Muscle fatigue monitoring method based on semi-supervised learning

Abstract

The invention discloses a muscle fatigue monitoring method based on semi-supervised learning, and belongs to the technical field of neuromuscular physiology and data processing. The problems that in an existing muscle fatigue monitoring method based on surface electromyogram signals, noise sensitivity is high, time-frequency domain feature extraction is insufficient, the model generalization ability is insufficient, and large-scale data labeling is excessively depended are solved. The method comprises the following steps: acquiring and preprocessing data through a surface myoelectricity sensor, constructing a self-supervised pre-training model containing a mask proxy task and a 1D CNN encoder, performing fine adjustment by using a semi-supervised learning framework enhanced by difference data, finally deploying the model, and outputting a muscle fatigue level through dynamic window slicing and confidence evaluation. The muscle fatigue monitoring ability can be effectively improved, the noise immunity of the model is enhanced, the potential structure of data is mined, fatigue features are accurately extracted, the generalization ability is improved, real-time monitoring, early warning and rehabilitation training optimization can be achieved, and powerful support is provided for exercise safety and rehabilitation treatment.

Description

Technical Field

[0001] This invention relates to the fields of neuromuscular physiology and data processing technology, and in particular to a method for monitoring muscle fatigue based on semi-supervised learning. Background Technology

[0002] Existing methods for monitoring muscle fatigue are often traditional, based on surface electromyography (sEMG) or inertial measurement units (IMUs). However, sEMG signals are susceptible to motion artifacts, power line interference, and environmental noise, resulting in low accuracy in feature extraction. Furthermore, the complex feature extraction and model calculations can affect the efficiency of real-time monitoring. IMU-based methods extract motion features (such as acceleration amplitude, angular velocity, and attitude angle) from accelerometer and gyroscope data to assess the impact of muscle fatigue on movement patterns. However, accelerometer and IMU data reflect changes in movement patterns rather than direct muscle activity, which may lead to inaccurate fatigue assessments. In addition, external vibrations and shocks can affect the accuracy of sensor data, making it difficult to adapt to complex real-world environments.

[0003] sEMG non-invasively acquires electrical signals from the muscle surface using patch-type electrode sensors. It allows for long-term, real-time monitoring of muscle activity without discomfort to the subject, making it suitable for continuous monitoring during daily activities, exercise training, or rehabilitation. However, sEMG signals often contain a large amount of redundant information, especially between adjacent channels where signals exhibit strong correlations. Traditional deep learning methods often fail to explicitly process this redundant information, potentially wasting computational resources and hindering effective model learning. Furthermore, while deep learning can automatically learn signal features, traditional deep learning models (such as fully connected neural networks and standard CNNs) struggle to directly capture the details of time and frequency domain information when processing sEMG, particularly at multi-scale features. This can prevent them from effectively extracting key features representing muscle fatigue, hindering effective identification and monitoring. Summary of the Invention

[0004] This invention provides a muscle fatigue monitoring method based on semi-supervised learning. It combines self-supervised pre-training and semi-supervised learning, relying on a large amount of unlabeled data and a small amount of labeled data to efficiently classify the collected raw data, thereby achieving muscle fatigue monitoring. Simultaneously, during self-supervised pre-training, a time-domain and frequency-domain consistency loss is employed, utilizing the complementary information provided by time-domain and frequency-domain features to comprehensively and efficiently extract key features representing muscle fatigue. In the semi-supervised learning, a consistency regularization method is introduced to ensure consistent output of the model on data with different perturbations, deeply exploring the latent structure of unlabeled data and learning more effective data representations. This addresses the problems of strong noise sensitivity, insufficient time-frequency domain feature extraction, insufficient model generalization ability, and over-reliance on large-scale labeled data in existing muscle fatigue monitoring methods based on surface electromyography (EMG) signals.

[0005] A muscle fatigue monitoring method based on semi-supervised learning, the method comprising the following steps:

[0006] S1. Collect raw data of the target muscle group using a surface electromyography sensor, and perform noise reduction preprocessing on the raw data;

[0007] S2. Construct a self-supervised pre-trained model that includes a masked proxy task and a 1D CNN encoder, and use the time-domain and frequency-domain consistency loss function for feature extraction training.

[0008] S3. A semi-supervised learning framework with differential data augmentation is used to fine-tune the self-supervised pre-trained model, and its parameters are optimized by fusing supervised classification loss and consistency regularization loss.

[0009] S4. Deploy the fine-tuned self-supervised pre-trained model as a real-time monitoring system, and output the muscle fatigue level through dynamic window slicing and confidence assessment.

[0010] Furthermore, S1 includes the following steps:

[0011] S11. Acquire signals from the target muscle group using a sensor with a bipolar electrode configuration;

[0012] S12. Power frequency interference is eliminated by using a 50 / 60Hz notch filter.

[0013]

[0014] Where, ω0=2πf0 / f s f0 is the power frequency, f0 is 50 / 60Hz, f s Where r is the sampling rate and r is the damping coefficient;

[0015] S13. A low-pass filter with a cutoff frequency of 500Hz is used to remove high-frequency noise.

[0016]

[0017] Where b and a are filter coefficients.

[0018] Furthermore, S2 includes the following steps:

[0019] S21. Randomly select a starting time point to generate a masking interval and construct a partially masked input signal:

[0020] S22. Use a 1D CNN encoder to extract features and a decoder to reconstruct the occluded region;

[0021] S23. Optimize the model by jointly calculating the time-domain reconstruction loss and the frequency-domain spectral consistency loss.

[0022] Furthermore, S21 includes the following steps:

[0023] S211. Construct the input layer: Accept the raw data, let the input data be...

