sEMG signal feature extraction and muscle fatigue level discrimination method based on artificial intelligence

Through multimodal sensors and composite artificial intelligence models, the problems of sEMG signal noise interference and personalized management are solved, efficient muscle fatigue monitoring and management are achieved, and exercise efficiency and safety are improved.

CN118266951BActive Publication Date: 2024-11-29QILU UNIVERSITY OF TECHNOLOGY (SHANDONG ACADEMY OF SCIENCES)
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202410395915.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-04-02
Publication Date
2024-11-29
Estimated Expiration
2044-04-02

AI Technical Summary

Technical Problem

The existing muscle fatigue monitoring methods based on sEMG signals have problems such as signal noise interference, limited accuracy, lack of in-depth analysis and personalized management suggestions, and it is difficult to meet the needs of improving exercise efficiency and preventing exercise injuries.

Method used

Multimodal sensors are used to capture sEMG signals and auxiliary data in real time, and noise is removed through adaptive filtering and bandpass filtering technology, and features are extracted by multi-scale entropy analysis and machine learning algorithms. Compound artificial intelligence models are built for muscle fatigue level discrimination and real-time management. suggestion.

Benefits of technology

Improve the accuracy and efficiency of sEMG signal processing, accurately capture dynamic changes in muscle fatigue, provide personalized and real-time muscle fatigue management strategies, significantly reduce sports injury reporting and improve training satisfaction.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118266951B_ABST
    Figure CN118266951B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for extracting sEMG signal features and distinguishing muscle fatigue levels based on artificial intelligence, comprising the following steps: S1, using a multimodal sensor device to capture sEMG signal data and auxiliary data of a user performing a specific physical activity in real time; S2, preprocessing the collected sEMG signal data and auxiliary data; S3, extracting features from the preprocessed sEMG signal data using multiscale entropy analysis and machine learning algorithms; S4, constructing and training a composite artificial intelligence model for distinguishing muscle fatigue levels based on the extracted features; S5, evaluating the user's muscle fatigue level in real time, and adjusting the exercise intensity and rest time arrangement according to the evaluation results. The present invention effectively improves the accuracy and efficiency of sEMG signal processing by adopting innovative signal preprocessing technology.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of muscle fatigue identification, and in particular to a method for extracting sEMG signal features and identifying muscle fatigue levels based on artificial intelligence. Background Art

[0002] In modern society, with the enhancement of health awareness and the development of sports science, people pay more and more attention to improving physical health and quality of life through sports. Especially among professional athletes and fitness enthusiasts, there is a high demand for improving exercise efficiency and preventing sports injuries. Muscle fatigue is a common physiological phenomenon during exercise. Its effective monitoring and management are of great significance for improving training effects and preventing sports injuries. sEMG (surface electromyography), as an important physiological signal that can reflect the state of muscle activity, is widely used in sports science, rehabilitation medicine and other fields. However, the existing muscle fatigue monitoring and management methods based on sEMG signals have some shortcomings and urgently need to be improved through innovative technologies.

[0003] First, the existing technology in sEMG signal processing is often interfered by factors such as signal noise and human motion artifacts, which affects the accuracy and efficiency of signal processing. Traditional signal preprocessing methods such as simple filtering technology are difficult to completely eliminate these interferences, resulting in limited accuracy in subsequent feature extraction and muscle fatigue discrimination. In addition, traditional methods rely more on experience and intuition in feature extraction, lack the means to deeply analyze the inherent complexity of sEMG signals, and are difficult to fully explore the information about muscle fatigue status contained in sEMG signals.

[0004] Secondly, regarding the identification of muscle fatigue levels, most existing technologies use simple threshold judgments or statistical models based on limited features. These methods fail to fully consider the complexity and individual differences of the muscle fatigue process, resulting in insufficient generalization and personalized adaptability of the identification results. In addition, existing technologies often lack real-time and personalization when providing muscle fatigue management advice, making it difficult to meet users' needs for timely and customized management strategies.

[0005] Therefore, how to provide a method for extracting sEMG signal features and distinguishing muscle fatigue levels based on artificial intelligence is an urgent problem that technicians in this field need to solve. Summary of the invention

[0006] One purpose of the present invention is to propose a method for extracting sEMG signal features and distinguishing muscle fatigue levels based on artificial intelligence. The present invention effectively improves the accuracy and efficiency of sEMG signal processing by adopting innovative signal preprocessing technology.

[0007] According to an embodiment of the present invention, a method for extracting sEMG signal features and distinguishing muscle fatigue levels based on artificial intelligence includes the following steps:

[0008] S1. Use multimodal sensor equipment to capture the sEMG signal data of users in real time when they are performing specific physical activities, and simultaneously collect auxiliary data such as ambient temperature, humidity, user's movement speed and heart rate;

[0009] S2. Preprocessing the collected sEMG signal data and auxiliary data, including removing power supply interference, motion artifacts and other noise that may affect signal quality, and filtering the signal using bandpass filtering technology;

[0010] S3. Use multiscale entropy analysis and machine learning algorithms to extract features from the preprocessed sEMG signal data, while considering the environmental and physiological parameters in the auxiliary data, and extract key features reflecting the muscle activity state from the sEMG signal data and auxiliary data. Multiscale entropy analysis is used to quantify the complexity changes of sEMG signals and capture the dynamic characteristics of muscle fatigue. The machine learning model further analyzes the dynamic characteristics, including time domain features, frequency domain features, and time-frequency domain features.

