Lower limb muscle fatigue assessment method and system based on multi-modal physiological signals

Abstract

The invention discloses a lower limb muscle fatigue assessment method and system based on multi-modal physiological signals, and relates to the technical field of biomedical engineering, target muscles are selected, sEMG signals, AUS signals and respiratory flux VE data are synchronously collected, and RPE is recorded; preprocessing various signals to remove interference and abnormal points; extracting sEMG time-frequency domain features and AUS muscle thickness features; fusing features and constructing vectors in a standardized manner; dividing five fatigue grades; and a data set is divided according to stratified sampling, a model is optimized through five-fold cross validation, and real-time features are input to output a fatigue level. Multi-modal physiological signals are fused, fatigue evaluation accuracy is improved, different fatigue stages are finely adapted, muscle load and energy consumption are reduced through personalized adaption and a dynamic adjustment strategy, rehabilitation training safety and comfort are improved, and efficient recovery of lower limb movement functions is assisted.

Description

Technical Field

[0001] This invention relates to the field of biomedical engineering technology, and in particular to a method and system for assessing lower limb muscle fatigue based on multimodal physiological signals. Background Technology

[0002] Lower limb motor dysfunction has become a significant health issue affecting the nation, with its causes concentrated in three main categories: stroke, age-related disability, and accidents. The number of stroke patients in my country has exceeded 15 million, and the incidence rate is rising annually. Over 80% of survivors experience symptoms such as muscle spasms, weakness, or even hemiplegia, severely impacting their ability to live independently. Data from the Seventh National Population Census shows that the elderly population aged 60 and above has reached 264 million, accounting for 18.7% of the total population. With the aging population, the number of elderly patients with declining lower limb motor function continues to increase. The "2023 Statistical Bulletin on the Development of the Cause of Persons with Disabilities" indicates that people with limb disabilities account for 49.11% of those receiving certified rehabilitation services, totaling 4.157 million, but rehabilitation institutions nationwide only have 266,000 staff, highlighting a significant imbalance between supply and demand for rehabilitation resources. Traditional manual rehabilitation therapy relies on therapist experience, is inefficient, and struggles to achieve personalized intervention, failing to meet the vast needs of patients. Robot-assisted rehabilitation technology has become a core direction for overcoming this predicament.

[0003] In the field of rehabilitation robotics, human-computer interaction based on biosignals is a key development trend, with non-invasive physiological signals such as surface electromyography (SEMG), electroencephalography (EEG), and ultrasound being widely used in robot control. Among these, SEMG reflects the electrical activity characteristics during muscle contraction, while A-mode ultrasound can accurately monitor changes in muscle morphology; the two are naturally complementary in information acquisition. However, existing research often focuses on single physiological signals, which has significant limitations: single SEMG signals are susceptible to power frequency interference and baseline drift, resulting in insufficient perception of muscle morphological changes; while single A-mode ultrasound signals can capture changes in muscle thickness, they cannot reflect the state of neural electrical activity. Furthermore, existing fatigue assessment algorithms have shortcomings: either they are limited in accuracy due to targeting a single research subject, or they require additional sensors, increasing system complexity, and they do not fully integrate subjective fatigue sensations with objective physiological indicators, resulting in incomplete assessment results.

[0004] As the mainstream rehabilitation robotic device, lower limb exoskeletons currently rely heavily on offline preset modes for human-computer interaction control, making it difficult to adapt to dynamically changing human-computer interaction scenarios. Existing fatigue management strategies simply replace movement once the fatigue threshold is reached, without grading and adapting according to the degree of fatigue, thus failing to meet the personalized rehabilitation needs of different fatigue stages. Furthermore, the stiffness control of traditional exoskeletons is mostly fixed, unable to dynamically adjust with muscle fatigue levels, leading to excessive muscle activation, excessive energy consumption, and potentially causing safety risks such as muscle spasms and falls. As rehabilitation needs upgrade towards precision, personalization, and safety, the shortcomings of existing technologies in multimodal signal fusion, precise fatigue assessment, and dynamic adaptive control are becoming increasingly apparent. There is an urgent need for an integrated solution that combines multimodal physiological signals to achieve precise muscle fatigue assessment and adaptive exoskeleton control. Summary of the Invention

[0005] This invention proposes a method and system for assessing lower limb muscle fatigue based on multimodal physiological signals, in order to solve the problems mentioned in the prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for assessing lower limb muscle fatigue based on multimodal physiological signals, comprising the following steps:

[0007] The multimodal physiological signal acquisition steps selected the tibialis anterior, lateral gastrocnemius, rectus femoris and biceps femoris of the lower limb as target muscles. The surface electromyography signal acquisition device and the A-mode ultrasound signal acquisition device were used to simultaneously acquire sEMG and AUS signals during muscle contraction. At the same time, respiratory flux (VE) data were acquired through the exercise metabolism analysis system, and the subject's fatigue perception was recorded in combination with the subjective fatigue perception scale (RPE).

[0008] The signal preprocessing steps include: using 20 to 500 Hz bandpass filtering and 50 Hz notch filtering to remove power frequency interference, baseline drift and high frequency noise from sEMG signals; performing time-varying gain compensation, bandpass filtering, envelope detection and logarithmic compression preprocessing on AUS signals; and using the three-standard-deviation principle to remove outliers from respiratory data.

[0009] In the feature extraction step, the time-domain features such as mean absolute value, root mean square value, and integral electromyography value are extracted from the preprocessed sEMG signal using the sliding window analysis method. The frequency-domain features such as median frequency and mean frequency are extracted using the short-time Fourier transform (STFT). The muscle thickness envelope features are extracted from the AUS signal using the brightness tracking method and the cross-correlation method. The features are then fused to obtain a multi-dimensional feature set.

[0010] The multimodal feature fusion step involves truncating the sEMG and AUS feature sequences based on the shortest feature sequence length, flattening the three-dimensional feature matrix into a two-dimensional vector, and constructing a fused feature vector after Z-score normalization to achieve complementary fusion of temporal, frequency, and morphological features.

[0011] The fatigue level classification steps are as follows: Based on the VE change pattern and RPE grading threshold, a joint criterion is established to classify muscle fatigue state into five levels: relaxed state, walking state, transitional state, fatigued state, and limit state. The VE range and RPE score interval corresponding to each level are also defined.

[0012] The evaluation model training and output steps involve dividing the fused feature vector and fatigue level label into training and test sets using stratified sampling. The machine learning model parameters are optimized through five-fold cross-validation, and the real-time fused features are input to achieve accurate evaluation and output of muscle fatigue level.

