A method and system for dynamic muscle force prediction using adaptive pre-processing of surface electromyography

Abstract

The present application relates to the field of artificial intelligence signal processing technology, specifically to a kind of self-adapting pre-processing surface myoelectric dynamic muscle force prediction method and system, the method includes the following steps: S1, signal acquisition and segmentation;S2, signal quality evaluation;S3, self-adapting pre-processing;S4, multi-domain feature fusion;S5, motion state recognition;S6, fatigue index construction and correction;S7, dynamic model fusion;S8, muscle force compensation output.The present application is driven by signal quality evaluation index adaptive pre-processing mechanism, according to the motion state and noise type dynamic adjustment band-pass filter cutoff frequency and denoising strategy, form pre-processing effect closed-loop feedback, both retain low-frequency effective myoelectric component in static scene, also effectively inhibit electrode slip artifact in dynamic movement, power frequency interference and non-stationary noise, solve the problem that fixed parameters are difficult to consider multiple scenarios.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence signal processing technology, specifically to an adaptive preprocessing method and system for predicting surface electromyography dynamic muscle strength. Background Technology

[0002] Surface electromyography (sEMG) is a bioelectrical signal that reflects the process of human muscle activity, collected by electrodes attached to the skin surface. It can characterize the activation state and contractile ability of the neuromuscular system in real time and is widely used in rehabilitation assessment, motion analysis, exoskeleton control and human-machine collaboration.

[0003] In existing technologies, muscle strength prediction typically involves extracting time-domain or frequency-domain features of surface electromyography (EMG) signals and combining them with regression models, support vector machines, or neural networks to establish a mapping relationship between EMG signals and muscle strength. This allows for muscle strength estimation under different motion scenarios using a fixed model structure and a fixed signal processing procedure.

[0004] However, in real-world dynamic motion scenarios, existing technologies have significant shortcomings:

[0005] Firstly, fixed bandpass filtering, fixed notch filtering, and fixed normalization are usually used. However, in dynamic motion, electrode displacement, muscle deformation, skin stretching, and changes in sweat can lead to a significant enhancement of low-frequency artifacts. Fixed parameters cannot simultaneously ensure the preservation of static effective components and the suppression of dynamic artifacts.

[0006] Secondly, a single prediction model is difficult to cover different states such as static isometric contraction, periodic dynamic motion, explosive rapid motion and fatigue sustained motion. The electromyography-muscle strength mapping relationship differs in various states, and the model is prone to prediction drift when switching states.

[0007] Third, muscle fatigue factors are usually only used as the output of classification results and do not participate deeply in the closed loop of state recognition correction, model weight adjustment and muscle strength compensation. As exercise time increases, the electromyographic signal spectrum drifts, resulting in the accumulation of prediction errors and insufficient long-term stability.

[0008] To address the aforementioned problems, we propose an adaptive preprocessing method and system for predicting dynamic muscle strength using surface electromyography. Summary of the Invention

[0009] The purpose of this invention is to provide an adaptive preprocessing method and system for predicting surface electromyography dynamic muscle strength, in order to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] An adaptive preprocessing method for predicting surface electromyography dynamic muscle strength includes the following steps:

[0012] S1. Signal Acquisition and Segmentation: Acquire multi-channel surface electromyography (EMG) signals of the target muscle group and segment the multi-channel EMG signals according to a preset sliding time window;

[0013] S2. Signal Quality Assessment: Calculate signal quality evaluation indicators for surface electromyography signals within each sliding time window. , It is obtained at least by weighting the effective signal-to-noise ratio, the correlation between adjacent channels, the stability of the root mean square value, and the proportion of power frequency noise energy.

[0014] S3, Adaptive Preprocessing: Based on And the initial motion state probability, adaptively adjust the signal preprocessing parameters, where, when When characterizing the enhancement of low-frequency motion artifacts, increase the low-frequency cutoff frequency of the bandpass filter; when When characterizing increased power frequency noise, enable notch filtering; when When characterizing artifacts of non-stationary motion, wavelet thresholding denoising or empirical mode decomposition reconstruction should be enabled;

[0015] S4. Multi-domain feature fusion: Extract the original electromyographic features, first-order differential electromyographic features, second-order differential electromyographic features, and third-order differential electromyographic features from the preprocessed surface electromyographic signal, and form a fused feature vector including time-domain features, frequency-domain features, time-frequency domain features, and nonlinear features;

[0016] S5. Motion State Recognition: Input the fused feature vector into the motion state recognition model, and output the state probability vectors corresponding to static isometric contraction state, periodic dynamic motion state, explosive rapid motion state, and fatigue continuous motion state;

[0017] S6. Construction and Correction of Fatigue Index: Construct a composite fatigue index based on the decrease in median frequency, the change in root mean square value, and sample entropy. And based on the composite fatigue index The probability of fatigue-prone motion states is corrected to obtain the corrected state probability vector.

[0018] S7. Dynamic Model Fusion: Based on the corrected state probability vector, and composite fatigue index The fusion weights of multiple muscle strength prediction models are determined, wherein the multiple muscle strength prediction models include at least a static muscle strength prediction model, a periodic dynamic muscle strength prediction model, a burst-speed muscle strength prediction model, and a fatigue time-series compensation model;

[0019] S8. Muscle Strength Compensation Output: The muscle strength estimation results from multiple muscle strength prediction models are weighted and fused according to the fusion weights to obtain a fused muscle strength prediction value, and then based on the composite fatigue index. Fatigue bias compensation is applied to the fusion muscle strength prediction value, and the final dynamic muscle strength prediction value or the percentage of maximum voluntary contractile force is output.

[0020] Preferably, in step S2, signal quality assessment, the signal quality evaluation index... The process is as follows:

[0021] Real-time detection of electromyographic signal amplitude to determine electrode detachment and amplifier saturation status;

[0022] The signal-to-noise ratio, root mean square stability, power frequency band energy proportion, and correlation coefficient between adjacent channels within a single window are calculated using the following formulas:

[0023]

[0024] in, Indicates the signal-to-noise ratio. Indicates the correlation between adjacent channels. Indicates root mean square stability. This indicates the proportion of power frequency noise energy. , , , These are the weighting coefficients;

[0025] After normalizing each indicator, a weighted fusion is performed to obtain the signal quality evaluation index (SQI).

[0026] according to The values ​​are used to classify quality levels, and corresponding graded exception handling mechanisms are executed. These graded exception handling mechanisms are as follows:

[0027] when When the quality falls below the first quality threshold, an alarm will be triggered, the electrodes will be recalibrated, or the signal will be reacquired.

[0028] when When the quality threshold is between the first and second quality thresholds, perform backup channel replacement or channel interpolation repair.

[0029] when When the quality exceeds the second quality threshold, the signal is allowed to proceed to the subsequent preprocessing and prediction process.

[0030] Preferably, in step S3, adaptive preprocessing, the process of adaptively adjusting the preprocessing parameters and the denoising strategy is as follows:

[0031] The output of the received signal quality monitoring and evaluation module The probability of motion state fed back by the motion state recognition module;

[0032] The low-frequency cutoff frequency of the bandpass filter is dynamically adjusted according to the motion state and the artifact intensity. In the static state, the effective low-frequency components are retained, and in the dynamic state, low-frequency motion artifacts are suppressed.

[0033] The dynamic adjustment rule for the low-frequency cutoff frequency is as follows:

[0034] Under static isometric contraction conditions, the low-frequency cutoff frequency is set to 10Hz~20Hz;

[0035] During periodic dynamic motion, the low-frequency cutoff frequency is adjusted to 20Hz~40Hz;

[0036] When strong motion artifacts are detected, the low-frequency cutoff frequency is further increased to 30Hz~50Hz;

[0037] The denoising strategy is dynamically switched according to the noise type: when power frequency noise is enhanced, a 50Hz notch filter is used; when non-stationary artifacts are enhanced, wavelet threshold denoising is used; and when high-frequency random noise is enhanced, empirical mode decomposition and reconstruction are used.

[0038] Recalculate after preprocessing This forms a closed-loop feedback adjustment mechanism until the signal quality meets the preset threshold.