[0011] S4. Based on the extracted features, a composite artificial intelligence model for muscle fatigue level discrimination is constructed and trained. The composite artificial intelligence model comprehensively analyzes the key features of sEMG signals and auxiliary data to determine whether the muscle is fatigued and the degree of fatigue;

[0012] S5. Evaluate the user's muscle fatigue level in real time and adjust exercise intensity and rest time arrangement based on the evaluation results.

[0013] Optionally, the S1 specifically includes the following steps:

[0014] S11, setting a multimodal sensor device, including a sEMG signal collector, a temperature sensor, a humidity sensor, a speedometer, and a heart rate monitor, for real-time monitoring and recording of physiological and environmental data of the user during specific physical activities;

[0015] S12. Collect comprehensive data for each time point t:

[0016] d final (t) = {d sEMG (t),d T (t),d H (t),d V (t),d HR (t)};

[0017] Among them, d sEMG (t) is the sEMG signal value at time point t, dT (t) is the ambient temperature at time t, d H (t) is the ambient humidity value at time point t, d V (t) is the user's motion speed at time point t, d HR (t) is the user's heart rate value at time t;

[0018] S13. Combine the comprehensive data of S12 to construct sEMG dataset D final , sEMG dataset D final The comprehensive data of all time points d final (t) Composition:

[0019] D final ={d final (1), d final (2), ..., d final (n)};

[0020] Where n is the total number of data points captured during a particular physical activity.

[0021] Optionally, the S2 specifically includes the following steps:

[0022] S21, for the final data set D final Noise is removed for each data type in the , and the specific operations include:

[0023] For sEMG signals sEMG (t), an adaptive filter is applied to remove power supply interference and motion artifacts. The adaptive filter dynamically adjusts its parameters according to the real-time characteristics of the signal to reduce the impact of noise. The adaptive filter uses the following formula for noise removal:

[0024]

[0025] in, is the noise estimated at time t, estimated by the minimum mean square error algorithm:

[0026]

[0027] Among them, w i (t) is the weight of the adaptive filter at time t, M is the order of the filter;

[0028] For ambient temperature, humidity, user movement speed, and heart rate data, a simple moving average filtering method is used to remove random noise caused by environmental changes or device errors:

[0029]

[0030] Among them, d X(t) represents any auxiliary data ambient temperature d T (t), humidity d H (t), user movement speed d V (t), or heart rate data d HR (t) is the value at time t, N is the size of the averaging window, d X (t) is the filtered data value;

[0031] S22, then the processed sEMG signal d sEMG (t) Use bandpass filtering to keep the useful frequency components of the signal:

[0032] Design a bandpass filter and choose a suitable lower limit frequency f L and upper frequency f H , in order to cover the important frequency range in the sEMG signal, f L Set to 20Hz, f H Set to 450Hz to filter out the typical frequencies of human muscle activity. Apply this bandpass filter f bp Onto the sEMG signal, use the Fast Fourier Transform to convert the signal to the frequency domain:

[0033] D sEMG (f) = FFT(d sEMG (t));

[0034] D′ sEMG (f) = D sEMG (f)·H(f);

[0035] d′ sEMG (t) = IFFT(D′ sEMG (f));

[0036] Among them, D sEMG (f) is the Fourier transform of the original sEMG signal, and H(f) is the frequency domain representation of the bandpass filter;

[0037] Apply filter characteristics to remove the L ,f H ] frequency components within the range, D' sEMG (f) is the representation of the sEMG signal in the frequency domain after bandpass filtering. It is converted back to the time domain by inverse Fourier transform to obtain the processed sEMG signal d' sEMG (t):

[0038]

[0039] S23, after preprocessing, all preprocessed data are recombined into a preprocessed comprehensive data set D preprocessed, which includes the noise-removed and filtered sEMG signal and auxiliary data:

[0040]

[0041] Optionally, S3 specifically includes the following steps:

[0042] S31, preprocessing the sEMG signal data d' sEMG (t) Perform multi-scale entropy analysis to quantify the complexity changes of sEMG signals at different time scales, reflecting the dynamic characteristics of muscle fatigue:

[0043] For each preprocessed sEMG signal sample, construct its multi-scale representation, and obtain time series of different scales by coarse-graining the signal. sEMG (t) is a coarse-grained time series at scale s The jth element is defined as:

[0044]

[0045] Among them, s represents the scale factor, which indicates the coarse-grained level of the time series considered in the multiscale entropy analysis. The larger the scale factor, the wider the range of the time series being averaged, which can capture the complexity changes of sEMG signals over a longer time range and understand the dynamic changes during muscle fatigue. It represents the jth element of the time series at scale s, which is obtained by coarse-graining the original sEMG signal. It represents the average activity intensity of the sEMG signal at a specific scale and analyzes the persistence and change trend of muscle activity.

[0046] For each time series of scale s When calculating the sample entropy SE(s,m,r), define the distance function d[x,y] as the maximum absolute difference between the two sequences (x,y), and calculate the ratio of the matching template vectors A(m,r,N):

[0047]

[0048] Among them, m represents the embedding dimension, which is used to construct the dimension of the template vector in the sample entropy calculation. It reflects the amount of historical information of the time series considered when calculating the sample entropy, which is related to the time correlation of the signal during muscle fatigue. r represents the similarity tolerance, which is the threshold used to judge whether two sequences are similar in the sample entropy calculation. It is a ratio based on the standard deviation of the sEMG signal, which reflects the sensitivity of signal fluctuations and is used to identify subtle changes caused by muscle fatigue. N is the length of the time series. A(m,r,N) represents the ratio of the number of similar template sequence logarithms to the number of total template sequence logarithms at scale m.

