CN114533089A - Lower limb action recognition and classification method based on surface electromyographic signals - Google Patents
Lower limb action recognition and classification method based on surface electromyographic signals Download PDFInfo
- Publication number
- CN114533089A CN114533089A CN202210161420.2A CN202210161420A CN114533089A CN 114533089 A CN114533089 A CN 114533089A CN 202210161420 A CN202210161420 A CN 202210161420A CN 114533089 A CN114533089 A CN 114533089A
- Authority
- CN
- China
- Prior art keywords
- signals
- electromyographic
- frequency
- electromyographic signals
- window
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- A—HUMAN NECESSITIES
- A61—MEDICAL OR VETERINARY SCIENCE; HYGIENE
- A61B—DIAGNOSIS; SURGERY; IDENTIFICATION
- A61B5/00—Measuring for diagnostic purposes; Identification of persons
- A61B5/24—Detecting, measuring or recording bioelectric or biomagnetic signals of the body or parts thereof
- A61B5/316—Modalities, i.e. specific diagnostic methods
- A61B5/389—Electromyography [EMG]
- A61B5/397—Analysis of electromyograms
Abstract
A lower limb action recognition and classification method based on surface electromyographic signals belongs to the technical field of electromyographic signal action recognition, and aims at the characteristics of high difficulty in a conventional electromyographic signal preprocessing process, complex processing process and nonlinearity of electroencephalogram signals. The method uses a Butterworth filter and a wave trap filter to carry out filtering processing, uses Hilbert transform to obtain an envelope curve of an electromyographic signal, uses overlapped sliding windows to segment the waveform of the electromyographic signal and extracts a characteristic value in the window. The method comprises the following steps: collecting electromyographic signals, preprocessing the electromyographic signals, segmenting the electromyographic signals through a sliding window, extracting characteristic values in the window, fusing the characteristics to form a characteristic vector, defining a BP neural network model structure, and training and identifying the neural network model. The invention improves the input of the neural network model from the angles of the pretreatment and the characteristic value extraction of the electromyographic signals, can simplify the complexity of the electromyographic signal pretreatment, and simultaneously improves the efficiency and the accuracy of the action recognition.
Description
Technical Field
The invention belongs to the technical field of electromyographic signal action recognition, and relates to a method for recognizing and classifying lower limb actions based on surface electromyographic signals. Compared with other electromyographic signal action identification methods, the method mainly solves the problem of complex preprocessing such as noise reduction and filtering in the identification process. According to the method, the filtered electromyographic signals are used for performing envelope analysis, the signals are processed by combining a sliding window, the signals in continuous time are segmented, the characteristic vectors are extracted for recognition by the method, and good classification accuracy is obtained on a data set.
Background
Surface electromyography (sEMG) signals are electrical fluctuations produced during muscle contraction, which are related to both the tissue physiological properties of the muscle itself and the nervous control system, reflecting the neuromuscular activity and functional status. Therefore, the electromyographic signals have been widely used in the research fields of the physiological medicine, the rehabilitation medicine, the sports medicine, etc., and become ideal control signals for driving the robot, controlling the movement of the prosthesis, and functional electrical stimulation. The sEMG has the advantages of non-invasiveness, no wound, simple operation and the like in measurement. Therefore, sEMG has important practical value in clinical medicine, rehabilitation medicine, sports science and the like, and sEMG is the only means which can enable a rehabilitation doctor to research the actual muscle function condition of a patient under a dynamic condition at present. In recent years, with the development of artificial intelligence, reliable technical means are provided for action recognition research based on electromyographic signals.
(1) The prior art has the following defects:
(2) the accuracy of the traditional identification method is closely related to the preprocessing process of the electromyographic signals, but the noise reduction process is a very difficult and complex process.
