CN114159079B - Multi-type muscle fatigue detection method based on feature extraction and GRU deep learning model - Google Patents

Abstract

The invention discloses a muscle fatigue detection method based on feature extraction and GRU deep learning models, which comprises the following steps: the method comprises the steps of 1, collecting surface electromyographic signals of a long-term back muscle group of a subject through a surface electromyographic sensor, dividing sample data, screening abnormal values, filtering and denoising, and setting classification labels according to fatigue limit; 2, extracting the cleaned data sample sliding window into a characteristic sequence of the shape [ s, c ], guiding the characteristic sequence into a GRU neural network for training, and setting sample sampling weight measures in the training process to solve the problem of unbalanced sample labels; and 3, adjusting the learning rate optimization model by using the verification set, selecting an optimal model by taking the accuracy of the verification set as a standard, and running the final model on the test set, wherein the fatigue detection of each final muscle area can reach an accuracy rate of more than 98%. The invention can overcome the limitation of the traditional single muscle detection method, carry out comprehensive fatigue detection on main muscle groups of the human body, and improve the detection accuracy.

Description

Multi-type muscle fatigue detection method based on feature extraction and GRU deep learning model

Technical Field

The invention relates to the technical field of physiological signal feature detection, in particular to a multi-type muscle fatigue detection method based on feature extraction and GRU deep learning model.

Background

The muscular system is an important part of the human body and provides power for various movements of the human body. However, after the muscles are kept tight for a long time or repeatedly operated, muscle fatigue can be generated, so that normal movement of a human body is affected, and even the muscles themselves can be damaged. Therefore, the accurate detection of the human muscle fatigue state is the basis of muscle fatigue relief and treatment, and has important kinematic and medical significance.

The surface myoelectricity (surface ElectroMyoGraphy, sEMG) signal is a weak current signal generated when muscle exercises, the change of the surface myoelectricity (surface ElectroMyoGraphy, sEMG) signal is related to factors such as the number of exercise units involved in the activities, the activity mode, the metabolic state and the like, the muscle activity state and the functional state can be accurately reflected in real time, and the surface myoelectricity (surface ElectroMyoGraphy, sEMG) signal has important practical value in the evaluation of muscle functions in the rehabilitation medical field and the fatigue judgment in the sports science and is mostly used for detecting the muscle fatigue of specific parts of a human body. Muscle fatigue is derived from a relatively complex physiological process, most researchers currently rely on experimental paradigms for analyzing muscle fatigue, extract different features for statistics and traditional machine learning analysis, and rely on a large number of manual operations in the early and late stages of the study, and lack real-time performance and accuracy.

The deep learning method adopts a multi-level neural network structure, can autonomously perform feature learning and hierarchical feature representation, and has the core of discarding links such as manual features in the traditional machine learning method. However, according to research, most of the existing deep learning methods are only used for fatigue detection of a specific muscle, so that the method has no universal applicability and simultaneously limits the real-time performance and accuracy of fatigue detection.

Disclosure of Invention

Aiming at the defect of lack of general applicability caused by single muscle in fatigue detection in the prior art, the invention provides a multi-type muscle fatigue detection method based on feature extraction and GRU deep learning model, so that more effective feature combination can be extracted, and multi-muscle fatigue detection is performed by combining with a GRU time sequence deep network, thereby improving general applicability and accuracy of detection.

The technical scheme adopted by the invention for solving the technical problems is as follows:

the invention discloses a multi-type muscle fatigue detection method based on feature extraction and GRU deep learning model, which is characterized by comprising the following steps:

s1: collecting surface electromyographic signal data of m muscles of a subject according to a sampling frequency f by a surface electromyographic detector, and respectively carrying out overlapping sliding window segmentation on the surface electromyographic signal data according to a window length u and a sliding step length step to generate a time sequence sample T of an ith muscle _i ＝[t ₁ ,t ₂ ,…,t _n ,…,t _N ]Wherein t is _n The potential value of the N-th surface electromyographic signal in the time sequence sample of the i-th muscle is represented, and N=u.f is the number of single sample signal points; i epsilon [1, m]；

S2: time series sample T for ith muscle _i Pretreatment and feature extraction:

s2.1: screening out time-series samples T of ith muscle _i Filtering and denoising the abnormal value in the sequence sample to obtain a preprocessed sequence sample T '' _i ；

S2.2: setting the window length to u ₁ For the time sequence samples T 'after pretreatment without overlapping' _i Sliding window and extracting s groups of characteristic sequences, each group comprising c different time domain and frequency domain characteristics, thereby generating a data structure as [ s, c ]]Characteristic sequence samples of the ith muscle of (2) are noted as