[0049] The sample entropy is calculated for each scale of the time series to quantify the complexity of the signal at that scale, where r is a multiple of the standard deviation, usually 0.1×SD, and SD is the standard deviation of the signal:

[0050]

[0051] S32, using machine learning algorithms to pre-process the sEMG signal data and auxiliary data d' T (t), d' H (t), d' V (t), d' HR (t) Feature extraction, including time domain features, frequency domain features, and time-frequency domain features, reflects the average level, fluctuation level, and peak intensity of muscle activity, and is used to determine whether the muscle is fatigued and the degree of fatigue:

[0052] Extract time domain features from sEMG signals:

[0053]

[0054] Extract frequency domain features from sEMG signals, frequency domain features F freq It is obtained by calculating the power spectral density PSD, where PSD(f) is:

[0055] PSD(f)=|FFT(d' sEMG (t))| 2 ;

[0056] Wherein, FFT stands for Fast Fourier Transform;

[0057] S33, combining the features extracted from the multi-scale entropy analysis and the machine learning algorithm to form a comprehensive feature vector for subsequent muscle fatigue level discrimination, the comprehensive feature vector F composite By combining the complexity features, time domain features, and frequency domain features obtained by multi-scale entropy analysis, freq And the auxiliary data features F considered aux constitute:

[0058] F composite = [SE(s,m,r),μ,σ 2 ,P max ,PSD(f),F aux ];

[0059] Among them, F aux Including the mean and change rate of ambient temperature and humidity.

[0060] Optionally, the S4 specifically includes the following steps:

[0061] S41. Build a composite artificial intelligence model that combines deep learning and machine learning technologies to comprehensively analyze the key features of sEMG signals and auxiliary data to determine whether muscles are fatigued and the degree of fatigue:

[0062] The input layer of the composite artificial intelligence model receives the comprehensive feature vector F composite , feature conversion is performed through multi-layer neural network layers, and the output calculation formula of each layer of neural network is:

[0063] a [l] =g [l] (W [l] (a [l-1] ⊙d [l-1] )+b [l] );

[0064] Among them, d [l-1] is the Dropout vector, which is used to randomly close some neurons in the l-1th layer, ⊙ represents the Hadamard product;

[0065] Add support vector machine for features extracted based on DNN and auxiliary data features F aux Classify the muscle fatigue level and use weighted cross entropy loss to deal with possible class imbalance:

[0066]

[0067] Where C is the number of categories, w i is the weight of category i, y i is the true label, is the predicted probability;

[0068] S42, prepare a training data set, including feature vectors F extracted from the actual measured sEMG signal and auxiliary data composite And the corresponding muscle fatigue level labels, the training data set is in the form of:

[0069]

[0070] Among them, y (i) is the muscle fatigue level label of the i-th sample, and N is the total number of samples;

[0071] S43. Use the training data set to train the composite artificial intelligence model and use cross-validation technology to optimize the model parameters:

[0072] Use a part of the training data to train the model, combined with the feature vector F compositeIncluding complexity features, time domain features, frequency domain features and auxiliary data features obtained by multi-scale entropy analysis, a weighted loss function is used to integrate the effects of different types of features on muscle fatigue level:

[0073]

[0074] Among them, α and β are adjustment terms used to balance the importance of classification error and regression error, w f is the weight of feature f, x i,f Yes F composite The fth eigenvalue of the i-th sample in , is the predicted value of the feature, y i,c is the real classification label, is the corresponding predicted probability;

[0075] The performance of the model was verified on another part of the data. For the prediction of muscle fatigue recovery time, the definition based on F composite A linear regression model of the features and L2 regularization is applied to prevent overfitting:

[0076]

[0077] in, is based on the comprehensive feature vector F composite The fatigue recovery time of the i-th sample predicted by the linear regression model, λ is the regularization parameter;

[0078] S44, the trained composite artificial intelligence model is based on the input feature vector F composite Predict muscle fatigue level.

[0079] Optionally, the S5 specifically includes the following steps:

[0080] S51, according to the composite feature vector F composite The evaluation results, i.e. the muscle fatigue level, dynamically adjust the exercise intensity and rest time suggestions. The evaluation results are divided into multiple levels, and each level corresponds to different exercise and rest suggestions:

[0081] Adjustment = f(Level fatigue );

[0082] Among them, Level fatigue is the muscle fatigue level assessed by the model, and Adjustment is a function of the exercise intensity and rest time adjustment recommendations based on the fatigue level;

[0083] S52, calculate the exercise intensity adjustment value Adjust intensity and rest time adjustment value Adjust rest, the adjustment value is based on fatigue level and user's personal health data, including previous exercise habits and health status:

[0084] Adjust intensity =g intensity (Level fatigue ,Data personal );

[0085] Adjust rest =g rest (Level fatigue ,Data personal );

[0086] Among them, g intensity and g rest It is a function that calculates the exercise intensity adjustment value and rest time adjustment value based on the muscle fatigue level and personal health data. personal Represents the user’s personal health data;

[0087] S53. Provide real-time feedback through a user interface, displaying muscle fatigue level assessment results and exercise intensity and rest time recommendations adjusted based on these results.