[0013] Furthermore, it also includes a multimodal feature weighted optimization step, designing an adaptive weighted fusion expression for features based on the differences in feature contribution at different fatigue stages.

[0014]

[0015] in To optimize the fused feature vector, The number of sEMG features. For the first sEMG feature weight coefficients For the first Standardized values ​​of sEMG features For the first Median frequency corresponding to each sEMG feature This represents the baseline median frequency of muscles at rest. This is the frequency difference adjustment coefficient. The number of AUS features. For the first AUS feature weight coefficients, For the first Standardized values ​​of AUS features, For the first The change in muscle thickness corresponding to each AUS feature. This represents the maximum change in muscle thickness.

[0016] Furthermore, it also includes evaluating the dynamic optimization steps of the model, introducing multi-model ensemble and majority voting decision-making mechanisms, and constructing the fatigue level decision optimization expression as follows:

[0017]

[0018] in The final fatigue level, The number of machine learning models integrated. For the first The confidence weight coefficients of each model For indicator functions when If the value is 1, then the value is 0; otherwise, the value is 0. For the first The output level of each model The historical decision-making influence coefficient. The level in historical data The number of correctly classified samples, This represents the total number of correctly classified samples in history.

[0019] Furthermore, it also includes a dynamic threshold adjustment step for fatigue levels. Based on the individual physiological baseline and exercise intensity of the subjects, the VE and RPE judgment thresholds for each fatigue level are corrected in real time. An individual baseline is established by collecting physiological parameters of the subjects in a resting state. Combined with intensity indicators such as speed and load during exercise, the threshold range of each level is adjusted by linear interpolation to make the fatigue classification more in line with individual physiological characteristics. At the same time, the threshold adjustment patterns of different individuals are recorded to establish a cross-individual threshold adaptation library.

[0020] Furthermore, it also includes a cross-individual fatigue feature adaptation step. To address the feature distribution shift caused by differences in muscle morphology and exercise habits among different individuals, a domain adaptive algorithm is used to transfer learning the fused features. By extracting common representations of fatigue features from different individuals, a domain-invariant feature space is constructed to eliminate evaluation bias caused by individual differences. At the same time, an individual feature bias compensation model is established to quickly adjust the evaluation model parameters based on the initial exercise data of new users.

[0021] Furthermore, it also includes real-time feedback of assessment results and model update steps, outputting fatigue assessment results to the exoskeleton control system in real time to provide data support for adaptive control; at the same time, it archives assessment data, actual fatigue state feedback and control effect data at fixed periods, and uses incremental learning algorithms to update assessment model parameters and feature weights, so that the model can continuously adapt to the dynamic changes of the subject's fatigue characteristics and improve the accuracy and reliability of long-term assessment.

[0022] Furthermore, it also includes a multi-scenario fatigue assessment adaptation step, which presets a scenario feature template library for different sports scenarios; by identifying the current sports scenario type, it calls the corresponding feature extraction parameters and evaluation model weights to adjust the judgment sensitivity of each fatigue level.

[0023] This invention also discloses a lower limb muscle fatigue assessment system based on multimodal physiological signals, comprising the following modules:

[0024] The multimodal signal acquisition module consists of a surface electromyography (SEMG) acquisition unit, an A-mode ultrasound acquisition unit, a metabolic data acquisition unit, and a subjective assessment input unit. The SEMG acquisition unit uses silver chloride electrode pads attached to the target muscle belly. The A-mode ultrasound acquisition unit is equipped with a wireless four-channel amplitude modulation device and an ultrasound coupling agent. The metabolic data acquisition unit uses a motor metabolism analysis system. The subjective assessment input unit provides a standardized RPE scale input interface. All units achieve synchronous acquisition through a unified trigger signal.

[0025] The signal preprocessing module is connected to the multimodal signal acquisition module and has a built-in sEMG signal filtering unit and AUS signal processing unit. The sEMG signal filtering unit performs bandpass filtering and notch filtering operations, the AUS signal processing unit completes time-varying gain compensation, bandpass filtering, envelope detection and logarithmic compression, and the metabolic data processing unit performs outlier removal and metabolic rate calculation.

[0026] The feature extraction module communicates with the signal preprocessing module and includes an sEMG feature extraction unit and an AUS feature extraction unit. The sEMG feature extraction unit uses the sliding window method and short-time Fourier transform (STFT) to extract time-domain and frequency-domain features. The AUS feature extraction unit extracts muscle thickness envelope features through brightness tracking and cross-correlation methods, and outputs a multi-dimensional original feature set.

[0027] The feature fusion module receives the raw feature set output by the feature extraction module, performs feature sequence truncation and Z-score standardization, and constructs a multimodal fusion feature vector through an adaptive weighted fusion algorithm to achieve complementary integration of different types of features.

[0028] The fatigue level classification module has a built-in VE and RPE joint criterion library, stores the judgment criteria corresponding to the five fatigue levels, receives metabolic data and subjective evaluation data, and outputs fatigue level labels for model training and validation.

[0029] The evaluation model module includes a model training unit and a real-time evaluation unit. The model training unit optimizes the machine learning model parameters through five-fold cross-validation. The real-time evaluation unit receives the fused feature vector output by the feature fusion module and calls the trained model to output the real-time fatigue level.

[0030] The dynamic feedback module connects the evaluation model module and the external exoskeleton control system, converting real-time fatigue level evaluation results into standardized control signals, while receiving exoskeleton operation status feedback data to provide a basis for model updates.

[0031] Furthermore, it also includes a signal synchronization calibration module, which connects to each acquisition unit of the multimodal signal acquisition module and has a built-in timestamp synchronization unit and spatial coordinate calibration unit. The timestamp synchronization unit adds a high-precision timestamp to the signal output by each acquisition unit and eliminates acquisition delay through a clock synchronization algorithm. The spatial coordinate calibration unit establishes a spatial correlation mapping of the signal acquisition position based on the installation position of the ultrasound transducer and electromyography electrode pads.

[0032] Furthermore, it also includes a data storage and analysis module, which consists of a historical database, a feature analysis unit, and a model optimization unit. The historical database stores the collected raw signals, preprocessed data, fused features, evaluation results, and feedback data in a categorized manner. The feature analysis unit regularly mines the feature change patterns under different fatigue levels and updates the feature weight coefficient library. The model optimization unit uses an incremental learning algorithm to continuously optimize and evaluate model parameters using archived data, while generating data statistical reports to provide data support for adjusting rehabilitation training programs.

