MI-EEG mixed time-frequency characterization method and characterization system based on CEEMDAN-CWT

Abstract

The invention discloses an MI-EEG mixed time frequency characterization method and characterization system based on CEEMDAN-CWT, and relates to the technical field of data processing. The problems that an existing time-frequency analysis method is large in calculation amount and high in calculation complexity are solved. The method comprises the steps that an MI-EEG data set is acquired and preprocessed; a CEEMDAN algorithm is improved, the preprocessed MI-EEG data set is decomposed by adopting the improved CEEMDAN algorithm, and a series of IMFs are obtained; adopting a sample entropy method to calculate a series of IMFs, and obtaining all eigenfunction sample entropies; all eigenfunction sample entropies are optimized by adopting a K-means clustering method, and IMFs with key information are obtained; performing CWT processing on IMFs with key information to obtain a series of time-frequency two-dimensional data; and adding a series of time-frequency two-dimensional data to obtain the time-frequency representation of the MI-EEG signal. The MI-EEG data processing method and device are suitable for MI-EEG data processing.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing. Background Art

[0002] Abstract—Brain-computer interface (BCI) is a technology that uses EEG signals to achieve human-computer interaction. Motor imagery is a commonly used EEG paradigm that has the advantage of active control and can be used in the fields of neurorehabilitation, prosthetic control, etc.

[0003] The processing of MI-EEG data has always been an important issue in the field of motor imagery-based brain-computer interfaces (MI-BCI). This is because in addition to the general EEG characteristics such as low signal-to-noise ratio, complex and diverse artifacts, and non-stationary signals, MI-EEG data also has unique challenges such as large individual differences and multi-task confusion. Therefore, feature extraction methods for EEG are often complex.

[0004] Compared with other BCI systems, the most critical and unique parts of the MI-BCI system are the signal processing and feature extraction module and the feature classification module. The signal processing and feature extraction module removes various artifacts and greatly improves the signal-to-noise ratio of the EEG signal. In the past, deep learning methods mainly extracted EEG signal features through time-frequency analysis. The feature classification module directly affects the accuracy of classification and thus determines the performance of the system. Choosing a suitable classifier is the key to system design. The design and optimization of the above two modules have an important impact on the performance and application effect of the system.

[0005] Time-frequency analysis is one of the important methods for EEG signal processing, which can effectively extract the features of EEG signals. However, the complexity and noise interference of MI-EEG signals themselves pose great challenges to feature extraction and classification. In order to improve the performance and usability of MI-BCI, it is necessary to design efficient and robust time-frequency representation methods. At present, many different time-frequency representation methods have been applied to the analysis and processing of MI-EEG signals, which can be mainly divided into the following categories:

[0006] 1. Methods based on space, time, and spectrum. This method comprehensively considers the information of MI-EEG signals in three dimensions: space, time, and spectrum, and uses the correlation or coherence between multi-channel data to enhance signal quality and feature expression capabilities. Common space, time, and spectrum methods include common spatial patterns, common spatial patterns of filter banks, and common sparse spectrum regression of filter banks. This type of method can effectively extract discriminative features and improve classification accuracy, but there are also some problems, such as overfitting risks, large amount of calculations, and multiple parameter selections.

[0007] 2. Methods based on modal decomposition, which decompose the MI-EEG signal into a series of intrinsic mode functions (IMFs), each of which represents an oscillation mode with different frequencies and amplitudes. Common modal decomposition methods include empirical mode decomposition, ensemble empirical mode decomposition, complete ensemble empirical mode decomposition, variational mode decomposition, etc. This type of method can adapt to non-stationary and nonlinear signals, but there are also some problems, such as mode aliasing, boundary effects, and high computational complexity.

[0008] 3. Based on the linear or quadratic time-frequency method, this method convolves or correlates the MI-EEG signal with some predefined or known functions to obtain a time-frequency diagram that reflects the energy distribution of the signal at different times and frequencies. Common linear or quadratic time-frequency methods include short-time Fourier transform, continuous wavelet transform, and Cui-Williams distribution. This type of method can intuitively display the time-frequency characteristics of the signal, but there are also some problems, such as limited time-frequency resolution, cross-term interference, and difficulty in parameter selection.