[0088] The beneficial effects of the present invention are:

[0089] (1) The present invention effectively improves the accuracy and efficiency of sEMG signal processing by adopting innovative signal preprocessing technology, including removing power supply interference, motion artifacts and other noise that may affect signal quality, and using bandpass filtering technology.

[0090] (2) The present invention extracts richer and deeper features from sEMG signals and auxiliary data by combining multi-scale entropy analysis and machine learning algorithms, including time domain features, frequency domain features, and time-frequency domain features, effectively capturing the dynamic changes in the muscle fatigue process. At the same time, the constructed composite artificial intelligence model comprehensively considers the key features of sEMG signals and auxiliary data, and can accurately determine whether the muscle is fatigued and the degree of fatigue. BRIEF DESCRIPTION OF THE DRAWINGS

[0091] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0092] Figure 1 This is a flow chart of a method for extracting sEMG signal features and distinguishing muscle fatigue levels based on artificial intelligence proposed by the present invention. DETAILED DESCRIPTION

[0093] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0094] refer to Figure 1 , a method for extracting sEMG signal features and distinguishing muscle fatigue levels based on artificial intelligence, comprising the following steps:

[0095] S1. Use multimodal sensor equipment to capture the sEMG signal data of users in real time when they are performing specific physical activities, and simultaneously collect auxiliary data such as ambient temperature, humidity, user's movement speed and heart rate;

[0096] In this implementation, S1 specifically includes the following steps:

[0097] S11, setting a multimodal sensor device, including a sEMG signal collector, a temperature sensor, a humidity sensor, a speedometer, and a heart rate monitor, for real-time monitoring and recording of physiological and environmental data of the user during specific physical activities;

[0098] S12. Collect comprehensive data for each time point t:

[0099] d final (t) = {d sEMG (t),d T (t),d H (t),d V (t),d HR (t)};

[0100] Among them, d sEMG (t) is the sEMG signal value at time point t, d T (t) is the ambient temperature at time t, d H (t) is the ambient humidity value at time point t, d V (t) is the user's motion speed at time point t, d HR (t) is the user's heart rate value at time t;

[0101] S13. Combine the comprehensive data of S12 to construct sEMG dataset D final , sEMG dataset D final The comprehensive data of all time points d final (t) Composition:

[0102] D final ={d final (1),d final (2),...,d final (n)};

[0103] Where n is the total number of data points captured during a particular physical activity.

[0104] S2. Preprocessing the collected sEMG signal data and auxiliary data, including removing power supply interference, motion artifacts and other noise that may affect signal quality, and filtering the signal using bandpass filtering technology;

[0105] In this implementation, S2 specifically includes the following steps:

[0106] S21, for the final data set D final Noise is removed for each data type in the , and the specific operations include:

[0107] For sEMG signals sEMG (t), an adaptive filter is applied to remove power supply interference and motion artifacts. The adaptive filter dynamically adjusts its parameters according to the real-time characteristics of the signal to reduce the impact of noise. The adaptive filter uses the following formula for noise removal:

[0108]

[0109] in, is the noise estimated at time t, estimated by the minimum mean square error algorithm:

[0110]

[0111] Among them, w i (t) is the weight of the adaptive filter at time t, M is the order of the filter;

[0112] For ambient temperature, humidity, user movement speed, and heart rate data, a simple moving average filtering method is used to remove random noise caused by environmental changes or device errors:

[0113]

[0114] Among them, d X (t) represents any auxiliary data ambient temperature d T (t), humidity d H (t), user movement speed d V (t), or heart rate data d HR (t) is the value at time t, N is the size of the averaging window, d X (t) is the filtered data value;

[0115] S22, then the processed sEMG signal d sEMG (t) Use bandpass filtering to keep the useful frequency components of the signal:

[0116] Design a bandpass filter and choose a suitable lower limit frequency f L and upper frequency f H , in order to cover the important frequency range in the sEMG signal, f L Set to 20Hz, f H Set to 450Hz to filter out the typical frequencies of human muscle activity. Apply this bandpass filter f bp Onto the sEMG signal, use the Fast Fourier Transform to convert the signal to the frequency domain:

[0117] D sEMG (f) = FFT(d sEMG (t));

[0118] D′ sEMG (f) = D sEMG (f)·H(f);

[0119] d′ sEMG (t) = IFFT(D′ sEMG (f));

[0120] Among them, D sEMG (f) is the Fourier transform of the original sEMG signal, and H(f) is the frequency domain representation of the bandpass filter;

[0121] Apply filter characteristics to remove the L , f H ] range, D′ sEMG (f) is the representation of the sEMG signal in the frequency domain after bandpass filtering. It is converted back to the time domain by inverse Fourier transform to obtain the processed sEMG signal d′ sEMG (t):

[0122]

[0123] S23, after preprocessing, all preprocessed data are recombined into a preprocessed comprehensive data set D preprocessed , which includes the noise-removed and filtered sEMG signal and auxiliary data:

[0124]

[0125] S3. Use multiscale entropy analysis and machine learning algorithms to extract features from the preprocessed sEMG signal data, while considering the environmental and physiological parameters in the auxiliary data, and extract key features reflecting the muscle activity state from the sEMG signal data and auxiliary data. Multiscale entropy analysis is used to quantify the complexity changes of sEMG signals and capture the dynamic characteristics of muscle fatigue. The machine learning model further analyzes the dynamic characteristics, including time domain features, frequency domain features, and time-frequency domain features.