(3) In the research of lower limb movement recognition based on muscle electrical signals, there are many aspects that need to be improved in terms of how to more accurately calculate the characteristic value of the muscle electrical signals, how to select a classifier, optimizing the performance of a pattern recognition classifier and the like. For example, in the study of the characteristics of the muscle electrical signal, the muscle electrical signal has the characteristics of instability and nonlinearity, which causes the problem of inconsistency between the classification model and the classification object, and the most basic physiological information of the muscle electrical signal is often ignored, so a more effective method is needed to extract hidden detailed information.
Disclosure of Invention
The invention aims to provide a novel method for carrying out envelope analysis based on electromyographic signals to identify lower limb actions aiming at the defects of the prior art. The method mainly simplifies the preprocessing process of the electromyographic signals, extracts signal information in an effective frequency range, and extracts envelope lines of the signals. Then, feature extraction and classification recognition are carried out on the basis by a sliding window segmentation method. Meanwhile, a four-layer BP neural network containing two hidden layers is adopted, fusion feature vectors of time domain features and frequency domain features of 6 channels are used as input, and the number of neurons of the input layer is consistent with the dimension of the feature vectors. The BP neural network has strong self-adaptive learning ability, adopts distributed memory to information, is not easy to lose the information, and has good fault tolerance and associative memory ability.
The technical scheme adopted by the invention for solving the technical problem comprises the following steps:
Step 2, preprocessing the surface electromyographic signals of the lower limbs: in order to increase the calculation speed, the electromyographic signals collected in the step 1 are subjected to Butterworth filtering to extract electromyographic signals in an effective frequency band, 50Hz and frequency multiplication interference signals thereof are removed by a wave trap, envelope curves of the filtered electromyographic signals are extracted by Hilbert transform, and the electromyographic signals are just convenient for operations such as rear sliding window segmentation processing, data normalization and the like.
Step 3, sliding window segmentation treatment: by adopting the overlapped sliding window method, the characteristic variables are extracted from the specific small window, so that the dimension can be reduced on one hand, and on the other hand, the characteristic variables can better reflect the characteristics of the section relative to the original data. The sliding window processing formula is as follows:
wherein k represents a signal channel; n represents the sliding window length; emg (i) represents the filtered electromyographic signals; f denotes the derivation of some eigenvalue function, such as MAV or MF.
And 4, calculating the feature extraction of the electromyographic signals. The non-stationary randomness of the electromyographic signals is not suitable for directly inputting the original data into the classifier for training and testing. In general, the greater the difference between feature vectors corresponding to different action types, the higher the recognition rate of the pattern classifier obtained based on this feature vector training. In the aspect of time domain analysis, the absolute mean value (MAV) can reflect the average intensity of the surface electromyogram signal of the segment and the muscle action intensity. In terms of frequency domain analysis, the Median Frequency (MF) may reflect the median value of the frequency of the electrical signal during muscle contraction. Therefore, the time domain feature absolute average value MAV and the median frequency MF in the frequency domain features are selected, and the time domain features and the frequency domain features are combined to carry out classification basis.
And 5, taking the absolute average value and the median frequency fusion vector as characteristic parameters. The time domain features are easy to obtain, the calculated amount is small, the time overhead is short, but the amplitude change is large, the volatility is strong, and the stability is poor. Then combining the frequency domain features can avoid a large amount of information loss in the time domain.
Step 6, defining a BP neural network model structure: and inputting the sample data after electromyographic signal processing into the model by using a BP neural network for learning, and training model parameters. The MAV, MF and fusion features in time domain and frequency domain signals are used as input data of a model, an input layer is set to be m (m is 60) neurons (6 channels), an output layer is provided with n (n is 6) nodes, and according to an empirical formula:
and setting the number q (q is 10) of hidden layer neurons.
Compared with the prior art, the invention has the beneficial effects that:
1. compared with the conventional method of classifying only by using filtered original electromyographic signals of the electromyographic signals by using an artificial neural network, the method integrates the noise reduction and filtering processes of the complicated original electromyographic signals, allows the noise of a certain frequency band to exist, carries out envelope analysis on the filtered electromyographic signals, and then carries out feature extraction by using a sliding window segmentation method aiming at the envelope, thereby achieving model training and classification recognition. The method comprises the steps of processing by using an envelope curve method, inputting a feature vector into a BP neural network to learn time domain and frequency domain information by using time domain features and frequency domain features and a fusion feature vector of the time domain and the frequency domain, and training to obtain the difference of the time domain features and the frequency domain features so as to classify different lower limb actions.
