CN114159079A - 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 model, which comprises the following steps: 1, acquiring surface electromyographic signals of long-term back muscle groups of a subject through a surface electromyographic sensor, segmenting sample data, screening abnormal values, filtering and denoising, and setting a classification label based on a fatigue limit; 2, extracting a characteristic sequence with the shape [ s, c ] from the cleaned data sample sliding window, importing the characteristic sequence into a GRU neural network for training, and setting a sample sampling weight measure in the training process to solve the problem of unbalanced sample labels; and 3, adjusting a 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 operating the final model on the test set, wherein the fatigue detection of each muscle area can reach the accuracy of more than 98 percent. The invention can overcome the limitation of the traditional single muscle detection method, carry out comprehensive fatigue detection on the main muscle group 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 models.

Background

The muscular system is an important component of the human body and provides power for various movements of the human body. However, the muscles can be kept tight for a long time or can generate muscle fatigue after repeated work, thereby affecting the normal movement of the human body and even causing damage to the muscles. Therefore, accurate detection of the muscle fatigue state of the human body is the basis of relieving and treating the muscle fatigue, and has important kinematic and medical significance.

Surface ElectroMyoGraphy (sEMG) signals are weak current signals generated during muscle movement, the changes of the sEMG signals are related to factors such as the number of motor units participating in the movement, the movement mode, the metabolic state and the like, the muscle movement state and the function state can be accurately reflected in real time, and the sEMG signals have important practical values in muscle function evaluation in the field of rehabilitation medicine and fatigue judgment in sports science and are mostly used for detecting muscle fatigue of specific parts of a human body. Muscle fatigue is derived from a relatively complex physiological process, most researchers at present generally rely on experimental paradigms when analyzing muscle fatigue, different features are extracted to carry out statistics and traditional machine learning analysis, the prior and later stages of the research all rely on a large number of manual operations, and instantaneity and accuracy are lacked.

The deep learning method adopts a multi-level neural network structure, can autonomously perform feature learning and hierarchical feature representation, and is most central in abandoning links such as manual features in the traditional machine learning method. However, according to research and development, most of the existing deep learning methods are only used for fatigue detection on a specific muscle, so that the method is not universally applicable, and meanwhile, the real-time performance and the accuracy of the fatigue detection are also limited.

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 a feature extraction and GRU deep learning model, so that more effective feature combinations can be extracted, and multi-muscle fatigue detection is carried out by combining a GRU time sequence deep network, thereby improving the general applicability and the accuracy of detection.

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

the invention relates to 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: acquiring surface electromyographic signal data of m muscles of a subject according to sampling frequency f by a surface electromyographic detector, and respectively carrying out sliding window segmentation on the surface electromyographic signal data in an overlapping manner according to window length u and sliding step length step to generate a time sequence sample T of the ith muscle_i＝[t₁,t₂,…,t_n,…,t_N]Wherein, t_nThe method comprises the steps of representing the potential value of an nth surface electromyogram signal in a time sequence sample of the ith muscle, wherein N-u-f is the number of single sample signal points; i is an element of [1, m ]]；

S2: time series sample T for ith muscle_iPreprocessing and characteristic extraction:

s2.1: time series sample T for screening out ith muscle_iThe abnormal value in the time sequence is filtered and denoised to obtain a preprocessed time sequence sample T'_i；

S2.2: setting window length to u₁The preprocessed time sequence samples T 'are processed without overlap'_iPerforming sliding window and extracting s groups of characteristic sequences, wherein each group comprises c different time domain and frequency domain characteristics, thereby generating a data structure of s, c]The characteristic sequence sample of the i-th muscle of (1), is recorded as

F_i＝[[f₁₁,f₁₂,…,f_1c],[f₂₁,f₂₂,…,f_2c],…,[f_a1,f_a2,…,f_ab,…,f_ac],…,[f_s1,f_s2,…,f_sc]](ii) a Wherein f is_abRepresents the b-th feature of the a-th feature sequence in the i-th muscle feature sequence sample, a E [1, s [ ]]，b∈[1,c]；