[0126] In this implementation, S3 specifically includes the following steps:

[0127] S31, preprocessing the sEMG signal data d′ sEMG (t) Perform multi-scale entropy analysis to quantify the complexity changes of sEMG signals at different time scales, reflecting the dynamic characteristics of muscle fatigue:

[0128] For each preprocessed sEMG signal sample, construct its multi-scale representation, and obtain time series of different scales by coarse-graining the signal. sEMG (t) is a coarse-grained time series at scale s The jth element is defined as:

[0129]

[0130] Among them, s represents the scale factor, which indicates the coarse-grained level of the time series considered in the multiscale entropy analysis. The larger the scale factor, the wider the range of the time series being averaged, which can capture the complexity changes of sEMG signals over a longer time range and understand the dynamic changes during muscle fatigue. It represents the jth element of the time series at scale s, which is obtained by coarse-graining the original sEMG signal. It represents the average activity intensity of the sEMG signal at a specific scale and analyzes the persistence and change trend of muscle activity.

[0131] For each time series of scale s When calculating the sample entropy SE(s, m, r), define the distance function d[x, y] as the maximum absolute difference between the two sequences (x, y), and calculate the ratio A(m, r, N) of the matching template vectors:

[0132]

[0133] Among them, m represents the embedding dimension, which is used to construct the dimension of the template vector in the sample entropy calculation, reflecting the amount of historical information of the time series considered when calculating the sample entropy, which is related to the time correlation of the signal during muscle fatigue. r represents the similarity tolerance, which is the threshold used to judge whether two sequences are similar in the sample entropy calculation. It is a ratio based on the standard deviation of the sEMG signal, reflecting the sensitivity of signal fluctuations, and is used to identify subtle changes caused by muscle fatigue. N is the length of the time series. A(m, r, N) represents the ratio of the number of similar template sequence logarithms to the number of total template sequence logarithms at scale m.

[0134] The sample entropy is calculated for each scale of the time series to quantify the complexity of the signal at that scale, where r is a multiple of the standard deviation, usually 0.1×SD, and SD is the standard deviation of the signal:

[0135]

[0136] S32, using machine learning algorithms to process the preprocessed sEMG signal data and auxiliary data d′ T (t), d′ H (t), d′V(t), d′ HR (t) Feature extraction, including time domain features, frequency domain features, and time-frequency domain features, reflects the average level, fluctuation level, and peak intensity of muscle activity, and is used to determine whether the muscle is fatigued and the degree of fatigue:

[0137] Extract time domain features from sEMG signals:

[0138]

[0139] Extract frequency domain features from sEMG signals, frequency domain features F freq It is obtained by calculating the power spectral density PSD, where PSD(f) is:

[0140] PSD(f)=|FFT(d′ sEMG (t))| 2 ;

[0141] Wherein, FFT stands for Fast Fourier Transform;

[0142] S33, combining the features extracted from the multi-scale entropy analysis and the machine learning algorithm to form a comprehensive feature vector for subsequent muscle fatigue level discrimination, the comprehensive feature vector F composite By combining the complexity features, time domain features, and frequency domain features obtained by multi-scale entropy analysis, freq And the auxiliary data features F considered aux constitute:

[0143] F composite = [SE(s, m, r), μ, σ 2 , P max ,PSD(f),F aux ];

[0144] Among them, F aux Including the mean and change rate of ambient temperature and humidity.

[0145] S4. Based on the extracted features, a composite artificial intelligence model for muscle fatigue level discrimination is constructed and trained. The composite artificial intelligence model comprehensively analyzes the key features of sEMG signals and auxiliary data to determine whether the muscle is fatigued and the degree of fatigue;

[0146] In this implementation, S4 specifically includes the following steps:

[0147] S41. Build a composite artificial intelligence model that combines deep learning and machine learning technologies to comprehensively analyze the key features of sEMG signals and auxiliary data to determine whether muscles are fatigued and the degree of fatigue:

[0148] The input layer of the composite artificial intelligence model receives the comprehensive feature vector F composite , feature conversion is performed through multi-layer neural network layers, and the output calculation formula of each layer of neural network is:

[0149] a [l] =g [l] (W [l] (a [l-1] ⊙d [l-1] )+b [l] );

[0150] Among them, d [l-1] is the Dropout vector, which is used to randomly close some neurons in the l-1th layer, ⊙ represents the Hadamard product;

[0151] Add support vector machine for features extracted based on DNN and auxiliary data features F aux Classify the muscle fatigue level and use weighted cross entropy loss to deal with possible class imbalance:

[0152]

[0153] Where C is the number of categories, w i is the weight of category i, y i is the true label, is the predicted probability;

[0154] S42, prepare a training data set, including feature vectors F extracted from the actual measured sEMG signal and auxiliary data composite And the corresponding muscle fatigue level labels, the training data set is in the form of:

[0155]

[0156] Among them, y (i) is the muscle fatigue level label of the i-th sample, and N is the total number of samples;

[0157] S43. Use the training data set to train the composite artificial intelligence model and use cross-validation technology to optimize the model parameters:

[0158] Use a part of the training data to train the model, combined with the feature vector F composite Including complexity features, time domain features, frequency domain features and auxiliary data features obtained by multi-scale entropy analysis, a weighted loss function is used to integrate the effects of different types of features on muscle fatigue level:

[0159]

[0160] Among them, α and β are adjustment terms used to balance the importance of classification error and regression error, w f is the weight of feature f, x i,f Yes F composite The fth eigenvalue of the i-th sample in , is the predicted value of the feature, y i,c is the real classification label, is the corresponding predicted probability;

[0161] The performance of the model was verified on another part of the data. For the prediction of muscle fatigue recovery time, the definition based on F composite A linear regression model of the features and L2 regularization is applied to prevent overfitting:

[0162]

[0163] in, is based on the comprehensive feature vector F composite The fatigue recovery time of the i-th sample predicted by the linear regression model, λ is the regularization parameter;

[0164] S44, the trained composite artificial intelligence model is based on the input feature vector F composite Predict muscle fatigue level.

