CN110555468A

CN110555468A - Electroencephalogram signal identification method and system combining recursion graph and CNN

Info

Publication number: CN110555468A
Application number: CN201910753679.4A
Authority: CN
Inventors: 王文波; 辜权; 狄奇; 喻敏; 陈贵词; 钱龙
Original assignee: Wuhan University of Science and Engineering WUSE
Current assignee: Wuhan University of Science and Engineering WUSE; Wuhan University of Science and Technology WHUST
Priority date: 2019-08-15
Filing date: 2019-08-15
Publication date: 2019-12-10

Abstract

The invention belongs to the technical field of medical treatment, and discloses an electroencephalogram signal identification method and system combining a recursion graph and CNN (computer network noise). The preprocessed electroencephalogram signal data are decomposed into intrinsic mode functions of different scales by utilizing empirical mode decomposition, and a multi-scale recursion graph of intrinsic mode components of each scale is calculated to obtain a level 1 characteristic; and taking the reconstructed multi-scale recursion map as the image characteristics of the left-hand EEG signal and the right-hand EEG signal, taking the multi-scale recursion map characteristics as the input of a convolutional neural network, carrying out classification and identification on the recursion map by using the convolutional neural network, and extracting the level 2 characteristics capable of better expressing the motor imagery EEG signal from the level 1 characteristics. The electroencephalogram signal recognition rate is high, and the electroencephalogram signal can be better recognized; the method and the device adopt a mutual information method to determine the delay time result more accurately. The invention adopts the ReLU activation function, and when the input is positive number, the problem of gradient saturation does not exist.

Description

electroencephalogram signal identification method and system combining recursion graph and CNN

Technical Field

The invention belongs to the technical field of medical treatment, and particularly relates to an electroencephalogram signal identification method and system combining a recursion diagram and a CNN.

background

currently, the current state of the art commonly used in the industry is such that:

A brain-computer interface (BCI) system realizes the establishment of a direct interaction channel with the external environment by studying the operation mode of current signals sent by the human brain in the cerebral cortex under the condition that the human brain is not supported by a peripheral nervous system. The key point for realizing the technology is how to effectively identify and classify various electroencephalogram signals generated by the brain by using a reasonable signal processing algorithm.

In recent years, the development trend of artificial intelligence is increasing, for example, IBM walton cognitive computation, google's AlphaGo robot, and a series of major events, so that more and more people are invested in the field of artificial intelligence, and various possibilities brought by artificial intelligence technology are dug in effort. At present, although artificial intelligence is rapidly developed in the fields of search engines, intelligent reading, computer vision, image classification, machine learning and the like, a large growing space still exists in the medical industry. If artificial intelligence can play a significant role in the field of medical health, the process of exploring diseases which cannot be cured at present, such as cancer, can be greatly accelerated. In recent years, the development of artificial intelligence based on human brain science has made a rapid progress, which gives people a new angle to know and reform the world. Brain-machine integration, that is, mining useful information from the nervous tissue of human brain, and using the information to build a model capable of performing human-machine interaction with the external environment, and the building of the model requires a clear understanding of the operation of human brain, which is also a crucial stage for the development of artificial intelligence in the future.

the human brain is a sponsor of all human activities, understands the functions, structures and actions of the human brain, explores the relation between electroencephalogram signals and human activities, forms deep understanding of behavior modes, thinking modes, consciousness, language expression ability and the like of human, and is an important challenge for human to know and understand the human. Meanwhile, the research and development of the brain-computer interface system can also greatly promote the development process of the AI technology and the human-computer interaction technology at the present stage.

The brain-computer interface system has more related fields, the difficult problems in the technical level are complex, and the problem is solved with certain difficulty, but in recent years, the improvement of computer hardware equipment and the development of related technologies provide a good technical support environment for the realization of the brain-computer interface system, so that more and more researchers and scholars are involved in the research and development of the brain-computer interface system. The research on brain-computer interfaces has been continued for many years, and the principles on brain-computer interfaces obtained from various studies are continuously accumulated, so that the research on brain-computer interface systems at the present stage is not so difficult. The brain-computer interface obtains the brain electrical signals sent by the brain on the basis of sufficiently understanding the operation mode of the human brain, so that the brain can control external equipment. The brain of the human body plays a role as a controller, and the acquired brain electrical signals are processed and interpreted by a computer so as to complete the conversion between brain signals and computer signals, thereby realizing the human-computer interaction.

the conventional BCI system realizes direct interaction of human brain to entity equipment by researching the change characteristics of EEG signals of cerebral cortex generated when a human body carries out various activities, and improves deep understanding of the human brain. Since the proposal of a brain-computer interface system, the field is rapidly developed in the exploration and experiment of researchers, and a series of important research results are obtained. A patient with lateral sclerosis typified by a famous physicist, a patient with disability due to an accident, and the like are important subjects for brain-computer interface application. The BCI system loses control over various functions of a human body to a certain extent, but the BCI system has the characteristic that direct interaction can be carried out with entity equipment only by the human brain, so that the BCI system can help the human brain to express information wanted in the brain of the user or directly communicate with the external environment.

in the twentieth century, Pfurtscheller and Guger developed and put into practical use BCI systems. In 2015, Haggag and Mohamed developed a single-channel EEG signal control device that could be more convenient for disabled people to operate a prosthetic limb. The patient whose hands can not normally move due to accidents directly controls the hand training device by means of the BCI system, and the rehabilitation training of daily activities can be completed only by controlling the entity equipment through the human brain.

however, the following problems also exist in the existing electroencephalogram identification technology:

(1) the signal-to-noise ratio of the electroencephalogram signal needs to be improved.

In the current brain-computer interface system research, in the signal acquisition stage, the acquired electroencephalogram signals contain more noise, so that real signals cannot be effectively retained, a better preprocessing technology and a better denoising algorithm need to be searched for to improve the signal-to-noise ratio of EEG, and the reliability of a BCI system is ensured.

(2) the system recognition rate needs to be improved.

The recognition rate of classification directly influences whether the BCI system can effectively respond to the real-time requirements of the user, and the higher the classification accuracy and speed are, the more correctly the BCI system can judge the actual intention of the user, and further send instructions to peripheral equipment so as to effectively control the peripheral equipment. And the classification accuracy of the system is improved, and the most important is a preprocessing stage and a feature extraction stage.

(3) Hardware technology is yet to be improved.

For the acquisition of electroencephalogram signals, at present, non-implanted signal acquisition is mainly adopted, and the method is more easily influenced by noise, so that the accuracy of the whole BCI system is reduced. If the hardware technology can be improved and promoted in the signal acquisition stage, and then the clean electroencephalogram signal can be obtained, the complexity and difficulty of denoising can be reduced.

(4) The problem of system self-adaptation.