[0024] X = [x1, x2, ..., x T ]∈R N×D

[0025] Where T is the time step, N is the number of samples, and D is the feature dimension of each sample;

[0026] S212. Construct the mask layer: Randomly select a starting time point t. start , satisfying 1≤t start ≤TL, where L is the length of the mask interval, resulting in a mask interval of [t] start ,t start +L), generate a mask matrix M∈{0,1} T×D , where each element m t,d :

[0027]

[0028] That is, the value within the mask range is 0, and the rest are 1.

[0029] The mask ratio is controlled by adjusting the mask interval length L. If the mask ratio is p... mask ,but

[0030] S213. Masked Input: Combine the mask matrix with the original data to generate a partially masked input X. masked :

[0031] X masked =X·M+[MASK]·(1-M)

[0032] [MASK] is a mask placeholder;

[0033] S22 includes the following steps:

[0034] S221. Constructing the encoder: Due to the obvious local temporal features of the original data, a 1D CNN Encoder is used as the encoder.

[0035] H = Encoder(X) masked ,θ e )

[0036] Initially, X masked As the initial input to the network:

[0037] H (0) =X masked

[0038] The output H of layer L (l) H is output from the previous layer (l-1) The result is obtained through convolution and nonlinear activation:

[0039] H (l) =RELU(Conv1D(H (l-1) W (l) ))

[0040] Among them, W (l) For convolution kernel weights,

[0041] After L layers of convolution, the final output of the encoder is:

[0042] H = H (L) ∈R T×D ;

[0043] S222. Constructing a decoder: The latent space representation is processed by the decoder to recover the signal waveform features within the unmasked region of the input signal, wherein the decoder only processes the feature vectors associated with the unmasked region.

[0044] Reconstructing the original data from latent feature H :

[0045]

[0046] S23 includes the following steps:

[0047] S231. Reconstruction Loss Function: A reconstruction loss function is constructed by calculating the prediction error of the masked region, forcing the self-supervised pre-trained model to learn the contextual features of the signal.

[0048] Time domain error :

[0049]

[0050] Frequency domain error :

[0051]

[0052] Spectral consistency loss:

[0053]

[0054] S232. Minimize the reconstruction loss using gradient descent and update the encoder and decoder parameters:

[0055]

[0056] Furthermore, S3 includes the following steps:

[0057] S31. Apply different data augmentations to the labeled data to generate dual-view input;

[0058] S32. Calculate the supervised classification loss for labeled samples;

[0059] S33. Enhance the consistency of predictions between views by using mean square error constraints;

[0060] S34. Dynamically adjust the combined weights of supervision loss and consistency loss.

[0061] Furthermore, S31 includes the following steps:

[0062] S311, Data Augmentation: For each labeled sample x, apply two different data augmentation operations to obtain two augmented versions:

[0063] x1 = A1(x), x2 = A2(x)

[0064] A1 and A2 are different data augmentation operations, belonging to the non-consistent time-domain transformation type, including but not limited to combinations of amplitude adjustment augmentation and time dimension manipulation augmentation, wherein:

[0065] The amplitude adjustment enhancements include adding Gaussian noise and random gain adjustment.

[0066] The time-dimensional adjustment enhancements include random cropping, time stretching, and cyclic translation.

[0067] S312, Forward Propagation: Calculate the prediction distributions of the two enhanced versions using the pre-trained encoder model:

[0068] p1=f(x1;θ), p2=f(x2;θ)

[0069] Where f(·θ) represents the encoder model, and p1 and p2 are the predicted class probability distributions;

[0070] In S32, the standard supervised loss is calculated using one version of the prediction and the true label y.

[0071]

[0072] Where I[y=c] is an indicator function, and c represents the category index;

[0073] In S33, the consistency loss is constructed using the mean square error between p1 and p2:

[0074]

[0075] S34 includes the following steps:

[0076] S341, Combined Total Loss: The final training loss is the weighted sum of the supervision loss and the consistency regularization loss.

[0077]

[0078] Here, λ is a hyperparameter used to balance the relative weights of the supervised task and the consistency regularization. λ is set to a small value in the early stage of training and then gradually increased in the later stage so that the self-supervised pre-trained model can first focus on learning the basic task and then gradually strengthen the robustness constraints.

[0079] S342, Backpropagation and Parameter Update: Using the gradient descent algorithm, backpropagation and parameter update are performed on the model parameters θ to minimize...

[0080] Furthermore, S4 includes the following steps:

[0081] S41. The preprocessed raw data is input into the trained model to obtain real-time classification results;

[0082] S42. Determine the fatigue level of muscles in a specific area based on the real-time classification results, and set a threshold. If the fatigue level reaches the threshold, trigger an early warning mechanism to prevent further muscle fatigue.

[0083] S43. Provide muscle fatigue feedback, feeding back the diagnostic results and actual sensations to the system, and regularly update the model and rule base.

[0084] Furthermore, the raw data is an sEMG signal.

[0085] A storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the above-described method.

[0086] A terminal includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method described above.

[0087] The beneficial effects of this invention are as follows: This invention provides a muscle fatigue monitoring method based on semi-supervised learning. First, through self-supervised learning, it learns effective feature representations on a large amount of unlabeled data and employs time-frequency consistency loss to capture the global characteristics of the signal, enhancing robustness to noise and perturbations and improving the model's ability to classify complex patterns (such as fatigue states). Second, by introducing consistency regularization into the semi-supervised learning and fine-tuning process, it can learn explicit supervisory signals from labeled data and force the model to output consistent values ​​for perturbed data, thereby uncovering the latent structure of unlabeled data, extracting additional information, and significantly improving generalization ability. Attached Figure Description

[0088] Figure 1 This is a flowchart of a muscle fatigue monitoring method based on semi-supervised learning according to the present invention. Detailed Implementation

[0089] 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.

