A hand movement intention recognition method based on multi-layer feature extraction of electromyographic signals

Abstract

The present invention discloses a hand movement intention recognition method based on multi-layer feature extraction of electromyographic signals, which obtains multiple electromyographic signals and corresponding hand movements A on the arm during hand movements; performs multi-layer feature extraction on the multiple electromyographic signals respectively, updates the time domain amplitude range of the electromyographic signals in real time, and extracts the number of signal zero crossing points that meet the intensity requirements according to the time domain intensity of the electromyographic signals in layers to obtain feature data; trains the feature data with the corresponding hand movements to obtain a prediction model. The multi-layer features extracted by the present invention provide frequency information of electromyographic signals at different signal intensities, thereby improving the accuracy of hand movement intention recognition.

Description

Technical Field

[0001] The invention belongs to the field of information, and in particular relates to a hand movement intention recognition method based on multi-layer feature extraction of electromyographic signals. Background Art

[0002] Surface electromyographic signals are a series of action potentials generated by each motor unit when the muscle is excited, which are superimposed on the surface of the skin. It is a non-stationary weak signal. In recent years, the research results of electromyographic signals have been applied in many fields such as clinical diagnosis, rehabilitation medicine, and biomechanics. Myoelectric control prosthesis is a typical application of electromyographic signals in the field of assisting the elderly and the disabled. It collects electromyographic signals on the surface of human muscles, obtains human movement intentions through feature extraction and pattern recognition, and then uses them for prosthetic movement control. This control mode has attracted widespread attention because of its natural movements and similarity to human operation habits. Domestic and foreign scholars have carried out a lot of research work to improve the accuracy of motion recognition. Most of the existing studies extract a variety of different electromyographic signal time domain, frequency domain or time-frequency domain features, combine them into a huge feature matrix, and then use them for motion recognition. Although this method can improve the accuracy of motion recognition, it leads to a sharp increase in the amount of calculation. For commercial prosthetic hands with limited processor computing power, this method is not applicable. Therefore, studying efficient feature extraction can highlight as much essential information related to the signal and the action as possible, which helps to improve the accuracy of hand movement intention recognition. Summary of the invention

[0003] Purpose of the invention: The present invention proposes a hand movement intention recognition method based on multi-layer feature extraction of electromyographic signals to improve the accuracy and stability of myoelectric prosthetic hand motion control.

[0004] Technical solution: The present invention provides a method for identifying hand movement intention based on multi-layer feature extraction of electromyographic signals, which specifically includes the following steps:

[0005] (1) Obtaining multiple electromyographic signals on the arm and the corresponding hand movements A when the hand moves;

[0006] (2) Perform multi-layer feature extraction on multiple EMG signals: update the time domain amplitude range of the EMG signals in real time and extract the number of signal zero crossing points that meet the intensity requirements according to the time domain intensity of the EMG signals to obtain feature data;

[0007] (3) Train the feature data and the corresponding hand movements to obtain a prediction model.

[0008] Furthermore, the multiple electromyographic signals described in step (1) are four electromyographic signals of the brachial flexor carpi, the brachioradialis, the extensor digitorum, and the extensor carpi ulnaris, and are recorded as EMG1, EMG2, EMG3, and EMG4, respectively.

[0009] Furthermore, the multi-channel electromyographic signals described in step (1) are obtained by electromyographic sensors respectively attached to the brachial flexor carpi, brachioradialis, extensor digitorum, and extensor carpi ulnaris of the user's arm.

[0010] Furthermore, the implementation process of step (2) is as follows:

[0011] For EMG1, update the maximum value of the electromyographic signal Max E1 and minimum value Min E1 , the update method is as follows:

[0012]

[0013]

[0014] Among them, EMG1_CURR is the voltage value of the electromyographic signal collected by the first electromyographic sensor at the current moment, Max E1 and Min E1 The initial values ​​of are all set to 0;

[0015] According to the preset feature extraction layer number N and the preset feature extraction window length W, where both N and W are integers greater than or equal to 2, the threshold TH(n) of each layer is calculated as follows:

[0016]

[0017] According to the threshold TH(n) of each layer, the characteristic value F(n,i) of the nth layer at the i-th sampling time is calculated as:

[0018]

[0019]

[0020] The features of EMG2, EMG3, and EMG4 electromyographic signals are extracted in the same way.

[0021] Furthermore, the prediction model in step (3) is a BP neural network model.

[0022] Beneficial effects: Compared with the prior art, the present invention has the following beneficial effects: the multi-layer features extracted by the present invention provide frequency information of electromyographic signals at different signal strengths, thereby improving the accuracy of hand movement intention recognition; compared with the existing methods of extracting electromyographic signal features through Fourier transform, wavelet transform, etc., the present invention has the characteristics of small computational complexity and low algorithm complexity, can effectively reduce the delay of feature extraction, is suitable for occasions without high-performance computing processing capabilities, and is particularly suitable for applications in commercial consumer-grade prosthetic hands; the hierarchical feature extraction proposed in the present invention sets a real-time update link for the electromyographic signal amplitude range, and dynamically adjusts the feature extraction threshold, which can effectively suppress the problem of unstable motion recognition caused by changes in the position of the electromyographic sensor and the skin during motion recognition; at the same time, this method of dynamically adjusting the electromyographic signal feature extraction threshold can also effectively suppress the influence of individual differences in electromyographic signals on motion recognition, thereby expanding the scope of application of the electromyographic signal motion recognition method for different people. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flow chart of the present invention;

[0024] Figure 2 This is a typical EMG waveform. A 500-ms window is used to extract features from the signal.