System adaptation mainly includes two types: the BCI system can identify the difference electroencephalogram signals generated by different main bodies, and the BCI system can identify the difference electroencephalogram signals generated by the same main body in different time and space. The BCI system needs to raise the adaptability requirements of both aspects.

in summary, the problems of the prior art are as follows:

the electroencephalogram signals collected by the existing electroencephalogram identification technology contain more noise, real signals cannot be effectively reserved, the system identification rate is not accurate enough, the hardware technology is not enough, and the system has the self-adaption problems that the difference electroencephalogram signals generated by different main bodies cannot be identified and the difference electroencephalogram signals generated by the same main body in different time and space cannot be identified.

the difficulty of solving the technical problems is as follows:

(1) The electroencephalogram signals are acquired in two modes, namely an implanted mode and a non-implanted mode, an electrode needs to be implanted into a cerebral cortex in the implanted mode, the method can cause damage to a subject, therefore, the electroencephalogram signals are acquired mainly in the non-implanted mode, the electrode is placed on the scalp of the subject to record signals in the non-implanted acquisition method, the method cannot cause damage to the subject, but the acquired signals are easily affected by noise, and therefore the accuracy of the whole BCI system is affected. How to improve and promote hardware technologies such as electroencephalogram caps, Emotiv electroencephalogram signal collectors and the like in a signal acquisition stage to obtain clean electroencephalogram signals is a problem.

(2) When electroencephalogram signals are collected, besides the influence of external factors such as the existence of noise around the electroencephalogram signals, the influence of noise generated by a subject, such as mental confusion and blinking, can cause artifacts in the collected electroencephalogram signals, and therefore, how to extract vectors capable of representing the electroencephalogram signals from original signals is also a key step of electroencephalogram signal identification.

(3) Because electroencephalogram signals have the characteristics of nonlinearity, instability and randomness, how to effectively extract and analyze the electroencephalogram signals is also the key point influencing the final conclusion, and the EEG feature extraction methods commonly used at present are a time domain analysis method, a frequency domain analysis method, a time frequency analysis method, a nonlinear analysis method, a common space mode method and a deep learning method.

(4) Finally, how to solve the problem of self-adaption of the system. The BCI system can identify the difference electroencephalogram signals generated by different subjects and the difference electroencephalogram signals of the same subject in different time and space.

The significance of solving the technical problems is as follows:

(1) BCI provides another mode for communicating and controlling with the outside for people, people can communicate with the outside through language and actions, but directly express ideas and control equipment through electroencephalogram signals, so that patients who think normally but have muscular atrophy can communicate with the outside to a certain extent, and the BCI can also help workers in special environments to provide auxiliary control, such as medical operations or special requirements in special environments.

(2) the influence of noise on the electroencephalogram signals is reduced, the identification accuracy for classifying the electroencephalogram signals can be improved, the classified identification rate directly influences whether the BCI system can effectively respond to the real-time requirements of the user or not, the higher the classification accuracy and speed are, the more correctly the BCI system can judge the actual intention of the user, and then the peripheral equipment can be instructed to effectively control the peripheral equipment.

(3) the improvement of the self-adaptability of the system can improve the generalization capability of the system, so that the system can really and effectively serve users and expand the service range. The recently-raised deep learning model adopted by the invention can improve the generalization capability of the system to a certain extent.

Disclosure of Invention

Aiming at the problems in the prior art, the invention provides an electroencephalogram signal identification method and system combining a recursion chart and a CNN.

the invention is realized in this way, a combined recursion diagram and CNN electroencephalogram signal identification method, which specifically comprises the following steps:

Decomposing the preprocessed electroencephalogram signal data into the intrinsic mode functions of different scales by utilizing empirical mode decomposition, and calculating a multi-scale recursion graph of the intrinsic mode components of all scales to obtain the level 1 characteristic.

And taking the reconstructed multi-scale recursion map as the image characteristics of the left-hand and right-hand EEG signals, taking the multi-scale recursion map characteristics as the input of a convolutional neural network, carrying out classification and identification on the recursion map by utilizing the advantage of the Convolutional Neural Network (CNN) in image processing, and extracting the level 2 characteristics capable of better expressing the motor imagery EEG signals from the level 1 characteristics.

Further, the electroencephalogram signal identification method combining the recursion map and the CNN specifically comprises the following steps:

the method comprises the following steps: dividing the collected data set into training data and testing data, training the training data as a training sample of the identification method, preprocessing the data, and classifying the processed EEG signals; and the test data is used as a final test for the identification accuracy of the identification method, and the trained classifier is used for classifying the test data and recording the result.

step two, in order to improve the classification accuracy, the invention carries out two times of feature extraction, firstly, multi-scale filtering processing is needed to be carried out on EEG signals, and more concise and effective distinguishing features are provided by utilizing a multi-scale recursion graph; secondly, the identification features with higher discrimination degree need to be further extracted from the recursive graph. Firstly, decomposing the motor imagery electroencephalogram signal into inherent mode functions of different scales by utilizing empirical mode decomposition on preprocessed electroencephalogram signal data, and calculating a multi-scale recursion graph of inherent mode components of each scale to obtain level 1 characteristics.

and step three, regarding the reconstructed multi-scale recursion diagram as the image characteristics of left and right hand EEG signals, regarding the multi-scale recursion diagram characteristics as the input of a convolutional neural network, classifying and identifying the recursion diagram by utilizing the advantage of the Convolutional Neural Network (CNN) in image processing, extracting the level 2 characteristics capable of better expressing motor imagery EEG signals from the level 1 characteristics, and debugging all parameters of the convolutional neural network, so that the neural network classifier achieves ideal precision.

Step four, applying the trained convolutional neural network to test data, and carrying out classification and identification on the test data; and performing first-stage feature extraction on the test electroencephalogram signal data by using a recursion graph method, and performing secondary feature extraction on the recursion graph of the test electroencephalogram signal data by using the trained convolutional neural network to obtain a final classification recognition result.

Further, in the first step, the data preprocessing specifically includes:

Training data: the method comprises the steps of carrying out down-sampling interception on training electroencephalogram data, selecting proper parameters by utilizing a mutual information method, carrying out filtering processing on a frequency domain and a space domain by using AR-CSP (AR-chip scale package) and then classifying processed EEG signals.

Test data: down-sampling and intercepting test electroencephalogram data, selecting appropriate parameters by using a mutual information method, performing filtering processing on a frequency domain and a spatial domain by using AR-CSP (AR-chip scale protocol), and classifying processed EEG (electroencephalogram) signals; and classifying the test data by using the trained classifier and recording the result.

Further, in the second step, the empirical mode decomposition specifically includes:

(1) Adding a white noise sequence n (t) to the original signal x (t) to obtain a signal with noise s (t), that is:

s(t)＝x(t)+n(t)

Where N (t) is white gaussian noise subject to N (0, σ ²).