[0039] Preferably, in step S4, the process of generating the fused feature vector in multi-domain feature fusion is as follows:

[0040] The preprocessed surface electromyography signals were subjected to first-order, second-order, and third-order differential processing to obtain multi-order differential signals;

[0041] One of the first-order differential electromyographic features is used to characterize the changes in the rate of change of electromyographic signals and the characteristics of muscle contraction speed.

[0042] Second-order differential electromyography features are used to characterize changes in electromyographic signal acceleration and muscle activation stability.

[0043] Third-order differential electromyography features are used to characterize the dynamic impact change trend of muscles and the abrupt change characteristics under high-dynamic motion conditions.

[0044] Time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear features were extracted from the original electromyography signal and the multi-order difference signal, respectively.

[0045] Various features are concatenated along fixed dimensions to form a multi-dimensional fused feature vector, which is then normalized and reconstructed before being output.

[0046] Preferably, in step S5, motion state recognition, the motion state recognition model is a CNN-LSTM hybrid network, a TCN-LSTM network, or a Transformer temporal classification network; the CNN-LSTM hybrid network extracts local spatial and dynamic change features of electromyography signals through convolutional layers, extracts time-series dynamic patterns through a long short-term memory network, and outputs probability values ​​for four types of motion states. Its working process is as follows:

[0047] It receives the fused feature vectors from a continuous sliding window as input;

[0048] Local spatial dynamic features at different scales are extracted using two convolutional layers;

[0049] Learn the temporal sequence variation patterns of electromyographic signals using a two-layer long short-term memory network;

[0050] Output the probability values ​​of four states: static isometric contraction, periodic dynamic motion, explosive rapid motion, and fatigue continuous motion, to form the initial state probability vector.

[0051] Preferably, in step S6, fatigue index construction and correction, the process of constructing the composite fatigue index and correcting the state probability is as follows:

[0052] Extract median frequency, root mean square value and sample entropy features from the fusion features, and calculate the degree of decrease in median frequency and the degree of change in root mean square value;

[0053] A composite fatigue index is constructed by integrating three types of features. The fatigue level is classified into three grades: mild, moderate, and severe.

[0054] Composite fatigue index Calculated using the following formula:

[0055]

[0056] in, Indicates the degree of decrease in median frequency. Indicates the degree of change in the root mean square value. Represents the normalized sample entropy; , , These are the weighting coefficients for the corresponding fatigue characteristic terms;

[0057] According to the composite fatigue index The probability of fatigue-prolonged motion state is weighted and corrected as follows:

[0058] The initial probability of fatigue-induced continuous motion state is weighted and corrected:

[0059]

[0060] in This represents the initial probability of a state of sustained motion due to fatigue. , For fatigue feedback weighting coefficients;

[0061] The probabilities of the four types of states are then normalized to obtain the corrected state probability vector.

[0062] Preferably, in step S7, the process of determining the fusion weights and outputting the fusion muscle strength prediction value in dynamic model fusion is as follows:

[0063] Four heterogeneous muscle strength prediction models are configured, each corresponding to one of the four motion states;

[0064] Based on the corrected state probabilities and the current window signal quality confidence level, the fusion weights of each model are calculated using the following formula:

[0065]

[0066] in, This is the corrected probability of the i-th type of state. This represents the confidence level or signal quality correction coefficient of the i-th type prediction model within the current window. The confidence coefficient is determined by the signal quality evaluation index. Determined in conjunction with historical prediction errors;

[0067] The muscle strength estimates output by each model are summed according to their weights to obtain the fused muscle strength prediction.

[0068] When the change in state probability exceeds a preset threshold, the fusion weights are dynamically updated to adapt to the motion state switching scenario.

[0069] Preferably, in step S8, the muscle strength compensation output, the multiple muscle strength prediction models specifically include:

[0070] Support vector regression model corresponding to static iso-length contraction state;

[0071] A long short-term memory network model corresponding to the periodic dynamic motion state;

[0072] Temporal convolutional networks, gated recurrent units, or fast regression models are used to address bursts of rapid motion.

[0073] A Transformer time-series compensation model corresponding to fatigue-prone continuous motion state.

[0074] An adaptive preprocessing surface electromyography dynamic muscle strength prediction system, comprising:

[0075] A surface electromyography (EMG) acquisition module is used to acquire multi-channel surface EMG signals of target muscle groups and process them in segments according to a preset sliding time window. The surface EMG acquisition module adopts a multi-channel flexible electrode array structure, including a flexible electrode array, an elastic band fixation structure, an analog front-end amplification circuit, a filtering circuit, an analog-to-digital conversion circuit, and a wireless transmission unit. The flexible electrode array uses flexible PCB electrodes to conform to the skin surface for acquiring EMG signals. The analog front-end amplification circuit adopts a differential amplification and high common-mode rejection ratio structure to amplify microvolt-level EMG signals.

[0076] The signal quality monitoring and evaluation module is used to calculate signal quality evaluation indicators for surface electromyography signals within each sliding time window. And perform hierarchical exception handling;

[0077] An adaptive preprocessing module is used to preprocess signals according to signal quality evaluation metrics. Based on the probability of the current motion state, the filtering parameters and denoising strategy are adaptively adjusted to dynamically preprocess the surface electromyography signal;

[0078] The multi-level dynamic electromyography feature fusion module is used to extract the original electromyography features and multi-level difference features of the preprocessed signal, and fuse multi-domain features to generate a fused feature vector.

[0079] The motion state recognition module is used to input the fused feature vector into the recognition model and output the initial state probability vectors corresponding to the four types of motion states.

[0080] The dynamic fatigue assessment module is used to construct a composite fatigue index. And based on the composite fatigue index Correct the initial state probability vector and output the corrected state probability vector;

[0081] The dynamic model fusion module is used to perform model fusion based on the corrected state probability vector and signal quality evaluation indicators. With composite fatigue index Determine the fusion weights of multiple muscle strength prediction models and output the fused muscle strength prediction value;

[0082] Muscle strength prediction module for use based on composite fatigue index Fatigue bias compensation is applied to the fusion muscle strength prediction value, and the final dynamic muscle strength prediction value or the percentage of maximum voluntary contractile force is output.

[0083] The control interface module is used to output the final dynamic muscle strength prediction results to external devices, enabling adaptive control of rehabilitation robots, exoskeleton robots, or human-robot collaborative robotic arms.

[0084] Preferably, in the muscle strength prediction module, the final dynamic muscle strength prediction value is calculated using the following formula:

[0085]

[0086] in To integrate predicted muscle strength values; The composite fatigue index; The trend of fatigue changes; As a signal quality evaluation indicator; These are the compensation coefficients obtained through calibration using the training set.

[0087] Compared with the prior art, the beneficial effects of the present invention are:

[0088] This invention employs an adaptive preprocessing mechanism driven by the Signal Quality Index (SQI). It dynamically adjusts the bandpass filter cutoff frequency and denoising strategy based on the motion state and noise type, forming a closed-loop feedback of preprocessing effect. This not only preserves the effective low-frequency electromyographic components in static scenarios but also effectively suppresses electrode slip artifacts, power frequency interference, and non-stationary noise in dynamic motion, thus solving the problem that fixed parameters cannot adequately address multiple scenarios.

[0089] This invention comprehensively characterizes the dynamic characteristics of muscle activation through multi-order difference and multi-domain feature fusion, accurately identifies the probability of four types of motion states by combining CNN-LSTM network, and adopts a heterogeneous model dynamic weighted fusion mechanism to replace a single fixed model, so that the optimal prediction model can be matched for static contraction, periodic motion, explosive motion and fatigue state, avoiding prediction drift when state switching.

[0090] The composite fatigue index of this invention participates in motion state probability correction, model fusion weight adjustment and final muscle force output compensation, forming a full-chain dynamic regulation of fatigue from state recognition to result output, effectively alleviating the error accumulation caused by spectrum drift during long-term exercise and significantly improving long-term prediction stability.