S3: set sample labels and create datasets:

s3.1: fatigue limit T recorded by experiment_boundarySetting a classification label according to the condition, and carrying out pretreatment on the time sequence sample T'_iIf t is_N＜T_boundaryThen set the characteristic sequence sample F of the ith muscle_iIs k₁(ii) a If t₁＜T_boundary＜t_NThen set up F_iIs k₂(ii) a If t₁＞T_boundaryThen set up F_iIs k₃To obtain a characteristic sequence sample F 'with a label'_iFurther obtaining a characteristic sequence sample set with a label

S3.2: sample F 'of tag-bearing signature sequence'_iDividing the characteristic sequence sample of each muscle into the ith training set S_train-iAnd the ith verification set S_val-iThereby mixing training sets of m kinds of muscles and constructing a total training set S_trainMixing validation sets of m muscles and constructing a total validation set S_val；

S4: according to the characteristic sequence sample set with the label

The quantity ratio of the medium 3 types of label samples is used for generating the sampling weight of the characteristic sequence sample of each muscle by using a weighted random sampling method and is used as a total training set S_trainSampling probability of corresponding characteristic sequence samples in the training process;

s5: build up of D layersF _ GRU neural network model composed of GRU units, and normalized total training set S'_trainInputting the size of each batch as bs into the F _ GRU neural network model for training, and after training, normalizing the total verification set S'_valVerifying the trained model according to the size of each batch as bs, continuously adjusting the learning rate lr by taking the accuracy ACC as an evaluation index, and stopping training when the accuracy ACC is not increased any more, so as to obtain a model with the highest retention accuracy as a finally trained F _ GRU neural network model;

s6: and carrying out fatigue detection on the surface electromyographic signal data to be tested by using the finally 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 characteristic sequence structure designed by the invention fully represents the sample by using the extracted characteristics, simultaneously reserves the characteristic change information of a single sample on the time dimension, and has stronger characteristic capability than that of extracting only one group of characteristics from the single sample;

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

3. from the view of different muscle types, the classification effect of the model obtained by training is not greatly different on different types of muscles, the classification effect is about 98 percent, the expression relation of the model learning commonality among different types of muscles is fully explained, and the universal applicability is realized;

4. in terms of performance, the method provided by the invention has the advantages that the time consumed for completing the task after extracting the features is shorter, the occupied memory is less, the comprehensive performance is stronger, and the method is suitable for detecting the task requirement in real time.

Drawings

Fig. 1 is a flowchart of a fatigue detection method based on feature extraction and GRU network of the present invention.

FIG. 2 is a schematic diagram of the data segmentation, feature extraction and label partitioning of the present invention.

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

Detailed Description

In the embodiment, a method for detecting multiple types of muscle fatigue based on feature extraction and a GRU deep learning model is characterized in that multiple muscle sEMG signals are collected as training data, effective time domain and frequency domain feature combinations are further designed and extracted as input, and a method for detecting the multiple muscle fatigue is designed by combining the principle characteristic of GRU deep learning network in time sequence detection. 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 sampling frequency f by a surface electromyographic detector, and respectively carrying out sliding window segmentation on the surface electromyographic signal data in an overlapping manner by using window length u and sliding step length step to generate a time sequence sample T of the ith muscle_i＝[t₁,t₂,…,t_n,…,t_N]Wherein, t_nThe method comprises the steps of representing the potential value of an nth surface electromyogram signal in a time sequence sample of the ith muscle, wherein N-u-f is the number of single sample signal points; i is an element of [1, m ]](ii) a In this embodiment, a subject performs a simulation action on a computer platform, 8 types of sEMG raw data of 30 subjects are collected, a sampling frequency f is 1000Hz, and data of 8 subjects are finally screened and retained; setting the window length u to be 3min and the sliding step length to be 5s, then N to be 180000, that is, one time sequence sample contains 180000 data points, and two adjacent samples overlap 175000 points, making full use of the original data, and dividing the sample points as shown in the data axis of fig. 2;