[0165] S5. Evaluate the user's muscle fatigue level in real time and adjust exercise intensity and rest time arrangement based on the evaluation results.

[0166] In this implementation, S5 specifically includes the following steps:

[0167] S51, according to the composite feature vector F composite The evaluation results, i.e. the muscle fatigue level, dynamically adjust the exercise intensity and rest time suggestions. The evaluation results are divided into multiple levels, and each level corresponds to different exercise and rest suggestions:

[0168] Adjustment = f(Level fatigue );

[0169] Among them, Level fatigue is the muscle fatigue level assessed by the model, and Adjustment is a function of the exercise intensity and rest time adjustment recommendations based on the fatigue level;

[0170] S52, calculate exercise intensity adjustment value AdjusL intensity and rest time adjustment value Adjust rest , the adjustment value is based on fatigue level and user's personal health data, including previous exercise habits and health status:

[0171] Adjust intensity =g intensity (Level fatigue ,Data personal );

[0172] Adjust rest =g rest (Level fatigue , Datd personal );

[0173] Among them, g intensity and g rest It is a function that calculates the exercise intensity adjustment value and rest time adjustment value based on the muscle fatigue level and personal health data. personal Represents the user’s personal health data;

[0174] S53. Provide real-time feedback through a user interface, displaying muscle fatigue level assessment results and exercise intensity and rest time recommendations adjusted based on these results.

[0175] Embodiment 1:

[0176] In a modern urban fitness center, a middle-aged man who has been engaged in IT work for a long time, called Mr. Li, decided to start regular fitness training to improve his health due to long hours of sitting in the office. In order to train more scientifically and efficiently and avoid muscle fatigue and injuries that may occur during exercise, Mr. Li chose an intelligent fitness auxiliary device that integrates the technology of the present invention.

[0177] In this scenario, the device collects muscle activity data through sEMG sensors installed on Mr. Li's main training muscle groups, and uses ambient temperature and humidity sensors and a heart rate monitor worn on Mr. Li's wrist to collect auxiliary data in real time. These data are wirelessly sent to the central processing unit in real time, where the technology described in the present invention is applied for analysis and processing.

[0178] During one training session, Mr. Li decided to do a high-intensity combination of aerobic exercise and strength training, including 30 minutes of treadmill exercise and a series of free weight training. In the first 10 minutes of treadmill exercise, the device analyzed sEMG signals and auxiliary data to show that Mr. Li's muscle activity was within the normal range and showed no obvious signs of fatigue. However, when the treadmill exercise was approaching 20 minutes, the system detected a significant increase in signs of fatigue in Mr. Li's leg muscles through the analysis of sEMG signals and auxiliary data by the composite artificial intelligence model.

[0179] Based on the analysis results of the composite artificial intelligence model, the system immediately issued a warning to Mr. Li and suggested adjusting the intensity of exercise, and also suggested adding 5 minutes of rest time before the next strength training. According to the system's advice, Mr. Li reduced the speed of the treadmill and took extra rest before the strength training.

[0180] In order to verify the effectiveness of the present invention, the fitness center compared the customer sports injury reports before and after the introduction of the present invention. Statistics show that before the introduction of this technology, there were an average of 150 reports of muscle strains or other related injuries caused by exercise per month. After the introduction of this technology, the number of reports of similar injuries decreased to an average of 3 per month, a decrease of 80.35%. At the same time, through comparative analysis, the customer's personal feeling report after training also showed a significant improvement, and the training satisfaction increased from the previous average of 75.48% to 95.95%.

[0181] In addition, to further demonstrate the beneficial effects of the present invention, the fitness center also specially recorded a three-month follow-up study of 200 participants (including Mr. Li). During the study, these participants all used the intelligent fitness auxiliary device integrating the technology of the present invention for daily training. The study showed that with the assistance of this technology, the average muscle recovery time of the participants was shortened from 48 hours when it was not used to 36 hours, and the training efficiency was improved by 25.34%. At the same time, the physical discomfort and fatigue reported by the participants after training were significantly reduced.

[0182] The present invention effectively improves the accuracy and efficiency of sEMG signal processing by adopting innovative signal preprocessing technology, including removing power supply interference, motion artifacts and other noises that may affect signal quality, and adopting bandpass filtering technology.

[0183] The present invention combines multi-scale entropy analysis and machine learning algorithms to extract richer and deeper features from sEMG signals and auxiliary data, including time domain features, frequency domain features, and time-frequency domain features, effectively capturing the dynamic changes in the muscle fatigue process. At the same time, the constructed composite artificial intelligence model comprehensively considers the key features of sEMG signals and auxiliary data, and can accurately determine whether the muscle is fatigued and the degree of fatigue.