[0033] Compared with existing technologies, the beneficial effects of this invention are:

[0034] Regarding the accuracy of fatigue assessment, this method innovatively integrates the time- and frequency-domain characteristics of surface electromyography (EMG) signals with muscle morphology characteristics from A-mode ultrasound, while also combining objective data on respiratory flux with a subjective fatigue perception scale to construct a multi-dimensional and multi-layered assessment system. Through feature-adaptive weighted fusion and multi-model integrated decision-making, it achieves accurate identification of different fatigue stages, compensating for the information limitations of single-signal assessments. The fatigue state is subdivided into five levels, which, compared to the traditional single-level model, better reflects the progressive physiological process of muscle fatigue, accurately capturing the dynamic changes from mild to extreme fatigue, and providing precise data support for subsequent adjustments to control strategies.

[0035] In terms of personalized adaptation capabilities, the system features dynamic threshold adjustment, cross-individual feature adaptation, and multi-scenario adaptation. By collecting resting physiological parameters of subjects to establish an individual baseline, and combining this with real-time adjustments to fatigue assessment criteria based on exercise intensity, it effectively eliminates assessment biases caused by individual differences. A feature template library is pre-set for different exercise scenarios, dynamically adjusting assessment parameters and model weights to ensure effective assessment results in various scenarios such as walking on flat ground, climbing stairs, and squatting. Cross-individual adaptation technology extracts common features through domain adaptive algorithms, enabling the model to adapt to new users with limited data, significantly expanding the applicable user base.

[0036] Regarding rehabilitation safety and comfort, the exoskeleton system dynamically generates personalized joint trajectories and adaptively adjusts impedance stiffness based on real-time fatigue assessment results. By optimizing the amplitude and phase parameters of the gait trajectory, it reduces muscle activation intensity; and dynamically increases assist stiffness according to fatigue level to compensate for decreased motor ability, reducing muscle load and energy consumption. This closed-loop control mode of "assessment-adjustment-feedback" effectively avoids the risks of muscle spasms and falls caused by over-fatigue, improving the comfort and safety of rehabilitation training and allowing patients to complete training in a safer state.

[0037] In terms of improving rehabilitation efficiency, the system optimizes muscle energy consumption, slows down the fatigue process, and extends effective training time. The dynamically adjusted exoskeleton assistance strategy precisely matches the patient's motor abilities, avoiding ineffective or insufficient assistance, helping patients better master the rhythm of movement, and promoting the recovery of motor function. Simultaneously, the data storage and analysis module archives the entire training process data, regularly mining the correlation between characteristic changes and rehabilitation effects, providing data support for the dynamic optimization of rehabilitation programs, continuously improving the targeting and effectiveness of rehabilitation training, and helping patients recover lower limb motor function more quickly. Overall, this method and system realize the transformation of lower limb rehabilitation from "experience-driven" to "data-driven," and from "uniform intervention" to "personalized adaptation," providing more precise, safe, and efficient rehabilitation assistance solutions for patients with lower limb motor dysfunction, and promoting the development of rehabilitation robot technology to a higher level. Attached Figure Description

[0038] Figure 1 This is a schematic block diagram of a method for assessing lower limb muscle fatigue based on multimodal physiological signals proposed in this invention.

[0039] Figure 2 This is a schematic block diagram of a system for assessing lower limb muscle fatigue based on multimodal physiological signals, as proposed in this invention.

[0040] Figure 3 A bar chart comparing the accuracy of fatigue assessment for different modes;

[0041] Figure 4 A biaxial line graph showing the changes in key characteristics under different fatigue levels;

[0042] Figure 5 Line graph comparing net respiratory metabolic rate under different control modes;

[0043] Figure 6 Grouped bar charts comparing muscle activation levels at different fatigue levels;

[0044] Figure 7 This is a line graph showing the change in exoskeleton stiffness coefficient with fatigue level. Detailed Implementation

[0045] 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.

[0046] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," and "counterclockwise," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0047] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified. Furthermore, the terms "installed," "connected," and "linked" should be interpreted broadly; for example, they may refer to a fixed connection, a detachable connection, or an integral connection; they may refer to a mechanical connection or an electrical connection; they may refer to a direct connection or an indirect connection through an intermediate medium; and they may refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances. The invention will now be described in further detail with reference to the accompanying drawings.

[0048] Reference Figures 1 to 7 A method for assessing lower limb muscle fatigue based on multimodal physiological signals, comprising the following steps:

[0049] The multimodal physiological signal acquisition steps selected the tibialis anterior, lateral gastrocnemius, rectus femoris and biceps femoris of the lower limb as target muscles. The surface electromyography signal acquisition device and the A-mode ultrasound signal acquisition device were used to simultaneously acquire sEMG and AUS signals during muscle contraction. At the same time, respiratory flux (VE) data were acquired through the exercise metabolism analysis system, and the subject's fatigue perception was recorded in combination with the subjective fatigue perception scale (RPE).

[0050] The signal preprocessing steps include: using 20 to 500 Hz bandpass filtering and 50 Hz notch filtering to remove power frequency interference, baseline drift and high frequency noise from sEMG signals; performing time-varying gain compensation, bandpass filtering, envelope detection and logarithmic compression preprocessing on AUS signals; and using the three-standard-deviation principle to remove outliers from respiratory data.

[0051] In the feature extraction step, the time-domain features such as mean absolute value, root mean square value, and integral electromyography value are extracted from the preprocessed sEMG signal using the sliding window analysis method. The frequency-domain features such as median frequency and mean frequency are extracted using the short-time Fourier transform (STFT). The muscle thickness envelope features are extracted from the AUS signal using the brightness tracking method and the cross-correlation method. The features are then fused to obtain a multi-dimensional feature set.

[0052] The multimodal feature fusion step involves truncating the sEMG and AUS feature sequences based on the shortest feature sequence length, flattening the three-dimensional feature matrix into a two-dimensional vector, and constructing a fused feature vector after Z-score normalization to achieve complementary fusion of temporal, frequency, and morphological features.

[0053] The fatigue level classification steps are based on the VE change pattern and RPE grading threshold to establish a joint criterion, and divide the muscle fatigue state into five levels: relaxed state, walking state, transitional state, fatigued state, and limit state, and clarify the VE range and RPE score range corresponding to each level.