[0091] This invention uses an SQI grading mechanism to automatically identify and handle abnormalities such as electrode detachment and poor contact, ensuring the reliability of input signals and making it widely adaptable to complex and dynamic application scenarios such as rehabilitation training, exoskeleton control, and human-machine collaboration. Attached Figure Description

[0092] Figure 1 This is a block diagram of the overall structure of the adaptive preprocessing surface electromyography dynamic muscle strength prediction system of the present invention;

[0093] Figure 2 This is an overall flowchart of the surface electromyography signal processing of the present invention;

[0094] Figure 3 This is a block diagram illustrating the principle of dynamic model switching and fusion in this invention.

[0095] Figure 4This is a flowchart illustrating the fatigue compensation mechanism of the present invention. Detailed Implementation

[0096] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0097] like Figure 1-4 As shown, an adaptive preprocessing method for predicting surface electromyography dynamic muscle strength includes the following steps:

[0098] S1. Signal Acquisition and Segmentation: Acquire multi-channel surface electromyography (EMG) signals of the target muscle group and segment the multi-channel EMG signals according to a preset sliding time window.

[0099] S2. Signal Quality Assessment: Calculate signal quality evaluation indicators for surface electromyography signals within each sliding time window. , It is obtained at least by weighting the effective signal-to-noise ratio, the correlation between adjacent channels, the stability of the root mean square value, and the proportion of power frequency noise energy.

[0100] Signal quality evaluation indicators The process is as follows:

[0101] Real-time detection of electromyographic signal amplitude to determine electrode detachment and amplifier saturation status;

[0102] The signal-to-noise ratio, root mean square stability, power frequency band energy proportion, and correlation coefficient between adjacent channels within a single window are calculated using the following formulas:

[0103]

[0104] in, Indicates the signal-to-noise ratio. Indicates the correlation between adjacent channels. Indicates root mean square stability. This indicates the proportion of power frequency noise energy. , , , These are the weighting coefficients;

[0105] After normalizing each indicator, a weighted fusion is performed to obtain the signal quality evaluation index. ;

[0106] according to The values ​​are used to classify quality levels, and corresponding graded exception handling mechanisms are executed. These graded exception handling mechanisms are as follows:

[0107] when When the quality falls below the first quality threshold, an alarm will be triggered, the electrodes will be recalibrated, or the signal will be reacquired.

[0108] when When the quality threshold is between the first and second quality thresholds, perform backup channel replacement or channel interpolation repair.

[0109] when When the quality exceeds the second quality threshold, the signal is allowed to proceed to the subsequent preprocessing and prediction process.

[0110] S3, Adaptive Preprocessing: Based on And the initial motion state probability, adaptively adjust the signal preprocessing parameters, where, when When characterizing the enhancement of low-frequency motion artifacts, increase the low-frequency cutoff frequency of the bandpass filter; when When characterizing increased power frequency noise, enable notch filtering; when When characterizing artifacts of non-stationary motion, wavelet thresholding denoising or empirical mode decomposition reconstruction should be enabled;

[0111] The process of adaptively adjusting preprocessing parameters and denoising strategies is as follows:

[0112] The output of the received signal quality monitoring and evaluation module The probability of motion state fed back by the motion state recognition module;

[0113] The low-frequency cutoff frequency of the bandpass filter is dynamically adjusted according to the motion state and the artifact intensity. In the static state, the effective low-frequency components are retained, and in the dynamic state, low-frequency motion artifacts are suppressed.

[0114] The dynamic adjustment rule for the low-frequency cutoff frequency is as follows:

[0115] Under static isometric contraction conditions, the low-frequency cutoff frequency is set to 10Hz~20Hz;

[0116] During periodic dynamic motion, the low-frequency cutoff frequency is adjusted to 20Hz~40Hz;

[0117] When strong motion artifacts are detected, the low-frequency cutoff frequency is further increased to 30Hz~50Hz;

[0118] The denoising strategy is dynamically switched according to the noise type: when power frequency noise is enhanced, a 50Hz notch filter is used; when non-stationary artifacts are enhanced, wavelet threshold denoising is used; and when high-frequency random noise is enhanced, empirical mode decomposition and reconstruction are used.

[0119] Recalculate after preprocessing This forms a closed-loop feedback adjustment mechanism until the signal quality meets the preset threshold.

[0120] S4. Multi-domain feature fusion: Extract the original electromyographic features, first-order differential electromyographic features, second-order differential electromyographic features, and third-order differential electromyographic features from the preprocessed surface electromyographic signal, and form a fused feature vector including time-domain features, frequency-domain features, time-frequency domain features, and nonlinear features;

[0121] The process of generating the fused feature vector is as follows:

[0122] The preprocessed surface electromyography signals were subjected to first-order, second-order, and third-order differential processing to obtain multi-order differential signals;

[0123] One of the first-order differential electromyographic features is used to characterize the changes in the rate of change of electromyographic signals and the characteristics of muscle contraction speed.

[0124] Second-order differential electromyography features are used to characterize changes in electromyographic signal acceleration and muscle activation stability.

[0125] Third-order differential electromyography features are used to characterize the dynamic impact change trend of muscles and the abrupt change characteristics under high-dynamic motion conditions.

[0126] Time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear features were extracted from the original electromyography signal and the multi-order difference signal, respectively.

[0127] Various features are concatenated along fixed dimensions to form a multi-dimensional fused feature vector, which is then normalized and reconstructed before being output.

[0128] S5. Motion State Recognition: Input the fused feature vector into the motion state recognition model, and output the state probability vectors corresponding to static isometric contraction state, periodic dynamic motion state, explosive rapid motion state, and fatigue continuous motion state;

[0129] The motion state recognition model is a CNN-LSTM hybrid network, a TCN-LSTM network, or a Transformer temporal classification network. The CNN-LSTM hybrid network extracts local spatial and dynamic features of electromyography signals through convolutional layers, extracts dynamic patterns of the time series through a long short-term memory network, and outputs probability values ​​for four types of motion states. Its working process is as follows:

[0130] It receives the fused feature vectors from a continuous sliding window as input;

[0131] Local spatial dynamic features at different scales are extracted using two convolutional layers;

[0132] Learn the temporal sequence variation patterns of electromyographic signals using a two-layer long short-term memory network;

[0133] Output the probability values ​​of four states: static isometric contraction, periodic dynamic motion, explosive rapid motion, and fatigue continuous motion, to form the initial state probability vector.

[0134] S6. Construction and Correction of Fatigue Index: Construct a composite fatigue index based on the decrease in median frequency, the change in root mean square value, and sample entropy. And based on the composite fatigue index The probability of fatigue-prone motion states is corrected to obtain the corrected state probability vector.

[0135] The process of constructing the composite fatigue index and correcting the state probability is as follows:

[0136] Extract median frequency, root mean square value and sample entropy features from the fusion features, and calculate the degree of decrease in median frequency and the degree of change in root mean square value;

[0137] A composite fatigue index is constructed by integrating three types of features. The fatigue level is classified into three grades: mild, moderate, and severe.

[0138] Composite fatigue index Calculated using the following formula:

[0139]

[0140] in, Indicates the degree of decrease in median frequency. Indicates the degree of change in the root mean square value. Represents the normalized sample entropy; , , These are the weighting coefficients for the corresponding fatigue characteristic terms;

[0141] According to the composite fatigue index The probability of fatigue-prolonged motion state is weighted and corrected as follows:

[0142] The initial probability of fatigue-induced continuous motion state is weighted and corrected:

[0143]

[0144] in This represents the initial probability of a state of sustained motion due to fatigue. , For fatigue feedback weighting coefficients;

[0145] The probabilities of the four types of states are then normalized to obtain the corrected state probability vector.

[0146] S7. Dynamic Model Fusion: Based on the corrected state probability vector, and composite fatigue index The fusion weights of multiple muscle strength prediction models are determined, wherein the multiple muscle strength prediction models include at least a static muscle strength prediction model, a periodic dynamic muscle strength prediction model, a burst-speed muscle strength prediction model, and a fatigue time-series compensation model;

[0147] The process of determining the fusion weights and outputting the fusion muscle strength prediction value is as follows:

[0148] Four heterogeneous muscle strength prediction models are configured, each corresponding to one of the four motion states;

[0149] Based on the corrected state probabilities and the current window signal quality confidence level, the fusion weights of each model are calculated using the following formula:

[0150]

[0151] in, This is the corrected probability of the i-th type of state. This represents the confidence level or signal quality correction coefficient of the i-th type prediction model within the current window. The confidence coefficient is determined by the signal quality evaluation index. Determined in conjunction with historical prediction errors;

[0152] The muscle strength estimates output by each model are summed according to their weights to obtain the fused muscle strength prediction.