[0184] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for extracting sEMG signal features and distinguishing muscle fatigue levels based on artificial intelligence, characterized in that: The following steps are involved: S1. Use multimodal sensor equipment to capture the sEMG signal data of users in real time when they are performing specific physical activities, and simultaneously collect auxiliary data such as ambient temperature, humidity, user's movement speed and heart rate; S2, preprocessing the collected sEMG signal data and auxiliary data, including removing power supply interference, motion artifacts and noise, and filtering the signal using bandpass filtering technology; S3. Use multiscale entropy analysis and machine learning algorithms to extract features from the preprocessed sEMG signal data, while considering the environmental and physiological parameters in the auxiliary data, and extract key features reflecting the muscle activity state from the sEMG signal data and auxiliary data. Multiscale entropy analysis is used to quantify the complexity changes of sEMG signals and capture the dynamic characteristics of muscle fatigue. The machine learning model further analyzes the dynamic characteristics, including time domain features, frequency domain features, and time-frequency domain features. S4. Based on the extracted features, a composite artificial intelligence model for muscle fatigue level discrimination is constructed and trained. The composite artificial intelligence model comprehensively analyzes the key features of sEMG signals and auxiliary data to determine whether the muscle is fatigued and the degree of fatigue; S5. Evaluate the user's muscle fatigue level in real time and adjust exercise intensity and rest time arrangement according to the evaluation results; The S1 specifically includes the following steps: S11, setting a multimodal sensor device, including a sEMG signal collector, a temperature sensor, a humidity sensor, a speedometer, and a heart rate monitor, for real-time monitoring and recording of physiological and environmental data of the user during specific physical activities; S12. Collect comprehensive data for each time point t: d final (t)={d sEMG (t),d T (t),d H (t),d V (t),d HR (t)}; Among them, d sEMG (t) is the sEMG signal value at time point t, d T (t) is the ambient temperature at time t, d H (t) is the ambient humidity value at time point t, d V (t) is the user's motion speed at time point t, d HR (t) is the user's heart rate value at time t; S13. Combine the comprehensive data of S12 to construct sEMG dataset D final , sEMG dataset D final The comprehensive data of all time points d final (t) Composition: D final ={d final (1),d final (2),...,d final (n)}; Where n is the total number of data points captured during a specific physical activity; The S2 specifically includes the following steps: S21, for the final data set D final Noise is removed for each data type in the , and the specific operations include: For sEMG signals sEMG (t), an adaptive filter is applied to remove power supply interference and motion artifacts. The adaptive filter dynamically adjusts its parameters according to the real-time characteristics of the signal to reduce the impact of noise. The adaptive filter uses the following formula for noise removal: in, is the noise estimated at time t, estimated by the minimum mean square error algorithm: Among them, w i (t) is the weight of the adaptive filter at time t, M is the order of the filter; For ambient temperature, humidity, user movement speed, and heart rate data, a simple moving average filtering method is used to remove random noise caused by environmental changes or device errors: Among them, d X (t) represents any auxiliary data ambient temperature d T (t), humidity d H (t), user movement speed d V (t), or heart rate data d HR (t) is the value at time t, N is the size of the averaging window, d′ X (t) is the data value after filtering; S22, then the processed sEMG signal d sEMG (t) Use bandpass filtering to keep the useful frequency components of the signal: Design a bandpass filter and choose a suitable lower limit frequency f L and upper frequency f H , in order to cover the important frequency range in the sEMG signal, f L Set to 20Hz, f H Set to 450Hz to filter out the typical frequencies of human muscle activity. Apply this bandpass filter F bp Onto the sEMG signal, use the Fast Fourier Transform to convert the signal to the frequency domain: D sEMG (f)=FFT(d sEMG (t)); D' sEMG (f)=D sEMG (f)·H(f); d' sEMG (t)=IFFT(D' sEMG (f)); Among them, D sEMG (f) is the Fourier transform of the original sEMG signal, and H(f) is the frequency domain representation of the bandpass filter; Apply filter characteristics to remove the L ,f H ] frequency components within the range, D' sEMG (f) is the representation of the sEMG signal in the frequency domain after bandpass filtering. It is converted back to the time domain by inverse Fourier transform to obtain the processed sEMG signal d' sEMG (t): S23, after preprocessing, all preprocessed data are recombined into a preprocessed comprehensive data set D preprocessed , which includes the noise-removed and filtered sEMG signal and auxiliary data: The S3 specifically includes the following steps: S31, preprocessing the sEMG signal data d' sEMG (t) Perform multi-scale entropy analysis to quantify the complexity changes of sEMG signals at different time scales, reflecting the dynamic characteristics of muscle fatigue: For each preprocessed sEMG signal sample, construct its multi-scale representation, and obtain time series of different scales by coarse-graining the signal. sEMG (t) is a coarse-grained time series at scale s The jth element is defined as: Among them, s represents the scale factor, which indicates the coarse-grained level of the time series considered in the multiscale entropy analysis. The larger the scale factor, the wider the range of the time series being averaged, which can capture the complexity changes of sEMG signals over a longer time range and understand the dynamic changes during muscle fatigue. It represents the jth element of the time series at scale s, which is obtained by coarse-graining the original sEMG signal. It represents the average activity intensity of the sEMG signal at a specific scale and analyzes the persistence and change trend of muscle activity. For each time series of scale s When calculating the sample entropy SE(s,m,r), define the distance function d[x,y] as the maximum absolute difference between the two sequences (x,y), and calculate the ratio of the matching template vectors A(m,r,N): Among them, m represents the embedding dimension, which is used to construct the dimension of the template vector in the sample entropy