[0009] 4. Methods based on pure time-frequency representation often find it difficult to strike a balance between the computational complexity of feature extraction and classification accuracy.

[0010] 5. Another method is to combine the modal decomposition method with the time-frequency analysis method to extract features from non-stationary MI-EEG signals. This method can improve the effect of feature extraction and reduce the amount of calculation. For example, the paper published in 2008: Motorimagery EEG detection by empirical mode decomposition. This method decomposes the signal into different IMFs, and then performs Hilbert transform on each IMF to extract time-frequency features. This method can effectively improve the accuracy of EEG classification. However, it is very sensitive to noise, resulting in poor feature extraction and difficulty in capturing the characteristics of non-stationary signals. Summary of the invention

[0011] The present invention solves the problems of large amount of calculation and high calculation complexity in the existing time-frequency analysis method.

[0012] To achieve the above object, the present invention provides the following solutions:

[0013] The present invention provides a MI-EEG mixed time-frequency characterization method based on CEEMDAN-CWT, the method comprising:

[0014] Step 1, obtaining a MI-EEG data set, and preprocessing the MI-EEG data set to obtain a preprocessed MI-EEG data set;

[0015] Step 2: Improve the CEEMDAN algorithm to obtain an improved CEEMDAN algorithm;

[0016] Step 3, using the improved CEEMDAN algorithm to perform CEEMDAN decomposition on each channel in each trail in the preprocessed MI-EEG dataset to obtain a series of IMFs;

[0017] Step 4: Use the sample entropy method to calculate a series of IMFs to obtain the sample entropy of all eigenfunctions;

[0018] Step 5: Using K-means clustering method to optimize the entropy of all eigenfunction samples to obtain IMFs with key information;

[0019] Step 6: Perform CWT processing on the IMFs with key information to obtain a series of time-frequency two-dimensional data;

[0020] Step 7: Add the series of time-frequency two-dimensional data to obtain the time-frequency representation of the MI-EEG signal.

[0021] Furthermore, in another preferred embodiment, the MI-EEG data set in the above step 1 is a binary classification data set.

[0022] Furthermore, in a preferred embodiment, the preprocessing in the above step 1 includes channel positioning and electrode selection, artifact removal and downsampling.

[0023] Furthermore, in another preferred embodiment, the channel positioning is specifically as follows: selecting three channels, C3, Cz and C4;

[0024] The electrode selection is specifically: selecting two electrodes C3 and C4.

[0025] Furthermore, in a preferred embodiment, the above-mentioned artifact removal specifically includes: using a FIR filter with a linear phase response characteristic to remove the artifact.

[0026] Furthermore, in a preferred embodiment, the downsampling is specifically: downsampling the data with a sampling rate of 50 Hz to 125 Hz.

[0027] Furthermore, in another preferred embodiment, the above step 2 is specifically as follows:

[0028] Set the threshold of the CEEMDAN algorithm to 1e-6 and the maximum number of iterations to 1000;

[0029] Set local extreme points to control the number and frequency range of IMFs.

[0030] The MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT described in the present invention can be fully implemented by computer software. Therefore, correspondingly, the present invention also provides a MI-EEG hybrid time-frequency characterization system based on CEEMDAN-CWT, and the system is:

[0031] A storage device for acquiring a MI-EEG data set, and preprocessing the MI-EEG data set to obtain the preprocessed MI-EEG data set;

[0032] A storage device for improving the CEEMDAN algorithm to obtain an improved CEEMDAN algorithm;

[0033] A storage device for performing CEEMDAN decomposition on each channel in each trail in the preprocessed MI-EEG data set using the improved CEEMDAN algorithm to obtain a series of IMFs;

[0034] A storage device for calculating a series of IMFs using a sample entropy method to obtain sample entropies of all eigenfunctions;

[0035] A storage device for optimizing the entropy of all eigenfunction samples by using a K-means clustering method to obtain IMFs with key information;

[0036] A storage device for performing CWT processing on the IMFs having key information to obtain a series of time-frequency two-dimensional data;

[0037] A storage device for adding the series of time-frequency two-dimensional data to obtain the time-frequency representation of the MI-EEG signal.

[0038] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, any one of the above-mentioned MI-EEG mixed time-frequency characterization methods based on CEEMDAN-CWT is executed.