[0153] When the change in state probability exceeds a preset threshold, the fusion weights are dynamically updated to adapt to the motion state switching scenario.

[0154] S8. Muscle Strength Compensation Output: The muscle strength estimation results from multiple muscle strength prediction models are weighted and fused according to the fusion weights to obtain a fused muscle strength prediction value, and then based on the composite fatigue index. Fatigue bias compensation is applied to the fusion muscle strength prediction value, and the final dynamic muscle strength prediction value or the percentage of maximum voluntary contractile force is output.

[0155] The multiple muscle strength prediction models specifically include:

[0156] Support vector regression model corresponding to static iso-length contraction state;

[0157] A long short-term memory network model corresponding to the periodic dynamic motion state;

[0158] Temporal convolutional networks, gated recurrent units, or fast regression models are used to address bursts of rapid motion.

[0159] A Transformer time-series compensation model corresponding to fatigue-prone continuous motion state.

[0160] An adaptive preprocessing surface electromyography (SEMG) dynamic muscle strength prediction system includes a SEMG acquisition module, a signal quality monitoring and evaluation module, an adaptive preprocessing module, a multi-level dynamic EEMG feature fusion module, a motion state recognition module, a dynamic fatigue assessment module, a dynamic model fusion module, a muscle strength prediction module, and a control interface module. Each module is connected sequentially according to the signal flow direction. At the same time, the results of motion state recognition and fatigue assessment are fed back to the preprocessing and model fusion stages to form a closed-loop control mechanism.

[0161] The surface electromyography (EMG) acquisition module is used to acquire multi-channel surface EMG signals of the target muscle group and process them in segments according to a preset sliding time window. The surface EMG acquisition module adopts a multi-channel flexible electrode array structure, including a flexible electrode array, an elastic band fixation structure, an analog front-end amplification circuit, a filtering circuit, an analog-to-digital conversion circuit, and a wireless transmission unit. The flexible electrode array uses flexible PCB electrodes to conform to the skin surface to acquire EMG signals. The analog front-end amplification circuit adopts a differential amplification and high common-mode rejection ratio structure to amplify microvolt-level EMG signals.

[0162] A flexible electrode array, constructed using flexible PCB electrodes, is used to contact the surface of human skin. Compared to traditional rigid electrodes, flexible electrodes better conform to the curvature of the human body, thereby reducing contact impedance changes caused by skin stretching during movement and improving signal stability. The electrode fixing structure employs an elastic strap structure, which reduces motion artifacts caused by electrode slippage during movement. The analog front-end amplification circuit includes a differential amplifier, an instrumentation amplifier, and a common-mode rejection circuit, which amplifies the weak electromyographic signals. Since the amplitude of surface electromyographic signals is typically only in the microvolt range, an amplification structure with high input impedance and a high common-mode rejection ratio is required to improve immunity to power frequency interference. Subsequently, an analog-to-digital converter circuit converts the analog electromyographic signals into digital signals, which are then transmitted to the subsequent processing system via Bluetooth, WiFi, or a serial port module.

[0163] The unique feature of this invention is the use of a multi-channel flexible array structure, which can not only collect raw surface electromyography (EMG) signals, but also acquire EMG distribution information at different spatial locations. This provides basic data support for subsequent analysis of first-order, second-order, and third-order EMG change characteristics, and improves the system's ability to perceive complex motion states.

[0164] The signal quality monitoring and evaluation module is used to calculate signal quality evaluation indicators for surface electromyography signals within each sliding time window. The system performs graded anomaly handling. Because the human body is easily affected by factors such as electrode detachment, electrode slippage, sweat interference, power frequency noise, and motion artifacts during dynamic movement, the original surface electromyography (EMG) signal is prone to distortion, drift, or abnormal fluctuations. Directly inputting low-quality signals into subsequent prediction models can easily lead to misjudgments of movement states and drift in muscle strength prediction results, thereby reducing the stability and reliability of the entire system.

[0165] This module takes the raw multi-channel surface electromyography (EMG) signal as input and first performs real-time detection of the EMG signal amplitude. When the EMG signal amplitude remains below a preset threshold:

[0166]

[0167] If the current electrode may detach or have poor contact, it is determined that the current electrode may be detached or have poor contact. If the amplitude of the electromyographic signal exceeds the preset upper limit threshold, it is determined that there may be motion artifact interference or amplifier saturation.

[0168] Subsequently, the system further calculates indicators such as signal-to-noise ratio (SNR), root-mean-square (RMS) stability, and power frequency band energy proportion to comprehensively evaluate the quality of the current electromyographic (EMG) signal. Specifically, the SNR assesses the ratio between effective EMG components and noise components; the RMS fluctuation reflects signal stability; and the power frequency band energy proportion determines the degree of power frequency interference. When the energy proportion in the 50Hz band exceeds a preset threshold, the system determines that the current signal has significant power frequency interference.

[0169] Constructing signal quality evaluation indicators based on multi-index fusion:

[0170]

[0171] in:

[0172] Normalized signal-to-noise ratio;

[0173] To normalize the correlation between adjacent channels;

[0174] For normalized RMS stability;

[0175] The normalized chemical frequency noise energy ratio;

[0176] Here are the weighting coefficients, where .

[0177] Note: It is a negative indicator, so a minus sign is used in front of it.

[0178] The system classifies the quality level of the current electromyographic signal based on the SQI value.

[0179] when If the current data is in the correct state, it will be allowed to enter the subsequent adaptive preprocessing module and muscle strength prediction module; otherwise, the system will execute an exception handling mechanism, including alarm prompts, electrode recalibration, automatic switching to backup channels, or re-acquisition, to avoid low-quality electromyographic signals affecting the subsequent prediction results.

[0180] Furthermore, this invention also determines the electrode contact status through correlation analysis between multi-channel electromyography signals. If the correlation coefficient between adjacent channels is lower than a preset threshold:

[0181]

[0182] This suggests that there may be electrode slippage or abnormal local contact in the corresponding channel.

[0183] An adaptive preprocessing module is used to preprocess signals according to signal quality evaluation metrics. Based on the probability of the current motion state, the filtering parameters and denoising strategy are adaptively adjusted to dynamically preprocess the surface electromyography (EMG) signal. Since the types of noise generated by the human body differ significantly under different motion states—for example, static contraction is mainly affected by power frequency noise, while dynamic motion is prone to low-frequency motion artifacts and electrode slippage interference—traditional fixed filtering parameters cannot simultaneously meet the signal processing needs of both static and dynamic scenarios, leading to a decrease in the accuracy of subsequent motion state recognition and muscle strength prediction.

[0184] This module takes the effective electromyographic signal output by the signal quality monitoring and evaluation module and the state probability result output by the motion state recognition module as input, and dynamically adjusts the filtering parameters and denoising strategy according to the current motion state to achieve adaptive preprocessing for different motion scenarios.

[0185] When the system identifies the current motion state as a static isometric contraction state, due to the scarcity of low-frequency motion artifacts, a lower low-frequency cutoff frequency is used for bandpass filtering to retain more effective low-frequency electromyographic information. The low-frequency cutoff frequency of the filter is then set as follows:

[0186]

[0187] When the system identifies the current state as a dynamic motion state and detects that the root mean square fluctuation exceeds a preset threshold:

[0188]

[0189] If a significant motion artifact is detected, the system automatically raises the low-frequency cutoff frequency to:

[0190]

[0191] To enhance the ability to suppress low-frequency motion artifacts.