(2) The noisy signal s (t) is empirically decomposed into a set of natural modal components IMF and a residual r _c (t), i.e.:

Wherein c is the number of IMF components.

(3) Repeating the step (1) and the step (2) for m times, wherein the white noise sequence amplitude filled in each time is different, namely:

(4) classifying the IMF generated by EMD treatment for m times according to layers, and then averaging to obtain the final IMF:

the gaussian white noise formula added in empirical mode decomposition should be:

wherein epsilon is the amplitude of the white Gaussian noise, N is the times of adding the white Gaussian noise, and epsilon _n represents the error of the added natural modal components of each order and the original signal.

Further, in step two, the calculating an appropriate embedding dimension and delay time of each intrinsic mode component by using a mutual information analysis method and a Cao-Liangyue method specifically includes:

(1) determining delay time by mutual information analysis:

For the time series { s _i }, p _s (s _i) is defined as the probability of occurrence of the variable s _i, and the information entropy of the system is the average information content of the variable s _i, which is abbreviated as entropy, and is defined as follows:

For two groups of signals { s _i, q _j }, let p _s,q (s _i, q _j) be the joint probability distribution of variables s _i and q _j, and then the joint entropy calculation formula is as follows:

Let [ S, Q ] ═ x (t), x (t + τ) ], then for the coupled system (S, Q), if S is known as S _i, then the uncertainty of Q is:

wherein P _q|s (q _j | s _i) is conditional probability.

If x is known at time T, then the uncertainty of x at time T + T is:

and tau takes different time delays in turn, and the mutual information is calculated:

I(τ)＝H(x)+H(x_τ)-H(x,x_τ)

(2) The Cao-Liangyue method determines the embedding dimension:

If the embedding dimension is m, any one vector point i in the reconstruction phase space can be represented as:

X_i＝{x(i),x(i+τ),x(i+2τ),…,x(i+(m-1)τ)}

Suppose that the nearest phase point of X _i is X _j, and the Euclidean distance between the two points is:

When the embedding dimension becomes m +1, the euclidean distance of the two points is:

the ratio of d _m+1 to d _m is defined as a (i, m), i.e.:

Defining:

Further, in step two, the recursive graph analysis method specifically includes:

Assuming that the original time sequence is { x ₁, x ₂, …, x _n }, the result after the phase space reconstruction according to the Takens' embedding theorem is:

X_i＝{x_i,x_i+τ,…,x_i+(m-1)τ}(i＝1,2,…,N)

Where N ═ N- (m-1) τ, τ is the delay time, and m is the embedding dimension; then the distance between any two vectors in these vector sets is:

d_ij＝‖X_i-Y_j‖

after selecting the threshold r, according to the formula:

R_ij＝Heaviside(r-d_ij)

obtaining a recursive matrix, wherein the Heaviside function is expressed as follows:

According to the above formula, points are drawn on a coordinate plane whose vertical axis and horizontal axis are the number of time series from the result of R _ij.

When R _ij is equal to 1, it represents that the euclidean distance between the two vectors X _i and X _j is less than the threshold R, the system is in a recursive state, and the corresponding position on the plane is displayed as a black dot, and when R _ij is equal to 0, it represents that the euclidean distance between the two vectors X _i and X _j is greater than the threshold R, the system is in a non-recursive state, and the corresponding position on the plane is displayed as a white dot.

further, in step three, the neural network specifically includes:

The convolutional neural network comprises an input layer, 3 convolutional layers, 3 pooling layers, a Flatten layer and an output layer, an activation function after convolution operation is a ReLU, a maximum pooling method is selected by the pooling operation, and a Sigmoid is selected by the classifier.

another object of the present invention is to provide an electroencephalogram signal identification system combining a recursion map and a CNN.

Another object of the present invention is to provide an information data processing terminal for implementing the electroencephalogram signal identification method of the joint recursion map and CNN.

It is another object of the present invention to provide a computer-readable storage medium, comprising instructions which, when run on a computer, cause the computer to perform the method for brain electrical signal identification of joint recursion maps and CNNs.

in summary, the advantages and positive effects of the invention are:

The average recognition rate of the two classes of motor imagery electroencephalogram signals by the two-stage feature extraction method combining the multi-scale recursion map and the convolution network is 0.6045, and is improved by about 3.35% compared with the average recognition rate of the first name in BCI competition IV, which shows that the method can perform effective classification and recognition on the two classes of motor imagery electroencephalogram signals.

Compared with a method which only adopts a multi-scale recursive graph analysis method or a neural network method, the two-stage feature extraction method based on the multi-scale recursive graph and the convolutional neural network provided by the invention has the advantages that the overall recognition rate of the motor imagery electroencephalogram signals is improved, the classification result and the first achievement in the BCI competition IV are improved by about 0.335, and meanwhile, the single recognition rate of each subject is slightly superior, so that the recognition rate comparison result shows that the two-stage feature extraction method based on the multi-scale recursive graph and the convolutional neural network has better electroencephalogram signal recognition performance to a certain extent.

the method and the device adopt a mutual information method to determine the delay time result more accurately. The invention adopts the ReLU activation function, when the input is positive number, the problem of gradient saturation does not exist, and the calculation speed is higher.

drawings

Fig. 1 is a flowchart of a joint recursion graph and CNN electroencephalogram signal identification method provided by an embodiment of the present invention.

fig. 2 is a flowchart of an integrated empirical mode decomposition algorithm provided by an embodiment of the present invention.

Fig. 3 is a schematic structural diagram of a convolutional neural network according to an embodiment of the present invention.

Fig. 4 is a schematic diagram of the basic components of a brain-computer interface provided in the embodiment of the present invention.

Fig. 5 is a schematic diagram of EEG signal acquisition provided by an embodiment of the invention.

fig. 6 is a schematic diagram of a classification and identification scheme of two types of motor imagery electroencephalogram signals provided by the embodiment of the invention.

fig. 7 is a schematic structural diagram of a LeNet-5 convolutional neural network provided in an embodiment of the present invention.

fig. 8 is a structure diagram of weight sharing according to an embodiment of the present invention.

Fig. 9 is a schematic diagram of a pooling operation provided by an embodiment of the present invention.

fig. 10 is a multi-scale recursive diagram of various IMF components provided by an embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described in detail with reference to the following embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

To solve the above problems, the present invention will be described in detail with reference to the accompanying drawings.

