold adjustment coefficient It calculates dynamic thresholds in real time and outputs decoding results.

[0119] Adaptive Updates: Set the update cycle Heaven, drift threshold Confidence threshold At the end of each cycle, the drift and confidence level are evaluated. If the conditions are met, a full parameter update is triggered, refreshing the baseline value.

[0120] Controlled experimental design and effect verification

[0121] To verify the inventive contribution of each module of the present invention and the synergistic effect of spatiotemporal decoupling, five control schemes were set up, using the same data acquisition, preprocessing procedures and experimental environment to ensure the uniqueness of variables.

[0122] 2.1 Control group setup

[0123] Control group 1 (traditional baseline protocol): wavelet soft threshold denoising (db4 wavelet, 5-level decomposition) + temporal statistical features + SVM fixed threshold classification, representing the current mainstream clinical protocol.

[0124] Control group 2 (temporal decoupling only): Fixed parameter standard VMD denoising + original temporal features + SVM fixed threshold classification, to verify the gain of single temporal denoising.

[0125] Control group 3 (spatial decoupling only): wavelet denoising + standard unconstrained NMF collaborative features + SVM fixed threshold classification, to verify the gain of single spatial collaborative decomposition.

[0126] Control group 4 (spatiotemporal decoupling without adaptation): Features are extracted using the spatiotemporal dual decoupling algorithm of this invention, combined with SVM fixed threshold classification, without dynamic threshold and periodic updates, and the incremental verification of dynamic threshold and adaptive module is separated.

[0127] Experimental group (complete method of this invention): dual adaptive spatiotemporal decoupling + dynamic threshold + neural remodeling adaptive update.

[0128] 2.2 Evaluation Index System

[0129] Quantitative indicators are set from three dimensions: signal quality, decoding accuracy, and long-term adaptability.

[0130] Signal denoising effect: signal-to-noise ratio (SNR), mean square error (MSE);

[0131] Real-time decoding performance: overall accuracy, false trigger rate, missed trigger rate, F1 score;

[0132] Long-term compatibility performance: The rate of accuracy decay in week 4 relative to week 1.

[0133] 2.3 Experimental Results and Comparative Analysis

[0134] (1) Comparison of signal denoising effects

[0135]

[0136] Results analysis: The signal-to-noise ratio of the experimental group was improved by 93% compared with the traditional baseline and by 35% compared with temporal decoupling only. Moreover, the signal-to-noise ratio improvement of temporal and spatial decoupling (5.92dB) was greater than the sum of the gains of temporal decoupling only (2.92dB) and spatial decoupling only (0.21dB), proving that temporal denoising provides high-quality input for spatial reconstruction, and the two produce synergistic gains rather than simple superposition.

[0137] (2) Comparison of real-time decoding performance

[0138]

[0139] Results analysis:

[0140] Temporal decoupling alone and spatial decoupling alone bring accuracy improvements of 5.57% and 6.89% respectively, while the combination of temporal and spatial decoupling brings an improvement of 14.16%, which is significantly greater than the sum of the gains of the two. This verifies the synergistic effect of the "temporal purification + spatial correction" dual decoupling architecture and achieves a technical effect of 1+1>2.

[0141] Compared with the control group, the experimental group achieved a further 4.75 percentage point increase in accuracy and a 41.8% decrease in false trigger rate, proving that the dynamic threshold mechanism effectively solved the misjudgment problem of fixed threshold.

[0142] The experimental group achieved an overall accuracy improvement of 18.91 percentage points compared to the traditional baseline, a 68.1% reduction in false trigger rate, and a 70.5% reduction in missed trigger rate, demonstrating a significant improvement in overall performance.

[0143] (3) Comparison of long-term adaptability performance

[0144]

[0145] Results analysis: The accuracy decay rate of the experimental group was only 2.08%, which was much lower than that of the other control groups. This proves that the adaptive update mechanism of the present invention can effectively track the electromyographic feature drift caused by neural remodeling in children, maintain high decoding accuracy, and greatly improve the applicability of long-term rehabilitation training.