s2: time series sample T for ith muscle_iPreprocessing and characteristic extraction:

s2.1: time series sample T for screening out ith muscle_iThe abnormal value in the time sequence is filtered and denoised to obtain a preprocessed time sequence sample T'_i(ii) a In this embodiment, the abnormal values to be screened out include null valuesAnd the ultra-limit value and the like, reserving several frequency bands of electromyographic signals by using Butterworth 5Hz high-pass filtering, and removing power frequency interference by using 50Hz band-stop filtering;

s2.2: setting window length to u₁The preprocessed time sequence samples T 'are processed without overlap'_iPerforming sliding window and extracting s groups of characteristic sequences, wherein each group comprises c different time domain and frequency domain characteristics, thereby generating a data structure of s, c]The characteristic sequence sample of the i-th muscle of (1), is recorded as

F_i＝[[f₁₁,f₁₂,…,f_1c],[f₂₁,f₂₂,…,f_2c],…,[f_a1,f_a2,…,f_ab,…,f_ac],…,[f_s1,f_s2,…,f_sc]](ii) a Wherein f is_abRepresenting the b-th feature of the a-th feature sequence in the feature sequence sample of the i-th muscle; a is in [1, s ]]，b∈[1,c]In this embodiment, u is set₁0.5s, 36 s, and 9 c, 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: the data shape is [36,9 ] generated based on the median frequency MDF and the average power frequency MNF of the Fourier transform, the median frequency IMDF and the average power frequency IMNF of the wavelet transform]As shown in fig. 2;

s3: set sample labels and create datasets:

s3.1: fatigue limit T recorded by experiment_boundarySetting a classification label according to the condition, and carrying out pretreatment on the time sequence sample T'_iIf t is_N＜T_boundaryThen set the characteristic sequence sample F of the ith muscle_iIs k₁(ii) a If t₁＜T_boundary＜t_NThen set up F_iIs k₂(ii) a If t₁＞T_boundaryThen set up F_iIs k₃To obtain a characteristic sequence sample F 'with a label'_iFurther obtaining a characteristic sequence sample set with a label

In this embodiment, k is set₁＝0、k₂＝1、k₃2, corresponding to a non-fatigue state, a fatigue transition state and a fatigue state respectively;

s3.2: sample F 'of tag-bearing signature sequence'_iDividing the characteristic sequence sample of each muscle into the ith training set S_train-iAnd the ith verification set S_val-iThereby mixing training sets of m kinds of muscles and constructing a total training set S_trainMixing validation sets of m muscles and constructing a total validation set S_val(ii) a In the embodiment, a training set and a verification set are divided according to the ratio of 8: 2;

s4: sample set based on tagged feature sequences

The quantity ratio of the medium 3 types of label samples is used for generating the sampling weight of the characteristic sequence sample of each muscle by using a weighted random sampling method and is used as a total training set S_trainSampling probability of corresponding characteristic sequence samples in the training process; in the embodiment, by means of a weightedRandomSampler tool in the pytorch, the sampled weight value of each sample is generated, and the problem of sample label imbalance is solved;

s5: constructing an F _ GRU neural network model consisting of D-layer GRU units, and normalizing the total training set S'_trainInputting the size of each batch as bs into an F _ GRU neural network model for training, and after training, normalizing the total verification set S'_valVerifying the trained model according to the size of each batch as bs, continuously adjusting the learning rate lr by taking the accuracy ACC as an evaluation index, and stopping training when the accuracy ACC is not increased any more, so as to obtain a model with the highest retention accuracy as a finally trained F _ GRU neural network model; in this embodiment, D is 2, the GRU unit structure is as shown in fig. 3, a two-layer F _ GRU network is constructed, the number of input samples per batch bs is set to 128, and the batch input data structure formed by this is [ bs, s, c] ＝[128,36,9]After reading the data samples, single channel normalization was performed using the InstanceNorm2d command in the pytorchNow, the initial learning rate lr is set to 0.01 for iterative training;

