Robust recognition method for EEG speech imagery with adjacent trial task aliasing measure

Abstract

The present invention provides a method for robust recognition of EEG speech imagery with aliasing measures of adjacent trial tasks. The steps of the present invention are as follows: 1. Multiple subjects collect EEG data under different speech imagery tasks. 2. Preprocessing and feature extraction. 3. Centralized processing. 4. Establish a machine learning model to realize the aliasing measure of speech imagery tasks between adjacent trials based on EEG data and the robust recognition of speech imagery. 5. Obtain the aliasing measure matrix, mapping matrix and "clean" low-rank data. 6. Speech imagery decoding is performed on the EEG data of the subject. The present invention takes into account the task aliasing phenomenon between adjacent trials in different types of speech imagery tasks, and provides a quantitative measurement method for the degree of aliasing. At the same time, the changing trend of the degree of aliasing over time can be explicitly displayed according to the aliasing measure matrix. In addition, the mapping matrix can measure the importance of different data features to decoding in the speech imagery decoding task.

Description

Technical Field

[0001] The invention belongs to the technical field of electroencephalogram signal processing, and in particular relates to an electroencephalogram speech imagery robust recognition method with adjacent trial task aliasing measure. Background Art

[0002] Speech refers to the ability or behavior of humans to express thoughts, feelings, and intentions through language. Speech can be conveyed through oral expression, written text, gestures, expressions, etc. Speech is one of the main tools for human communication and communication. It helps people share information, build relationships, solve problems, etc. Verbal imagination is the construction of various images, scenes, concepts or emotions through thinking and imagination with the help of language ability. When we hear a story, description or poem, the brain will automatically generate corresponding images or scenes in our mind based on the words heard. This ability enables people to experience, understand and share various feelings, ideas and experiences through language. Verbal imagination is the basis of literature, art and communication. In the brain, language processing involves multiple areas, including Broca's area and temporal gyrus, which play an important role in the production and understanding of speech. When an individual engages in verbal imagination, the brain will produce bioelectric signals related to verbal imagination behavior - electroencephalography (EEG). In particular, there are stronger related electrical signals in the brain areas involved in language processing. We can use EEG to decode verbal imagination. The field of verbal imagination decoding involves a variety of technologies and methods, including natural language understanding, machine learning, neural networks, etc. Researchers are trying to enable computers to recognize and decode people's expressed thoughts, emotions, and intentions from text or speech. Since EEG signals originate from brain bioelectricity, they are not easy to disguise. Therefore, EEG decoding of verbal imagination based on EEG signals can better reflect the individual's actual verbal imagination task type.

[0003] How to decode the type of speech imagery task of an individual based on EEG is an important goal of speech imagery decoding research. Based on EEG speech imagery data, it often contains a variety of data that interfere with speech imagery decoding. An obvious interference is that between two adjacent trials of different speech imagery tasks, the data of the latter trial often contains the data of the speech imagery type of the previous trial. Because when the brain contacts two adjacent tasks, the "image" of the previous task will still remain briefly in the next task, resulting in the aliasing phenomenon of the superposition of two speech imagery task data for a certain period of time. We refer to this phenomenon as "task aliasing". This aliasing phenomenon will be particularly obvious in the early stage of the next trial. However, how to explicitly quantify the degree of aliasing from a mathematical perspective and explicitly express the changing trend of the degree of aliasing is an urgent problem to be solved. Secondly, in addition to the influence of aliased data, signals including electrooculography, electromyography and other bioelectricity will also interfere with decoding. However, a large number of existing methods take measures to pre-process the data first, remove the interference data based on experience, and then use the processed data to decode speech imagery. This method mainly uses experience to denoise the data, which makes the denoising of the data too rough. The real solution data is likely to be lost during the processing, and the effect of removing interference noise is not obvious. A robust decoding method that can denoise more accurately has become a research hotspot in the field of speech imagery decoding. Summary of the invention