[0039] The present invention also provides a computer device, which includes a memory and a processor, wherein a computer program is stored in the memory, and when the processor runs the computer program stored in the memory, the processor executes any one of the above-mentioned MI-EEG mixed time-frequency characterization methods based on CEEMDAN-CWT.

[0040] The beneficial effects of the present invention are:

[0041] 1. The present invention provides a MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT, which improves the CEEMDAN algorithm by setting the threshold or maximum number of iterations of the CEEMDAN algorithm, thereby controlling the convergence of the CEEMDAN algorithm, and controlling the number and frequency range of IMFs by setting local extreme points. By reducing the number of channels, using sample entropy and K-means clustering methods, the amount of calculation and the complexity of the algorithm are reduced, and the efficiency of the modal decomposition time-frequency characterization method is further improved, and the feature separability is improved.

[0042] 2. The present invention provides a MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT, which improves the binary classification accuracy by 8% compared with the CWT time-frequency characterization method, has a smaller standard deviation on 9 subjects, and significantly improves the accuracy and stability of MI-EEG classification.

[0043] The present invention is applicable to the processing of MI-EEG data. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 is a flow chart of the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT described in Implementation Method 1;

[0045] Figure 2 is a two-channel time series data diagram of a single Trail after preprocessing as described in Implementation 11;

[0046] Figure 3 is a schematic diagram of IMFs after CEEMDAN decomposes single-trail single-channel MI-EEG data according to the eleventh embodiment;

[0047] Figure 4 Schematic diagram of sample entropy of each IMFs of single-Trail MI-EEG single-channel data according to Implementation Example 11;

[0048] Figure 5 It is the co-IMFs formed by applying K-means clustering to the sample entropy of each IMFs of the Trail single-channel EEG as described in the eleventh embodiment and then superimposing and mixing;

[0049] Figure 6 It is a time-frequency diagram of the three co-IMFs based on the Morlet wavelet CWT transformation described in the eleventh embodiment;

[0050] Figure 7 (a) is a CEEMDAN-CWT MI-EEG time-frequency representation diagram of a single trail of co-IMF0, co-IMF1, and co-IMF2 components among the multiple co-IMFs components described in Embodiment 11;

[0051] Figure 7 (b) is a CEEMDAN-CWT MI-EEG time-frequency representation diagram of a single trail of the co-IMF0 component among the multiple co-IMFs components described in Embodiment 11;

[0052] Figure 7 (c) is a CEEMDAN-CWT MI-EEG time-frequency representation diagram of a single trail of co-IMF0 and co-IMF2 components among the multiple co-IMFs components described in Embodiment 11;

[0053] Figure 7 (d) is a CEEMDAN-CWT MI-EEG time-frequency representation diagram of a single trail of the co-IMF1 component among the multiple co-IMFs components described in Embodiment 11;

[0054] Figure 7 (e) is a CEEMDAN-CWT MI-EEG time-frequency representation diagram of a single trail of co-IMF1 and co-IMF2 components among the multiple co-IMFs components described in Embodiment 11;

[0055] Figure 8 (a) is a binary classification accuracy curve diagram based on CEEMDAN-CWT representation according to the eleventh embodiment;

[0056] Figure 8 (b) is the binary classification Loss curve based on CEEMDAN-CWT representation described in the eleventh embodiment;

[0057] Fig. 9 This is a summary table of the MI-EEGNet binary classification results based on the BCIC 2b dataset described in Implementation Example 11. DETAILED DESCRIPTION

[0058] The specific implementation mode of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples will help those skilled in the art to further understand the present invention, but are not intended to limit the present invention in any form. It should be noted that, for those of ordinary skill in the art, several changes and improvements can also be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.

[0059] Implementation method 1. See Figure 1 This embodiment describes a mixed time-frequency characterization method for MI-EEG based on CEEMDAN-CWT. The method is as follows:

[0060] Step 1, obtaining a MI-EEG data set, and preprocessing the MI-EEG data set to obtain a preprocessed MI-EEG data set;

[0061] Step 2: Improve the CEEMDAN algorithm to obtain an improved CEEMDAN algorithm;

[0062] Step 3, using the improved CEEMDAN algorithm to perform CEEMDAN decomposition on each channel in each trail in the preprocessed MI-EEG dataset to obtain a series of IMFs;

[0063] Step 4: Use the sample entropy method to calculate a series of IMFs to obtain the sample entropy of all eigenfunctions;

[0064] Step 5: Using K-means clustering method to optimize the entropy of all eigenfunction samples to obtain IMFs with key information;

[0065] Step 6: Perform CWT processing on the IMFs with key information to obtain a series of time-frequency two-dimensional data;

[0066] Step 7: Add the series of time-frequency two-dimensional data to obtain the time-frequency representation of the MI-EEG signal.