[0090] Reference Figure 1 As shown, a muscle fatigue monitoring method based on semi-supervised learning includes the following steps:

[0091] S1. Collect raw data of the target muscle group using a surface electromyography sensor, and perform noise reduction preprocessing on the raw data;

[0092] S2. Construct a self-supervised pre-trained model that includes a masked proxy task and a 1D CNN encoder, and use the time-domain and frequency-domain consistency loss function for feature extraction training.

[0093] S3. A semi-supervised learning framework with differential data augmentation is used to fine-tune the self-supervised pre-trained model, and its parameters are optimized by fusing supervised classification loss and consistency regularization loss.

[0094] S4. Deploy the fine-tuned self-supervised pre-trained model as a real-time monitoring system, and output the muscle fatigue level through dynamic window slicing and confidence assessment.

[0095] Specifically, in terms of feature extraction and model robustness, this invention constructs a self-supervised pre-trained model that includes a masked proxy task and a 1DCNN encoder, and utilizes a time-domain and frequency-domain consistency loss function for feature extraction training, thus fully mining the key features in electromyography (EMG) signals. After noise reduction preprocessing, the raw data is input into the self-supervised pre-trained model. During this process, the model learns the contextual features of the signal, capturing not only the global characteristics of the signal but also enhancing its robustness to noise and disturbances. This enables the model to accurately extract features representing muscle fatigue when faced with complex EMG signals.

[0096] Regarding data utilization and model generalization ability, this invention employs a semi-supervised learning framework with differential data augmentation to fine-tune the self-supervised pre-trained model. By applying different data augmentations to labeled data to generate dual-view inputs, the supervised classification loss of the labeled samples is calculated. Simultaneously, mean squared error constraints are used to enhance the predictive consistency between views, and the combined weights of the supervised loss and consistency loss are dynamically adjusted. This approach allows the model to learn explicit supervisory signals from labeled data while simultaneously forcing the model to output consistent results on perturbed data, uncovering the latent structure of unlabeled data, and extracting additional information. This significantly improves the model's generalization ability, enabling it to more stably and accurately determine muscle fatigue states in various scenarios.

[0097] In practical applications, deploying the finely tuned self-supervised pre-trained model as a real-time monitoring system, and outputting muscle fatigue levels through dynamic window slicing and confidence assessment, demonstrates significant practical value. In sports scenarios, it can monitor athletes' muscle fatigue in real time, triggering an early warning mechanism when fatigue levels reach a threshold, effectively preventing injuries due to over-fatigue and providing strong protection for sports safety. In rehabilitation training scenarios, it helps medical staff understand patients' muscle fatigue during recovery, adjust rehabilitation training plans based on this data, and thus improve sports rehabilitation and training effects, compensating for the shortcomings of existing technologies in muscle state analysis and demonstrating significant clinical application value.

[0098] Furthermore, S1 includes the following steps:

[0099] S11. Select a high-precision, low-noise sensor, place electrodes according to the target muscle group (such as the waist and back), ensure good contact, and use a bipolar electrode configuration to reduce common-mode noise;

[0100] S12. Power frequency interference is eliminated by using a 50 / 60Hz notch filter.

[0101]

[0102] Where, ω0=2πf0 / f sf0 is the power frequency, f0 is 50 / 60Hz, f s Where r is the sampling rate and r is the damping coefficient;

[0103] S13. A low-pass filter with a cutoff frequency of 500Hz is used to remove high-frequency noise.

[0104]

[0105] Where b and a are filter coefficients.

[0106] Specifically, this embodiment focuses on the raw data acquisition and preprocessing stages. From sensor selection to interference cancellation and noise filtering, each step works closely together, playing a crucial role in improving the accuracy and reliability of muscle fatigue monitoring. In terms of signal acquisition, a bipolar electrode configuration is used to acquire signals from the target muscle group. This configuration has unique advantages. Bipolar electrodes can effectively reduce common-mode interference and more accurately capture the electrical signals generated by the target muscle group. Since muscle electrical signals are easily interfered with by the surrounding environment and other bioelectrical signals during propagation, the bipolar electrode configuration can highlight the target signal and suppress interference signals through differential measurement, providing a reliable raw data foundation for subsequent accurate analysis of muscle status.

[0107] To eliminate power frequency interference, a 50 / 60Hz notch filter is used. Power frequency interference is widespread in electromyography (EMG) signal acquisition environments, such as common mains power interference. If not effectively removed, these interference signals will mix into the EMG signal, severely affecting signal quality and the accuracy of feature extraction. By using a specific notch filter, utilizing its precise frequency response characteristics, 50 / 60Hz interference signals can be specifically attenuated, making the acquired EMG signal purer and ensuring that subsequent analysis is not affected by power frequency interference, thus improving signal reliability and stability.

[0108] To remove high-frequency noise, a low-pass filter with a cutoff frequency of 500Hz was used. In electromyography (EMG) signals, high-frequency noise can originate from various factors, such as internal noise of electronic devices and electromagnetic radiation in the environment. This high-frequency noise increases signal complexity and interferes with the extraction of muscle fatigue-related features. The 500Hz low-pass filter effectively blocks high-frequency noise while preserving the effective low-frequency EMG signals. This removes unnecessary interference while maximizing the preservation of signal features related to muscle fatigue, providing a higher-quality signal source for subsequent accurate analysis of muscle fatigue levels.