[0004] The present invention aims at the problem that the existing EEG-based speech imagination decoding method lacks adjacent trial aliasing measurement and is insufficient in robustness, and proposes an EEG speech imagination robust recognition method with adjacent trial task aliasing measurement. On the one hand, the present invention measures the aliasing degree of adjacent trial speech imagination tasks through the aliasing measurement matrix E, and quantitatively gives the data aliasing degree between different speech imagination task types from a mathematical perspective; on the other hand, by solving the aliasing measurement matrix E, the noise in the original EEG data can be eliminated to obtain a "clean" low-rank data matrix D, and then the speech imagination classification task is performed based on the data D under the robustness framework-nuclear norm, so that a strong robust speech imagination classification method can be obtained. Data denoising and speech imagination classification are carried out simultaneously in one module, so the noise of the data can be removed more accurately, and the denoising effect is better and more obvious. In addition, the present invention also has the function of measuring the importance of different features in the speech imagination task. Compared with traditional methods, the present invention has more obvious application prospects.

[0005] In a first aspect, the present invention provides a method for robust recognition of EEG speech imagery with adjacent trial task aliasing measure, which comprises the following steps:

[0006] Step 1: The subjects collect EEG data under different speech imagery tasks to obtain the subjects' EEG data and their corresponding speech imagery task labeling matrix.

[0007] Step 2: EEG data obtained in step 1 Perform preprocessing and feature extraction.

[0008] Step 3: Centralize the EEG data X′ obtained in step 2 to obtain a sample set X.

[0009] Step 4: Construct the objective function of the speech imagery task aliasing measure between adjacent trials as shown in formula (1):

[0010]

[0011] In formula (1), D is the “clean” low-rank data matrix that has a direct effect on speech imagery recognition; E is the aliasing measure matrix generated by the speech imagery adjacent trial task; P is the mapping matrix; Y is the speech imagery task labeling matrix; α, β and γ are regularization parameters, ||·|| * represents the nuclear norm of the matrix. ||·||2,1 represents the l 2，1 norm;||·|| 1，2 l represents the matrix 1，2 norm; for any matrix A, its l 1，2 Norm a ij are the elements of the matrix A.

[0012] Step 5: According to the EEG sample set X and the objective function shown in formula (1), the “clean” low-rank data matrix D, aliasing metric matrix E and mapping matrix P corresponding to the subject are obtained.

[0013] Step 6: Collect EEG signals from the subjects, pre-process, extract features and centralize the EEG signals to obtain EEG feature X t , use one or more of the “clean” low-rank data D, aliasing metric matrix E, and mapping matrix P obtained in step 5, and the EEG feature X t The combination and decoding are performed to obtain the subject's speech imagination type when the EEG signal is collected.

[0014] Preferably, after step 5 is executed, the aliasing value for measuring the degree of aliasing of the subject when collecting adjacent trial data is obtained based on the aliasing measurement matrix E as follows:

[0015] Calculate the aliasing value of each observation corresponding to the aliasing measure matrix E Size, as shown in formula (2):

[0016]

[0017] Among them, ||·||2 represents the l2 norm of the vector. j represents the j-th column of the aliasing measure matrix E, and j represents the sampling sequence number of the speech imagination EEG data.

[0018] Aliasing Value The larger the value is, the greater the influence of the previous speech imagination behavior on the current speech imagination behavior when the subject is performing speech imagination; the aliasing value The changes over time indicate the trend of the residual degree of the precursor imagery task in the subjects' brains over time.

[0019] Preferably, after step 5 is executed, the importance of each feature of the sample is measured according to the mapping matrix P. The specific process is as follows:

[0020] Calculate the importance value ξ(i) of the vector of each feature of the mapping matrix, as shown in formula (3):

[0021] ξ(i)=||p i ||2 Formula (3)

[0022] p i Represents the i-th row of the mapping matrix P.