F _i ＝[[f ₁₁ ,f ₁₂ ,…,f _1c ],[f ₂₁ ,f ₂₂ ,…,f _2c ],…,[f _a1 ,f _a2 ,…,f _ab ,…,f _ac ],…,[f _s1 ,f _s2 ,…,f _sc ]]The method comprises the steps of carrying out a first treatment on the surface of the Wherein f _ab The b-th feature of the a-th feature sequence in the sample of the feature sequence representing the i-th muscle, a.epsilon.1, s]，b∈[1,c]；

S3: setting a sample tag and creating a dataset:

s3.1: fatigue limit T recorded in experiments _boundary For the time sequence samples T 'after pretreatment according to the setting classification labels' _i If t _N ＜T _boundary Then the characteristic sequence of the ith muscle is setSample F _i The label of (2) is k ₁ The method comprises the steps of carrying out a first treatment on the surface of the If t ₁ ＜T _boundary ＜t _N Setting F _i The label of (2) is k ₂ The method comprises the steps of carrying out a first treatment on the surface of the If t ₁ ＞T _boundary Setting F _i The label of (2) is k ₃ Thereby obtaining a characteristic sequence sample F 'with a label' _i Thereby obtaining a characteristic sequence sample set with a label

S3.2: sample of tagged signature sequence F' _i The characteristic sequence sample of each muscle is divided into an ith training set S _train-i And the ith verification set S _val-i Thereby mixing training sets of m muscles and constructing a total training set S _train Mix the validation sets of m muscles and construct the total validation set S _val ；

S4: from the labeled feature sequence sample set

The ratio of the number of 3 kinds of label samples in the system is used for generating the sampling weight of the characteristic sequence sample of each muscle by using a weighted random sampling method, and the sampling weight is used as a total training set S _train Sampling probability of corresponding feature sequence samples in the training process;

s5: F_GRU neural network model composed of D-layer GRU units is constructed, and the normalized total training set S 'is obtained' _train Inputting bs into the F_GRU neural network model according to the size of each batch for training, and normalizing the total verification set S 'after training' _val Verifying the model obtained by training according to bs as the size of each batch, continuously adjusting the learning rate lr by taking the accuracy ACC as an evaluation index, and stopping training when the accuracy ACC is not improved any more, so as to obtain the model with the highest retention accuracy as a final trained F_GRU neural network model;

s6: and performing fatigue detection on the surface electromyographic signal data to be tested by using the final trained F_GRU neural network model, and outputting a fatigue state corresponding to the classification label.

Compared with the prior art, the invention has the beneficial effects that:

1. the feature sequence structure designed by the invention fully characterizes the sample by using the extracted features, and simultaneously retains the feature change information of a single sample in the time dimension, so that the feature sequence structure has stronger feature characterization capability than that of extracting only one group of features of the single sample;

2. in the classification effect, the classification accuracy rate of the designed characteristic sequence data structure matched with the GRU network method for completing the muscle fatigue task can reach more than 98%, and the classification accuracy rate is superior to that of the traditional statistical analysis and machine learning classification method;

3. from different muscle types, the classification effect of the model obtained by training is not greatly different on different types of muscles, and the classification effect is about 98%, so that the model fully shows that the model learns the common expression relationship among different types of muscles, and the model has general applicability;

4. in terms of performance, the method disclosed by the invention has the advantages of shorter time consumption, less occupied memory and stronger comprehensive performance for completing tasks after extracting and extracting the features, and is suitable for detecting task requirements in real time.

Drawings

FIG. 1 is a flow chart of a method of fatigue detection based on feature extraction and GRU network of the present invention.

FIG. 2 is a schematic diagram of data segmentation, extracted features and label partitioning in accordance with the present invention.

FIG. 3 is a schematic block diagram of the GRU unit of the invention.

Detailed Description

In the embodiment, the multi-type muscle fatigue detection method based on the feature extraction and the GRU deep learning model is characterized in that a plurality of muscle sEMG signals are collected as training data, effective time domain and frequency domain feature combinations are further designed and extracted to be used as input, and the principle characteristics of GRU deep learning network in time sequence detection are combined, so that the multi-muscle fatigue detection method is designed, and the GRU network model trained by using a unique sample structure and more effective feature combinations has the advantages of strong universal applicability and high accuracy when performing multi-muscle fatigue detection tasks. Specifically, as shown in fig. 1, the steps are as follows:

s1: collecting surface electromyographic signal data of m muscles of a subject according to a sampling frequency f by a surface electromyographic detector, and respectively carrying out overlapping sliding window segmentation on the surface electromyographic signal data according to a window length u and a sliding step length step to generate a time sequence sample T of an ith muscle _i ＝[t ₁ ,t ₂ ,…,t _n ,…,t _N ]Wherein t is _n The potential value of the N-th surface electromyographic signal in the time sequence sample of the i-th muscle is represented, and N=u.f is the number of single sample signal points; i epsilon [1, m]The method comprises the steps of carrying out a first treatment on the surface of the In the embodiment, the subjects perform simulation actions on a computer platform, 8 sEMG original data of 30 subjects are collected, the sampling frequency f=1000 Hz, and finally 8-person data are screened and reserved; setting the window length u=3min and the sliding step length step=5s, wherein n=180000, namely one time sequence sample comprises 180000 data points, two adjacent samples overlap 175000 points, the original data is fully utilized, and the sample point division is shown as a data axis in fig. 2;

s2: time series sample T for ith muscle _i Pretreatment and feature extraction:

s2.1: screening out time-series samples T of ith muscle _i Filtering and denoising the abnormal value in the sequence sample to obtain a preprocessed sequence sample T '' _i The method comprises the steps of carrying out a first treatment on the surface of the In the embodiment, the screened abnormal values comprise null values, super-limit values and the like, the Butterworth 5Hz high-pass filtering is used for reserving a plurality of frequency bands of the electromyographic signals, and the 50Hz band-stop filtering is used for removing power frequency interference;

s2.2: setting the window length to u ₁ For the time sequence samples T 'after pretreatment without overlapping' _i Sliding window and extracting s groups of characteristic sequences, each group comprising c different time domain and frequency domain characteristics, thereby generating a data structure as [ s, c ]]Characteristic sequence samples of the ith muscle of (2) are noted as

F _i ＝[[f ₁₁ ,f ₁₂ ,…,f _1c ],[f ₂₁ ,f ₂₂ ,…,f _2c ],…,[f _a1 ,f _a2 ,…,f _ab ,…,f _ac ],…,[f _s1 ,f _s2 ,…,f _sc ]]The method comprises the steps of carrying out a first treatment on the surface of the Wherein f _ab A b-th feature of the a-th feature sequence in the sample of feature sequences representing the i-th muscle; a E [1, s ]]，b∈[1,c]In the present embodiment, u is set ₁ =0.5 s, s=36, c=9, i.e. each set of sequences contains 5 time domain features: zero crossing rate ZC, average rectification value ARV, root mean square myoelectric value RMS, average absolute value MAV, sign change slope SSC and 4 frequency domain features: median frequency MDF and average power frequency MNF based on fourier transform, median frequency IMDF and average power frequency IMNF based on wavelet transform, generating a data shape of [36,9 ]]As shown in fig. 2;

s3: setting a sample tag and creating a dataset:

s3.1: fatigue limit T recorded in experiments _boundary For the time sequence samples T 'after pretreatment according to the setting classification labels' _i If t _N ＜T _boundary Setting the characteristic sequence sample F of the ith muscle _i The label of (2) is k ₁ The method comprises the steps of carrying out a first treatment on the surface of the If t ₁ ＜T _boundary ＜t _N Setting F _i The label of (2) is k ₂ The method comprises the steps of carrying out a first treatment on the surface of the If t ₁ ＞T _boundary Setting F _i The label of (2) is k ₃ Thereby obtaining a characteristic sequence sample F 'with a label' _i Thereby obtaining a characteristic sequence sample set with a label

In the present embodiment, k is set ₁ ＝0、k ₂ ＝1、k ₃ =2, corresponding to the non-fatigue state, the fatigue transition state, and the fatigue state, respectively;

s3.2: sample of tagged signature sequence F' _i The characteristic sequence sample of each muscle is divided into an ith training set S _train-i And the ith verification set S _val-i Thereby mixing training sets of m muscles and constructing a total training set S _train Mix the validation sets of m muscles and construct the total validation set S _val The method comprises the steps of carrying out a first treatment on the surface of the In the embodiment, the training set and the verification set are divided in a ratio of 8:2;

s4: based on tagged signature sequencesSample set

The ratio of the number of 3 kinds of label samples in the system is used for generating the sampling weight of the characteristic sequence sample of each muscle by using a weighted random sampling method, and the sampling weight is used as a total training set S _train Sampling probability of corresponding feature sequence samples in the training process; in the embodiment, the weighted value of each sampled sample is generated by means of the weightedrandom sampler tool in the pytorch, so that the problem of unbalanced sample labels is solved;