[0054] The evaluation model training and output steps involve dividing the fused feature vector and fatigue level label into training and test sets using stratified sampling. The machine learning model parameters are optimized through five-fold cross-validation, and the real-time fused features are input to achieve accurate evaluation and output of muscle fatigue level.

[0055] This invention also includes a multimodal feature weighted optimization step, which designs an adaptive weighted fusion expression for features based on the differences in feature contribution at different fatigue stages.

[0056]

[0057] in To optimize the fused feature vector, The number of sEMG features. For the first sEMG feature weight coefficients For the first Standardized values ​​of sEMG features For the first Median frequency corresponding to each sEMG feature This represents the baseline median frequency of muscles at rest. This is the frequency difference adjustment coefficient. The number of AUS features. For the first AUS feature weight coefficients, For the first Standardized values ​​of AUS features, For the first The change in muscle thickness corresponding to each AUS feature. This represents the maximum change in muscle thickness. By dynamically adjusting feature weights, the contribution of effective features in fatigue assessment is enhanced, interference from noisy features is suppressed, and the relevance of feature representation is improved.

[0058] This invention also includes a dynamic optimization step for the evaluation model, introduces multi-model integration and majority voting decision-making mechanisms, and constructs an optimization expression for fatigue level decision-making as follows:

[0059]

[0060] in The final fatigue level, The number of machine learning models integrated. For the first The confidence weight coefficients of each model For indicator functions when If the value is 1, then the value is 0; otherwise, the value is 0. For the first The output level of each model The historical decision-making influence coefficient. The number of correctly classified samples of rank k in the historical data. This represents the total number of correctly classified samples in history. By integrating the outputs of multiple models with historical decision-making experience, the risk of misclassification by a single model is reduced, improving the stability and accuracy of fatigue level assessment and adapting to the differences in fatigue characteristics among different individuals.

[0061] This invention also includes a dynamic threshold adjustment step for fatigue levels. Based on the individual physiological baseline and exercise intensity of the subject, the VE and RPE judgment thresholds for each fatigue level are corrected in real time. An individual baseline is established by collecting physiological parameters of the subject at rest. Combined with intensity indicators such as speed and load during exercise, a linear interpolation method is used to adjust the threshold range for each level, making the fatigue classification more closely match individual physiological characteristics. Simultaneously, the threshold adjustment patterns of different individuals are recorded to establish a cross-individual threshold adaptation library, shortening the adaptation period for new users and improving the universality of the assessment method.

[0062] This invention also includes a cross-individual fatigue feature adaptation step. Addressing the feature distribution shift caused by differences in muscle morphology and movement habits among individuals, a domain adaptive algorithm is employed to perform transfer learning on the fused features. By extracting common representations of fatigue features from different individuals, a domain-invariant feature space is constructed to eliminate evaluation bias caused by individual differences. Simultaneously, an individual feature bias compensation model is established to quickly adjust the evaluation model parameters based on the initial movement data of new users, enabling the model to achieve accurate evaluation with limited data and expanding the applicability of the method.

[0063] This invention also includes real-time feedback of assessment results and model updates, outputting fatigue assessment results to the exoskeleton control system in real time to provide data support for adaptive control. Simultaneously, assessment data, actual fatigue state feedback, and control effect data are archived at fixed intervals. An incremental learning algorithm is used to update the assessment model parameters and feature weights, enabling the model to continuously adapt to the dynamic changes in the subject's fatigue characteristics, thereby improving the accuracy and reliability of long-term assessments.

[0064] This invention also includes a multi-scenario fatigue assessment adaptation step, which pre-sets a scenario feature template library for different exercise scenarios such as walking on flat ground, climbing stairs, and squatting. By identifying the current exercise scenario type, the corresponding feature extraction parameters and assessment model weights are called to adjust the sensitivity of each fatigue level determination. For example, in high-intensity exercise scenarios, the response speed for determining fatigue state and limit state is improved, and in low-intensity rehabilitation scenarios, the accuracy of distinguishing between relaxed state and walking state is optimized, ensuring the effectiveness of assessment results in different scenarios.

[0065] This invention discloses a lower limb muscle fatigue assessment system based on multimodal physiological signals, comprising the following modules:

[0066] The multimodal signal acquisition module consists of a surface electromyography (SEMG) acquisition unit, an A-mode ultrasound acquisition unit, a metabolic data acquisition unit, and a subjective assessment input unit. The SEMG acquisition unit uses silver chloride electrode pads attached to the target muscle belly. The A-mode ultrasound acquisition unit is equipped with a wireless four-channel amplitude modulation device and an ultrasound coupling agent. The metabolic data acquisition unit uses a motor metabolism analysis system. The subjective assessment input unit provides a standardized RPE scale input interface. All units achieve synchronous acquisition through a unified trigger signal.

[0067] The signal preprocessing module is connected to the multimodal signal acquisition module and has a built-in sEMG signal filtering unit and AUS signal processing unit. The sEMG signal filtering unit performs bandpass filtering and notch filtering operations, the AUS signal processing unit completes time-varying gain compensation, bandpass filtering, envelope detection and logarithmic compression, and the metabolic data processing unit performs outlier removal and metabolic rate calculation.

[0068] The feature extraction module communicates with the signal preprocessing module and includes an sEMG feature extraction unit and an AUS feature extraction unit. The sEMG feature extraction unit uses the sliding window method and short-time Fourier transform (STFT) to extract time-domain and frequency-domain features. The AUS feature extraction unit extracts muscle thickness envelope features through brightness tracking and cross-correlation methods, and outputs a multi-dimensional original feature set.

[0069] The feature fusion module receives the raw feature set output by the feature extraction module, performs feature sequence truncation and Z-score standardization, and constructs a multimodal fusion feature vector through an adaptive weighted fusion algorithm to achieve complementary integration of different types of features.

[0070] The fatigue level classification module has a built-in VE and RPE joint criterion library, stores the judgment criteria corresponding to the five fatigue levels, receives metabolic data and subjective evaluation data, and outputs fatigue level labels for model training and validation.

[0071] The evaluation model module includes a model training unit and a real-time evaluation unit. The model training unit optimizes the machine learning model parameters through five-fold cross-validation. The real-time evaluation unit receives the fused feature vector output by the feature fusion module and calls the trained model to output the real-time fatigue level.

[0072] The dynamic feedback module connects the evaluation model module and the external exoskeleton control system, converting real-time fatigue level evaluation results into standardized control signals, while receiving exoskeleton operation status feedback data to provide a basis for model updates.