[0109] Furthermore, S2 includes the following steps:

[0110] S21. Randomly select a starting time point to generate a masking interval and construct a partially masked input signal:

[0111] S22. Use a 1D CNN encoder to extract features and a decoder to reconstruct the occluded region;

[0112] S23. Optimize the model by jointly calculating the time-domain reconstruction loss and the frequency-domain spectral consistency loss.

[0113] Specifically, in this embodiment, in signal processing and feature learning, a masked interval is generated by randomly selecting a starting time point and a partially masked input signal is constructed. This operation simulates the possible signal loss or interference in actual monitoring. In this way, the model is forced to learn how to reconstruct the masked area using the unmasked portion of the signal, thereby prompting it to delve deeper into the contextual features of the signal. This not only enhances the model's understanding of the overall signal structure but also improves its ability to handle incomplete or interfered signals, enabling it to stably extract muscle fatigue-related features even in complex environments.

[0114] A 1D CNN encoder was used to extract features, fully leveraging the advantages of 1D CNNs in processing time-series data. Since the raw electromyography (EMG) data has significant local temporal features, the 1D CNN encoder can effectively capture these features through convolutional operations, transforming them into more representative feature vectors. The decoder's process of reconstructing the occluded region further validates and optimizes the features extracted by the encoder. Based on the feature vector output by the encoder, the decoder recovers the signal waveform features within the unoccluded region of the input signal. This process not only verifies the effectiveness of the encoder's feature extraction but also enables the model to learn a more complete and robust feature representation, thereby more accurately identifying muscle fatigue states.

[0115] A joint computation of temporal reconstruction loss and frequency-domain spectral consistency loss optimizes the model, constraining it from both time and frequency domains. The temporal reconstruction loss calculates the prediction error of the masked region, enabling the model to better learn the time-series features of the signal in the time domain. The frequency-domain spectral consistency loss, from a frequency domain perspective, ensures the model can accurately reconstruct the signal in the frequency domain as well. The combination of these two methods fully utilizes the complementary information provided by time and frequency domain features, comprehensively improving the model's ability to learn electromyographic signal features. Models trained in this way can more accurately capture the changing characteristics of muscle fatigue states in both the time and frequency domains, thereby improving the accuracy and reliability of muscle fatigue monitoring.

[0116] Furthermore, S21 includes the following steps:

[0117] S211. Construct the input layer: Accept raw data (such as sEMG signals), Let the input data be...

[0118] X = [x1, x2, ..., x T ]∈R N×D

[0119] Where T is the time step, N is the number of samples, and D is the feature dimension of each sample (e.g., the length of the sEMG signal sequence).

[0120] S212. Construct the mask layer: Randomly select a starting time point t. start , satisfying 1≤t start ≤TL, where L is the length of the mask interval, resulting in a mask interval of [t] start ,t start +L), generate a mask matrix M∈{0,1} T×D , where each element m t,d :

[0121]

[0122] That is, the value within the mask range is 0, and the rest are 1.

[0123] The mask ratio is controlled by adjusting the mask interval length L. If the mask ratio is p... mask ,but

[0124] S213. Masked Input: Combine the mask matrix with the original data to generate a partially masked input X. masked :

[0125] X masked =X·M+[MASK]·(1-M)

[0126] Where [MASK] is a mask placeholder (such as zero-padding or Gaussian noise, etc.);

[0127] S22 includes the following steps:

[0128] S221. Constructing the encoder: Due to the obvious local temporal features of the original data, a 1D CNN Encoder is used as the encoder.

[0129] H = ENcoder(X) masked ,θ e )

[0130] Initially, X masked As the initial input to the network:

[0131] H (0) =X masked

[0132] The output H of layer L (l) H is output from the previous layer (l-1) The result is obtained through convolution and nonlinear activation:

[0133] H(l) =RELU(Conv1D(H (l-1) W (l) ))

[0134] Among them W (l) For convolution kernel weights,

[0135] After L layers of convolution, the final output of the encoder is:

[0136] H = H (L) ∈R T×D ;

[0137] S222. Constructing a decoder: The latent space representation is processed by the decoder to recover the signal waveform features within the unmasked region of the input signal, wherein the decoder only processes the feature vectors associated with the unmasked region.

[0138] Reconstructing the original data from latent feature H :

[0139]

[0140] S23 includes the following steps:

[0141] S231. Reconstruction Loss Function: A reconstruction loss function is constructed by calculating the prediction error of the masked region, forcing the self-supervised pre-trained model to learn the contextual features of the signal.

[0142] Time domain error :

[0143]

[0144] Frequency domain error :

[0145]

[0146] Spectral consistency loss:

[0147]

[0148] S232. Minimize the reconstruction loss using gradient descent and update the encoder and decoder parameters:

[0149]

[0150] Specifically, in this embodiment, the raw data is defined in a specific format during the construction of the input layer, clarifying the data structure and dimensions, and providing a clear and standardized input format for subsequent model processing. This standardized definition enables the model to process data more efficiently, reduces errors and uncertainties in the data processing process, and ensures that the model can quickly and accurately extract relevant information when receiving raw electromyography data.

[0151] The construction and related operations of the masking layer—namely, randomly selecting a starting time point to determine the masking interval, generating the masking matrix, and controlling the masking ratio—provide diverse data samples for model learning. By adjusting the masking interval length to control the masking ratio, the model can be trained under different levels of signal occlusion, enhancing its adaptability to various signal loss situations. This helps the model learn more comprehensive signal features, improving its robustness to complex and incomplete signals. Therefore, in actual muscle fatigue monitoring, even when encountering partial signal loss or interference, it can accurately determine muscle status.