s5: F_GRU neural network model composed of D-layer GRU units is constructed, and the normalized total training set S 'is obtained' _train Inputting bs into the F_GRU neural network model according to the size of each batch for training, and normalizing the total verification set S 'after training' _val Verifying the model obtained by training according to bs as the size of each batch, continuously adjusting the learning rate lr by taking the accuracy ACC as an evaluation index, and stopping training when the accuracy ACC is not improved any more, so as to obtain the model with the highest retention accuracy as a final trained F_GRU neural network model; in this embodiment, d=2, the structure of the GRU unit is shown in fig. 3, a dual-layer f_gru network is constructed, the number of input samples per batch bs=128 is set, and the batch input data structure thus constructed is [ bs, s, c ]] ＝[128,36,9]After reading the data sample, performing single-channel normalization, and performing iterative training by using an InstanceNorm2d command in a pytorch, wherein the initial learning rate lr=0.01 is set;

s6: performing fatigue detection on surface electromyographic signal data to be tested by utilizing the final trained F_GRU neural network model, and outputting a fatigue state corresponding to the classification label; in the step S6.1-S6.2 shown in fig. 1, in the present example, surface electromyographic signal data is recorded every 3 periods in real-time detection, filtering and denoising are performed on the data collected at the current moment, and feature extraction is performed, in the same step S2, the processed feature samples are normalized and then input into the GRU optimal model obtained in step S5, and fatigue detection results of eight muscles are output, namely the final result.

Claims (1)

1. A multi-type muscle fatigue detection method based on feature extraction and GRU deep learning model is characterized by comprising the following steps:

s1: collecting surface electromyographic signal data of m muscles of a subject according to a sampling frequency f by a surface electromyographic detector, and respectively carrying out overlapping sliding window segmentation on the surface electromyographic signal data according to a window length u and a sliding step length step to generate a time sequence sample T of an ith muscle _i ＝[t ₁ ,t ₂ ,…,t _n ,…,t _N ]Wherein t is _n The potential value of the N-th surface electromyographic signal in the time sequence sample of the i-th muscle is represented, and N=u.f is the number of single sample signal points; i epsilon [1, m]；

S2: time series sample T for ith muscle _i Pretreatment and feature extraction:

s2.1: screening out time-series samples T of ith muscle _i Filtering and denoising the abnormal value in the sequence sample T after pretreatment _i ′；

S2.2: setting the window length to u ₁ For the time sequence sample T after pretreatment without overlapping _i ' sliding window and extracting s sets of feature sequences, each set comprising c different time-domain, frequency-domain features, thereby generating a data structure of [ s, c ]]Characteristic sequence samples of the ith muscle of (2) are noted as

F _i ＝[[f ₁₁ ,f ₁₂ ,…,f _1c ],[f ₂₁ ,f ₂₂ ,…,f _2c ],…,[f _a1 ,f _a2 ,…,f _ab ,…,f _ac ],…,[f _s1 ,f _s2 ,…,f _sc ]]The method comprises the steps of carrying out a first treatment on the surface of the Wherein f _ab The b-th feature of the a-th feature sequence in the sample of the feature sequence representing the i-th muscle, a.epsilon.1, s]，b∈[1,c]；

S3: setting a sample tag and creating a dataset:

s3.1: fatigue limit T recorded in experiments _boundary For the time sequence sample T after pretreatment according to the setting of the classification label _i ' if t _N ＜T _boundary Setting the characteristic sequence sample F of the ith muscle _i The label of (2) is k ₁ The method comprises the steps of carrying out a first treatment on the surface of the If t ₁ ＜T _boundary ＜t _N Setting F _i The label of (2) is k ₂ The method comprises the steps of carrying out a first treatment on the surface of the If t ₁ ＞T _boundary Setting F _i The label of (2) is k ₃ Thereby obtaining a characteristic sequence sample F with a label _i ' and further obtaining a characteristic sequence sample set with a label

S3.2: sample F of the signature sequence with the tag _i The feature sequence sample of each muscle in' is divided into an ith training set S _train-i And the ith verification set S _val-i Thereby mixing training sets of m muscles and constructing a total training set S _train Mix the validation sets of m muscles and construct the total validation set S _val ；

S4: from the labeled feature sequence sample set

The ratio of the number of 3 kinds of label samples in the system is used for generating the sampling weight of the characteristic sequence sample of each muscle by using a weighted random sampling method, and the sampling weight is used as a total training set S _train Sampling probability of corresponding feature sequence samples in the training process;

s5: F_GRU neural network model composed of D-layer GRU units is constructed, and the normalized total training set S 'is obtained' _train Inputting bs into the F_GRU neural network model according to the size of each batch for training, and normalizing the total verification set S 'after training' _val Verifying the model obtained by training according to bs as the size of each batch, continuously adjusting the learning rate lr by taking the accuracy ACC as an evaluation index, and stopping training when the accuracy ACC is not improved any more, so as to obtain the model with the highest retention accuracy as a final trained F_GRU neural network model;

s6: and performing fatigue detection on the surface electromyographic signal data to be tested by using the final trained F_GRU neural network model, and outputting a fatigue state corresponding to the classification label.

Images