[0067] In practical application, if Figure 1 As shown, the MI-EEG data set is preprocessed to avoid signal loss or deformation caused by excessive signal preprocessing, and to ensure the reliability and accuracy of the signal. The CEEMDAN algorithm is improved to adjust the parameters of the CEEMDAN algorithm. By adjusting the parameters of the CEEMDAN algorithm, the number and quality of IMFs are controlled, and the number of IMFs is not less than 5. The improved CEEMDAN algorithm is used to perform CEEMDAN decomposition on each channel in each trail in the preprocessed MI-EEG data set to obtain a series of IMFs; wherein, the algorithm processing flow of CEEMDAN is as follows:

[0068] (1) Initialization: Set the number of noise sequences N, the amplitude of the noise sequence a, the original signal x(t), and the number of iterations n = 0;

[0069] (2) Add different white noises to the original signal to obtain N disturbance signals: x n (t) = x(t) + w n (t);

[0070] where w n (t) has a mean of 0 and a variance of a 2 White noise sequence.

[0071] (3) For each disturbance signal x n (t) Perform classical EMD decomposition to obtain a set of intrinsic mode functions (IMFs) and residuals r n (t), that is

[0072]

[0073] where c n,k (t) is the kth IMF of the nth disturbance signal, K is the number of IMFs,

[0074] (4) For each IMF, the arithmetic mean of all disturbance signals is calculated to obtain the final IMF.

[0075]

[0076] in is the kth final IMF.

[0077] (5) Take the arithmetic mean of all disturbance signals for each residual to obtain the final residual, that is,

[0078]

[0079] (6) Determine whether the final residual satisfies monotonicity or is small enough. If yes, stop decomposition; if not, set n = n + 1, replace the original signal with the final residual, and repeat steps (2) to (6).

[0080] (7) Output the final decomposition result, that is,

[0081]

[0082] The sample entropy method is used to calculate a series of IMFs to obtain the sample entropy of all eigenfunctions; wherein, the sample entropy is a method to measure the complexity of time series, which is used to reflect the irregularity and unpredictability of time series. The K-means clustering method is used to optimize the sample entropy of all eigenfunctions to obtain IMFs with key information; wherein, K-means clustering is an unsupervised machine learning method that divides the data set into several similar subsets, called clusters. The goal of K-means clustering is to make the distance between data points in the same cluster as small as possible, and the distance between different clusters as large as possible.

[0083] Due to the non-stationarity of EEG signals, when applying traditional time-frequency analysis methods, they are easily affected by noise and modal interference. The CEEMDAN-CWT provided in this embodiment is a hybrid time-frequency analysis method based on modal decomposition, which combines the advantages of CEEMDAN and CWT analysis methods to better describe the time and frequency dynamic characteristics of EEG signals with nonlinear and non-stationary characteristics.

[0084] This embodiment provides a MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT, which improves the CEEMDAN algorithm by setting the threshold or maximum number of iterations of the CEEMDAN algorithm, thereby controlling the convergence of the CEEMDAN algorithm, and controlling the number and frequency range of IMFs by setting local extreme points. By reducing the number of channels and using sample entropy and K-means clustering methods, the amount of calculation and the complexity of the algorithm are reduced, further improving the efficiency of the modal decomposition time-frequency characterization method and improving the feature separability.

[0085] Implementation method 2. This implementation method is to illustrate the MI-EEG data set of step 1 in the MI-EEG mixed time-frequency characterization method based on CEEMDAN-CWT described in implementation method 1. The MI-EEG data set adopts a binary classification data set.