[0192] The system dynamically switches between different signal denoising strategies based on noise characteristics. When the system detects that the energy proportion of the power frequency band exceeds the threshold, a notch filter is activated to suppress 50Hz power frequency noise. When dynamic motion artifact enhancement is detected, wavelet threshold denoising is used to process non-stationary noise. When high-frequency random noise enhancement is detected, empirical mode decomposition (EMD) is used to adaptively decompose and reconstruct the signal, thereby improving the quality of electromyographic signals in complex motion scenarios.

[0193] Furthermore, this invention establishes a preprocessing effect feedback mechanism. After completing preprocessing, the system recalculates the signal quality indicators. When the processed signal quality evaluation value meets a preset threshold: If the signal is in the correct state, it is allowed to enter the subsequent multi-domain feature extraction module; otherwise, the system will automatically readjust the filtering parameters or re-execute the denoising process, thus forming a closed-loop adaptive preprocessing mechanism.

[0194] Finally, the system normalizes the processed electromyographic signals to reduce amplitude differences between different individuals, different acquisition locations, and different muscle activation intensities, thereby improving the generalization ability of subsequent motion state recognition and muscle strength prediction models.

[0195] A multi-level dynamic electromyography (EMG) feature fusion module is used to extract the original EMG features and multi-level differential features of the preprocessed signal, and fuse multi-domain features to generate a fused feature vector. Since traditional EMG analysis methods typically use only single time-domain or frequency-domain features for modeling, they are insufficient to fully describe the time-varying characteristics and nonlinear changes in muscle activation patterns during complex dynamic movements. Therefore, this invention further introduces a multi-level dynamic signal analysis mechanism and a multi-domain feature fusion mechanism to improve the system's ability to identify complex motion states and fatigue states.

[0196] This module takes the adaptively preprocessed surface electromyography (EMG) signal as input. It first performs multi-order differential processing on the original EMG signal to obtain information on dynamic changes in EMG. The first-order differential signal is defined as follows:

[0197]

[0198] It is used to characterize the rate of change of electromyographic signals and the characteristics of changes in muscle contraction speed.

[0199] The second-order difference signal is defined as:

[0200]

[0201] It is used to characterize changes in the acceleration of electromyographic signals and the stability characteristics of muscle activation.

[0202] A third-order difference signal is defined as:

[0203]

[0204] It is used to characterize the dynamic impact change trend of muscles and the abrupt change characteristics under highly dynamic motion conditions.

[0205] Subsequently, the system extracts time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear features from the original electromyographic signal and the multi-order differential signal, respectively.

[0206] Among them, the time-domain features include root mean square (RMS), mean absolute value (MAV), waveform length (WL), and zero-crossing rate (ZC), which are used to describe the characteristics of muscle activation intensity and amplitude changes.

[0207] Frequency domain features, including median frequency (MDF), mean frequency (MNF), and power spectral density (PSD), are used to describe the spectral drift characteristics under muscle fatigue conditions.

[0208] Time-frequency domain features, including the short-time Fourier transform spectrum (STFT) and wavelet packet energy distribution, are used to describe the non-stationary changes in electromyographic signals during dynamic motion.

[0209] Nonlinear features, including sample entropy (SampEn), detrended fluctuation analysis (DFA), and Lempel-Ziv complexity, are used to describe the complex dynamic behavior of the neuromuscular system.

[0210] A multi-level dynamic electromyography feature fusion module constructs a multi-dimensional dynamic electromyography feature vector:

[0211]

[0212] in:

[0213] Represents the time-domain eigenvectors;

[0214] Represents the frequency domain eigenvector;

[0215] Represents the time-frequency domain eigenvectors;

[0216] Represents a nonlinear eigenvector;

[0217] This represents a multi-order difference dynamic feature vector.

[0218] Finally, the system normalizes and reconstructs the dimensions of the fused feature vectors and outputs them to the subsequent motion state recognition module and dynamic muscle strength prediction module to improve the system's ability to model complex dynamic motion and fatigue states.

[0219] The motion state recognition module is used to input the fused feature vector into the recognition model and output the initial state probability vectors corresponding to the four types of motion states. Since the surface electromyography signals have significantly different time-domain, frequency-domain, and dynamic change characteristics under different motion states, accurate identification of the current motion state is of great significance for improving the adaptability and robustness of the subsequent muscle strength prediction model.

[0220] This module takes the feature vector output by the multi-order dynamic electromyography feature fusion module as input. The input features include multi-dimensional dynamic information such as original surface electromyography signal features, first-order differential dynamic features, second-order dynamic change features, and fatigue features. Among them, the original electromyography features are used to characterize muscle activation intensity, the first-order dynamic features are used to reflect changes in muscle contraction speed, the second-order dynamic features are used to describe muscle activation stability, and the fatigue features are used to characterize changes in muscle fatigue state.

[0221] This invention classifies human movement states into the following categories:

[0222] S1 represents the static isometric contraction state;

[0223] S2 represents a periodic dynamic motion state;

[0224] S3 indicates a burst of rapid motion;

[0225] S4 indicates a state of fatigue during continuous exercise.

[0226] The system employs a CNN-LSTM hybrid network to construct a motion state recognition model. Specifically, the convolutional neural network (CNN) is used to extract local spatial features and local dynamic change features of the electromyographic (EMG) signal; the long short-term memory network (LSTM) is used to extract the time-series variation patterns and dynamic temporal characteristics of the EMG signal.

[0227] Furthermore, the present invention employs a length of:

[0228]

[0229] A sliding time window is used to segment the electromyographic features for input. The CNN part consists of two convolutional layers with kernel sizes of:

[0230] 3×3 and 5×5 are used to extract local dynamic features at different scales; the LSTM part adopts a two-layer temporal network structure to learn the dynamic law of motion state changing over time.

[0231] Finally, the network outputs a motion state probability vector:

[0232]

[0233] in:

[0234] Represents the probability of a static state;

[0235] Represents the probability of dynamic motion states;

[0236] Indicates the probability of an explosive motion state;

[0237] This represents the probability of a fatigued state of motion.

[0238] The system determines the current motion state based on the maximum state probability.

[0239] when When the current motion state is determined to belong to the corresponding state category, the corresponding state information is sent to the dynamic model selection module.

[0240] Furthermore, this invention introduces a fatigue feedback correction mechanism. The system dynamically corrects the state probability based on the fatigue index output by the fatigue assessment module.

[0241]

[0242] in:

[0243] This represents the corrected state probability;

[0244] Indicates fatigue index;

[0245] This represents the fatigue feedback weighting coefficient.

[0246] Through the fatigue feedback mechanism, the system can dynamically correct the motion state recognition results under fatigue conditions, thereby improving the stability and accuracy of motion state recognition in complex motion environments.

[0247] Finally, the corrected state probability results are output to the dynamic model selection module to drive the dynamic switching and fusion of subsequent muscle strength prediction models.

[0248] The dynamic fatigue assessment module is used to construct a composite fatigue index. And based on the composite fatigue index The initial state probability vector is corrected, and the corrected state probability vector is output. Since human muscles exhibit significant spectral drift, changes in electromyographic energy, and a decrease in the complexity of neuromuscular control during continuous contraction, accurate assessment of muscle fatigue state is crucial for improving the accuracy of dynamic muscle strength prediction.

[0249] This module takes the frequency domain features, time domain features, and nonlinear features output by the multi-level dynamic electromyography feature fusion module as input, and constructs a composite fatigue index to achieve a comprehensive assessment of muscle fatigue.

[0250] Traditional methods typically use only the degree of median frequency (MDF) decline as a fatigue criterion, but a single frequency domain index is insufficient to accurately describe the muscle fatigue process under complex dynamic motion environments. Therefore, this invention further integrates spectral drift characteristics, electromyographic energy change characteristics, and nonlinear complexity characteristics to establish a composite fatigue index model:

[0251]

[0252]

[0253]

[0254] in:

[0255] Indicates fatigue index;

[0256] Indicates the degree of decrease in median frequency;

[0257] Indicates the degree of change in the root mean square value;

[0258] Represents sample entropy;

[0259] , , This represents the weighting coefficients corresponding to different fatigue characteristics.

[0260] Among them, the decrease in median frequency is used to characterize the characteristic of the electromyographic spectrum shifting towards lower frequencies during muscle fatigue; the change in root mean square value is used to reflect the change in muscle activation energy; and the sample entropy is used to describe the characteristic of decreased complexity of the neuromuscular system.