calculation. It reflects the amount of historical information of the time series considered when calculating the sample entropy, which is related to the time correlation of the signal during muscle fatigue. r represents the similarity tolerance, which is the threshold used to judge whether two sequences are similar in the sample entropy calculation. It is a ratio based on the standard deviation of the sEMG signal, which reflects the sensitivity of signal fluctuations and is used to identify subtle changes caused by muscle fatigue. N is the length of the time series. A(m,r,N) represents the ratio of the number of similar template sequence logarithms to the number of total template sequence logarithms at scale m. The sample entropy is calculated for each scale of the time series to quantify the complexity of the signal at that scale, where r is a multiple of the standard deviation, usually 0.1×SD, and SD is the standard deviation of the signal: S32, using machine learning algorithms to pre-process the sEMG signal data and auxiliary data d' T (t,d' H (t), d' V (t), d' HR (t) Feature extraction, including time domain features, frequency domain features, and time-frequency domain features, reflects the average level, fluctuation level, and peak intensity of muscle activity, and is used to determine whether the muscle is fatigued and the degree of fatigue: Extract time domain features from sEMG signals: Extract frequency domain features from sEMG signals, frequency domain features F freq It is obtained by calculating the power spectral density PSD, where PSD(f) is: PSD(f)=|FFT(d' sEMG (t))| 2 ; Wherein, FFT stands for Fast Fourier Transform; S33, combining the features extracted from the multi-scale entropy analysis and the machine learning algorithm to form a comprehensive feature vector for subsequent muscle fatigue level discrimination, the comprehensive feature vector F composite By combining the complexity features, time domain features, and frequency domain features obtained by multi-scale entropy analysis, freq And the auxiliary data features F considered aux constitute: F composite =[SE(s,m,r),μ,σ 2 ,P max ,PSD(f),F aux ]; Among them, F aux Including the mean and change rate of ambient temperature and humidity; The S4 specifically comprises the following steps: S41. Construct a composite artificial intelligence model combining deep learning and machine learning technologies to comprehensively analyze the key features of sEMG signals and auxiliary data to determine whether muscles are fatigued and the degree of fatigue: The input layer of the composite artificial intelligence model receives the comprehensive feature vector F composite , feature conversion is performed through multi-layer neural network layers, and the output calculation formula of each layer of neural network is: a [l] =g [l] (W [l] (a [l-1] ⊙d [l-1] )+b [l] ); Among them, d [l-1] is the Dropout vector, which is used to randomly close some neurons in the l-1th layer, ⊙ represents the Hadamard product; Add support vector machine for features extracted based on DNN and auxiliary data features F aux Classify the muscle fatigue level and use weighted cross entropy loss to deal with possible class imbalance: Where C is the number of categories, w i is the weight of category i, y i is the true label, is the predicted probability; S42, prepare a training data set, including feature vectors F extracted from the actual measured sEMG signal and auxiliary data composite And the corresponding muscle fatigue level labels, the training data set is in the form of: Among them, y (i) is the muscle fatigue level label of the i-th sample, and N is the total number of samples; S43. Use the training data set to train the composite artificial intelligence model and use cross-validation technology to optimize the model parameters: Use a part of the training data to train the model, combined with the feature vector F composite Including complexity features, time domain features, frequency domain features and auxiliary data features obtained by multi-scale entropy analysis, a weighted loss function is used to integrate the effects of different types of features on muscle fatigue level: Among them, α and β are adjustment terms used to balance the importance of classification error and regression error, w f is the weight of feature f, x i,f Yes F composite The fth eigenvalue of the i-th sample in , is the predicted value of the feature, y i,c is the real classification label, is the corresponding predicted probability; The performance of the model was verified on another part of the data. For the prediction of muscle fatigue recovery time, the definition based on F composite A linear regression model of the features and L2 regularization is applied to prevent overfitting: in, is based on the comprehensive feature vector F composite The fatigue recovery time of the i-th sample predicted by the linear regression model, λ is the regularization parameter; S44, the trained composite artificial intelligence model is based on the input feature vector F composite Predict muscle fatigue levels; The S5 specifically includes the following steps: S51, according to the composite feature vector F composite The evaluation results, i.e. the muscle fatigue level, dynamically adjust the exercise intensity and rest time suggestions. The evaluation results are divided into multiple levels, and each level corresponds to different exercise and rest suggestions: Adjustment=f(Level fatigue ); Among them, Level fatigue is the muscle fatigue level assessed by the model, and Adjustment is a function of the exercise intensity and rest time adjustment recommendations based on the fatigue level; S52, calculate the exercise intensity adjustment value Adjust intensity and rest time adjustment value Adjust rest , the adjustment value is based on fatigue level and user's personal health data, including previous exercise habits and health status: Adjust intensity =g intensity (Level fatigue ,Data personal ); Adjust rest =g rest (Level fatigue ,Data personal ); Among them, g intensity and g rest It is a function that calculates the exercise intensity adjustment value and rest time adjustment value based on the muscle fatigue level and personal health data. personal Represents the user’s personal health data; S53. Provide real-time feedback through a user interface, displaying muscle fatigue level assessment results and exercise intensity and rest time recommendations adjusted based on these results.

Citation Information

Patent Citations

  • Method and system for predicting muscle fatigue in real time based on wearable device

    CN115985463A

  • Muscle fatigue recognition method fusing electromyographic signal multi-fractal and LSTM (Long Short Term Memory)

    CN116244584A