[0086] In practical application, the MI-EEG data set of this embodiment selects a binary classification data set: BCIC IV 2b, which is composed of MI-EEG signals of 9 healthy subjects. The data is collected using 3 EEG channel devices with a sampling frequency of 250Hz. The MI-EEG signals in this data set are processed by high-pass filtering (0.5Hz) and low-pass filtering (100Hz) to remove irrelevant signals such as DC offset and high-frequency noise of the electrode.

[0087] Implementation method three. This implementation method is an example of the preprocessing of step one in the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT described in implementation method one. The preprocessing includes channel positioning and electrode selection, artifact removal and downsampling.

[0088] In practical applications of this embodiment, preprocessing is used to avoid signal loss or deformation caused by excessive signal preprocessing, and to ensure the reliability and accuracy of the signal. Preprocessing includes channel positioning and electrode selection, artifact removal and downsampling. Among them, channel positioning and electrode selection not only relax the requirements on the number of leads of the device during EEG acquisition, but also greatly simplify the complexity of data processing, improve the signal-to-noise ratio, and reduce the amount of algorithm calculation in the feature extraction stage. The purpose of artifact removal is to remove artifacts and improve the signal-to-noise ratio. Downsampling can retain the information of the target frequency band, avoid the occurrence of aliasing effects, and thus maintain the reliability of the signal.

[0089] Embodiment 4. This embodiment is an example of channel positioning and electrode selection in the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT described in Embodiment 3. The channel positioning is specifically: selecting three channels, C3, Cz and C4;

[0090] The electrode selection is specifically: selecting two electrodes C3 and C4.

[0091] In actual application, this embodiment selects three channels, C3, Cz and C4, for the pure time-frequency method, and selects two electrode channels C3 and C4 for mapping the motor-related cortex of the brain. The C3 and C4 electrodes are selected as the target brain area. These electrodes are located near the midline of the scalp surface, just covering the position of the motor-related cortex, and have high signal quality and relatively stable spatial distribution. The selection of two specific channels not only relaxes the requirements for the number of leads of the device during the EEG acquisition process, but also greatly simplifies the complexity of data processing, improves the signal-to-noise ratio of the signal, and reduces the amount of algorithm calculation in the feature extraction stage.

[0092] Embodiment 5. This embodiment is an example of artifact removal in the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT described in Embodiment 3. The artifact removal specifically includes: using an FIR filter with a linear phase response characteristic to remove artifacts.

[0093] In actual application of this embodiment, the artifact removal is specifically: using an FIR filter with a linear phase response characteristic to remove artifacts. The purpose is to remove artifacts and improve the signal-to-noise ratio. If the ICA-independent component analysis algorithm is used, the effective components will be mistakenly identified or eliminated, and the number of channels required for the device is high. Because the ICA-independent component analysis algorithm needs to correctly identify independent components, the algorithm that automatically identifies independent components will mistakenly identify or eliminate effective components; manual identification of independent components requires a lot of expert knowledge. Moreover, the ICA-independent component analysis algorithm belongs to the blind source analysis method. In order to accurately remove artifacts, the number of channels required for the device is high. Therefore, a filter is used to remove artifacts. The filter is implemented using an FIR-finite impulse response filter with a linear phase response characteristic, which can achieve zero-phase filtering without signal delay. The 4Hz to 38Hz bandpass filter achieves the filtering of most artifacts: heartbeat artifacts at around 0.5Hz, electrooculographic artifacts at around 2Hz, power frequency interference at 50Hz, and electromyographic artifacts at higher frequencies. The data retained the following frequency bands involved in motor imagery: theta band (4-8Hz), alpha band (8-13Hz), beta band (13-30Hz) and part of the low gam-ma band (30-38Hz).

[0094] Implementation method 6. This implementation method is to illustrate the downsampling in the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT described in Implementation method 3. The downsampling is specifically: downsampling the data with a sampling rate of 50 Hz to 125 Hz.

[0095] In practical applications, this implementation method downsamples the data with a sampling rate of 250 Hz to 125 Hz. This reduces the amount of data and computational complexity, improves processing speed, and reduces the storage requirements of the device. By filtering the data through a filter before sampling, the information of the target frequency band can be retained, and the occurrence of aliasing effects can be avoided, thereby maintaining the reliability of the signal.