The electroencephalogram signal identification method combining the recursion graph and the CNN provided by the embodiment of the invention specifically comprises the following steps:

As shown in fig. 1, the method for identifying an electroencephalogram signal by combining a recursion diagram and a CNN provided by the embodiment of the present invention specifically includes the following steps:

s101, dividing the acquired data set into training data and testing data, and preprocessing the data respectively.

s102, performing empirical mode decomposition on two types of motor imagery electroencephalogram signals, namely training data and test data respectively to obtain inherent modal components of different scales; meanwhile, a mutual information analysis method and a Cao-Liangyue method are used for calculating the appropriate embedding dimension and delay time of each inherent modal component to carry out spatial reconstruction, a recursion graph of each inherent modal component is obtained, a recursion graph analysis method is used for carrying out feature extraction, and the level 1 feature is obtained.

s103, training by taking the level 1 recursive graph characteristics of the training data set as the input of the convolutional neural network, and debugging each parameter of the convolutional neural network.

S104, applying the trained convolutional neural network to test data, and carrying out classification and identification on the test data; and performing feature extraction on the data of the tested brain electrical signals by using a recursion graph method, and performing feature extraction on the recursion graph of the data of the tested brain electrical signals by using the trained convolutional neural network to obtain a classification recognition result.

in step S101, the data preprocessing provided in the embodiment of the present invention specifically includes:

as shown in fig. 2, in step S102, the empirical mode decomposition provided by the embodiment of the present invention specifically includes:

s(t)＝x(t)+n(t)

Where N (t) is white gaussian noise subject to N (0, σ ²).

Wherein c is the number of IMF components.

In step S102, the calculating an appropriate embedding dimension and delay time of each eigenmode component by using the mutual information analysis method and the Cao-Liangyue method provided by the embodiment of the present invention specifically includes:

(1) Determining delay time by mutual information analysis:

Wherein P _q|s (q _j | s _i) is conditional probability.

if x is known at time T, then the uncertainty of x at time T + T is:

I(τ)＝H(x)+H(x_τ)-H(x,x_τ)

(2) The Cao-Liangyue method determines the embedding dimension:

X_i＝{x(i),x(i+τ),x(i+2τ),…,x(i+(m-1)τ)}

the ratio of d _m+1 to d _m is defined as a (i, m), i.e.:

Defining:

In step S102, the recursive graph analysis method provided in the embodiment of the present invention specifically includes:

X_i＝{x_i,x_i+τ,…,x_i+(m-1)τ}(i＝1,2,…,N)

d_ij＝‖X_i-Y_j‖

After selecting the threshold r, according to the formula:

R_ij＝Heaviside(r-d_ij)

As shown in fig. 3, in step S103, the neural network provided in the embodiment of the present invention specifically includes:

The present invention will be further described with reference to the following specific examples.

Example 1:

1. Brain-computer system.

1.1 brain-computer interface system.

the brain-computer interface system has more related fields, the difficult problems in the technical level are complex, and the problem is solved with certain difficulty, but in recent years, the improvement of computer hardware equipment and the development of related technologies provide a good technical support environment for the realization of the brain-computer interface system, so that more and more researchers and scholars are involved in the research and development of the brain-computer interface system.

1.1.1 brain-computer interface.

The term "brain" does not merely refer to the consciousness or thought of a person, but represents the brain tissue structure of all the living things on earth. While a "machine" characterizes all external devices capable of computing information. The primitive form of the brain-computer interface system is formed just because there is some correlation between the two, i.e. whether the direct connection to the "brain" can be established using the information processing advantages of the "computer". The research on brain-computer interfaces has been continued for many years, and the principles on brain-computer interfaces obtained from various studies are continuously accumulated, so that the research on brain-computer interface systems at the present stage is not so difficult. The brain-computer interface obtains the brain electrical signals sent by the brain on the basis of sufficiently understanding the operation mode of the human brain, so that the brain can control external equipment. The brain of the human body plays a role as a controller, and the acquired brain electrical signals are processed and interpreted by a computer so as to complete the conversion between brain signals and computer signals, thereby realizing the human-computer interaction.

1.1.2 principles and compositions of brain-computer interfaces.

Analysis shows that the current activity of the human nervous system changes due to external stimulation or when the human brain generates motor consciousness. The current change of the nervous system indicates the limb or thinking change of the human body to some extent, the constantly changed signals can be captured in a proper mode to be used as characteristic signals for identifying the motion of the human body, and according to the characteristics of different display modes of each type of signals, the consciousness, idea or thought of the human brain is converted into an instruction capable of driving entity equipment by using a computer, so that the human brain can directly interact with external equipment, and a brain-computer interface system is realized.

fig. 4 is a block diagram of the basic components of the brain-computer. The brain-computer interface system identifies the signal characteristics in the cerebral cortex, converts the signal characteristics into a signal mode which can be identified by a computer after special processing, and delivers the signal mode to the computer for characteristic identification and processing.

(1) And (5) signal acquisition.

The signal acquisition is mainly to screen out special signals capable of indicating the intention of the human brain from all activity signals sent by the cerebral cortex of the human body and acquire the signals. The current acquisition means are mainly implantable electrodes and non-implantable electrodes. Electrode implantation has made great progress internationally, and foreign researchers have placed several metal electrodes on a very thin plastic layer and laid a wire electrode material with high sensitivity so that the electrodes can reach the brain area smoothly. The method can quickly capture activity signals sent by the cerebral cortex of the human, and simultaneously can transmit feedback information to the brain of the human in time, so that the disabled can operate various auxiliary external devices in the method due to the advantage. In China, the characteristics of non-implantation type mainly researched are that a head-wearing type electroencephalogram scanner is used for capturing activity signals of cerebral cortex. The non-implanted signal acquisition method is high in safety and simple to operate, and is an enthusiastic method for BCI researchers.

(2) And (6) signal processing.

1) Pretreatment: the method mainly filters noise and reduces the influence of the noise on the subsequent electroencephalogram signal identification.

2) Feature extraction: the method mainly screens out signal characteristics capable of representing brain will from various EEG signals sent from cerebral cortex, and classifies and identifies the key characteristics by using a classifier.

3) and (4) feature classification: the method mainly modifies the parameters of the classifier, improves the recognition rate of the classifier, enables a brain-computer interface system to distinguish EEG signals representing various brain intentions, and outputs the distinguishing result in the form of a computer command.

(3) A peripheral control device.

and according to the processing result of the EEG signal by the computer, the peripheral equipment is directly operated.

(4) and (6) feeding back.

when the human brain interacts with the entity equipment, the entity equipment needs to feed back the control condition to the human brain, so that the human brain can adjust the EEG in real time according to the feedback condition, and direct interaction of the human brain on the entity equipment is realized.

2. And (4) motor imagery experimental data and preprocessing.

2.1 motor imagery electroencephalogram experimental data set.