[0152] The method of generating the input after masking cleverly combines the mask matrix with the original data, preserving some of the original signal information while introducing mask placeholders. This design allows the model to better learn how to extract features and reconstruct signals using limited information during training, further enhancing the model's understanding of signal context. By learning the relationship between masked and unmasked regions, the model can uncover deeper signal features, thereby more accurately capturing feature patterns related to muscle fatigue.

[0153] The 1D CNN encoder, through multiple layers of convolution and nonlinear activation operations, can progressively extract deep local temporal features from the original data. These features are crucial for identifying muscle fatigue states. The decoder, based on the latent features output by the encoder, processes only the feature vectors associated with the unmasked regions to reconstruct the original data. This not only improves reconstruction efficiency but also allows the model to focus on key features, further optimizing feature representation and thus enhancing the model's ability to identify muscle fatigue states.

[0154] In the model optimization phase, a specific reconstruction loss function is constructed, and the encoder and decoder parameters are updated using the gradient descent algorithm, enabling the model to optimize towards more accurate signal reconstruction. The calculation of time-domain and frequency-domain errors measures the model's reconstruction performance from both dimensions, with spectral consistency loss combining the two to ensure good reconstruction results in both dimensions. By continuously adjusting the model parameters, the model can learn more effective signal feature representations, improving its ability to capture muscle fatigue features.

[0155] Furthermore, S3 includes the following steps:

[0156] S31. Apply different data augmentations to the labeled data to generate dual-view input;

[0157] S32. Calculate the supervised classification loss for labeled samples;

[0158] S33. Enhance the consistency of predictions between views by using mean square error constraints;

[0159] S34. Dynamically adjust the combined weights of supervision loss and consistency loss.

[0160] Specifically, this embodiment applies different data augmentations to the labeled data to generate dual-view inputs, enriching the data diversity for model training. Amplitude adjustment augmentations, such as adding Gaussian noise and random gain adjustment, simulate signal interference and intensity changes that may occur during actual acquisition. Temporal dimension manipulation augmentations, such as random pruning, time stretching, and cyclic translation, alter the time-series characteristics of the signal. These operations enable the model to access data in various forms and learn different data features, preventing the model from over-relying on a single data pattern during training, thereby improving the model's adaptability to complex and variable electromyographic signals.

[0161] Calculating the supervised classification loss for labeled samples provides the model with a clear learning objective. By comparing the prediction of one of the enhanced versions with the true label to calculate the standard supervised loss, the model has a correct direction to follow during the learning process. Based on the features and label information of the labeled data, it can quickly learn key features related to muscle fatigue, improving the accuracy of the model's classification of muscle fatigue. This supervised learning approach enables the model to effectively acquire knowledge from labeled data, accelerating the model's convergence speed and improving training efficiency.

[0162] By enhancing the consistency of predictions across views through mean squared error constraints, the potential value of unlabeled data is fully explored. A consistency loss is constructed using the mean squared error between the prediction distributions of two different enhanced versions of data, forcing the model to output consistent results for data with different perturbations. This enables the model to learn the underlying structure and common features of the data, enhancing its robustness and generalization ability. Even when faced with unseen data, the model can make more accurate judgments based on the learned underlying structure, thereby improving the reliability of the model in monitoring muscle fatigue states in different scenarios.

[0163] The combined weights of the supervision loss and consistency loss are dynamically adjusted to balance the different needs during the model learning process. In the early stages of training, a smaller consistency loss weight is set, allowing the model to focus on learning basic tasks and grasping the main characteristics of muscle fatigue from labeled data, thus building a preliminary classification model. As training progresses, the consistency loss weight is gradually increased to strengthen robustness constraints, enabling the model to better utilize information from unlabeled data while maintaining accuracy. This further optimizes model performance and improves the stability and reliability of the model in monitoring muscle fatigue in complex environments.

[0164] Furthermore, in the semi-supervised learning part, consistency regularization is introduced. Consistency regularization requires the model to maintain consistent prediction results in the face of small perturbations in the input data or different data augmentations. For the same labeled sample x, we generate two variants x1 and x2 using two different data augmentation methods, and require the model to predict the two as similarly as possible.

[0165] S31 includes the following steps:

[0166] S311, Data Augmentation: For each labeled sample x (and its label y), apply two different data augmentation operations to obtain two augmented versions:

[0167] x1 = A1(x), x2 = A2(x)

[0168] A1 and A2 are different data augmentation operations, belonging to the non-consistent time-domain transformation type, including but not limited to combinations of amplitude adjustment augmentation and time dimension manipulation augmentation, wherein:

[0169] The amplitude adjustment enhancements include adding Gaussian noise and random gain adjustment.

[0170] The time-dimensional adjustment enhancements include random cropping, time stretching, and cyclic translation.

[0171] S312, Forward Propagation: Calculate the prediction distributions of the two enhanced versions using the pre-trained encoder model:

[0172] p1=f(x1;θ), p2=f(x2;θ)

[0173] Where f(·θ) represents the encoder model, and p1 and p2 are the predicted class probability distributions;

[0174] In S32, the standard supervised loss (using cross-entropy loss for a classification task as an example) is calculated using the prediction of one of the versions (e.g., x1) and the true label y.

[0175]

[0176] Where I[y=c] is an indicator function, and c represents the category index;

[0177] In S33, the consistency loss is constructed using the mean squared error (MSE) between p1 and p2:

[0178]

[0179] S34 includes the following steps:

[0180] S341, Combined Total Loss: The final training loss is the weighted sum of the supervision loss and the consistency regularization loss.