[0073] This invention also includes a signal synchronization calibration module, which connects to each acquisition unit of the multimodal signal acquisition module and incorporates a timestamp synchronization unit and a spatial coordinate calibration unit. The timestamp synchronization unit adds a high-precision timestamp to the signals output by each acquisition unit and eliminates acquisition delay through a clock synchronization algorithm. The spatial coordinate calibration unit establishes a spatial correlation mapping of the signal acquisition positions based on the installation positions of the ultrasound transducer and electromyography electrode pads, ensuring the spatiotemporal matching of sEMG signals and AUS signals and improving the accuracy of multimodal fusion.

[0074] This invention also includes a data storage and analysis module, which consists of a historical database, a feature analysis unit, and a model optimization unit. The historical database stores the collected raw signals, preprocessed data, fused features, evaluation results, and feedback data in a categorized manner. The feature analysis unit periodically mines the feature change patterns under different fatigue levels and updates the feature weight coefficient library. The model optimization unit uses an incremental learning algorithm to continuously optimize and evaluate model parameters using archived data, while generating data statistical reports to provide data support for adjusting rehabilitation training programs.

[0075] The following two examples further illustrate the specific implementation of this system:

[0076] Example 1

[0077] Muscle fatigue assessment and exoskeleton fitting for lower limb rehabilitation training in healthy individuals

[0078] This embodiment is applied to lower limb rehabilitation training scenarios for healthy individuals. It addresses the needs of sports enthusiasts and those in the postoperative recovery period for lower limb muscle fatigue monitoring and scientific training. By accurately assessing fatigue status through multimodal physiological signals, it links with exoskeleton devices to provide personalized assistance, avoiding muscle damage caused by overtraining and improving the scientific nature and safety of training.

[0079] During the multimodal physiological signal acquisition phase, the tibialis anterior, lateral gastrocnemius, rectus femoris, and biceps femoris were selected as target muscles. The surface electromyography (EMG) acquisition unit used silver chloride electrodes, which were attached to the muscle belly along the muscle fiber direction to ensure close contact with the skin, with a sampling rate set to 2000 Hz. The A-mode ultrasound acquisition unit was equipped with a wireless four-channel device; the transducer surface was coated with ultrasound coupling agent and attached to the corresponding skin area of ​​the target muscle, with a sampling rate of 20 Hz. The exercise metabolic analysis system simultaneously acquired respiratory rate and tidal volume, calculating respiratory flux (VE). All acquisition units were synchronized through a unified trigger signal to avoid data timing deviations, and a standardized RPE scale was provided for subjects to record their subjective fatigue feelings in real time.

[0080] In the signal preprocessing stage, surface electromyography (EMG) signals are subjected to 20-500Hz bandpass filtering to remove high-frequency noise and low-frequency baseline drift, and 50Hz notch filtering to suppress power frequency interference. A-mode ultrasound signals undergo time-varying gain compensation, bandpass filtering, Hilbert transform envelope detection, and moving average smoothing sequentially. Respiratory data are processed using a three-standard-deviation principle to remove abnormal fluctuations, ensuring data reliability. The preprocessed signals are then subjected to a high-precision timestamp via a signal synchronization calibration module to establish a spatial correlation mapping, ensuring spatiotemporal matching between surface EMG and A-mode ultrasound signals.

[0081] In the feature extraction stage, surface electromyography (EMG) signals were used with a 50ms sliding window and a 25ms overlap rate to extract temporal features such as mean absolute value, root mean square (RMS), and integral EMG values. Short-time Fourier transform was used to extract frequency domain features such as median frequency and mean frequency. For A-mode ultrasound signals, the muscle surface boundary was located using brightness tracking, and time offset was corrected using cross-correlation. After fusion, muscle thickness envelope features were extracted. In the feature fusion stage, the shortest feature sequence length was used as a benchmark for truncation. After Z-score standardization, an adaptive weighting algorithm was used to dynamically adjust the feature weights of the two signal types, enhancing the contribution of effective features.

[0082] In the fatigue level classification stage, a joint criterion was established based on VE changes and RPE grading: a gradual increase in VE and an RPE of 6 to 8 indicated a relaxed state; an RPE of 9 to 10 indicated a leisurely state; an RPE of 11 to 14 and a significant increase in VE indicated a transitional state; VE exceeding 43 L / min and an RPE of 15 to 18 indicated a fatigued state; and an RPE ≥ 19 indicated a limiting state. The evaluation model employed a support vector machine algorithm, using stratified sampling to divide the training and test sets, five-fold cross-validation to optimize parameters, and a short 32ms analysis window combined with majority voting to improve the evaluation accuracy to meet real-time control requirements.

[0083] During the exoskeleton control and linkage phase, the system dynamically generates personalized joint trajectories based on real-time fatigue levels. By fusing standard templates and fatigue-adapted trajectories with weighted coefficients, it adjusts the joint range of motion and angular velocity. The variable stiffness impedance control strategy adjusts the stiffness coefficients according to the fatigue level; the higher the fatigue level, the greater the assist stiffness, compensating for decreased mobility. The dynamic feedback module transmits the evaluation results to the exoskeleton actuators in real time and simultaneously receives equipment operating status data, providing a basis for model updates. The data storage and analysis module categorizes and archives raw signals, feature data, and evaluation results. The feature analysis unit periodically mines fatigue feature change patterns and updates the feature weight coefficient library.

[0084] Table 1 Comparison of the effects of rehabilitation training on healthy individuals

[0085] Evaluation index Performance of traditional single signal evaluation mode Performance of the multi-modal evaluation system of the present application Fatigue evaluation accuracy 70-75 91-94 Muscle activation peak value 0.7-0.8 0.45-0.55 Net respiratory metabolic rate relative value 1.0 0.7-0.75 Fatigue process delay time minutes 0 15-20

[0086] Table 1 clearly demonstrates the advantages of this invention in rehabilitation training for healthy individuals. Traditional single-signal assessment methods rely on single data from surface electromyography or ultrasound, resulting in low accuracy and an inability to comprehensively reflect muscle status. This leads to excessive muscle activation, high energy consumption, and no fatigue-delaying effect. This invention significantly improves fatigue assessment accuracy through multimodal signal fusion and adaptive weighted optimization. It dynamically adjusts the exoskeleton assistance strategy, significantly reducing peak muscle activation and energy consumption, effectively delaying the fatigue process. Synchronous calibration and spatiotemporal matching technologies ensure data reliability, while majority voting and model optimization enhance assessment stability, providing precise data support for scientific training in healthy individuals and avoiding excessive fatigue injuries.