[0096] Implementation VII. This implementation is to illustrate step 2 in the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT described in Implementation 1 by way of example, and the specific step 2 is:

[0097] Set the threshold of the CEEMDAN algorithm to 1e-6 and the maximum number of iterations to 1000;

[0098] Set local extreme points to control the number and frequency range of IMFs.

[0099] In practical applications, this implementation method controls the quantity and quality of IMFs by adjusting the parameters of the CEEMDAN algorithm, controls the convergence of the CEEMDAN algorithm by setting a threshold or a maximum number of iterations, and controls the quantity and frequency range of IMFs by setting local extreme points.

[0100] Embodiment 8. This embodiment provides a MI-EEG hybrid time-frequency characterization system based on CEEMDAN-CWT, the system is:

[0101] A storage device for acquiring a MI-EEG data set, and preprocessing the MI-EEG data set to obtain the preprocessed MI-EEG data set;

[0102] A storage device for improving the CEEMDAN algorithm to obtain an improved CEEMDAN algorithm;

[0103] A storage device for performing CEEMDAN decomposition on each channel in each trail in the preprocessed MI-EEG data set using the improved CEEMDAN algorithm to obtain a series of IMFs;

[0104] A storage device for calculating a series of IMFs using a sample entropy method to obtain sample entropies of all eigenfunctions;

[0105] A storage device for optimizing the entropy of all eigenfunction samples by using a K-means clustering method to obtain IMFs with key information;

[0106] A storage device for performing CWT processing on the IMFs having key information to obtain a series of time-frequency two-dimensional data;

[0107] A storage device for adding the series of time-frequency two-dimensional data to obtain the time-frequency representation of the MI-EEG signal.

[0108] Embodiment 9. This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT as described in any one of Embodiments 1 to 7 is executed.

[0109] Embodiment 10. This embodiment provides a computer device, which includes a memory and a processor, wherein the memory stores a computer program. When the processor runs the computer program stored in the memory, the processor executes the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT as described in any one of Embodiments 1 to 7.

[0110] Implementation 10. This implementation is to conduct an experimental analysis of the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT described in any one of Implementation 1 to Implementation 7. First, complete the visualization data of each link of the CEEMDAN-CWT MI-EEG time-frequency characterization method; secondly, build the MI-EEGNet network with image input under the TensorFlow framework based on the Benchmark network EEGNet with time-series EEG data input; finally, compare and verify the classification accuracy of different time-frequency characterization methods.

[0111] Specifically, first complete the visualization data of each link of the CEEMDAN-CWT MI-EEG time-frequency representation method, such as Figures 2 to 7 As shown; among them, Figure 2 It shows the two-channel time series data of a single Trail after preprocessing; Figure 3 It shows the six IMFs obtained by CEEMDAN decomposing a single-channel EEG of the preprocessed Trail; Figure 4 The sample entropy of each IMFs of the single-channel EEG of the Trail is shown; Figure 5 The K-means clustering based on the sample entropy of each IMFs of the single-channel EEG of the Trail is shown, followed by superposition and mixing of three co-IMFs; Figure 6The three co-IMFs were transformed by CWT based on morlet wavelet to obtain time-frequency diagrams. After adding the co-IMFs CWT results at each time-frequency point, a single-channel single-Trail time-frequency diagram was obtained. Then, the two-channel (C3, C4) time-frequency diagrams were superimposed and spliced ​​to obtain the CEEMDAN-CWT MI-EEG time-frequency representation diagram of a single Trail, as shown in the figure. Figure 7 shown.

[0112] Secondly, based on the EEGNet skeleton of the Banchmark network with time-series EEG data input, the input of the MI-EEG representation image and the convolution kernel matching are performed under the TensorFlow framework to build the MI-EEGNet network with images as input.

[0113] Finally, the classification accuracy of different time-frequency representation methods was compared and verified. Figure 8 As shown, the subjects were classified independently and the relevant data of each subject was trained separately: each network was trained and tested by 10-fold cross validation, and 70% and 30% of each subject's data were divided into training set and test set. Fig. 9 As shown in the figure, for the BCIC 2b dataset containing 9 subjects, the CEEMDAN (coIMF0+coIMF2) and CWT hybrid representation with a standard deviation of (±5.5) has the highest average accuracy of 78.13% for the MI-EEG time-frequency representation images classified by MI-EEGNet. The average accuracy of the pure time-frequency method CWT method is low, and the standard deviation of the subjects is high.