[0025] Figure 3 For Figure 2 The waveform diagram shown is obtained by performing traditional zero-crossing point extraction on the electromyographic signal;

[0026] Figure 4 For Figure 2 The waveform diagram obtained after performing 2-layer feature extraction on the electromyographic signal shown;

[0027] Figure 5 For Figure 2 The waveform diagram obtained after 3-layer feature extraction of the electromyographic signal shown;

[0028] Figure 6 For Figure 2 The waveform diagram shown is obtained after 5-layer feature extraction of the electromyographic signal. DETAILED DESCRIPTION

[0029] The present invention is further described in detail below with reference to the accompanying drawings.

[0030] The present invention provides a hand movement intention recognition method based on multi-layer feature extraction of electromyographic signals, including signal acquisition, multi-layer feature extraction, prediction model training and movement intention recognition. Figure 1 As shown, the specific steps include:

[0031] Step 1: Take the first to fourth electromyographic sensors, which are respectively used to use the surface electromyographic signals of the brachial flexor carpi, brachioradialis, extensor digitorum, and extensor carpi ulnaris on the arm, and synchronously record the electromyographic signals and corresponding hand movements collected by the first to fourth electromyographic sensors when the wearer makes a hand movement, and record them as EMG1, EMG2, EMG3, EMG4, and A, respectively.

[0032] Step 2: Perform multi-layer feature extraction on the four electromyographic signals collected in step 1. Taking the multi-layer feature extraction of EMG1 as an example, the specific steps are as follows:

[0033] Step 2.1: Update the maximum value of the electromyographic signal Max E1 and minimum value Min E1 , the update method is as follows:

[0034]

[0035]

[0036] Among them, EMG1_CURR is the voltage value of the electromyographic signal collected by the first electromyographic sensor at the current moment, Max E1 and Min E1 The initial values ​​of are all set to 0;

[0037] Step 2.2: According to the preset feature extraction layer number N and the preset feature extraction window length W, where N is an integer greater than or equal to 2, the threshold TH(n) of each layer is calculated as follows:

[0038]

[0039] Step 2.3: According to the threshold values ​​TH(n) of each layer obtained in step 2.2, the characteristic value F(n,i) of the nth layer at the i-th sampling time is calculated as:

[0040]

[0041]

[0042] Wherein, j is a certain sampling time. When the data at sampling time j in the EMG1 sequence data has a different sign from the data at time (j-1), and the absolute value of the difference between the two is greater than TH(n), it is recorded as 1.

[0043] Step 3: Train the feature data obtained in step 2 with the corresponding hand movements to obtain a prediction model.

[0044] The prediction model is BP neural network, and other types of prediction models can also be used.

[0045] The first to fourth myoelectric sensors are respectively attached to the brachial flexor carpi, brachioradialis, extensor digitorum, and extensor carpi ulnaris of the user's arm. Therefore, only four myoelectric sensors need to be attached to the brachial flexor carpi, brachioradialis, extensor digitorum, and extensor carpi ulnaris of the wearer's arm, and the user can perform hand movement intention recognition after randomly making several target gestures.

[0046] Figure 2 This is a typical EMG waveform. A 500-ms window is used to extract features from the signal. Figures 3 to 6 The waveforms are obtained by performing traditional zero-crossing feature extraction on the electromyographic signal, the waveform obtained by performing 2-layer feature extraction, the waveform obtained by performing 3-layer feature extraction, and the waveform obtained by performing 5-layer feature extraction. Figures 3 to 6 It can be seen that the results of hierarchical feature extraction can not only reflect the results of traditional zero-crossing feature extraction, but also provide more information. The more layers of decomposition, the more information can be provided.

Claims

1. A hand movement intention recognition method based on multi-layer feature extraction of electromyographic signals, characterized in that: The following steps are involved: (1) Obtaining multiple electromyographic signals on the arm and the corresponding hand movements A when the hand moves; (2) Perform multi-layer feature extraction on multiple EMG signals: update the time domain amplitude range of the EMG signals in real time and extract the number of signal zero crossing points that meet the intensity requirements according to the time domain intensity of the EMG signals to obtain feature data; (3) training the feature data and the corresponding hand movements to obtain a prediction model; The multi-channel electromyographic signals described in step (1) are four-channel electromyographic signals of the brachial flexor carpi, brachioradialis, extensor digitorum, and extensor carpi ulnaris, and are recorded as EMG1, EMG2, EMG3, and EMG4 respectively; The implementation process of step (2) is as follows: For EMG1, update the maximum value of the electromyographic signal Max E1 and minimum value Min E1 , the update method is as follows: Among them, EMG1_CURR is the voltage value of the electromyographic signal collected by the first electromyographic sensor at the current moment, Max E1 and Min E1 The initial values ​​of are all set to 0; According to the preset feature extraction layer number N and the preset feature extraction window length W, where N and W are both integers greater than or equal to 2, the threshold TH(n) of each layer is calculated as follows: According to the threshold TH(n) of each layer, the characteristic value F(n,i) of the nth layer at the i-th sampling time is calculated as: The features of EMG2, EMG3, and EMG4 electromyographic signals are extracted in the same way.

2. The hand movement intention recognition method based on multi-layer feature extraction of electromyographic signals according to claim 1 is characterized in that: The multi-channel electromyographic signals described in step (1) are obtained by electromyographic sensors respectively attached to the brachial flexor carpi, brachioradialis, extensor digitorum, and extensor carpi ulnaris of the user's arm.

3. The hand movement intention recognition method based on multi-layer feature extraction of electromyographic signals according to claim 1 is characterized in that: The prediction model in step (3) is a BP neural network model.

Images