[0023] Preferably, the specific process of obtaining the "clean" low-rank data matrix D in step 5 is as follows:

[0024] Fix other variables to update variable D and rewrite equation (4) as follows:

[0025]

[0026] make Formula (5) is written as follows:

[0027]

[0028] make The update rule of D obtained by linear algebraic transformation is shown in formula (6):

[0029]

[0030] Among them, k is the number of iterative updates, is the update parameter. The final update rule of D is obtained by using the singular value threshold operator method as shown in formula (7):

[0031]

[0032] By fixing other variables to update a single variable, we get the update rule for variable K:

[0033]

[0034] in,

[0035] Preferably, the update rules of Lagrange multipliers Y1-Y3 are as follows:

[0036] Y1=Y1+μ(XDE) Formula (9)

[0037] Y2=Y2+μ(K-Y+D T P) Formula (10)

[0038] Y3=Y3+μ(K-RH) Formula (11)

[0039] Preferably, the speech imagination categories include a variety of different words or sentences.

[0040] Preferably, in step 6, the EEG feature X is combined with the projection matrix P, and a nearest neighbor (NN) classifier is used to calculate the number of correctly classified points to complete speech imagery decoding.

[0041] Preferably, in step 1, the method for the subject to perform the speech imagination task is: using speech imagination materials to induce speech imagination in the subject; the speech imagination materials include pictures, audio or video that induces speech imagination.

[0042] Preferably, the pretreatment process in step 2 is as follows:

[0043] 2-1. Downsample the EEG data recorded at 1024 Hz to a sampling rate of 256 Hz, then filter out noise and artifacts related to myoelectric activity using a fifth-order Butterworth bandpass filter with a frequency range of 1 Hz to 70 Hz, extract the state characteristics of the EEG signal in different frequency bands, and use short-time Fourier transform to calculate the differential entropy of the EEG signal as the sample matrix;

[0044] 2-2. Perform short-time Fourier transform on EEG data with a time window of 4 seconds and no overlap to extract differential entropy features As shown in formula (12);

[0045]

[0046] In formula (12), is the input sample matrix, is the element in the input sample matrix; is the probability density function, and the EEG data obtained after processing is X′;

[0047] Updated differential entropy features As shown in (13);

[0048]

[0049] In formula (13), is the standard deviation of the probability density function; is the expectation of the probability density function.

[0050] In a second aspect, the present invention provides a computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the memory stores the computer program; the processor executes the aforementioned EEG speech imagery robust recognition method with adjacent trial task aliasing measure.

[0051] In a third aspect, the present invention provides a readable storage medium storing a computer program; when the computer program is executed by a processor, it is used to implement the aforementioned EEG speech imagery robust recognition method with adjacent trial task aliasing measure.

[0052] The present invention has the following beneficial effects:

[0053] 1. The aliasing measure matrix obtained by the present invention can quantitatively examine the degree of speech imagination aliasing between different speech imagination task types in adjacent trials, that is, the degree of influence of the previous trial task on the next trial. At the same time, it also explicitly gives the changing trend of the aliasing degree.

[0054] 2. The mapping matrix obtained by the present invention can measure the importance of different features in the speech imagination task.

[0055] 3. The speech imagery decoding method proposed in the present invention unifies the decoding process and the data denoising process into one stage, and simultaneously applies the robustness framework of the nuclear norm, thereby finally obtaining a highly robust speech imagery decoding method. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic diagram of the overall framework of the present invention;

[0057] Figure 2 A schematic diagram of the aliasing measure of adjacent trials in the speech imagery task of the present invention;

[0058] Figure 3 A schematic diagram of the mapping matrix in the present invention for measuring the importance of different features in a speech imagination task. DETAILED DESCRIPTION

[0059] The present invention will be further described below in conjunction with the accompanying drawings.

[0060] The present invention solves the key problem that there is task aliasing and insufficient decoding robustness in EEG signals between adjacent trials of different speech imagination task types in speech imagination decoding based on the following starting point: We believe that in speech imagination recognition, between adjacent trials of different speech imagination task types, the previous speech imagination will remain in the next speech imagination for a short time, resulting in an aliasing phenomenon in which two speech imagination task data are superimposed for a certain period of time, and it is very important to quantitatively measure the degree of aliasing and explicitly show the changing trend of the degree of aliasing from the perspective of data. The present invention measures the degree of speech imagination aliasing between adjacent trials through the aliasing measurement matrix E, and can explicitly give the changing trend of this aliasing. In addition, it is also a very important research goal in the field of speech imagination to eliminate other noise data that interfere with the speech imagination task, including aliasing interference, and finally perform a robust speech imagination decoding. The present invention unifies data noise elimination and speech imagination decoding into one stage, and provides a strong robust decoding model based on the nuclear norm. The above two processes are of great significance to speech imagination decoding.