[0261] Furthermore, this invention classifies the degree of muscle fatigue according to the magnitude of the composite fatigue index.

[0262] when At that time, it was determined to be a state of mild fatigue;

[0263] when At that time, it was determined to be a state of moderate fatigue;

[0264] when At that time, it was determined to be a state of severe fatigue.

[0265] Furthermore, this invention establishes a fatigue feedback correction mechanism. When the system detects a continuous increase in fatigue level, it dynamically compensates and corrects the muscle strength prediction results.

[0266]

[0267]

[0268]

[0269] in:

[0270] The initial probability of fatigue-prone continuous motion state output by the motion state recognition model;

[0271] This is the corrected, unnormalized probability of fatigue-prone motion state.

[0272] It is a composite fatigue index, representing the current absolute fatigue level of the muscles;

[0273] The change in the fatigue index represents the trend of fatigue (the rate of aggravation / relief).

[0274] , , where is the fatigue feedback weighting coefficient, and are pre-calibrated model parameters, which respectively control the correction magnitude of "absolute fatigue level" and "fatigue change trend" on the state probability.

[0275] Through the aforementioned fatigue compensation mechanism, the system can dynamically reduce the prediction error caused by muscle output decay under fatigue conditions, thereby improving the stability and accuracy of muscle strength prediction under long-term continuous exercise conditions.

[0276] Finally, the fatigue assessment results will be simultaneously output to the motion state recognition module and the dynamic model selection module, for use in motion state probability correction and dynamic switching of prediction models.

[0277] The dynamic model fusion module is used to perform model fusion based on the corrected state probability vector and signal quality evaluation indicators. With composite fatigue index The fusion weights of multiple muscle strength prediction models are determined, and the fused muscle strength prediction value is output. Since the dynamic characteristics of electromyographic signals in the human body vary significantly under different movement states, traditional fixed single-model prediction methods are difficult to apply simultaneously to multiple scenarios such as static contraction, dynamic movement, and fatigue. Therefore, this invention proposes a dynamic model fusion mechanism based on motion state probability-driven fusion.

[0278] This module takes the state probability vector output by the motion state recognition module and the multi-level dynamic electromyographic features as input, and adaptively calls different muscle strength prediction models according to the dynamic characteristics of electromyography corresponding to different motion states.

[0279] in:

[0280] Under static isometric contraction, the electromyographic signal changes relatively steadily. Therefore, a support vector regression (SVR) model is used to predict muscle strength in order to improve the regression stability under small sample conditions.

[0281] During periodic dynamic motion, electromyographic signals exhibit significant time-series dynamic changes. Therefore, a long short-term memory network (LSTM) is used for dynamic temporal modeling.

[0282] Under sustained exercise conditions of fatigue, the long-term temporal dependence of electromyographic signals increases. Therefore, the Transformer temporal compensation model is used to model and correct the long-term fatigue change trend.

[0283] Furthermore, instead of directly selecting a single model using the traditional hard switching method, this invention constructs a dynamic weighted fusion mechanism based on motion state probabilities.

[0284] The motion state recognition module outputs a state probability vector:

[0285]

[0286] in:

[0287] Represents the probability of a static state;

[0288] Represents the probability of dynamic motion states;

[0289] Indicates the probability of an explosive motion state;

[0290] This represents the probability of a fatigued state of motion.

[0291] Correspondingly, different prediction models output muscle strength prediction results:

[0292]

[0293] in:

[0294] express Model output;

[0295] express Model output;

[0296] express The output of the compensation model.

[0297] The system dynamically weights and fuses the outputs of multiple models based on state probabilities:

[0298]

[0299] Model fusion formula:

[0300]

[0301] in:

[0302]

[0303] It is the model confidence level, the current window signal quality, or the reciprocal of the historical error.

[0304] Through the above probability-driven fusion mechanism, the system can dynamically adjust the output weights of different models according to the current motion state, thereby avoiding the prediction drift problem caused by traditional single-model prediction in complex motion scenarios.

[0305] Furthermore, when the system detects a significant change in state probability:

[0306]

[0307] In this case, the model fusion weights are dynamically updated to enhance the system's ability to adapt to rapid changes in motion state.

[0308] Finally, the dynamically fused muscle strength prediction results are output to the muscle strength prediction module, and further combined with the fatigue compensation mechanism to generate the final muscle strength prediction value.

[0309] Final fatigue compensation formula:

[0310]

[0311] Muscle strength prediction module for use based on composite fatigue index The module compensates for fatigue bias in the fused muscle strength predictions and outputs the final dynamic muscle strength prediction or the percentage of maximum voluntary contractile force. This module serves as the final output module of the entire system, enabling continuous muscle strength estimation during dynamic human movement.

[0312] This module takes as input the feature vector output by the multi-order dynamic electromyography feature fusion module, the state probability results output by the motion state recognition module, and the fatigue index output by the fatigue assessment module. The input features include multi-dimensional dynamic electromyography features such as root mean square (RMS), median frequency (MDF), wavelet packet energy, first-order dynamic change features, second-order dynamic change features, and fatigue index.

[0313] The system first dynamically calls the corresponding muscle strength prediction model based on the state probability results output by the dynamic model fusion module. Specifically:

[0314] In the static isometric shrinkage state, a support vector regression (SVR) model is used for prediction;

[0315] Temporal muscle strength modeling was performed using a Long Short-Term Memory (LSTM) network under periodic dynamic motion conditions.

[0316] The Transformer compensation model is used to dynamically correct long-term fatigue drift under fatigued motion conditions.

[0317] Furthermore, the system dynamically fuses the outputs of multiple models using state probabilities to obtain initial muscle strength predictions:

[0318]

[0319] in:

[0320] This indicates the initial muscle strength prediction results;

[0321] , , These represent the probabilities of different motion states;

[0322] , , These represent the output results of different models.

[0323] Subsequently, the system dynamically compensates and corrects the initial muscle strength prediction results based on the composite fatigue index output by the fatigue assessment module.

[0324]

[0325] in:

[0326] This indicates the final muscle strength prediction result;

[0327] This represents the initial predicted muscle strength value;

[0328] Indicates the composite fatigue index;

[0329] This represents the fatigue compensation coefficient.

[0330] Through a dynamic fatigue compensation mechanism, the system can reduce prediction drift caused by muscle fatigue during long-term continuous exercise, thereby improving the stability of dynamic muscle strength prediction.

[0331] Furthermore, the present invention employs a sliding time window method for continuous muscle strength prediction.

[0332] The system has a length of:

[0333] The time window is used to continuously segment the electromyographic signal and output continuous muscle strength values ​​or MVC percentage results in real time:

[0334]

[0335] in:

[0336] Indicates the percentage of maximum voluntary contractile force;

[0337] This indicates the final muscle strength prediction result;

[0338] This represents the reference value for the maximum voluntary contractile force.

[0339] Furthermore, this invention uses a grip strength meter or force sensor to collect real muscle strength values ​​as model training labels. The electromyography (EMG) signal sampling system and the grip strength meter sampling system are time-aligned using a unified clock synchronization mechanism to ensure the synchronization consistency between EMG features and real muscle strength labels.

[0340] Ultimately, the system outputs continuous dynamic muscle strength prediction results in real time, and can be further used in application scenarios such as exoskeleton control, rehabilitation robot control, and human-machine collaborative assisted control.

[0341] The control interface module outputs the final dynamic muscle strength prediction results to external devices, enabling adaptive control of rehabilitation robots, exoskeleton robots, or human-robot collaborative robotic arms. The interface supports multiple communication protocols and is adaptable to various application scenarios, including rehabilitation robots, exoskeleton robots, and human-robot collaborative robotic arms.

[0342] For example, in rehabilitation robot scenarios, the system can dynamically adjust the resistance and assist force of rehabilitation training based on real-time muscle strength and fatigue status; in exoskeleton control scenarios, the system can adjust the assist ratio of the exoskeleton in real time based on muscle strength prediction results; in human-machine collaboration scenarios, the system can quickly respond to coordinated actions based on burst motion prediction and automatically reduce the workload based on fatigue status.