[0114] In summary, the MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT described in this embodiment uses CEEMDAN to perform modal decomposition on MI-EEG based on the power of the four frequency bands of EEG. Therefore, performing CWT on MI-EEG after decomposing the modality based on CEEMDAN helps to obtain a better MI-EEG time-frequency characterization. At the same time, it can be concluded that only taking the coIMF0 and coIMF2 components has better characterization ability and can reduce the amount of calculation in the CWT stage, which verifies the optimization method based on sample entropy and K-means clustering.

[0115] The results of experimental verification show that the binary classification accuracy of the optimized CEEMDAN-CWT method is 8% higher than that of the CWT time-frequency representation method, and has a smaller standard deviation on 9 subjects, which can significantly improve the accuracy and stability of MI-EEG classification.

[0116] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In the description of the present invention, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0117] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, without contradiction.

[0118] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may be subject to various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. The MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT is characterized by: The method is: Step 1, obtaining a MI-EEG data set, and preprocessing the MI-EEG data set to obtain a preprocessed MI-EEG data set; Step 2: Improve the CEEMDAN algorithm to obtain an improved CEEMDAN algorithm; Step 3, using the improved CEEMDAN algorithm to perform CEEMDAN decomposition on each channel in each trail in the preprocessed MI-EEG dataset to obtain a series of IMFs; Step 4: Use the sample entropy method to calculate a series of IMFs to obtain the sample entropy of all eigenfunctions; Step 5: Using K-means clustering method to optimize the entropy of all eigenfunction samples to obtain IMFs with key information; Step 6: Perform CWT processing on the IMFs with key information to obtain a series of time-frequency two-dimensional data; Step 7: Add the series of time-frequency two-dimensional data to obtain the time-frequency representation of the MI-EEG signal.

2. The MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT according to claim 1, characterized in that: The MI-EEG data set in step 1 is a binary classification data set.

3. The MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT according to claim 1, characterized in that: The preprocessing in step 1 includes channel positioning and electrode selection, artifact removal and downsampling.

4. The MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT according to claim 3, characterized in that: The channel positioning is specifically as follows: selecting three channels, C3, Cz and C4; The electrode selection is specifically: selecting two electrodes C3 and C4.

5. The MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT according to claim 3, characterized in that: The artifact removal specifically includes: using a FIR filter with a linear phase response characteristic to remove artifacts.

6. The MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT according to claim 3, characterized in that: The downsampling is specifically: downsampling the data with a sampling rate of 50 Hz to 125 Hz.

7. The MI-EEG hybrid time-frequency characterization method based on CEEMDAN-CWT according to claim 1, characterized in that: The step 2 is specifically as follows: Set the threshold of the CEEMDAN algorithm to 1e-6; the maximum number of iterations to 1000; Set local extreme points to control the number and frequency range of IMFs.

8. The MI-EEG hybrid time-frequency characterization system based on CEEMDAN-CWT is characterized by: The system is: A storage device for acquiring a MI-EEG data set, and preprocessing the MI-EEG data set to obtain the preprocessed MI-EEG data set; A storage device for improving the CEEMDAN algorithm to obtain an improved CEEMDAN algorithm; A storage device for performing CEEMDAN decomposition on each channel in each trail in the preprocessed MI-EEG data set using the improved CEEMDAN algorithm to obtain a series of IMFs; A storage device for calculating a series of IMFs using a sample entropy method to obtain sample entropies of all eigenfunctions; A storage device for optimizing the entropy of all eigenfunction samples by using a K-means clustering method to obtain IMFs with key information; A storage device for performing CWT processing on the IMFs having key information to obtain a series of time-frequency two-dimensional data; A storage device for adding the series of time-frequency two-dimensional data to obtain the time-frequency representation of the MI-EEG signal.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the MI-EEG mixed time-frequency characterization method based on CEEMDAN-CWT according to any one of claims 1 to 7 is executed.

10. A computer device, characterized in that: The device includes a memory and a processor, wherein a computer program is stored in the memory. When the processor runs the computer program stored in the memory, the processor executes the CEEMDAN-CWT-based MI-EEG mixed time-frequency characterization method according to any one of claims 1 to 7.

Images