[0061] like Figure 1 As shown in FIG, a robust EEG speech imagery recognition method with adjacent trial task aliasing measure is provided, and the specific steps are as follows:

[0062] Step 1: Based on the speech imagination material (in this embodiment, pictures, audio or video containing obvious speech imagination induction are used), collect the EEG data of the subject when inducing the corresponding speech imagination.

[0063] The EEG signals of N subjects were collected in V different environments or times, and N·V groups of EEG data were obtained. The data volume of each group of data was d*n, where d is the dimension of each group of data and n is the number of time-related EEG data samples obtained in one collection. A group of data includes EEG data at multiple moments obtained in one collection. The same subject corresponds to the tested task category in the test task, and each subject completes the test of all the tested task categories.

[0064] In this embodiment, the task category to be tested is the verbal imagination category of the testee, including the need to drink water, the need to eat, the need to watch TV, and the need to go to the toilet. In addition, the recognized verbal imagination type can also be letters with different pronunciations; the letters with different pronunciations are mapped to specific tasks by imagining them.

[0065] "Different environments or times" means that when the same subject collects V sets of EEG data, the time periods may be distributed on different dates; the experimental contents of these different batches are the same, but their states on different dates may be different; this reflects that the essential characteristics of EEG will not change significantly with the change of time or experimental scene. These differences in time or different experimental scenes are the differentiation part.

[0066] Step 2: Preprocess and extract features from the EEG data collected in step 1. The EEG data recorded at 1024Hz is sampled at a sampling rate of 256Hz, and then filtered through a fifth-order Butterworth bandpass filter with a frequency of 1Hz to 70Hz to remove noise and artifacts related to activities such as electrooculography and electromyography. Signal processing technology is used to extract the state characteristics of the EEG signal in different frequency bands. Specifically, differential entroy (DE) is used as a feature to calculate the EEG signal. The differential entropy of is used as the sample matrix Where d is the number of sample features and n is the total number of samples. The preprocessing process is as follows:

[0067] 2-1. Downsample the EEG data to 256 Hz and then perform band-pass filtering to a range of 1 Hz to 70 Hz.

[0068] 2-2. Perform short-time Fourier transform on EEG data with a time window of 4 seconds and no overlap to extract differential entropy features As shown in the following formula (14);

[0069]

[0070] In formula (14), is the input sample matrix, is the element in the input sample matrix; is the probability density function, and the processed EEG data is represented by X′;

[0071] Updated differential entropy features As shown in the following formula (15);

[0072]

[0073] In formula (15), is the standard deviation of the probability density function; is the expectation of the probability density function.

[0074] It can be seen that the differential entropy feature is essentially the logarithmic form of the power spectrum density feature, that is, The preprocessing of EEG signals aims to improve the signal-to-noise ratio, thereby improving the preprocessing effect of the data and reducing interference.

[0075] Step 3: Centralize the preprocessed sample matrix X′. First, calculate the center matrix I n is the n-dimensional identity matrix, 1 n is an n-dimensional column vector whose element values ​​are all 1. The data set after centralization is X=X′H.

[0076] Step 4. In the present invention, considering the measure of speech imagery aliasing, the aliasing measure matrix E is embedded into the speech imagery decoding model. At the same time, aliasing interference is essentially the "residue" of the previous speech imagery task between adjacent trials of speech imagery tasks, so this part of the aliasing data should meet the characteristics of the previous trial, that is, the samples are correlated with each other and the samples are independent of each other. Therefore, the aliasing measure matrix is ​​l 1，2 norm constraint. This norm can make the aliasing metric matrix correlated within samples and sparse between samples. At the same time, in order to minimize the impact of noise on speech image decoding, the nuclear norm is used to constrain the data and label mapping. The nuclear norm uses the singular values ​​of the matrix to measure the distance between samples and labels instead of l F The norm is based on the Euclidean distance measurement, so the nuclear norm is more robust for decoding noisy speech imagery data. Taking the above two aspects into consideration and unifying data denoising and decoding into one module, the EEG speech imagery robust decoding model with adjacent trial task aliasing measure is obtained as shown in the following equation (16):