[0181]

[0182] Here, λ is a hyperparameter used to balance the relative weights of the supervised task and the consistency regularization. λ is set to a small value in the early stage of training and then gradually increased in the later stage so that the self-supervised pre-trained model can first focus on learning the basic task and then gradually strengthen the robustness constraints.

[0183] S342. Backpropagation and Parameter Update: Using gradient descent algorithms (such as Adam), backpropagation and parameter update are performed on the model parameters θ to minimize...

[0184] Specifically, in this embodiment, the data augmentation step employs a combination of amplitude adjustment augmentation and time dimension manipulation augmentation to provide a rich and diverse range of data samples for model training. Adding Gaussian noise simulates unavoidable environmental interference in actual monitoring, while random gain adjustment covers intensity fluctuations that may occur during signal acquisition. These two amplitude adjustment augmentations enable the model to adapt to changes in signal amplitude, improving its robustness to noise. Time dimension manipulation augmentations, such as random pruning, time stretching, and cyclic translation, alter the time-series structure of the signal, allowing the model to learn to identify muscle fatigue characteristics from different time scales and sequences. This enhances the model's adaptability to changes in the temporal characteristics of muscle activity, helps capture more comprehensive muscle fatigue-related information, avoids overfitting, and improves the model's generalization ability.

[0185] During forward propagation, the pre-trained encoder model is used to calculate the prediction distributions of the two augmented versions. This operation fully utilizes the feature representation capabilities learned by the pre-trained model. By predicting different augmented data, the model can further explore the relationship between the new features after data augmentation and muscle fatigue state based on the existing pre-training knowledge, providing richer information for subsequent loss calculation and model optimization, making the model's predictions more accurate and reliable.

[0186] In terms of loss calculation, a standard supervised loss is calculated using one version of the prediction and the true label, providing a clear learning direction for the model. The supervised loss enables the model to adjust its parameters to minimize prediction error based on known muscle fatigue state labels, thereby accurately learning feature patterns related to muscle fatigue and improving the model's classification accuracy for muscle fatigue states. Simultaneously, a consistency loss is constructed using the mean squared error between the two enhanced prediction distributions, forcing the model to make consistent predictions for data with different perturbations. This encourages the model to learn the inherent structure and invariant features of the data, uncovering potential information in unlabeled data, enhancing the model's stability and robustness, and enabling it to maintain good performance even when faced with different changes in electromyography (EMG) signals.

[0187] Regarding the dynamic adjustment of the combined weights of supervised loss and consistency loss, a smaller consistency loss weight in the early stages of training allows the model to focus on learning basic muscle fatigue features from labeled data, establishing a preliminary effective classification model. As training progresses, the consistency loss weight is gradually increased, guiding the model to utilize the latent structural information of unlabeled data to further optimize the model's feature representation and predictive capabilities. This dynamic adjustment strategy ensures that the model can converge quickly and learn key features in the early stages of training, while fully utilizing unlabeled data in the later stages to improve the model's generalization performance, making the model more stable and accurate in muscle fatigue monitoring tasks.

[0188] Furthermore, S4 includes the following steps:

[0189] S41. The raw data (such as sEMG signal) is preprocessed and then input into the trained model to obtain real-time classification results;

[0190] S42. Determine the fatigue level of muscles in a specific area based on the real-time classification results, and set a threshold. If the fatigue level reaches the threshold, trigger an early warning mechanism to prevent further muscle fatigue.

[0191] S43. Provide muscle fatigue feedback, feeding back the diagnostic results and actual sensations to the system, and regularly update the model and rule base.

[0192] Specifically, this embodiment focuses on the practical application of muscle fatigue monitoring. From data input to model to obtain results, to judging the degree of fatigue and issuing warnings based on the results, and then to feedback optimization, a complete and efficient closed-loop system is formed, which effectively promotes the transition of muscle fatigue monitoring technology from theoretical research to practical application.

[0193] The raw data is preprocessed and then input into a trained model to obtain real-time classification results. This process enables real-time monitoring of muscle fatigue. In scenarios such as sports training and rehabilitation, timely understanding of muscle fatigue is crucial. By processing the collected electromyographic signals in real time, the model can quickly provide classification results for muscle fatigue, offering immediate information to athletes, patients, and related professionals. This allows them to adjust exercise or treatment plans based on the current muscle state, avoiding injuries caused by over-fatigue and improving training and rehabilitation outcomes.

[0194] Based on real-time classification results, the system assesses the fatigue level of specific muscle groups and sets thresholds to trigger an early warning mechanism, providing crucial protection against muscle fatigue. The threshold setting is based on in-depth research into the development patterns of muscle fatigue; when muscle fatigue reaches the threshold, the early warning mechanism is rapidly activated. This is significant in sports scenarios. For example, during high-intensity training or competition, athletes may neglect their muscle fatigue due to focus on training or competition. The early warning mechanism can promptly remind them to stop exercising or adjust the intensity, effectively preventing sports injuries such as muscle strains and stress fractures, protecting athletes' health, and extending their athletic careers. In rehabilitation training scenarios, overtraining can affect the rehabilitation process or even cause secondary injuries for patients. The early warning mechanism can help medical staff adjust rehabilitation plans in a timely manner, ensuring that rehabilitation training is conducted safely and effectively.

[0195] Providing feedback on muscle fatigue, feeding back diagnostic results and actual sensations to the system, and regularly updating the model and rule base are crucial steps in improving the accuracy and adaptability of muscle fatigue monitoring. Actual sensation feedback provides valuable real-world evidence for model optimization, as an individual's subjective experience of muscle fatigue may contain information the model has not yet captured. Integrating this feedback into the system and regularly updating the model and rule base allows the model to continuously learn and adapt to the characteristics of muscle fatigue in different individuals and scenarios, thus continuously optimizing monitoring effectiveness. Over time and with data accumulation, the model will become increasingly accurate and better suited to practical application needs.