The experimental Data set adopted by the invention is from a Data-sets 2a Data set in the 2008 international BCI competition IV provided by Austria science and technology university. The data 9 shows that the experimenters (a01 to a09) respectively perform two different types of EEG signal motor imagery tasks, namely, left-handed and right-handed motor imagery tasks. A single experiment divides the motor imagery task into four parts: the first stage is a Fixation cross stage, and in the first 2s after the experiment begins, the subject needs to be quiet and waits for a screen prompt; the second stage is a Cue stage, when 2s is finished, a left-right direction arrow is displayed on a screen, the arrow lasts for 1s, and the subject prepares to perform a corresponding motor imagery task according to the screen display; the third stage is a Motor image stage, when the 3 rd s begins, a subject needs to perform a Motor imagery task according to the arrow prompt for 3 seconds; the fourth phase is the Break phase, where the subject stops the motor imagery task, relaxes appropriately, while preparing for the next set of experiments. Each time period contained 48 single experiments, 12 per class. In the whole experiment process, the system continuously collects and stores the electroencephalogram signals, and divides a plurality of data into 288 training samples and 288 testing samples.

2.2 two kinds of motor imagery electroencephalogram signal identification schemes.

Two types of motor imagery electroencephalogram signal identification schemes are shown in fig. 6.

in the training mode, down-sampling and intercepting are carried out on electroencephalogram data, appropriate parameters are selected by utilizing a mutual information method, then filtering processing is carried out on a frequency domain and a spatial domain by using AR-CSP, then the processed EEG signals are classified, and in the classification process, all parameters of a classifier need to be debugged, so that the result is more referential.

In the test mode, preprocessing the electroencephalogram data according to the training mode, then classifying the data for test by using the trained classifier, and recording the result.

And 3, reconstructing a recursive graph based on the EMD electroencephalogram signal phase space.

3.1 empirical mode decomposition principle.

fig. 2 is a flowchart of the empirical mode decomposition algorithm, and the specific steps are as follows:

s(t)＝x(t)+n(t)#(3.1)

N (t) is white gaussian noise subject to N (0, σ ²).

wherein c is the number of IMF components.

wherein epsilon is the amplitude of the white Gaussian noise, N is the number of times of adding the white Gaussian noise, epsilon _n represents the error between the added natural modal components of each order and the original signal, and when N is 100-300 and epsilon is 0.01-0.5 times of the standard deviation of the selected signal, the white noise error is reduced to the minimum.

3.2 phase space reconstruction.

For general time series, the investigation of the properties is mainly focused in the time domain, and for some time series with chaotic properties, the investigation of the properties needs to be performed in a multidimensional space due to the randomness and the dispersion of the motion. The phase space reconstruction means that one-dimensional or multidimensional time series and one-to-one corresponding connections of coordinate points in the phase space are established, so that features which cannot be obtained by a conventional method are mined by utilizing the connections. The chaotic time sequence in the chaotic system refers to a one-dimensional or multidimensional time sequence generated by the chaotic system. Assuming that the time series is yes, the structural properties of the attractor are included in the time series.

in the beginning of the 20 th century and the 80 s, Packard et al proposed a derivative reconstruction method and a coordinate delay reconstruction method for the phase space reconstruction of time series. The two methods provide better theoretical reference for the phase space reconstruction research at the time, so that the theory about the phase space reconstruction at the time is endless. Some researchers found that the derivative reconstruction method can make the phase space reconstruction result unstable due to noise, so that the derivative reconstruction method is used in few experiments in the later research process.

The coordinate delay method formula is as follows:

y(i)＝{x(i),…,x(i+(d-1)τ)},1≤i≤n-(d-1)τ#(3.6)

Where d represents the embedding dimension and τ represents the time delay. It can be seen from the formula that, when reconstructing the d-dimensional phase space, the time delay of the time sequence needs to be adjusted continuously, and the dimension of the time sequence needs to be kept within a certain range.

in the early 80 s of the 20 th century, Takens proposed the embedding theorem: for an infinitely long, noiseless one-dimensional scalar time sequence { x (i): i ═ 1, …, n } of the d 'dimensional chaotic attractor, a d-dimensional embedded phase space can be found in the topological invariant sense, as long as the dimension d is more than or equal to 2d' + 1.

From the above theorem, it can be seen that the phase space reconstruction involves two indexes: i.e. embedding dimension and delay time. Both indexes are not verified in experiments, and in practical application, a time sequence is accompanied by certain noise interference, so that two parameters of phase space reconstruction need to select proper values according to practical situations. At present, the statements on how to determine these two indicators are roughly divided into two categories: one is to consider that there is no correlation between the two, so they choose to determine the delay time according to the characteristics of the time series, and then choose a suitable method or theory to determine the embedding dimension according to the actual situation. The second is to consider that there is some correlation between the two. Therefore, when reconstructing the phase space, they will take the delay time and the embedding dimension into consideration, and then proceed to the next operation after the determination is made.

3.2.1 determination of delay time.

On the one hand, if the delay time is too small, the vector space

y(i)＝{x(i),…,x(i+(d-1)τ)},1≤i≤n-(d-1)τ#(3.7)

The two coordinate components x (i + j τ) and x (i + (j +1) τ) in the above equation become indistinguishable due to proximity, resulting in the two coordinate components not exhibiting reasonable independence; on the other hand, when the delay time is too long, the change of the two states before and after the time delay is too large, two coordinate components are completely independent, the information is very complex, and the chaotic attractor is excessively separated. Thus, by determining a reasonable delay time, the phase space trajectories can be separated within a reasonable range.

(1) Mean shift method

Rosenstein M T, Collins J, De Luca C J propose an Average Displacement method (AD) to determine the time delay:

wherein M ═ N- (M-1) × τ, N is the sequence length.

the average shift method, due to its characteristics, requires more experimental data for determining the delay time, and the result is often accompanied by larger error, so it is rarely used

(2) Method of auto-correlation function

For the chaotic sequence x (1), x (2), …, x (n), the autocorrelation function is:

Where τ is a time interval, R (τ) represents the degree of correlation at two times, and the larger the autocorrelation function is, the stronger the correlation between the function x (i) and the delay time τ is, when R (τ) falls to (1-e ^-1) of the initial value R (0), the time τ at this time can be considered as a delay time for reconstructing the phase space.

The autocorrelation function method cannot be effectively popularized to high-dimensional data due to the characteristics of the autocorrelation function method, and therefore, the autocorrelation function method is rarely used in actual operation.

(3) Mutual information method.

in the middle of the 80's of the 20 th century, Fraser and Swinney proposed Mutual Information Method (Mutual Information Method) — for a time series { s _i }, defining p _s (s _i) as the probability of occurrence of a variable s _i, the Information entropy of a system is the average Information content of the variable s _i, which is simply called entropy, and is defined as follows:

wherein P _q|s (q _j | s _i) is conditional probability.