[0077]

[0078] In formula (16), is the EEG speech imagery dataset after preprocessing and centering, d represents the dimension of the sample, n represents the number of samples; D is the noise-free "clean" low-rank data matrix in the dataset that has a direct effect on speech imagery recognition; E is the aliasing measure matrix generated by adjacent trial tasks of speech imagery in the dataset, that is, aliasing data; P is the mapping matrix; Y is the speech imagery task label matrix corresponding to the sample; α, β and γ are regularization parameters; ||·|| * It represents the nuclear norm of the matrix, also known as the sum of the core singular values ​​of the matrix. For a matrix A, its nuclear norm can be expressed as where σ i are the singular values ​​of matrix A, and r is the rank of the matrix. ||·|| 2，1 l represents the matrix 2，1 norm, n, m represent the number of rows and columns of the matrix, a ij Represents the element in the i-th row and j-th column of matrix A. ||·|| 1，2 l represents the matrix 1，2 norm, l 1，2 The norm is designed based on the characteristics of aliased data: "mutual correlation within samples and mutual independence between samples", so that the aliasing measurement matrix E can meet the characteristics of aliased data that the samples are correlated within and sparse between samples.

[0079] Step 5: Optimize the objective function as shown in formula (16), that is, obtain the update rules of the three variables D, E and P of the model design. Overall, it can be performed under the structure of alternating iterative updates, that is, fix all other variables for the variable to be optimized to obtain the solution of the variable to be optimized.

[0080] 5-1. Introducing slack variables R, H so that K = RH, equation (16) is equivalent to the following equation (17):

[0081]

[0082] Where X represents the speech imagery EEG dataset after preprocessing and centering. ||·|| F l represents the matrix F Norm, for matrix A, its nuclear norm can be expressed as

[0083] 5-2. The augmented Lagrange multiplier method of formula (17) is:

[0084]

[0085] Among them, Y1~Y3 are Lagrange multipliers, and μ is the penalty factor;

[0086] 5-3. By fixing other variables and updating a single variable, we can get the update rule of variable R:

[0087]

[0088] in Represents the identity matrix.

[0089] 5-4. By fixing other variables and updating a single variable, we can get the update rule of variable H:

[0090]

[0091] 5-5. By fixing other variables and updating a single variable, we can get the update rule of variable H:

[0092] P=μ(2βG+μDD T ) -1 DQ1 Formula (21)

[0093] in, G is a diagonal matrix, and the diagonal elements in G are denoted by g ii , P i represents the i-th row of the matrix P;

[0094] 5-6. By fixing other variables and updating a single variable, we can get the update rule of variable E:

[0095]

[0096] in, represents the identity matrix, q j represents the j-th column of Q2, E(:, i) represents the j-th column of matrix E;

[0097] 5-7. Fix other variables to update variable D. At this time, equation (18) can be written as equation (23):

[0098]

[0099] make Formula (24) can be written as follows

[0100]

[0101] make Using some linear algebraic transformations, we can get the update rule of D as (25)

[0102]

[0103] Among them, k is the number of iterative updates, is the update parameter. Using the singular value thresholding (SVT) method, the final update rule of D can be obtained as shown in formula (26):

[0104]

[0105] The singular value thresholding (SVT) method is a method for matrix recovery and sparse reconstruction. The algorithm is based on nuclear norm regularization. By performing singular value decomposition on the data matrix, the singular values ​​below the set threshold are set to 0, and then the reconstructed data matrix is ​​obtained based on the new singular value matrix and the left and right singular value vector matrices:

[0106] 5-8. By fixing other variables and updating a single variable, we can get the update rule of variable K:

[0107]

[0108] in,

[0109] 5-9. The update rules of Lagrange multipliers Y1~Y4 are as shown in equations (28) to (30):

[0110] Y1=Y1+μ(XDE) Formula (28)

[0111] Y2=Y2+μ(K-Y+D T P) Formula (29)

[0112] Y3=Y3+μ(K-RH) Formula (30)

[0113] Step 6: Collect the EEG data of the subjects and perform centralized processing to obtain the EEG feature X t ; Use the projection matrix P and EEG feature X learned in step 5 t The nearest neighbor (NN) classifier is used to calculate the number of correctly classified points to complete the speech imagery decoding of the sample set. Noise data including aliasing data is considered in the decoding process, and a strong robust decoding method is used, so the accuracy of speech imagery decoding can be better improved.