[0196] Furthermore, the raw data is an sEMG signal.

[0197] A storage medium storing a computer program, characterized in that the computer program, when executed by a processor, implements the above-described method.

[0198] Specifically, this embodiment focuses on the storage medium, where the stored computer program, when executed, enables a muscle fatigue monitoring method, bringing several significant advantages. From the perspective of technology inheritance and reuse, the storage medium acts as a treasure trove of knowledge, storing advanced algorithms and processes for muscle fatigue monitoring in the form of a computer program. This makes the technology no longer limited to specific development environments or one-off use scenarios. Researchers can quickly apply this muscle fatigue monitoring method in different projects and on different devices, as long as there is a corresponding processor capable of reading the storage medium and executing the program. This greatly saves time and costs associated with repetitive development, accelerating the promotion and application of the technology in various fields. In terms of device compatibility and scalability, the existence of the storage medium allows the muscle fatigue monitoring technology to easily adapt to various terminal devices. Whether it is professional medical monitoring equipment or wearable sports monitoring gear, as long as it has the corresponding storage medium reading function and a suitable processor, it can carry the computer program. This greatly expands the application scope of muscle fatigue monitoring technology, meeting the needs of different users in different scenarios. For example, sports enthusiasts can use smart bracelets to monitor their muscle fatigue in real time during exercise using programs stored on the storage medium. Medical professionals can also utilize this technology in hospital rehabilitation equipment to more accurately assess patients' muscle recovery. From a data security and stability perspective, the storage medium provides a stable storage environment for computer programs. Compared to data temporarily stored in memory, program data on the storage medium is less susceptible to damage from unexpected power outages, system failures, etc. This ensures the stability and reliability of the muscle fatigue monitoring method during execution, preventing inaccurate monitoring results due to data loss or program errors. Even if the device experiences a sudden restart, as long as the storage medium is intact, the program can continue to run normally, ensuring the continuity of muscle fatigue monitoring and providing users with stable and reliable monitoring services.

[0199] A terminal includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method described above.

[0200] Specifically, the terminal described in this embodiment possesses a memory, a processor, and an executable computer program, offering numerous significant advantages in muscle fatigue monitoring applications. Regarding hardware collaboration, the memory, processor, and computer program work closely together. The memory is responsible for stably storing the program code and large amounts of data required for muscle fatigue monitoring, including training model data and rule base information, ensuring secure data storage and readily available access during device operation. The processor, with its powerful computing capabilities, quickly executes the computer program, efficiently processing the collected raw electromyographic data. This collaboration enables the terminal to rapidly complete a series of complex tasks such as data preprocessing, model calculation, and result analysis, ensuring real-time and accurate muscle fatigue monitoring. For example, in sports settings, it can promptly assess athletes' muscle condition, providing strong support for event security. In terms of system integration and ease of use, this terminal integrates all the key elements required for muscle fatigue monitoring. Without the need for additional complex external devices or cumbersome system environments, users can easily achieve muscle fatigue monitoring simply by operating this terminal device. This integrated design not only facilitates user operation but also lowers the technological application threshold, enabling wider adoption of muscle fatigue monitoring technology. Whether for professional medical institutions, sports training teams, or ordinary fitness enthusiasts, this terminal can be conveniently used for muscle fatigue monitoring, meeting the monitoring needs of different user groups in various scenarios. Regarding data management and continuous optimization, the terminal can autonomously manage the data generated during the monitoring process. On one hand, it stores the collected raw data and analysis results in an orderly manner, facilitating subsequent review and comparative analysis, providing a data foundation for long-term tracking of muscle state changes. On the other hand, by regularly updating the computer program, the algorithm and model for muscle fatigue monitoring can be continuously optimized, improving the accuracy and reliability of monitoring. Simultaneously, the terminal can further improve its rule base based on user feedback and newly collected data, achieving self-optimization and adapting to changes in different users and scenarios, providing higher-quality and more personalized services for muscle fatigue monitoring.

[0201] This invention discloses a muscle fatigue monitoring method based on semi-supervised learning. First, it employs self-supervised learning to learn effective feature representations on a large amount of unlabeled data and uses time-frequency consistency loss to capture the global characteristics of the signal, enhancing robustness to noise and perturbations and improving the model's ability to classify complex patterns (such as fatigue states). Second, it introduces consistency regularization into the semi-supervised learning and fine-tuning process. This not only learns explicit supervisory signals from labeled data but also forces the model to output consistent values ​​for perturbed data, uncovering the latent structure of unlabeled data, extracting additional information, and significantly improving generalization ability.

[0202] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A muscle fatigue monitoring method based on semi-supervised learning, characterized in that, The method includes the following steps: S1. Collect raw data of the target muscle group using a surface electromyography sensor, and perform noise reduction preprocessing on the raw data; S2. Construct a self-supervised pre-trained model that includes a masked proxy task and a 1D CNN encoder, and use the time-domain and frequency-domain consistency loss function for feature extraction training. S3. A semi-supervised learning framework with differential data augmentation is used to fine-tune the self-supervised pre-trained model, and its parameters are optimized by fusing supervised classification loss and consistency regularization loss. S4. Deploy the fine-tuned self-supervised pre-trained model as a real-time monitoring system, and output the muscle fatigue level through dynamic window slicing and confidence assessment.