If x is known at time T, then the uncertainty of x at time T + T is:

I(τ)＝H(x)+H(x_τ)-H(x,x_τ)#(3.14)

the mutual information method is currently the most commonly used method, and although the determination process of the delay time is more complicated, the result obtained by applying the mutual information method to the chaotic time sequence is more accurate.

3.2.2 determination of embedding dimension.

(1) geometric invariant method:

According to the Takens embedding theory, in the phase space reconstruction, a part of the geometric invariants can present the characteristic of completely no correlation with the embedding dimension when the phase space dimension is far larger than the minimum embedding dimension. According to this theory, when performing the phase space reconstruction, an embedding dimension that enables these geometric invariants to no longer change can be selected as the reconstructed phase space dimension.

(2) False approach point method.

The False Nearest point method (FNN) was proposed by Kennel M B, Brown R, Abarbanel H D I in 1992. From the geometric point of view, the chaotic time series is the projection of the track of the high-dimensional phase space chaotic motion in a one-dimensional space, and in the process of the projection, the track of the chaotic motion is distorted. Two points which are not adjacent in the high-dimensional phase space become two adjacent points when projected on the one-dimensional space, namely the false proximity point. The phase space reconstruction is to restore the track of the chaotic motion from the chaotic time sequence, and the track of the chaotic motion is opened along with the increase of the embedded dimension number, and the false near points are gradually removed, so that the track of the whole chaotic motion is restored, which is the principle of the false near points.

X_i＝{x(i),x(i+τ),x(i+2τ),…,x(i+(m-1)τ)}#(3.15)

if d _m+1 is far from d _m, it indicates that the two points are false proximity points.

(3) cao algorithm

To improve the false approach point method, professor Cao Liangyue proposes the Cao algorithm. In the process of phase space reconstruction, the Cao algorithm does not need to consider embedding dimension, and the specific calculation process is as follows:

The ratio of d _m+1 to d _m is defined as a (i, m), i.e.:

Wherein d _m+1 and d _m are as defined in formulas (3.17) and (3.16).

Defining:

3.3 recursive graph analysis.

3.3.1 recursive graph approach.

In the early 40 s of the 20 th century, Monk excavated the recursion existing in various places in the universe by observing various phenomena in the universe based on the theory of the recursion property ubiquitous in the ecosystem. At the end of the 80's 20 th century, Eckmann et al first proposed a recursive graph analysis method capable of graphing the basic recursive characteristics of a phase space, which exhibited the structural characteristics of a nonlinear time series with a unique perspective, so that significant breakthroughs were made in the research on the recursive characteristics of a high-dimensional phase space at that time. In fact, for some special systems, it is usually possible to describe them using multidimensional vectors, and the recursion can be determined by the distance between these multidimensional vectors, since the system is multi-dimensionally vectorized and then corresponds to each coordinate point in the phase space. However, this integrity also complicates recursive graph analysis. In the early 90 s of the 20 th century, Zbilut and Webber proposed a recursive quantitative analysis method. The method is insensitive to noise and only requires that the signal has piecewise stationarity.

At present, the theory about the recursive graph is already many, which also promotes the application process of the method in real life. In 2006, Yan used a recursive graph approach to analyze the dynamics of a sound sequence. Chen and Yang use a recursive graph to identify and classify myocardial infarction diseases. Nguyen et al use a recursive graph to analyze the infant's respiratory signal to achieve the same step analysis of the signal data. The above-mentioned example shows that the recursive graph method can effectively improve the development process of the BCI system.

3.3.2 recursive graph.

X_i＝{x_i,x_i+τ,…,x_i+(m-1)τ}(i＝1,2,…,N)#(3.21)

Where N ═ N- (m-1) τ, τ is the delay time, and m is the embedding dimension. Then the distance between any two vectors in these vector sets is:

d_ij＝‖X_i-Y_j‖#(3.22)

After selecting the threshold r, according to the formula:

R_ij＝Heaviside(r-d_ij)#(3.23)

According to the above formula, points are drawn on a coordinate plane with the vertical axis and the horizontal axis as time series numbers according to the result of R _ij, when R _ij is 1, the Euclidean distance between two vectors of X _i and X _j is smaller than a threshold value R, the system is in a recursive state, the corresponding position on the plane is displayed as a black point, when R _ij is 0, the Euclidean distance between two vectors of X _i and X _j is larger than the threshold value R, the system is in a non-recursive state, and the corresponding position on the plane is displayed as a white point.

4, extracting features based on the multi-scale recursive graph and the convolutional neural network.

4.1 convolutional neural networks.

A convolutional neural network is a Hierarchical Model whose input may be image, audio, or video data. Convolutional neural networks progressively extract more efficient features from raw data by stacking a set of Convolution (Convolution) operations, Non-linear Activation functions (Non-linear Activation functions) and Pooling (Pooling) operations layer by layer. Compared with classical artificial neural networks such as BP neural network, limited Boltzmann machine, Hopfiled neural network and SOFM network, the convolutional neural network has better recognition effect on images, and is often used in the fields of handwritten number recognition, face recognition, doorplate recognition and the like. LeNet-5 is a classical convolutional neural network. The structure of the LeNet-5 convolutional neural network is shown in FIG. 7, and the network structure comprises 7 layers except the input layer, namely two convolutional layers, two pooling layers, two full-link layers and one output layer, wherein each layer is provided with weights of different scales.

the convolutional networks that are popular today vary greatly in both depth and size. For example, in 2015, the residual neural network proposed by microsoft team reaches a scale of more than 1000 layers.

4.1.1 convolution layer.

One feature of convolutional layers is local connectivity. In the conventional classical artificial neural network, in order to enable the neural network to effectively recognize images, scientific researchers connect each pixel point of the images with neurons as features, so that the training of the neural network becomes complex due to excessive parameters between layers, and in the convolutional neural network containing convolutional layers, only part of information of the image pixels is needed in the middle layer due to convolution operation, which is equivalent to the fact that each neuron is connected with only part of pixels of corresponding images, so that the number of weights can be greatly reduced, and the training efficiency is accelerated.

Another feature of convolutional layers is weight sharing. For example, an image is processed and input as a convolution layer in a numerical matrix manner, and at this time, the convolution kernel performs convolution calculation on different areas of the numerical matrix of the image according to parameters set manually, so as to extract features corresponding to the features included in the image. As shown in FIG. 8, the weight sharing process of convolutional layer makes only 3 groups of weights different.

4.1.2 activate function.

according to neurology, activation of a neuron requires that the signal weight delivered by the dendrite to which it is attached be greater than a threshold. The function of the activation function is therefore to determine whether each neuron in the neural network is above a threshold. In a convolutional neural network, the activation function should have the following properties:

(1) Non-linear.

(2) continuous or micro. Gradient descent method requirements.

(3) The range is not saturated.

(4) monotonicity.