[0114] After step 7 and step 5 are executed, the aliasing degree measure of adjacent trial aliasing data is obtained as follows (31). A definition provided by the present invention is shown as follows Figure 2 .

[0115] Calculate the aliasing value for each observation corresponding to the aliasing measure matrix Size, as shown in formula (31):

[0116]

[0117] Among them, ||·||2 represents the l2 norm of the vector. For a vector a, a i Represents the jth element of a vector. j It represents the j-th column of the aliased data E, and j represents the sequence number of the speech imagination EEG data of the subject collected in chronological order.

[0118] By measuring the aliasing data between adjacent trials, we can quantitatively examine from a data level the degree of influence of the preceding speech imagination task on the current task between adjacent trials when the human brain performs a speech imagination task, that is, the changing trend of the residual degree of the preceding imagination task in the brain over time.

[0119] After the mapping matrix P is obtained in step 8 and step 5, the importance of each feature of the sample can be measured according to the mapping matrix. The specific process is as follows: Figure 3 .

[0120] Calculate the importance value ξ(i) of the vector of each feature of the mapping matrix, as shown in formula (32):

[0121] ξ(i)=||p i ||2 Formula (32)

[0122] p i Represents the i-th row of the mapping matrix P.

[0123] Through the above method, we considered the task aliasing phenomenon between adjacent trials in different types of speech imagination tasks, and gave a quantitative measurement method for the degree of aliasing. At the same time, we can explicitly show the trend of the degree of aliasing over time based on the aliasing measurement matrix. In addition, according to the mapping matrix in the invention, the importance of different data features to decoding in the speech imagination decoding task can be measured. Finally, denoising and decoding are unified into one module to make the denoising process more accurate, and a highly robust speech imagination decoding method is obtained by combining the nuclear norm. The method proposed in the present invention has obvious advantages in human speech imagination decoding.

Claims

1. A robust EEG speech imagery recognition method with adjacent trial task aliasing measure, characterized by: The following steps are involved: Step 1: The subjects collect EEG data under different speech imagery tasks to obtain the EEG data of the subjects and their corresponding speech imagery task labeling matrix; Step 2: EEG data obtained in step 1 The sample matrix as input is preprocessed and feature extracted to obtain the processed EEG data X′; Step 3: Centralize the EEG data X′ obtained in step 2 to obtain a sample set X; Step 4: Construct the objective function of the speech imagery task aliasing measure between adjacent trials as shown in formula (1): In formula (1), D is the low-rank data matrix that has a direct effect on speech imagery recognition; E is the aliasing measure matrix generated by the speech imagery adjacent trial task; P is the mapping matrix; Y is the speech imagery task labeling matrix; α, β and γ are regularization parameters, ||·|| * represents the nuclear norm of the matrix; ||·|| 2,1 l represents the matrix 2,1 norm;||·|| 1,2 l represents the matrix 1，2 norm; for any matrix A, its l 1，2 Norm a ij is the element of matrix A; Step 5: According to the EEG sample set X and the objective function shown in formula (1), the low-rank data matrix D, the aliasing measure matrix E and the mapping matrix P corresponding to the subject are obtained; Step 6: Collect EEG signals from the subjects, pre-process, extract features and centralize the EEG signals to obtain EEG feature X t , use one or more of the low-rank data D, aliasing metric matrix E and mapping matrix P obtained in step 5, and the EEG feature X t The combination and decoding are performed to obtain the subject's speech imagination type when the EEG signal is collected.