[0146] The above comparative experiments fully demonstrate that the spatiotemporal dual decoupling architecture proposed in this invention is not a simple combination of two algorithms, but rather produces significant synergistic technical effects through progressive decoupling of "signal-level denoising - feature-level correction". Combining four major innovations—adaptive parameter optimization, prior-guided sparse reconstruction, dynamic threshold discrimination, and neural remodeling update—it systematically solves the industry pain points of decoding residual electromyography in polio patients, and has outstanding substantive features and significant progress.

[0147] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for decoding motor intent of paralyzed lower limbs based on spatiotemporal decoupling of residual electromyographic signals, characterized in that: Includes the following steps: S1. Collect multi-channel surface electromyography (EMG) signals of the target muscle group of the paralyzed lower limb and obtain the raw EMG matrix after preprocessing; S2, Adaptive decoupling in time dimension: Constrained variational mode decomposition with dual adaptive modality number and penalty factor is adopted to separate motion-related signals and background noise from the original electromyography matrix channel by channel, and the highly correlated modal components are selected to reconstruct the motion-related electromyography sequence; S3. Spatial Dimension Collaborative Reconstruction and Decoupling: Using the standard muscle synergy basis matrix of healthy individuals as a prior template, a non-negative matrix decomposition model with prior constraints and L1 sparse regularization is constructed. The muscle synergy basis and activation coefficient matrix are solved, and abnormal compensatory synergy components are removed to obtain the corrected muscle synergy activation matrix. S4. Dynamic Threshold Motion Intent Decoding: Based on the corrected muscle coactivation matrix, spatiotemporal fusion features are extracted, a probability discrimination model is constructed through kernel density estimation, and the discrimination threshold is dynamically adjusted in combination with real-time signal fluctuations to output the lower limb motion intent decoding results. S5. Adaptive Update of Neural Remodeling: Periodically assess the drift of electromyographic features and decoding confidence, and automatically update the temporal decoupling parameters, spatial reconstruction parameters and dynamic threshold model parameters to adapt to the neural remodeling process of the child.

2. The method for decoding the motor intent of a paralyzed lower limb based on spatiotemporal decoupling of residual electromyographic signals according to claim 1, characterized in that: S2 includes the following steps: S21. Parameter Initialization: Calculate the initial signal-to-noise ratio (SNR) 0 of the original single-channel signal according to the formula. The initial value of the secondary penalty factor is adaptively set. The lower the signal-to-noise ratio, the larger the penalty factor, thereby improving the frequency resolution capability. Initialize the number of modes Set energy threshold Frequency band overlap threshold ; S22. Construct a variational model with correlation constraints, with the objective function as follows: ， Constraints: ， In the formula, This is a single-channel raw electromyography signal. For the first One intrinsic mode component For the first The center frequency of each mode; As a pre-calibrated reference template for movement-evoked electromyography, Let Pearson's cross-correlation coefficient be used. These are the modal weighting coefficients. The threshold for overall correlation constraint; S23. The constrained problem is transformed into an unconstrained point problem by using the Alternating Direction Multiplier Method (ADMM), and the modal components and center frequencies are iteratively updated until convergence is obtained. All modal components under the value; S24. Adaptive determination of the optimal number of modes: Calculation of the remaining signal The ratio of the energy of the original signal to the total energy of the signal. and the relative interval between the center frequencies of adjacent modes. .like or ,but Self-incrementing Return to step S22 to continue decomposition; if both conditions are met simultaneously, stop the iteration and determine the optimal number of modes. ; S25. Modal screening and reconstruction: Calculate the cross-correlation coefficient with the reference template component by component, and retain coefficients greater than the correlation threshold. The modal components are superimposed and reconstructed to form a single-channel motion-related electromyography (EMG) signal; after processing all channels, they are combined to form a motion-related EMG matrix. .