s6: carrying out fatigue detection on the surface electromyographic signal data to be tested by using the finally trained F _ GRU neural network model, and outputting a fatigue state corresponding to the classification label; as shown in steps S6.1-S6.2 in fig. 1, in the real-time detection, surface electromyogram signal data is recorded every 3 cycles, filtering, denoising and feature extraction are performed on the data acquired at the current time, in the same step S2, the processed feature samples are normalized and input into the GRU optimal model obtained in step S5, and fatigue detection results of eight muscles are output, that is, final results are obtained.

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: acquiring surface electromyographic signal data of m muscles of a subject according to sampling frequency f by a surface electromyographic detector, and respectively carrying out sliding window segmentation on the surface electromyographic signal data in an overlapping manner according to window length u and sliding step length step to generate a time sequence sample T of the ith muscle_i＝[t₁,t₂,…,t_n,…,t_N]Wherein, t_nThe method comprises the steps of representing the potential value of an nth surface electromyogram signal in a time sequence sample of the ith muscle, wherein N-u-f is the number of single sample signal points; i is an element of [1, m ]]；

S2: time series sample T for ith muscle_iPreprocessing and characteristic extraction:

s2.1: time series sample T for screening out ith muscle_iThen filtering and denoising the abnormal value to obtain a preprocessed time sequence sample T_i′；

S2.2: setting window length to u₁The preprocessed time-series samples T are processed without overlap_i' sliding window and extracting s groups of characteristic sequences, each group containing c different time domain and frequency domain characteristics, thereby generating a data structure of [ s, c]The characteristic sequence sample of the i-th muscle of (1), is recorded as

F_i＝[[f₁₁,f₁₂,…,f_1c],[f₂₁,f₂₂,…,f_2c],…,[f_a1,f_a2,…,f_ab,…,f_ac],…,[f_s1,f_s2,…,f_sc]](ii) a Wherein f is_abRepresents the b-th feature of the a-th feature sequence in the i-th muscle feature sequence sample, a E [1, s [ ]]，b∈[1,c]；

S3: set sample labels and create datasets:

s3.1: fatigue limit T recorded by experiment_boundaryFor setting classification labels according to the pre-processed time sequence samples T_i', if t_N＜T_boundaryThen set the characteristic sequence sample F of the ith muscle_iIs k₁(ii) a If t₁＜T_boundary＜t_NThen set up F_iIs k₂(ii) a If t₁＞T_boundaryThen set up F_iIs k₃Thereby obtaining a characteristic sequence sample F with a label_i', to obtain a sample set of tagged feature sequences

S3.2: sample F of characteristic sequence with label_i' the characteristic sequence samples of each muscle are divided into the ith training set S_train-iAnd the ith verification set S_val-iThereby mixing training sets of m kinds of muscles and constructing a total training set S_trainMixing validation sets of m muscles and constructing a total validation set S_val；

S4: according to the characteristic sequence sample set with the label

The quantity ratio of the medium 3 types of label samples is used for generating the sampling weight of the characteristic sequence sample of each muscle by using a weighted random sampling method and is used as a total training set S_trainSampling probability of corresponding characteristic sequence samples in the training process;

s5: constructing F _ GRU neural network composed of D-layer GRU unitsModel, and normalized total training set S'_trainInputting the size of each batch as bs into the F _ GRU neural network model for training, and after training, normalizing the total verification set S'_valVerifying the trained model according to the size of each batch as bs, continuously adjusting the learning rate lr by taking the accuracy ACC as an evaluation index, and stopping training when the accuracy ACC is not increased any more, so as to obtain a model with the highest retention accuracy as a finally trained F _ GRU neural network model;

s6: and carrying out fatigue detection on the surface electromyographic signal data to be tested by using the finally trained F _ GRU neural network model, and outputting a fatigue state corresponding to the classification label.

Images