[0087] Example 2

[0088] Assessment and Personalized Rehabilitation Training of Lower Limb Muscle Fatigue in Stroke Patients

[0089] This embodiment is applied to the rehabilitation training scenario of patients with lower limb motor dysfunction after stroke. The patients have symptoms such as muscle weakness and spasticity, low muscle tolerance, and are prone to safety risks due to fatigue. It is necessary to improve the rehabilitation effect through accurate fatigue assessment and personalized exoskeleton assistance while ensuring safety.

[0090] During the multimodal physiological signal acquisition phase, the electrode pads and ultrasound transducer placement were adjusted to avoid areas prone to spasticity, taking into account the patient's muscle atrophy and spasticity characteristics. The electrode pads were made of low-sensitivity materials to minimize skin irritation. The surface electromyography (EMG) acquisition unit had a sampling rate of 2000 Hz, and the A-mode ultrasound acquisition unit had a sampling rate of 20 Hz. The exercise metabolism analysis system simplified the operation process, facilitating patient cooperation in collecting respiratory data. The subjective assessment input unit provided voice prompts to assist patients in accurately completing the Respiratory Physical Examination (RPE) scale. All acquisition units achieved temporal and spatial synchronization through a signal synchronization calibration module, eliminating signal deviations caused by involuntary limb movements.

[0091] In the signal preprocessing stage, considering the low signal-to-noise ratio and high interference of the patient's surface electromyography (EMG) signals, enhanced filtering is implemented by adding an adaptive noise suppression algorithm to the existing bandpass and notch filtering. For A-mode ultrasound signal preprocessing, a signal enhancement step is added to improve the accuracy of muscle boundary recognition. Respiratory data is further smoothed using a sliding window method to reduce fluctuations caused by irregular breathing. The preprocessed data is then transmitted to the feature extraction module, while the data storage and analysis module backs up the original data in real time.

[0092] In the feature extraction stage, the sliding window length of the surface electromyography (EMG) signal was adjusted to 80ms with an overlap rate of 25ms. Time-domain features, including mean absolute value, root mean square (RMS), and integral EMG values, were extracted. Frequency-domain features, including median and mean frequencies, were extracted using short-time Fourier transform. A-mode ultrasound signals were optimized with brightness tracking and cross-correlation fusion algorithms to adapt to the small range of muscle thickness variation in patients. In the feature fusion stage, cross-individual fatigue feature adaptation technology was introduced to extract common characteristics of fatigue features from different patients, constructing a domain-invariant feature space. This space was then used to quickly adjust model parameters based on initial motion data from new patients, eliminating assessment bias caused by individual differences.

[0093] During the fatigue level classification phase, individual baselines are established based on patients' resting physiological parameters. Linear interpolation is used to dynamically adjust the VE and RPE thresholds for each level, taking into account the intensity of rehabilitation training. For example, the VE thresholds for transitional and fatigue states are appropriately lowered for patients with weaker muscle strength. The assessment model employs an incremental learning algorithm, periodically updating parameters using archived data to adapt to the dynamic changes in muscle function during the patient's rehabilitation process. A majority voting method is also used to reduce the risk of misclassification.

[0094] In the multi-scenario adaptation phase, a feature template library is pre-set for two core rehabilitation scenarios: walking on flat ground and climbing stairs. The system identifies the current scenario type through environmental sensors, calls the corresponding feature extraction parameters and model weights, optimizes the recognition accuracy of relaxed and strolling states in the walking on flat ground scenario, and improves the response speed of transitional and fatigued states in the climbing stairs scenario. The exoskeleton control strategy is tailored to the characteristics of muscle weakness in patients. In the initial stage, the assistive stiffness is increased, and the range of motion of the joints is gradually adjusted as the fatigue level increases to avoid excessive joint movement that may cause spasms.

[0095] The dynamic feedback module reduces data transmission latency, ensuring that assessment results drive exoskeleton adjustments in real time. Administrators use statistical reports generated by the data storage and analysis module to dynamically optimize rehabilitation training programs. The signal synchronization calibration module continuously corrects spatiotemporal deviations, ensuring the accuracy of multimodal data fusion. The feature analysis unit mines the correlation between patient fatigue characteristics and rehabilitation outcomes, providing a basis for personalized program adjustments.

[0096] Table 2 Comparison of the effects of rehabilitation training on stroke patients

[0097] Evaluation index Performance of traditional rehabilitation evaluation mode Performance of the multi-modal evaluation system of the present application Fatigue evaluation accuracy 65-70 88-92 Rehabilitation training comfort score 5-6 8-9 Incidence of adverse events 18-22 3-5 Lower limb motor function recovery rate relative value 1.0 1.5-1.8

[0098] Table 2 data highlights the application value of this invention in stroke patient rehabilitation scenarios. Traditional rehabilitation assessment methods rely on manual observation and single signals, resulting in low accuracy and an inability to accurately capture patient fatigue levels, leading to poor training comfort, frequent adverse events such as muscle spasms and falls, and low rehabilitation efficiency. This invention achieves accurate assessment of patient fatigue levels through cross-individual adaptation and dynamic threshold adjustment. Personalized exoskeleton assistance strategies reduce muscle load and safety risks, improving training comfort. Multi-scenario adaptation and incremental learning technology adapt to the patient's rehabilitation process, and data-driven solution optimization accelerates motor function recovery. Synchronous calibration and low-sensitivity acquisition design ensure signal reliability and patient tolerance, providing stroke patients with a safe, efficient, and personalized rehabilitation assistance solution. (Reference) Figure 3 This figure visually demonstrates the core advantages of the multimodal signal fusion method of this invention, overcoming the industry pain point of the one-sidedness of traditional single-modal assessment. A single sEMG modality can only capture muscle electrical activity characteristics, is easily affected by power frequency interference and baseline drift, and has an accuracy of only 41%-45%, failing to comprehensively reflect muscle fatigue status. A single AUS modality focuses on muscle morphological changes, and although its accuracy is better than sEMG (64%-68%), it lacks neural electrical activity information support, resulting in a single assessment dimension. This invention integrates the time-domain and frequency-domain features of sEMG with the muscle thickness envelope features of AUS, achieving complementary advantages between the two signal types through Z-score standardization and adaptive weighted fusion. The multimodal fusion accuracy is stable at 91%-94%, significantly surpassing single-modality assessment, and further optimizes the misclassification problem through majority voting, meeting the assessment accuracy requirements of real-time control and providing accurate data support for subsequent exoskeleton adaptive control.