[0343] Example 1 (Example of dynamic prediction of upper limb grip strength):

[0344] This system utilizes an 8-channel flexible surface electromyography (EMG) electrode array to acquire EMG signals from three target muscle groups—the forearm flexors, biceps brachii, and triceps brachii—at a sampling rate of 1000Hz. Signal quality is first verified, and then dynamic adaptation to motion states is applied using 20-450Hz bandpass filtering and 50Hz power frequency notch filtering. Time-domain, frequency-domain, and nonlinear features are extracted using a 200ms window and a 50ms step size. A CNN-LSTM hybrid network is used to identify motion states and simultaneously calculate the muscle fatigue index in real time. Based on a dynamic switching prediction model according to work conditions, the system ultimately outputs continuous muscle strength values. The percentage was determined using data tags calibrated with a hand grip dynamometer. , , And the drift error after fatigue is used as the basis for performance evaluation.

[0345] Parameter details:

[0346] Target muscles: forearm flexors, biceps brachii, triceps brachii

[0347] Acquisition hardware: 8-channel flexible surface electromyography electrode array

[0348] Sampling frequency: 1000Hz

[0349] Filtering parameters: Bandpass 20Hz-450Hz, 50Hz power frequency notch.

[0350] Analysis window: Window length 200ms, sliding step 50ms

[0351] Recognition Model: CNN-LSTM Hybrid Neural Network

[0352] Calibration equipment: hand grip dynamometer

[0353] Output content: continuous muscle strength values, percentage

[0354] Evaluation indicators: , , And drift error after fatigue.

[0355] Example 2 (Example of Human-Robot Collaborative Robotic Arm Assisted Control):

[0356] This invention can be applied to human-machine collaborative robotic arm systems. It leverages real-time muscle strength, movement intent probability, and muscle fatigue status obtained through electromyography (EMG) signal analysis to achieve adaptive and precise control of the robotic arm and human-machine collaborative safety protection. Based on the aforementioned EMG signal acquisition, feature extraction, movement state recognition, muscle strength prediction, and fatigue assessment algorithms, the system acquires real-time muscle strength values, MVC percentages, rapid movement burst probability, and fatigue persistence probability of the forearm flexor muscles, biceps, and triceps. Combining these with multi-dimensional human physiological and movement status data, it dynamically adjusts the robotic arm's impedance parameters, assist force, response speed, and control mode.

[0357] The specific control logic is as follows: When a decrease in human muscle strength and an increase in the fatigue index are predicted, it indicates a decline in the human muscle's load-bearing capacity and increased fatigue. The system proactively increases the robotic arm's assistance force and assistance ratio to effectively share the human's workload. When human muscle strength continues to decline and fatigue intensifies, the system reduces the frequency of robotic arm control command updates, slows down the work pace, and switches to a safety mode when necessary to avoid the risk of loss of control and injury during human-machine collaboration. When the probability of a sudden, rapid human movement is detected to be increasing, it is predicted that the user intends to perform sudden, rapid work. The system immediately increases the robotic arm's motion response speed to ensure synchronized and flexible human-machine action coordination. When the probability of a sustained state of human fatigue increases, the system simultaneously reduces the work training load and increases the robotic arm's assistance ratio, significantly reducing the burden on human muscles. While stably completing collaborative tasks, it effectively alleviates muscle fatigue and avoids overexertion, achieving precise linkage between algorithm recognition results and actual robotic arm control actions, significantly improving the safety, adaptability, and comfort of human-machine collaboration.

[0358] Example 3 (Example of continuous training with a rehabilitation robot):

[0359] This solution can be applied to continuous muscle strength training scenarios using rehabilitation robots. A single training session is set to last several tens of minutes. In the early stages of training, the muscles are in a non-fatigued state. As the training progresses, the muscles gradually enter a state of accumulated fatigue. During the training process, a fixed prediction model and the dynamic fusion model of this invention are used simultaneously to conduct muscle strength prediction comparison experiments. The prediction error fluctuations after muscle fatigue occur are continuously monitored and recorded. Comparison of the prediction results of the two models verifies that, compared to the traditional fixed model, the dynamic fusion model of this invention can adapt to the physiological signal changes caused by muscle fatigue, effectively suppressing the problem of muscle strength prediction value drift. Even when the muscles are in a state of continuous fatigue, it can still maintain higher prediction accuracy, meeting the detection and evaluation needs of long-term rehabilitation training.

[0360] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting dynamic muscle strength using surface electromyography with adaptive preprocessing, characterized in that, Includes the following steps: S1. Signal Acquisition and Segmentation: Acquire multi-channel surface electromyography (EMG) signals of the target muscle group and segment the multi-channel EMG signals according to a preset sliding time window; S2. Signal Quality Assessment: Calculate signal quality evaluation indicators for surface electromyography signals within each sliding time window. , It is obtained at least by weighting the effective signal-to-noise ratio, the correlation between adjacent channels, the stability of the root mean square value, and the proportion of power frequency noise energy. S3, Adaptive Preprocessing: Based on And the initial motion state probability, adaptively adjust the signal preprocessing parameters, where, when When characterizing the enhancement of low-frequency motion artifacts, increase the low-frequency cutoff frequency of the bandpass filter; when When characterizing increased power frequency noise, enable notch filtering; when When characterizing artifacts of non-stationary motion, wavelet thresholding denoising or empirical mode decomposition reconstruction should be enabled; S4. Multi-domain feature fusion: Extract the original electromyographic features, first-order differential electromyographic features, second-order differential electromyographic features, and third-order differential electromyographic features from the preprocessed surface electromyographic signal, and form a fused feature vector including time-domain features, frequency-domain features, time-frequency domain features, and nonlinear features; S5. Motion State Recognition: Input the fused feature vector into the motion state recognition model, and output the state probability vectors corresponding to static isometric contraction state, periodic dynamic motion state, explosive rapid motion state, and fatigue continuous motion state; S6. Construction and Correction of Fatigue Index: Construct a composite fatigue index based on the decrease in median frequency, the change in root mean square value, and sample entropy. And based on the composite fatigue index The probability of fatigue-prone motion states is corrected to obtain the corrected state probability vector. S7. Dynamic Model Fusion: Based on the corrected state probability vector, and composite fatigue index The fusion weights of multiple muscle strength prediction models are determined, wherein the multiple muscle strength prediction models include at least a static muscle strength prediction model, a periodic dynamic muscle strength prediction model, a burst-speed muscle strength prediction model, and a fatigue time-series compensation model; S8. Muscle Strength Compensation Output: The muscle strength estimation results from multiple muscle strength prediction models are weighted and fused according to the fusion weights to obtain a fused muscle strength prediction value, and then based on the composite fatigue index. Fatigue bias compensation is applied to the fusion muscle strength prediction value, and the final dynamic muscle strength prediction value or the percentage of maximum voluntary contractile force is output.

2. The adaptive preprocessing surface electromyography dynamic muscle strength prediction method according to claim 1, characterized in that, In step S2, signal quality assessment, the signal quality evaluation indicators... The process is as follows: Real-time detection of electromyographic signal amplitude to determine electrode detachment and amplifier saturation status; The signal-to-noise ratio, root mean square stability, power frequency band energy proportion, and correlation coefficient between adjacent channels within a single window are calculated using the following formulas: in, Indicates the signal-to-noise ratio. Indicates the correlation between adjacent channels. Indicates root mean square stability. This indicates the proportion of power frequency noise energy. , , , These are the weighting coefficients; After normalizing each indicator, a weighted fusion is performed to obtain the signal quality evaluation index. ; according to The values ​​are used to classify quality levels, and corresponding graded exception handling mechanisms are executed. These graded exception handling mechanisms are as follows: when When the quality falls below the first quality threshold, an alarm will be triggered, the electrodes will be recalibrated, or the signal will be reacquired. when When the quality threshold is between the first and second quality thresholds, perform backup channel replacement or channel interpolation repair. when When the quality exceeds the second quality threshold, the signal is allowed to proceed to the subsequent preprocessing and prediction process.