2. In order to keep human motion information in non-stationary electromyographic signals as much as possible, the invention divides the signal change process within a period of time into a plurality of tiny signal segments, each signal segment can only have about 0.1-0.2s, the small signal segment can be regarded as a relatively stable signal, and then signal detection and feature extraction of each action are carried out aiming at the small segment of process. And (4) selecting a sliding window method with overlapping, extracting characteristic variables from a specific small window, and sequentially sliding one by one. And through the moving average processing, all channels can basically realize simplification of the preprocessed electromyographic signals, and meanwhile, the extracted features are also sufficiently used for classification judgment, so that the method has a good classification effect.
Drawings
FIG. 1 is a schematic flow chart of an embodiment of the present invention;
FIG. 2 is a sEMG graph of rectus femoris muscle walking on level ground with electromyography acquisition as used in the present invention;
fig. 3 is a schematic diagram of a sliding window segmentation method for processing sEMG according to the present invention;
FIG. 4 is a signal diagram of the present invention with sEMG feature extracted for neural network input;
Detailed Description
The specific flow of the invention is shown in fig. 1, and the following detailed description of the specific embodiments is provided.
Step 2, electromyographic signal preprocessing: in order to increase the calculation speed, the electromyographic signals collected in the step 1 are subjected to Baster filtering to extract the electromyographic signals in an effective frequency band, 50Hz and frequency multiplication interference signals thereof are removed by a wave trap, envelope lines of the filtered electromyographic signals are extracted by Hilbert transform, and the electromyographic signals are just extracted to facilitate the operations of sliding window processing, data normalization and the like.
Step 3, sliding window segmentation treatment: referring to fig. 3, a sliding window method with an overlap is employed. The window size N and window increment value M are primarily determined, and then a sliding window processing formula is utilized, as follows:
wherein k represents a signal channel; n represents the sliding window length; emg (i) represents the filtered electromyographic signal; f denotes the derivation of some eigenvalue function.
Step 4, feature extraction of the electromyographic signals: in the aspect of time domain analysis, the time domain analysis mainly comprises an absolute mean value (MAV), a root mean square value (RMS), an absolute value integral, a zero crossing point number, a variance, a Willison amplitude, a time sequence model of an EMG signal, an EMG histogram and the like. In the aspect of frequency domain analysis, the main analysis method is to perform fast fourier transform on sEMG signals to obtain frequency spectra or power spectra of the sEMG signals, which can reflect changes of the sEMG signals in different frequency components, so that the sEMG changes can be reflected well in the frequency dimension, and the common methods include mean frequency (MPF), cepstral coefficient and Median Frequency (MF) of the electromyogram power spectra. The absolute mean value (MAV) reflects the average intensity of the electrical signals of the surface muscle and the intensity of the muscle action. The Median Frequency (MF) may reflect the median value of the frequency of the electrical signal during muscle contraction.
And 5, performing feature fusion, and fusing the two feature values of the absolute average value and the median frequency into a feature parameter.
Step 6, defining a BP neural network model structure: inputting the preprocessed sample data into a BP neural network model for learning, continuously training model parameters, performing maximum iteration for 1500 times, and performing minimum gradient change of 1 × 10-10。
Example 1:
1. acquiring electromyographic signals: according to the muscle distribution map of the lower limbs of the human body and the corresponding relation between the lower limb actions and the muscles, the types of the muscles for acquiring the electromyographic signals are determined as follows: rectus femoris, biceps femoris, semitendinosus, tibialis anterior, lateral gastrocnemius, and medial gastrocnemius. The muscle corresponding channels are respectively: channel 1, channel 2, channel 3, channel 4, channel 5, channel 6. In the test, 15 Chinese testees (7 male and 8 female) are selected, the acquisition frequency is 2000Hz, and the acquisition actions are walking for 1 minute at 4km/h, running for 1 minute at 8km/h, squatting for 1 minute, ascending for 1 minute at 5km/h, descending for 1 minute at 5km/h and ascending and descending for 5 times. Muscle electrical signals of 6 channel muscles of 15 subjects were collected, respectively.