(5) Approximately linear at the origin.

currently, the commonly used activation functions only have a part of the above properties, and some commonly used activation functions are described as follows:

1) sigmoid function, whose expression is:

Since the gradient of the function becomes small when the input is slightly far away from the coordinate origin, the optimization of the weight is not facilitated, and the function output is not centered on 0, so that the weight updating efficiency is reduced, and the like, and the application is less.

(2) the Tanh function, whose expression is:

the Tanh function is a hyperbolic tangent function, the output interval is (-1,1), and the function image takes 0 as the center.

(3) The ReLU function, whose expression is:

F(x)＝max(0,x)#(4.3)

the ReLU function is an activation function which is widely applied at present, and compared with a Sigmoid function and a Tanh function, the ReLU function has the advantages of no problem of gradient saturation when the input is positive, high calculation speed and the like. However, the ReLU function is not a 0-centered function.

(4) the PReLU function, whose expression is:

F(x)＝max(ax,x)#(4.4)

The PReLU function is an improvement over the ReLU function, which can be activated even if the input is negative.

(5) ELU function, whose expression is:

The ELU function is also an improved version of the ReLU function, and compared with the ReLU function, the ELU function can be activated when the input is negative, and the output has certain anti-jamming capability.

4.1.3 pooling layer.

The main role of the pooling layer is to perform feature mapping on the feature map. The most common pooling operations currently used are two: a maximum pooling method, namely, taking the maximum value of the sub-region of the feature map as a feature mapping result, which is a method frequently selected by the current convolutional neural network pooling layer; another method is average pooling, which is to take the average of the sub-regions of the feature map as the result of the feature mapping, and is rarely used because of its poor performance in practice.

as shown in fig. 9, a 4 × 4 signature is input, and pooling operation is performed on each of the sub-regions 2 × 2 therein with a step size of 2, so that each sub-region does not overlap.

4.1.4 fully connected layers.

In the convolutional neural network, the fully connected layer mainly establishes the relation of the local characteristics extracted by the front hidden layer to the test sample so as to classify the characteristics. Typically, the fully-connected layer is implemented primarily by convolution operations: when the current layer is a convolutional layer and the result shows convolution with the height h and the width w, the fully-connected layer converts the convolutional layer into a global convolution corresponding to the fully-connected layer; when the front layer is a fully connected layer, the fully connected layer of the back layer converts it into a convolution with a convolution kernel of 1 × 1. The matrix-vector product is a fully-connected core operation:

y＝Wx#(4.6)

4.2 two-stage feature extraction method based on multi-scale recursion graph and convolutional neural network

combining the characteristics of a multi-scale recursion graph and a convolution neural network, carrying out two-stage feature extraction on two types of motor imagery signals: firstly, performing Empirical Mode Decomposition (EMD) on preprocessed data, and calculating a recursion graph of intrinsic mode components (IMFs) of all scales; and then, taking the recursive graph of each scale as an input of a convolutional neural network, and carrying out classification and identification on the recursive graph by using the convolutional neural network. The motor imagery electroencephalogram signal feature extraction and classification steps of the two types of motor imagery tasks are as follows:

(1) And selecting proper embedding dimension and delay time, and performing feature extraction on the preprocessed electroencephalogram signal data by using a recursive graph analysis method. Firstly, EMD decomposition is carried out on two types of motor imagery electroencephalogram signals respectively to obtain IMFs, a recursion graph is calculated for each IMF, and therefore the level 1 characteristic is obtained.

(2) Training by taking the level 1 recursive graph characteristics of a training data set as input of a CNN, debugging various parameters of the CNN network to ensure that the network precision is ensured, wherein the convolutional neural network comprises an input layer, 3 convolutional layers, 3 pooling layers, a Flatten layer and an output layer, activating function dimension ReLU after convolution operation, and finally a classifier selects Sigmoid.

(3) And applying the trained convolutional neural network to the test data, and carrying out classification and identification on the test data. And (3) performing feature extraction on the tested electroencephalogram data according to the mode of (1), and performing another feature extraction on a recursion graph of the tested electroencephalogram data by using the trained convolutional neural network to obtain a classification recognition result.

5 experiment and result analysis.

5.1 parameter selection.

In the 1 st level feature extraction, the proper embedding dimension and delay time are selected for each inherent modal component IMF, so that the feature extraction precision of the motion image signal can be effectively improved. According to the content of 3.3, the delay time and the embedding dimension of each IMF are obtained according to a mutual information analysis method and a Cao-Liangyue method as follows:

TABLE 1 embedding dimension and delay time for each IMF component

during the 2 nd level feature extraction, proper convolutional layers and pooling layer numbers and an activation function need to be selected to prevent the overfitting phenomenon during network training. The structure of the convolutional neural network for performing the 2 nd level feature extraction this time is shown in fig. 3. The convolutional neural network comprises 3 layers of convolutional layers and 3 layers of pooling layers, wherein an active function selects a ReLU function which is widely applied at present, a maximum pooling method is selected in a pooling operation, and a Sigmoid classifier is selected in a final classifier.

5.2 the brain wave signal recursion graph construction based on EMD decomposition.

After the motor imagery electroencephalogram signals are decomposed through an empirical mode, a group of IMFs with different scales can be obtained, each IMF is subjected to phase space reconstruction, a recursion graph is calculated, and then a multi-scale recursion graph of the EEG signals can be obtained. Taking a Data-sets 2a Data set in international BCI competition IV in 2008 as an example, selecting a left hand of a tester with the number of A01 and a right hand of a tester with the number of A02 as a group respectively, obtaining the delay time and the embedding dimension of each IMF according to a mutual information analysis method and a Cao-Liangyue method, and then calculating a multi-scale recursion diagram of each IMF, wherein the result is shown in FIG. 10. Fig. 10 visually presents the running states of different motor imagery electroencephalograms, but the information amount of the sequencing recursion diagram is large, so that the sequencing recursion diagram is inconvenient to be directly used for classification, the image is used as the input of a convolutional neural network, and the secondary features of the motor imagery electroencephalograms are extracted through supervised training.

wherein: FIG. 10a shows a recursive diagram of IMF1, with left-handed and right-handed electroencephalograms from left to right. FIG. 10b is a recursive diagram of IMF2, showing left-handed and right-handed brain electrical signals from left to right. FIG. 10c shows a recursive diagram of IMF3, with left-handed and right-handed electroencephalograms from left to right. FIG. 10d shows a recursive diagram of IMF4, with left-handed and right-handed electroencephalograms from left to right.

5.3 analysis of results.

and (3) taking the Data-sets 2a Data set as an EEG signal classification object, and testing a classification recognition result of a two-stage feature extraction method based on a multi-scale recursion graph and a convolutional neural network. Table 2 is a table for comparing the classification recognition rates, and the first three rows in the table are the test results of the first three Data sets 2a of International BCI Competition IV in 2008.