2. The method for robust EEG speech imagery recognition with adjacent trial task aliasing measure according to claim 1, characterized in that: Based on the aliasing measure matrix E obtained in step 5, the aliasing value for measuring the degree of aliasing when the subject collects adjacent trial data is obtained as follows: Calculate the aliasing value of each observation corresponding to the aliasing measure matrix E Size, as shown in formula (2): Among them, ||·||2 represents the l2 norm of the vector; e j represents the j-th column of the aliasing measure matrix E, where j represents the sampling sequence number of the speech imagery EEG data; Aliasing Value The larger the value is, the greater the influence of the previous speech imagination behavior on the current speech imagination behavior when the subject is performing speech imagination; the aliasing value The changes over time indicate the trend of the residual degree of the precursor imagery task in the subjects' brains over time.

3. The method for robust EEG speech imagery recognition with adjacent trial task aliasing measure according to claim 1, characterized in that: The importance of each feature of the sample is measured according to the mapping matrix P obtained in step 5. The specific process is as follows: Calculate the importance value ξ(i) of the vector of each feature of the mapping matrix, as shown in formula (3): ξ(i) = ||p i ||2 Equation (3) p i Represents the i-th row of the mapping matrix P.

4. The method for robust recognition of EEG speech imagery with adjacent trial task aliasing measure according to claim 1, characterized in that: The specific process of obtaining the low-rank data matrix D in step 5 is as follows: Fix other variables to update variable D. The expression for updating variable D is: Where μ is the penalty factor; ||·|| F l represents the matrix F norm; Y1 and Y2 are Lagrange multipliers; K = RH; R and H are slack variables; α is the regularization parameter; make Formula (4) is written as follows: make The update rule of D obtained by linear algebraic transformation is shown in formula (6): Among them, k is the number of iterative updates, To update the parameters; the final update rule of D is obtained by using the singular value threshold operator method as shown in formula (7) By fixing other variables to update a single variable, we get the update rule for variable K: in, Y3 is the Lagrange multiplier.

5. The method for robust EEG speech imagery recognition with adjacent trial task aliasing measure according to claim 4, characterized in that: The update rules of Lagrange multipliers Y1~Y3 are as follows: Y1=Y1+μ(XDE) Formula (9) Y2 = Y2 + μ(K - Y + D T P) Equation (10) Y3=Y3+μ(K-RH) Formula (11).

6. The method for robust EEG speech imagery recognition with adjacent trial task aliasing measure according to claim 1, characterized in that: In step 6, the EEG sample set X is combined with the projection matrix P, and the nearest neighbor classifier is used to calculate the number of correctly classified points to complete the speech imagery decoding.

7. The method for robust EEG speech imagery recognition with adjacent trial task aliasing measure according to claim 1, characterized in that: In step 1, the method for the subject to perform the speech imagination task is: using speech imagination materials to induce speech imagination in the subject; the speech imagination materials include pictures, audio or video that induces speech imagination.

8. The method for robust EEG speech imagery recognition with adjacent trial task aliasing measure according to claim 1, characterized in that: The preprocessing process in step 2 is as follows: 2-1. Downsample the EEG data recorded at 1024 Hz to a sampling rate of 256 Hz, then filter out noise and artifacts related to myoelectric activity using a fifth-order Butterworth bandpass filter with a frequency range of 1 Hz to 70 Hz, extract the state characteristics of the EEG signal in different frequency bands, and use short-time Fourier transform to calculate the differential entropy of the EEG signal as the sample matrix; 2-2. Perform short-time Fourier transform on EEG data with a time window of 4 seconds and no overlap to extract differential entropy features As shown in formula (12); In formula (12), is the input sample matrix, is the element in the input sample matrix; is the probability density function, and the processed EEG data is represented by X′; Updated differential entropy features As shown in (13); In formula (13), is the standard deviation of the probability density function; is the expectation of the probability density function.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: The memory stores a computer program; the processor executes the EEG speech imagery robust recognition method with adjacent trial task aliasing measure as described in any one of claims 1-8.

10. A readable storage medium storing a computer program; characterized in that: When the computer program is executed by a processor, it is used to implement the EEG speech imagery robust recognition method with adjacent trial task aliasing measure as described in any one of claims 1 to 8.

Images