[0099] Reference Figure 4This figure clearly presents the synergistic changes of multimodal features during the muscle fatigue process, providing a physiological basis for fatigue level classification. As the fatigue level escalates from the relaxed state to the extreme state, the median frequency of sEMG continuously decreases (45-98Hz). This change is related to the physiological mechanisms of reduced muscle fiber conduction velocity and continuous recruitment of high-threshold motor units, reflecting a decline in the regulatory efficiency of the neuromuscular system. Meanwhile, AUS muscle thickness gradually increases (21-31mm), stemming from metabolic swelling of the muscle during fatigue, consistent with the positive correlation between muscle thickness changes and fatigue level confirmed by Alvarez's team. The inverse synergistic changes of these two features confirm the scientific validity of multimodal fusion—sEMG reflects electrophysiological activity, while AUS captures morphological changes. The combination of the two achieves a comprehensive characterization of the fatigue state, providing quantitative support for the fine classification of the five levels and avoiding the one-sidedness of single-feature assessment.

[0100] Reference Figure 5 This figure highlights the significant effect of the adaptive control strategy of this invention in reducing energy consumption and delaying fatigue. Without the exoskeleton, muscles bear the full exercise load, and with each trial, the metabolic rate continuously increases (4.1-7.8 W / kg), leading to rapid fatigue accumulation. Traditional exoskeletons use a fixed control mode, providing only passive assistance, which cannot adapt to changes in muscle fatigue levels, resulting in limited reduction in metabolic rate (3.9-7.4 W / kg), and a continued rapid increase in the later stages. This invention, based on real-time fatigue assessment results, dynamically adjusts joint trajectory and impedance stiffness—reducing joint range of motion and increasing assistance stiffness as fatigue deepens, effectively compensating for decreased exercise capacity. Under adaptive control mode, the metabolic rate remains at a low level (3.3-6.1 W / kg) with a gradual increase, significantly delaying the non-linear spike in metabolic rate, extending the effective rehabilitation training time, while reducing muscle load and improving training comfort and safety.

[0101] Reference Figure 6 This figure visually illustrates the optimization effect of the system on muscle load, solving the problem of excessive muscle activation in traditional rehabilitation training. Without an exoskeleton, muscles must independently support body weight and complete movements. As fatigue levels increase, peak muscle activation continuously rises (0.71-0.80), and high activation easily leads to muscle spasms, over-fatigue, and even the risk of falls. This invention dynamically generates personalized joint trajectories and uses variable stiffness impedance control to precisely match the assistance intensity according to fatigue levels—the more severe the fatigue, the greater the assistance stiffness, effectively sharing the muscle load. Under system control, peak muscle activation is stabilized at 0.46-0.55, a reduction of 21%-37% compared to without an exoskeleton, and the increase with fatigue levels is gradual. Low activation not only reduces energy consumption but also reduces the risk of muscle injury, allowing patients to complete rehabilitation training in a safer and more comfortable state, especially suitable for stroke patients and other individuals with low muscle tolerance.

[0102] refer to Figure 7 This diagram illustrates the scientific basis of the variable stiffness impedance control strategy of this invention, solving the problem that traditional exoskeletons with fixed stiffness cannot adapt to changes in fatigue state. Traditional exoskeletons, with their fixed stiffness design, cannot respond to the decline in motor ability caused by muscle fatigue, and are prone to insufficient assistance or excessive impact. This invention dynamically adjusts the stiffness coefficient based on real-time fatigue levels, gradually increasing from 8.2 N·m / rad in the relaxed state to 24.1 N·m / rad in the extreme state, exhibiting an exponential growth trend, which better matches the physiological characteristics of non-linear decay in output capacity after muscle fatigue. The stiffness is lower in the relaxed state, preserving the patient's voluntary movement space; the stiffness is significantly increased in the fatigue and extreme states, enhancing the assistance effect, compensating for insufficient muscle strength, and preventing joint spasms caused by excessive movement. This dynamically adaptable stiffness adjustment not only ensures the effectiveness of rehabilitation training but also improves the safety and comfort of human-computer interaction, allowing the exoskeleton to adapt to the personalized needs of different fatigue stages throughout the entire process.

[0103] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A lower limb muscle fatigue assessment method based on multi-modal physiological signals, characterized in that, Comprising the following steps: A multi-modal physiological signal acquisition step, selecting the tibialis anterior muscle, lateral gastrocnemius muscle, rectus femoris muscle, and biceps femoris muscle as target muscles, using a surface electromyography signal acquisition device and an A-mode ultrasonic signal acquisition device to synchronously acquire sEMG signals and AUS signals during muscle contraction, simultaneously acquiring ventilation data VE through a metabolic analysis system, and recording the fatigue perception of the subjects through a subjective fatigue perception scale RPE; A signal preprocessing step, using 20-500 Hz band-pass filtering and 50 Hz notch filtering to remove power frequency interference, baseline drift, and high-frequency noise for sEMG signals, and performing time-varying gain compensation, band-pass filtering, envelope detection, and logarithmic compression preprocessing for AUS signals, and using a three-sigma rule to remove abnormal points for respiratory data; A feature extraction step, using a sliding window analysis method to extract time-domain features such as average absolute value, root mean square, and integrated electromyography value for preprocessed sEMG signals, using short-time Fourier transform STFT to extract frequency-domain features such as median frequency and mean frequency, using brightness tracking and cross-correlation methods to extract muscle thickness envelope features for AUS signals, and fusing to obtain a multi-dimensional feature set; A multi-modal feature fusion step, truncating sEMG and AUS feature sequences based on the shortest feature sequence length, flattening a three-dimensional feature matrix into a two-dimensional vector, and constructing a fusion feature vector after Z-score standardization to achieve complementary fusion of time-domain, frequency-domain, and morphological features; A fatigue level classification step, establishing a joint criterion based on the VE variation law and RPE classification threshold, dividing muscle fatigue states into five levels: relaxed state, walking state, transition state, fatigue state, and limit state, and dividing the corresponding VE range and RPE score interval for each level; An evaluation model training and output step, dividing the training set and test set according to stratified sampling of the fusion feature vector and fatigue level label, optimizing machine learning model parameters through five-fold cross-validation, and inputting real-time fusion features to achieve accurate evaluation and output of muscle fatigue levels.