2. The muscle fatigue monitoring method based on semi-supervised learning according to claim 1, characterized in that, S1 includes the following steps: S11. Acquire signals from the target muscle group using a sensor with a bipolar electrode configuration; S12. Power frequency interference is eliminated by using a 50 / 60Hz notch filter. Where, ω0=2πf0 / f s f0 is the power frequency, f0 is 50 / 60Hz, f s Where r is the sampling rate and r is the damping coefficient; S13. A low-pass filter with a cutoff frequency of 500Hz is used to remove high-frequency noise. Where b and a are filter coefficients.

3. The muscle fatigue monitoring method based on semi-supervised learning according to claim 2, characterized in that, S2 includes the following steps: S21. Randomly select a starting time point to generate a masking interval and construct a partially masked input signal: S22. Use a 1D CNN encoder to extract features and a decoder to reconstruct the occluded region; S23. Optimize the model by jointly calculating the time-domain reconstruction loss and the frequency-domain spectral consistency loss.

4. The muscle fatigue monitoring method based on semi-supervised learning according to claim 3, characterized in that, S21 includes the following steps: S211. Construct the input layer: Accept the raw data, let the input data be... X=[x1,x2,…,x T ]∈R N×D Where T is the time step, N is the number of samples, and D is the feature dimension of each sample; S212. Construct the mask layer: Randomly select a starting time point t. start , satisfying 1≤t start ≤TL, where L is the length of the mask interval, resulting in a mask interval of [t] start ,t start +L), generate a mask matrix M∈{0,1} T×D , where each element m t,d : That is, the value within the mask range is 0, and the rest are 1. The mask ratio is controlled by adjusting the mask interval length L. If the mask ratio is p... mask ,but S213. Masked Input: Combine the mask matrix with the original data to generate a partially masked input X. masked : X masked =X·M+[MASK]·(1-M) [MASK] is a mask placeholder; S22 includes the following steps: S221. Constructing the encoder: Due to the obvious local temporal features of the original data, a 1D CNN Encoder is used as the encoder. H=Encoder(X masked ,θ e ) Initially, X masked As the initial input to the network: H (0) =X masked The output H of layer L (l) Output H from the previous layer (l-1) The result is obtained through convolution and nonlinear activation: A (l) =RELU(Conv1D(H (l-1) ,W (l) )) Among them, W (l) For convolution kernel weights, After L layers of convolution, the final output of the encoder is: H=H (L) ∈R T×D ; S222. Constructing a decoder: The latent space representation is processed by the decoder to recover the signal waveform features within the unmasked region of the input signal, wherein the decoder only processes the feature vectors associated with the unmasked region. Reconstructing the original data from latent feature H S23 includes the following steps: S231. Reconstruction Loss Function: A reconstruction loss function is constructed by calculating the prediction error of the masked region, forcing the self-supervised pre-trained model to learn the contextual features of the signal. Time domain error Frequency domain error Spectral consistency loss: S232. Minimize the reconstruction loss using gradient descent and update the encoder and decoder parameters:

5. The muscle fatigue monitoring method based on semi-supervised learning according to claim 4, characterized in that, S3 includes the following steps: S31. Apply different data augmentations to the labeled data to generate dual-view input; S32. Calculate the supervised classification loss for labeled samples; S33. Enhance the consistency of predictions between views by using mean square error constraints; S34. Dynamically adjust the combined weights of supervision loss and consistency loss.

6. The muscle fatigue monitoring method based on semi-supervised learning according to claim 5, characterized in that, S31 includes the following steps: S311, Data Augmentation: For each labeled sample x, apply two different data augmentation operations to obtain two augmented versions: x1 = A1(x), x2 = A2(x) A1 and A2 are different data augmentation operations, belonging to the non-consistent time-domain transformation type, including but not limited to combinations of amplitude adjustment augmentation and time dimension manipulation augmentation, wherein: The amplitude adjustment enhancements include adding Gaussian noise and random gain adjustment. The time-dimensional adjustment enhancements include random cropping, time stretching, and cyclic translation. S312, Forward Propagation: Calculate the prediction distributions of the two enhanced versions using the pre-trained encoder model: p1=f(x1;θ), p2=f(x2;θ) Where f(·θ) represents the encoder model, and p1 and p2 are the predicted class probability distributions; In S32, the standard supervised loss is calculated using one version of the prediction and the true label y. Where I[y=c] is an indicator function, and c represents the category index; In S33, the consistency loss is constructed using the mean square error between p1 and p2: S34 includes the following steps: S341, Combined Total Loss: The final training loss is the weighted sum of the supervision loss and the consistency regularization loss. Here, λ is a hyperparameter used to balance the relative weights of the supervised task and the consistency regularization. λ is set to a small value in the early stage of training and then gradually increased in the later stage so that the self-supervised pre-trained model can first focus on learning the basic task and then gradually strengthen the robustness constraints. S342, Backpropagation and Parameter Update: Using the gradient descent algorithm, backpropagation and parameter update are performed on the model parameters θ to minimize...

7. The muscle fatigue monitoring method based on semi-supervised learning according to claim 6, characterized in that, S4 includes the following steps: S41. The preprocessed raw data is input into the trained model to obtain real-time classification results; S42. Determine the fatigue level of muscles in a specific area based on the real-time classification results, and set a threshold. If the fatigue level reaches the threshold, trigger an early warning mechanism to prevent further muscle fatigue. S43. Provide muscle fatigue feedback, feeding back the diagnostic results and actual sensations to the system, and regularly update the model and rule base.

8. The muscle fatigue monitoring method based on semi-supervised learning according to claim 7, characterized in that, The raw data is an sEMG signal.

9. A storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-8.

10. A terminal, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-8.

Images