TABLE 2 Classification identification rate comparison table

The improved filter bank CSP is expanded to multiple classes by adopting an OVR mode in the 1 st competition, and the used classifier is a naive Bayes Parzen window classifier. And in the 2 nd competition, OVO-CSP is adopted to extract features, then the LDA method is used for further reducing dimension, and a Bayesian classifier is used for classification. In the 3 rd competition, CSP is adopted for feature extraction, and three groups of two-layer binary tree multi-class classifiers are constructed by using SVM as a classifier for classification.

as can be seen from Table 2, the average value of the overall recognition rate of all the subjects obtained by the method of the present invention is 0.6045, which is about 3.35% higher than the 1 st match performance and about 8.8% higher than the 2 nd match performance.

Compared with a method which only adopts a multi-scale recursive graph analysis method or a neural network method, the two-stage feature extraction method based on the multi-scale recursive graph and the convolutional neural network provided by the invention has the advantages that the overall recognition rate of the motor imagery electroencephalogram signals is improved, the classification result and the first achievement in the BCI competition IV are improved by about 0.335, meanwhile, the single recognition rate of each subject is also slightly superior, and the recognition rate comparison result in the table 2 shows that the two-stage feature extraction method based on the multi-scale recursive graph and the convolutional neural network has better electroencephalogram signal recognition performance to a certain extent.

in the above embodiments, the implementation may be wholly or partially realized by software, hardware, firmware, or any combination thereof. When used in whole or in part, may be implemented in a computer program product comprising one or more computer instructions. When loaded or executed on a computer, cause the flow or functions according to embodiments of the invention to occur, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer instructions may be stored on a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, from one website site, computer, server, or data center to another website site, computer, server, or data center via wire (e.g., coaxial cable, fiber optics, Digital Subscriber Line (DSL), or wireless (e.g., infrared, wireless, microwave, etc.)). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device including one or more available media integrated servers, data centers, and the like. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., Solid State Disk (SSD)), among others.

the above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents and improvements made within the spirit and principle of the present invention are intended to be included within the scope of the present invention.

Claims

1. the electroencephalogram signal identification method based on the joint recursion graph and CNN is characterized by comprising the following steps:

Decomposing the motor imagery electroencephalogram signal into inherent modal functions of different scales, and calculating a multi-scale recursion graph of inherent modal components of each scale to obtain a level 1 characteristic;

And taking the reconstructed multi-scale recursion map as an image feature of the EEG signal, taking the multi-scale recursion map feature as an input of a convolutional neural network, carrying out classification and identification on the recursion map by using the convolutional neural network, and extracting a level 2 feature for expressing the motor imagery EEG signal from the level 1 feature.

2. The method for electroencephalogram signal identification of a joint recursion graph and CNN as claimed in claim 1, wherein the method for electroencephalogram signal identification of a joint recursion graph and CNN specifically comprises the following steps:

Dividing an acquired data set into training data and testing data, and respectively preprocessing the data;

Performing empirical mode decomposition on two types of motor imagery electroencephalogram signals, namely training data and test data respectively to obtain inherent modal components of different scales; meanwhile, calculating a proper embedding dimension and delay time of each inherent modal component by using a mutual information analysis method and a Cao-Liangyue method to perform spatial reconstruction to obtain a recursion graph of each inherent modal component, and performing feature extraction by using a recursion graph analysis method to obtain a level 1 feature;

Step three, training by taking the level 1 recursive graph characteristics of the training data set as the input of the convolutional neural network, and debugging each parameter of the convolutional neural network;

step four, applying the trained convolutional neural network to test data for classification and identification; and performing feature extraction on the test electroencephalogram signal data by using a recursion graph method, and performing feature extraction on the recursion graph of the test electroencephalogram signal data by using the trained convolutional neural network again to obtain a classification recognition result.

3. The method for electroencephalogram signal identification by combining a recursion map and a CNN as claimed in claim 2, wherein in the first step, the data preprocessing specifically comprises:

Training data: down-sampling and intercepting training electroencephalogram data, selecting appropriate parameters by using a mutual information method, performing filtering processing on a frequency domain and a spatial domain by using AR-CSP (auto-correlation-ranging chip scale package), and classifying processed EEG (electroencephalogram) signals;

Test data: down-sampling and intercepting test electroencephalogram data, selecting appropriate parameters by using a mutual information method, performing filtering processing on a frequency domain and a spatial domain by using AR-CSP (auto-correlation-ranging chip scale package), and classifying processed EEG (electroencephalogram) signals; and classifying the test data by using the trained classifier and recording the result.

4. The electroencephalogram signal identification method combining the recursion map and the CNN as set forth in claim 2, wherein in the second step, the empirical mode decomposition specifically includes:

s(t)＝x(t)+n(t)；

Where N (t) is white gaussian noise subject to N (0, σ ²);

Wherein c is the number of IMF components;

5. the method for electroencephalogram signal identification by combining a recursion map and CNN as set forth in claim 2, wherein in the second step, the calculating an appropriate embedding dimension and delay time for each intrinsic mode component by using a mutual information analysis method and a Cao-Liangyue method specifically comprises:

(1) Determining delay time by mutual information analysis:

wherein, P _q|s (q _j | s _i) is conditional probability;

If x is known at time T, then the uncertainty of x at time T + T is:

I(τ)＝H(x)+H(x_τ)-H(x,x_τ)；

(2) the Cao-Liangyue method determines the embedding dimension:

X_i＝{x(i),x(i+τ),x(i+2τ),…,x(i+(m-1)τ)}；

the ratio of d _m+1 to d _m is defined as a (i, m), i.e.:

Defining:

6. the electroencephalogram signal identification method combining the recursion map and the CNN as set forth in claim 2, wherein in the second step, the recursion map analysis method specifically comprises:

X_i＝{x_i,x_i+τ,…,x_i+(m-1)τ}(i＝1,2,…,N)；

d_ij＝‖X_i-Y_j‖；

after selecting the threshold r, according to the formula:

R_ij＝Heaviside(r-d_ij)；

Drawing points on a coordinate plane with the vertical axis and the horizontal axis as the time series number according to the result of R _ij;

7. The electroencephalogram signal identification method combining the recursion map and the CNN as set forth in claim 2, wherein in step three, the neural network specifically comprises:

8. a joint recursion graph and CNN electroencephalogram signal identification system of the joint recursion graph and CNN electroencephalogram signal identification method of claim 1.

9. an information data processing terminal for implementing the electroencephalogram signal identification method combining the recursion map and the CNN as set forth in any one of claims 1 to 7.

10. a computer-readable storage medium comprising instructions which, when executed on a computer, cause the computer to perform the method of joint recursion graph and CNN brain electrical signal identification according to any one of claims 1-7.