2. Preprocessing electromyographic signals: manually removing interference of signals outside a frequency band of 20-450 Hz, performing Butterworth band-pass filtering on the electromyographic signals acquired in the step 1, extracting the electromyographic signals of 20-450 Hz, removing interference signals of 50Hz and frequency multiplication thereof by a trap filter, extracting an envelope curve of the filtered electromyographic signals by using Hilbert transform, and performing positive processing and normalization operation on the extracted electromyographic signals of the envelope curve.
3. Sliding window segmentation processing: by using a sliding window method with overlap, a window increment 12 is selected, and characteristic variables are extracted by sliding processing of signals in a time period one by one in a window with the size of 120 windows. The sliding window processing formula is as follows:
wherein k represents a signal channel; n represents the sliding window length; emg (i) represents the filtered electromyographic signal; f denotes the determination of a function of characteristic values
Step 4, calculating the feature extraction of the electromyographic signals: absolute mean and median frequency. Wherein: the formula for calculating the absolute average is defined as follows:
wherein Ns is the total number of sEMG channels, i is the number of sampling points of the original signal, and xi is the value of the myoelectric signal of a certain signal.
The formula for the median frequency is:
where fm is the median frequency to be found, p (f) is the electromyographic power spectrum, and the power spectrum of sEMG can be obtained by fast fourier transform.
And 5, performing feature fusion on the absolute average value and the median frequency to form a fusion feature vector.
Step 6, defining a BP neural network model structure: inputting the sample data after electromyographic signal processing into a BP neural network model for learning, wherein the model adopts 2 hidden layers, the number of neurons in the hidden layers is respectively 12 and 8, the maximum iteration time is 1500 times, and the minimum gradient change is 1 multiplied by 10-10。
And 7, inputting the test sample into the trained BP neural network model, and calculating the accuracy of motion classification and identification.
The above embodiments are merely illustrative of the present invention and are not intended to limit the present invention. It will be understood by those skilled in the art that various combinations, modifications and equivalents of the technical solutions of the present invention may be made without departing from the scope of the technical solutions of the present invention, and the technical solutions of the present invention are intended to be covered by the claims of the present invention.
Claims (4)
1. A lower limb action recognition and classification method based on surface electromyogram signals is characterized by comprising the following steps:
step 1, collecting electromyographic signals of the surface of the lower limb: collecting lower limb electromyographic signals of a subject by adopting a 6-channel electrode cap, wherein the sampling rate is 2000 Hz;
step 2, preprocessing the surface electromyographic signals of the lower limbs: performing Barter filtering on the electromyographic signals acquired in the step 1 to extract the electromyographic signals in an effective frequency band, removing 50Hz and frequency-doubling interference signals by using a wave trap, extracting envelope lines of the filtered electromyographic signals by using Hilbert transform, righting the electromyographic signals to facilitate sliding window processing and data normalization;
step 3, determining the window size and window increment of the sliding window: extracting characteristic variables from a window by adopting an overlapped sliding window method; the sliding window processing formula is as follows:
wherein k represents a signal channel; n represents the sliding window length; emg (i) represents the filtered electromyographic signal; f is to obtain a certain characteristic value function, namely MAV or MF;
step 4, extracting the characteristics of the electromyographic signals; in the aspect of time domain analysis, the absolute average value MAV reflects the average intensity and muscle action intensity of the surface electromyographic signal of the section; in the aspect of frequency domain analysis, the median frequency MF can reflect the intermediate value of the frequency of an electric signal in the process of muscle contraction;