2. The method of claim 1, wherein, Further comprising a multi-modal feature weighting optimization step, designing a feature adaptive weighted fusion expression for different fatigue stage feature contribution differences as in To optimize the fused feature vector, The number of sEMG features. For the first sEMG feature weight coefficients For the first Standardized values ​​of sEMG features For the first Median frequency corresponding to each sEMG feature This represents the baseline median frequency of muscles at rest. This is the frequency difference adjustment coefficient. The number of AUS features. For the first AUS feature weight coefficients, For the first Standardized values ​​of AUS features, For the first The change in muscle thickness corresponding to each AUS feature. This represents the maximum change in muscle thickness.

3. The method of claim 1, wherein, Further comprising an evaluation model dynamic optimization step, introducing multi-model integration and majority voting decision mechanisms, and constructing a fatigue level decision optimization expression as wherein is the final fatigue level, is the number of integrated machine learning models, is the confidence weight coefficient of the th model, is the indicator function that takes 1 when and 0 otherwise, is the output level of the th model, is the historical decision impact coefficient, is the number of correctly classified samples of level in the historical data, is the total number of correctly classified samples in the historical data.

4. The method of claim 1, wherein, Further comprising a fatigue level dynamic threshold adjustment step, based on the individual physiological baseline and exercise intensity of the subjects, real-time correcting the VE and RPE determination thresholds of each fatigue level; by collecting the physiological parameters of the subjects in a resting state to establish an individual baseline, combining intensity indicators such as speed and load during exercise, and using linear interpolation to adjust the threshold range of each level, the fatigue division is more in line with individual physiological characteristics; at the same time, record the threshold adjustment law of different individuals, and establish a cross-individual threshold adaptation library.

5. The method of claim 1, wherein, It also includes a cross-individual fatigue feature adaptation step. In order to solve the problem of feature distribution deviation caused by different individual muscle morphology and exercise habit, a domain adaptation algorithm is used for transfer learning of the fused features; by extracting the common representation of different individual fatigue features, a domain invariant feature space is constructed to eliminate the evaluation deviation caused by individual differences; at the same time, an individual feature deviation compensation model is established, and the evaluation model parameters are quickly adjusted according to the initial exercise data of the new user.

6. The method of claim 1, wherein, It also includes an evaluation result real-time feedback and model updating step. The fatigue evaluation results are output to the exoskeleton control system in real time to provide data support for adaptive control; at the same time, the evaluation data, actual fatigue state feedback and control effect data are archived at a fixed period, and the evaluation model parameters and feature weights are updated using an incremental learning algorithm to make the model continuously adapt to the dynamic changes of the fatigue features of the subjects, improving the accuracy and reliability of long-term evaluation.

7. The method of claim 1, wherein, It also includes a multi-scene fatigue evaluation adaptation step. For different exercise scenes, a scene feature template library is preset; By recognizing the current exercise scene type, the corresponding feature extraction parameters and evaluation model weights are called to adjust the sensitivity of each fatigue level.

8. The multi-modal physiological signal based lower limb muscle fatigue assessment system according to any one of claims 1-7, characterized in that, It includes the following modules: A multi-modal signal acquisition module composed of a surface electromyography acquisition unit, an A-mode ultrasound acquisition unit, a metabolic data acquisition unit and a subjective evaluation input unit. The surface electromyography acquisition unit uses silver chloride electrode sheets to adhere to the target muscle belly. The A-mode ultrasound acquisition unit is equipped with a wireless four-channel amplitude modulation device and an ultrasonic coupling agent. The metabolic data acquisition unit uses a sports metabolic analysis system. The subjective evaluation input unit provides a standardized RPE scale input interface. Each unit synchronously acquires through a unified trigger signal; A signal preprocessing module connected to the multi-modal signal acquisition module, with an sEMG signal filtering unit and an AUS signal processing unit built-in. The sEMG signal filtering unit performs band-pass filtering and notch filtering operations. The AUS signal processing unit completes time-varying gain compensation, band-pass filtering, envelope detection and logarithmic compression. The metabolic data processing unit performs outlier rejection and metabolic rate calculation; A feature extraction module in communication with the signal preprocessing module, containing an sEMG feature extraction unit and an AUS feature extraction unit. The sEMG feature extraction unit uses a sliding window method and a short-time Fourier transform (STFT) to extract time-domain and frequency-domain features. The AUS feature extraction unit extracts muscle thickness envelope features through brightness tracking and cross-correlation methods, and outputs a multi-dimensional original feature set; A feature fusion module that receives the original feature set output by the feature extraction module, performs feature sequence truncation and Z-score standardization processing, and constructs a multi-modal fusion feature vector through an adaptive weighted fusion algorithm to realize complementary integration of different types of features; A fatigue level division module with a VE and RPE joint criterion library built-in, storing the determination criteria corresponding to the five fatigue levels, receiving metabolic data and subjective evaluation data, and outputting fatigue level labels for model training and verification. The evaluation model module includes a model training unit and a real-time evaluation unit. The model training unit optimizes the parameters of the machine learning model through five-fold cross-validation. The real-time evaluation unit receives the fused feature vector output by the feature fusion module and calls the trained model to output real-time fatigue levels. The dynamic feedback module connects the evaluation model module and the external exoskeleton control system, converts the real-time fatigue level evaluation results into standardized control signals, and receives exoskeleton operation state feedback data to provide a basis for model updating.

9. The lower limb muscle fatigue assessment system based on multi-modal physiological signals according to claim 8, characterized in that, The signal synchronization calibration module connects each acquisition unit of the multi-modal signal acquisition module and includes a timestamp synchronization unit and a spatial coordinate calibration unit. The timestamp synchronization unit adds high-precision timestamps to the signals output by each acquisition unit and eliminates acquisition delays through a clock synchronization algorithm. The spatial coordinate calibration unit establishes a spatial correlation mapping of signal acquisition positions based on the installation positions of the ultrasonic transducer and the electromyography electrode.

10. The lower limb muscle fatigue assessment system based on multi-modal physiological signals according to claim 8, characterized in that, The data storage and analysis module includes a historical database, a feature analysis unit, and a model optimization unit. The historical database stores the collected raw signals, preprocessed data, fused features, evaluation results, and feedback data. The feature analysis unit regularly mines the feature change rules under different fatigue levels, updates the feature weight coefficient library, and uses an incremental learning algorithm to continuously optimize the evaluation model parameters using archived data while generating data statistical reports to provide data support for rehabilitation training scheme adjustments.

Images