3. The method for decoding the motor intent of a paralyzed lower limb based on spatiotemporal decoupling of residual electromyographic signals according to claim 2, characterized in that: S3 includes the following steps: S31. Construction of a Health Synergy Prior Template: Multi-channel electromyography data of the lower limbs of healthy individuals are collected in advance, and a standard muscle synergy basis matrix is ​​extracted through standard nonnegative matrix decomposition. As a priori template, in The muscle synergy dimension is determined by a cumulative variance contribution rate ≥ 90%. S32. Construct a nonnegative matrix factorization objective function with prior constraints and sparse regularization: ， In the formula, the first term is the reconstruction error term, which measures the fitting accuracy of matrix decomposition; the second term is the prior constraint term, which guides the coordinating basis to move closer to the healthy and normal movement pattern and suppresses the abnormal compensatory mode; the third term is the L1 sparse regularization term, which conforms to the physiological characteristic of "only a few coordinating modules are activated in a single movement" and at the same time improves the distinguishability of the coordinating mode. These are the prior constraint weighting coefficients. These are the sparse regularization coefficients; S33. Solve the above objective function using the multiplicative iteration rule. The iteration formula is as follows: ， After iterating until the objective function converges, the optimized collaborative basis matrix is ​​obtained. With activation coefficient matrix ; S34. Abnormal Cooperative Component Removal: Calculate the cosine similarity between the cooperative basis vector and the standard cooperative basis vector column by column, and filter out components with a similarity greater than a set threshold. By identifying the effective co-activation components and removing the abnormal co-activation components corresponding to the compensatory patterns, the corrected muscle co-activation matrix is ​​finally obtained. .

4. The method for decoding the motor intent of a paralyzed lower limb based on spatiotemporal decoupling of residual electromyographic signals according to claim 3, characterized in that: S4 includes the following steps: S41. Feature Extraction: Set a sliding time window and step size, and extract features within each time window. Extract time-domain features (mean, root mean square, waveform length) and frequency-domain features (average power frequency, median frequency), and concatenate them to form a spatiotemporal fusion feature vector. ; S42. Probability Density Fitting: For each type of motion intention training sample, a Gaussian kernel density estimation is used to fit the probability density function of the corresponding feature vector. ， In the formula, The number of training samples for this type. For core bandwidth, The Gaussian kernel function; S43. Dynamic Threshold Determination: Calculating the current feature vector based on Bayes' theorem. Belongs to the Posterior probability of motion intention Set adaptive discrimination threshold: ， In the formula, As the initial baseline threshold, The variance of the feature vectors within the current sliding window. This is the threshold adjustment coefficient. The greater the signal fluctuation and the higher the variance, the higher the threshold should be to suppress false triggering due to noise; when the signal is stable, the threshold should be lowered to reduce missed transmissions. When the maximum posterior probability is greater than the dynamic threshold of the corresponding category, the corresponding motion intention is output; otherwise, it is determined to be a static state.

5. The method for decoding the motor intent of a paralyzed lower limb based on spatiotemporal decoupling of residual electromyographic signals according to claim 4, characterized in that: S5 includes the following steps: S51, Index Calculation: Based on length... The sliding time window is used as the update period to calculate the average electromyographic amplitude within the current window. Average amplitude of the reference period Relative drift: ， Simultaneously calculate the average decoding confidence within the current window. The confidence level is defined as the mean of the maximum posterior probability corresponding to the decoded output. S52. Update trigger mechanism: Set drift threshold With confidence threshold ,when and When this occurs, a model update is triggered: the correlation screening threshold for time decoupling is re-optimized. With penalty factor The prior constraint weights and sparse coefficients of spatial reconstruction are determined, and the kernel density estimation model and dynamic threshold parameters are refitted based on the newly added effective samples. S53, Benchmark Refresh: After the update is completed, the electromyographic features and decoding results of the current cycle are stored in the benchmark library, and the benchmark amplitude is updated. This serves as a reference benchmark for the next update, enabling continuous iterative optimization of the model as the neural remodeling process of the child progresses.

Images