step 5, combining the time domain characteristics and the frequency domain characteristics to avoid the loss of the detail information of the original data, and fusing the absolute average value and the median frequency as characteristic parameters;
step 6, defining a BP neural network model structure: inputting the preprocessed sample data into a BP neural network model for learning, continuously training model parameters, wherein the maximum iteration times are 1500 times, and the minimum gradient change is 1 multiplied by 10-10。
2. The method for identifying and classifying lower limb actions based on surface electromyography signals according to claim 1, wherein the electromyography signal preprocessing in step 2 is specifically as follows:
filtering the surface myoelectric signal by using a Butterworth filter and a wave trap to remove signals with specific frequencies in the signal, and reserving the original frequency of the signal; decomposing the electromyographic signals into three frequency bands of less than 20Hz, 20-450 Hz and more than 450Hz, extracting the electromyographic signals of 20-450 Hz by using Butterworth band-pass filtering, and then removing 50Hz and frequency multiplication interference signals thereof by using a wave trap; extracting electromyographic signals within a frequency band of 20-450 Hz, extracting envelope lines of the signals by Hilbert transform, turning over energy signals of a negative half shaft of a longitudinal shaft to a positive half shaft to finish positive selection, and finally obtaining preprocessed electromyographic signals.
3. The method for identifying and classifying lower limb movements based on surface electromyography signals according to claim 1, wherein the sliding window segmentation process in step 3 is as follows:
performing sliding window processing on the lower limb surface electromyographic signals in the acquisition time, wherein N represents the length of a sliding window, M represents window increment, adopting the fixed window length and the window increment to sequentially slide the signals in the time period one by one, and calculating the characteristic value of the signals in each window;
through analyzing different sliding window sizes and sample variances under window increment, the suggested window size is 50T < N <150T, wherein T is the sampling period of the surface myoelectricity acquisition system and is in unit of millisecond; the window size is fixed, and the size of M is in the range of 8T-15T.
4. The method for recognizing and classifying lower limb movements based on surface electromyographic signals according to claim 1, wherein the step 4 of calculating the characteristic parameters of the electromyographic signals comprises the following steps: absolute mean and median frequency characteristics, as follows:
the adopted characteristic parameters are the absolute average value MAV, the median frequency MF and the fusion characteristics of the absolute average value MAV and the median frequency MF, and the single characteristic and the fusion characteristics of the absolute average value MAV and the median frequency MF are respectively used for fusion to form fusion characteristics for identification; using the sliding window segmentation method in step 3, wherein: the formula for calculating the absolute average is defined as follows:
where N is the sliding window size, i is the number of sampling points of the original signal, xiIs the ith electromyographic signal value in the segmented data;
the formula for the median frequency is:
wherein, fmThe median frequency to be solved, P (f) the electromyographic power spectrum, and the power spectrum of the sEMG is obtained through fast Fourier transform.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210161420.2A CN114533089A (en) | 2022-02-22 | 2022-02-22 | Lower limb action recognition and classification method based on surface electromyographic signals |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210161420.2A CN114533089A (en) | 2022-02-22 | 2022-02-22 | Lower limb action recognition and classification method based on surface electromyographic signals |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN114533089A true CN114533089A (en) | 2022-05-27 |
Family
ID=81678320
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210161420.2A Pending CN114533089A (en) | 2022-02-22 | 2022-02-22 | Lower limb action recognition and classification method based on surface electromyographic signals |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114533089A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115114962A (en) * | 2022-07-19 | 2022-09-27 | 歌尔股份有限公司 | Control method and device based on surface electromyogram signal and wearable device |