3. The adaptive preprocessing surface electromyography dynamic muscle strength prediction method according to claim 1, characterized in that, In step S3, adaptive preprocessing, the process of adaptively adjusting the preprocessing parameters and denoising strategy is as follows: The output of the received signal quality monitoring and evaluation module The probability of motion state fed back by the motion state recognition module; The low-frequency cutoff frequency of the bandpass filter is dynamically adjusted according to the motion state and the artifact intensity. In the static state, the effective low-frequency components are retained, and in the dynamic state, low-frequency motion artifacts are suppressed. The dynamic adjustment rule for the low-frequency cutoff frequency is as follows: Under static isometric contraction conditions, the low-frequency cutoff frequency is set to 10Hz~20Hz; During periodic dynamic motion, the low-frequency cutoff frequency is adjusted to 20Hz~40Hz; When strong motion artifacts are detected, the low-frequency cutoff frequency is further increased to 30Hz~50Hz; The denoising strategy is dynamically switched according to the noise type: when power frequency noise is enhanced, a 50Hz notch filter is used; when non-stationary artifacts are enhanced, wavelet threshold denoising is used; and when high-frequency random noise is enhanced, empirical mode decomposition and reconstruction are used. Recalculate after preprocessing This forms a closed-loop feedback adjustment mechanism until the signal quality meets the preset threshold.

4. The adaptive preprocessing surface electromyography dynamic muscle strength prediction method according to claim 1, characterized in that, In step S4, the process of generating the fused feature vector in multi-domain feature fusion is as follows: The preprocessed surface electromyography signals were subjected to first-order, second-order, and third-order differential processing to obtain multi-order differential signals; One of the first-order differential electromyographic features is used to characterize the changes in the rate of change of electromyographic signals and the characteristics of muscle contraction speed. Second-order differential electromyography features are used to characterize changes in electromyographic signal acceleration and muscle activation stability. Third-order differential electromyography features are used to characterize the dynamic impact change trend of muscles and the abrupt change characteristics under high-dynamic motion conditions. Time-domain features, frequency-domain features, time-frequency-domain features, and nonlinear features were extracted from the original electromyography signal and the multi-order difference signal, respectively. Various features are concatenated along fixed dimensions to form a multi-dimensional fused feature vector, which is then normalized and reconstructed before being output.

5. The adaptive preprocessing method for predicting surface electromyography dynamic muscle strength according to claim 1, characterized in that, In step S5, motion state recognition, the motion state recognition model is a CNN-LSTM hybrid network, a TCN-LSTM network, or a Transformer temporal classification network. The CNN-LSTM hybrid network extracts local spatial and dynamic change features of electromyography signals through convolutional layers, extracts dynamic patterns of time series through a long short-term memory network, and outputs probability values ​​for four types of motion states. Its working process is as follows: It receives the fused feature vectors from a continuous sliding window as input; Local spatial dynamic features at different scales are extracted using two convolutional layers; Learn the temporal sequence variation patterns of electromyographic signals using a two-layer long short-term memory network; Output the probability values ​​of four states: static isometric contraction, periodic dynamic motion, explosive rapid motion, and fatigue continuous motion, to form the initial state probability vector.

6. The adaptive preprocessing method for predicting surface electromyography dynamic muscle strength according to claim 5, characterized in that, In step S6, fatigue index construction and correction, the process of constructing the composite fatigue index and correcting the state probability is as follows: Extract median frequency, root mean square value and sample entropy features from the fusion features, and calculate the degree of decrease in median frequency and the degree of change in root mean square value; A composite fatigue index is constructed by integrating three types of features. The fatigue level is classified into three grades: mild, moderate, and severe. Composite fatigue index Calculated using the following formula: in, Indicates the degree of decrease in median frequency. Indicates the degree of change in the root mean square value. Represents the normalized sample entropy; , , These are the weighting coefficients for the corresponding fatigue characteristic terms; According to the composite fatigue index The probability of fatigue-prolonged motion state is weighted and corrected as follows: The initial probability of fatigue-induced continuous motion state is weighted and corrected: in This represents the initial probability of a state of sustained motion due to fatigue. , For fatigue feedback weighting coefficients; The probabilities of the four types of states are then normalized to obtain the corrected state probability vector.

7. The adaptive preprocessing method for predicting surface electromyography dynamic muscle strength according to claim 5, characterized in that, In step S7, the process of determining the fusion weights and outputting the fusion muscle strength prediction value in dynamic model fusion is as follows: Four heterogeneous muscle strength prediction models are configured, each corresponding to one of the four motion states; Based on the corrected state probabilities and the current window signal quality confidence level, the fusion weights of each model are calculated using the following formula: in, This is the corrected probability of the i-th type of state. This represents the confidence level or signal quality correction coefficient of the i-th type prediction model within the current window. The confidence coefficient is determined by the signal quality evaluation index. Determined in conjunction with historical prediction errors; The muscle strength estimates output by each model are summed according to their weights to obtain the fused muscle strength prediction. When the change in state probability exceeds a preset threshold, the fusion weights are dynamically updated to adapt to the motion state switching scenario.

8. The adaptive preprocessing method for predicting surface electromyography dynamic muscle strength according to claim 5, characterized in that, In step S8, the muscle strength compensation output, the multiple muscle strength prediction models specifically include: Support vector regression model corresponding to static iso-length contraction state; A long short-term memory network model corresponding to the periodic dynamic motion state; Temporal convolutional networks, gated recurrent units, or fast regression models are used to address bursts of rapid motion. A Transformer time-series compensation model corresponding to fatigue-prone continuous motion state.

9. A surface electromyography dynamic muscle strength prediction system with adaptive preprocessing, characterized in that, include: The surface electromyography (EMG) acquisition module is used to acquire multi-channel surface EMG signals of the target muscle group and process them in segments according to a preset sliding time window. The surface electromyography (EMG) acquisition module adopts a multi-channel flexible electrode array structure, including a flexible electrode array, an elastic band fixation structure, an analog front-end amplifier circuit, a filter circuit, an analog-to-digital conversion circuit, and a wireless transmission unit; the flexible electrode array uses flexible PCB electrodes to conform to the skin surface to acquire EMG signals. The analog front-end amplifier circuit adopts a differential amplification and high common-mode rejection ratio structure to amplify microvolt-level electromyographic signals; The signal quality monitoring and evaluation module is used to calculate signal quality evaluation indicators for surface electromyography signals within each sliding time window. And perform hierarchical exception handling; An adaptive preprocessing module is used to preprocess signals according to signal quality evaluation metrics. Based on the probability of the current motion state, the filtering parameters and denoising strategy are adaptively adjusted to dynamically preprocess the surface electromyography signal; The multi-level dynamic electromyography feature fusion module is used to extract the original electromyography features and multi-level difference features of the preprocessed signal, and fuse multi-domain features to generate a fused feature vector. The motion state recognition module is used to input the fused feature vector into the recognition model and output the initial state probability vectors corresponding to the four types of motion states. The dynamic fatigue assessment module is used to construct a composite fatigue index. And based on the composite fatigue index Correct the initial state probability vector and output the corrected state probability vector; The dynamic model fusion module is used to perform model fusion based on the corrected state probability vector and signal quality evaluation indicators. With composite fatigue index Determine the fusion weights of multiple muscle strength prediction models and output the fused muscle strength prediction value; Muscle strength prediction module for use based on composite fatigue index Fatigue bias compensation is applied to the fusion muscle strength prediction value, and the final dynamic muscle strength prediction value or the percentage of maximum voluntary contractile force is output. The control interface module is used to output the final dynamic muscle strength prediction results to external devices, enabling adaptive control of rehabilitation robots, exoskeleton robots, or human-robot collaborative robotic arms.

10. The adaptive preprocessing surface electromyography dynamic muscle strength prediction system according to claim 9, characterized in that, In the muscle strength prediction module, the final dynamic muscle strength prediction value is calculated using the following formula: in To integrate predicted muscle strength values; The composite fatigue index; The trend of fatigue changes; As a signal quality evaluation indicator; These are the compensation coefficients obtained through calibration using the training set.

Images