| CN115778408A (en) * | 2022-12-10 | 2023-03-14 | 福州大学 | Step-by-step electromyographic signal activity segment detection method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101987048A (en) * | 2009-08-03 | 2011-03-23 | 深圳先进技术研究院 | Artificial limb control method and system thereof |
| CN109522810A (en) * | 2018-10-22 | 2019-03-26 | 上海师范大学 | A kind of myoelectric limb hand gesture identification method based on community vote mechanism |
-
2022
- 2022-02-22 CN CN202210161420.2A patent/CN114533089A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101987048A (en) * | 2009-08-03 | 2011-03-23 | 深圳先进技术研究院 | Artificial limb control method and system thereof |
| CN109522810A (en) * | 2018-10-22 | 2019-03-26 | 上海师范大学 | A kind of myoelectric limb hand gesture identification method based on community vote mechanism |
Non-Patent Citations (2)
| Title |
|---|
| KATHARINA FUCHS A等: "Speech/Non-Speech Detection for Electro-Larynx Speech Using EMG", 《IN PROCEEDINGS OF THE INTERNATIONAL CONFERENCE ON BIO-INSPIRED SYSTEMS AND SIGNAL PROCESSING》, pages 1 - 3 * |
| MD. REZWANUL AHSAN等: "Electromygraphy (EMG) Signal based Hand Gesture Recognition using Artificial Neural Network (ANN)", 《2011 4TH INTERNATIONAL CONFERENCE ON MECHATRONICS (ICOM)》, pages 2 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115114962A (en) * | 2022-07-19 | 2022-09-27 | 歌尔股份有限公司 | Control method and device based on surface electromyogram signal and wearable device |
| CN115778408A (en) * | 2022-12-10 | 2023-03-14 | 福州大学 | Step-by-step electromyographic signal activity segment detection method |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Chen et al. | Adaptive real-time identification of motor unit discharges from non-stationary high-density surface electromyographic signals | |
| CN111012341B (en) | Artifact removal and electroencephalogram signal quality evaluation method based on wearable electroencephalogram equipment | |
| CN204931634U (en) | Based on the depression evaluating system of physiologic information | |
| CN114533089A (en) | Lower limb action recognition and classification method based on surface electromyographic signals | |
| CN110598676B (en) | Deep learning gesture electromyographic signal identification method based on confidence score model | |
| CN111860410A (en) | Myoelectric gesture recognition method based on multi-feature fusion CNN | |
| CN114052744B (en) | Electrocardiosignal classification method based on impulse neural network | |
| CN108681685A (en) | A kind of body work intension recognizing method based on human body surface myoelectric signal | |
| CN107822629A (en) | The detection method of extremity surface myoelectricity axle | |
| CN114532993B (en) | Automatic detection method for electroencephalogram high-frequency oscillation signals of epileptic patients | |
| Ahmed et al. | EMG Signal classification for detecting neuromuscular disorders | |
| Tello et al. | Towards sEMG classification based on Bayesian and k-NN to control a prosthetic hand | |
| CN114159079A (en) | Multi-type muscle fatigue detection method based on feature extraction and GRU deep learning model | |
| Kaur et al. | Comparison of the techniques used for segmentation of EMG signals | |
| CN116910464A (en) | Myoelectric signal prosthetic hand control system and method | |
| Xu et al. | ART2 neural network for surface EMG decomposition | |
| CN116784860A (en) | Electrocardiosignal characteristic extraction system based on morphological heart beat template clustering | |
| CN114098768B (en) | Cross-individual surface electromyographic signal gesture recognition method based on dynamic threshold and EasyTL | |
| CN112932508B (en) | Finger activity recognition system based on arm electromyography network | |
| Ling-Ling et al. | Electromyographic movement pattern recognition based on random forest algorithm | |
| Wahab et al. | Analysis and classification of forearm muscles activities during gripping using EMG signals | |
| Altamirano | EMG pattern prediction for upper limb movements based on wavelet and Hilbert-Huang transform | |
| Kholdorov et al. | Analysis of methods and algorithms for feature extraction of biosignals of muscle activity | |
| Li et al. | Improving the myoelectric motion classification performance by feature filtering strategy | |
| CN113625882B (en) | Myoelectric gesture recognition method based on sparse multichannel correlation characteristics |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |