Electroencephalogram feature optimization extraction method based on deep learning model

Abstract

The invention discloses an electroencephalogram feature optimization extraction method based on a deep learning model, relates to the technical field of electroencephalogram analysis, and improves the emotion recognition precision. According to the method, detection signal fragments of different standard emotion electroencephalogram data are mutually matched to obtain characteristic emotion signal fragments, a plurality of characteristic fragments are traversed in each standard emotion electroencephalogram data through the characteristic emotion signal fragments, and the space-time correlation weight of each characteristic fragment relative to the corresponding standard emotion electroencephalogram data is obtained; establishing a multi-dimensional feature space-time matrix under various emotional stimulations, establishing emotional association trees corresponding to the multi-dimensional feature space-time matrix, and sequentially matching and connecting the emotional association trees corresponding to different emotional stimulations to obtain a full-emotional feature forest; and converting newly generated standard emotion electroencephalogram data into a detection feature sub-matrix sequence, inputting the detection feature sub-matrix sequence into the full-emotion feature forest, and further outputting an emotion proportion associated with the corresponding standard emotion electroencephalogram data.

Description

Technical Field

[0001] The present invention relates to the technical field of electroencephalogram analysis, and specifically to a method for optimizing the extraction of electroencephalogram features based on a deep learning model. Background Art

[0002] Emotion recognition is of great significance in multiple fields, such as mental health monitoring, intelligent human-computer interaction, market research, etc. Traditional emotion recognition methods are mainly based on external manifestations such as facial expressions and speech intonation. There are certain limitations in traditional methods. For example, they are easily affected by factors such as subjective disguise and environmental interference, resulting in low recognition accuracy.

[0003] With the development of electroencephalogram technology, using electroencephalogram (EEG) signals for emotion recognition has become a research hotspot. Electroencephalogram signals can directly reflect the electrical activities of the brain and more accurately reflect the emotional state of an individual. However, current emotion recognition based on EEG signals still faces many challenges. On the one hand, the original EEG signals contain a large amount of noise and interference. How to effectively preprocess them to extract valuable information is a key issue.

[0004] On the other hand, EEG signals have complex spatio-temporal characteristics. How to accurately capture and analyze these characteristics and establish an effective emotion recognition model is an important problem in improving the accuracy and reliability of emotion recognition. Existing methods have deficiencies in dealing with the spatio-temporal characteristics of EEG signals, resulting in the need to improve the effect of emotion recognition. Therefore, a method for optimizing the extraction of electroencephalogram features based on a deep learning model is provided. Summary of the Invention

[0005] In order to solve the above technical problems, the purpose of the present invention is to provide a method for optimizing the extraction of electroencephalogram features based on a deep learning model.

[0006] In order to achieve the above purpose, the present invention provides the following technical solutions:

[0007] A method for optimizing the extraction of electroencephalogram features based on a deep learning model, comprising the following steps:

[0008] Step S1: Set up an electroencephalogram acquisition channel to collect a number of original emotion electroencephalogram signals under various emotion stimuli, and after standardizing the original emotion electroencephalogram signals, obtain standard emotion electroencephalogram data;

[0009] Step S2: Set up a feature extraction network. The feature extraction network is provided with multiple feature extraction kernels. Then, randomly select multiple initial feature extraction points in each standard emotion electroencephalogram data. Then, the feature extraction kernels spread from the initial feature extraction points to both sides to extract detection signal segments, and match the detection signal segments of different standard emotion electroencephalogram data to generate emotion signal segments;

[0010] Step S3: Denote each emotional signal segment as a characteristic emotional signal segment, traverse several characteristic segments in each standard emotional EEG data through the characteristic emotional signal segments, obtain the spatio-temporal correlation weights of each characteristic segment relative to the standard emotional EEG data where it is located, and then establish a multi-dimensional characteristic spatio-temporal matrix under various emotional stimuli;

[0011] Step S4: Establish an emotional correlation tree corresponding to the multi-dimensional characteristic spatio-temporal matrix, then sequentially match and connect the emotional correlation trees corresponding to different emotional stimuli to obtain a full-emotion characteristic forest, convert the latest generated standard emotional EEG data into a sequence of detection characteristic sub-matrices and input them into the full-emotion characteristic forest, and then output the emotional proportion associated with the corresponding standard emotional EEG data.

[0012] Further, the acquisition process of the original emotional EEG signals includes:

[0013] Install a 32-channel EEG acquisition device on the heads of multiple subjects, set the signal acquisition frequency of the 32-channel EEG acquisition device, with a bandwidth of 0.5 - 45 Hz, and then collect the original emotional EEG signals of the subjects under different emotional stimuli through the 32-channel EEG acquisition device.

[0014] Further, the process of standardizing the original emotional EEG signals includes:

[0015] Set m time nodes, obtain the mean and standard deviation of each original emotional EEG signal, and then perform standard processing on each original emotional EEG signal through the Z-score standardization method. Connect the standardized values at each time node in chronological order to obtain the standard emotional EEG data corresponding to the original emotional EEG signal, where m is a natural number greater than 100.

[0016] Further, the extraction process of the detection signal segments includes:

[0017] The feature extraction network is set with n feature extraction kernels. Input each standard emotional EEG data into the feature extraction network at the same time. Then the feature extraction network randomly selects n initial feature extraction points in the standard emotional EEG data, where n is a natural number greater than 50;

[0018] Set the unit diffusion length, make each feature extraction kernel attach to the initial feature extraction point in turn, and then start from both ends of each initial feature extraction point at the same time. Each feature extraction kernel intercepts the detection signal segments with the unit diffusion length from both ends with the initial feature extraction point it attaches to as the center.

[0019] Further, the process of matching the detection signal segments of different standard emotional EEG data includes:

[0020] Match the detection signal segments of the standard emotion EEG data under different emotion stimuli with each other. If the similarity between two detection signal segments is above 90%, then eliminate the corresponding initial feature extraction points of the two detection signal segments, and randomly set a new initial feature extraction point on the corresponding standard emotion EEG data;

[0021] If the similarity between two detection signal segments is 90% or below, then retain the two detection signal segments and continue with the next matching of detection signal segments;

[0022] After all the detection signal segments have completed mutual matching, match the detection signal segments that correspond to the same emotion stimulus but come from different standard emotion EEG data among the retained detection signal segments, and set a similarity threshold;

[0023] If the similarity between two detection signal segments is greater than or equal to the similarity threshold, then set the same feature annotation for the two detection signal segments, otherwise do nothing;

[0024] Eliminate the positions of the initial feature extraction points corresponding to each detection signal segment without a feature annotation, and randomly select a new feature extraction point. For the detection signal segments with a feature annotation, eliminate the positions of the corresponding initial feature extraction points, and extract a detection signal segment with a two - times unit diffusion length centered on the initial feature extraction point;

[0025] First, for different standard emotion EEG data under different emotion stimuli, if there is a detection signal segment with a length of two - times unit diffusion length or more, and in the secondary matching result, there is a shorter detection signal segment with a similarity above 90%, then eliminate the initial feature extraction point corresponding to the shorter detection signal segment and reset the initial feature extraction point;

[0026] If in the secondary matching result, the similarity with an equal - length or longer detection signal segment is above 90%, then roll back the two corresponding detection signal segments to a detection signal segment with a one - times unit diffusion length, and record the rolled - back detection signal segment as an emotion signal segment;

[0027] Repeat the previous detection signal segment matching process for the detection signal segment with a one - times unit diffusion length. For the subsequent detection signal segments with three, four, …, i times unit diffusion lengths, adopt the matching process for the detection signal segment with a two - times unit diffusion length, where i is a natural number greater than 3.

[0028] Further, whenever a new initial feature extraction point falls within the already recorded emotional signal segment, and with the initial feature extraction point as the center, a detection signal segment with a unit diffusion length is completely within the emotional signal segment, then one initial feature extraction point is permanently reduced from the corresponding standard emotional EEG data. Repeat the above detection signal segment matching process until there are no initial feature extraction points in the standard emotional EEG data.

[0029] Further, the process of establishing the multi-dimensional feature spatio-temporal matrix includes:

[0030] When the matching between the standard emotional EEG data in step S2 is completed, mark the emotional signal segments in each standard emotional EEG data as characteristic emotional signal segments, and match the characteristic emotional signal segments with each other in the standard emotional EEG data under the same emotional stimulus. Piecewise fuse the characteristic emotional signal segments with a similarity greater than or equal to the similarity threshold.

[0031] Then, traverse each standard emotional EEG data through the characteristic emotional signal segments, and mark several characteristic segments in each standard emotional EEG data according to the traversal results.

[0032] Obtain the query matrix, key matrix, and value matrix of each characteristic segment through the multi-head self-attention mechanism method. Denote the entire standard emotional EEG data as the brain region, and then obtain the spatio-temporal correlation weight between the corresponding characteristic segment and the brain region according to the query matrix, key matrix, and value matrix.

[0033] Integrate the spatio-temporal correlation weights of the same kind of characteristic segments at each occurrence time sequence to generate the corresponding time sequence weight matrix. At the same time, select the occurrence time sequence when the spatio-temporal correlation weight of each characteristic segment is the largest, and overlap each characteristic segment in turn and transform it into the corresponding multi-dimensional feature spatio-temporal matrix.

[0034] Further, the process of establishing the full-emotion feature forest:

[0035] Divide the multi-dimensional feature spatio-temporal matrix into several feature sub-matrices, set a feature tree node for each feature sub-matrix, and input each characteristic segment and the corresponding time sequence weight matrix into the feature tree node.

[0036] At the same time, according to the spatio-temporal positions of each feature sub-matrix in the multi-dimensional feature spatio-temporal matrix, connect each feature tree node in turn to obtain the emotion correlation tree corresponding to the corresponding type of emotion stimulus.

[0037] Match the feature tree nodes in the emotion correlation trees under different types of emotion stimuli with each other, and set the correlation similarity threshold. It should be noted that the correlation similarity threshold is less than 0.9.

[0038] If the similarity between the feature fragments in two feature tree nodes is greater than or equal to the associated similarity threshold, the corresponding two feature tree nodes are connected to each other, otherwise no operation is performed;

[0039] Repeat the above feature tree node matching and connection process to obtain the full emotion feature forest.

[0040] Furthermore, the process of obtaining the emotion proportion of the standard emotion EEG data to be analyzed includes:

[0041] Obtain standard emotional EEG data to be analyzed, traverse the characteristic fragments under various emotional stimuli to the standard emotional EEG data, and obtain characteristic fragment sequences under multiple emotions according to the traversal results;

[0042] Convert each feature fragment sequence into a detection feature submatrix sequence, annotate the detection feature submatrix with the corresponding occurrence time sequence, and input the detection feature submatrix sequence into the full emotion feature forest;

[0043] Starting from the first detection feature submatrix according to the appearance time sequence, it is synchronously compared with each emotion association tree in the full emotion feature forest in turn, and then the appearance time sequence of each detection feature submatrix is ​​counted to retrieve the corresponding time sequence weight matrix, and the time sequence weight value of each detection feature submatrix in different emotion association trees is obtained;

[0044] The total values ​​of the temporal weights of the feature submatrix sequences in different emotion association trees are statistically detected, the total values ​​of the temporal weights of different emotion association trees are accumulated, and the proportions of various emotions are obtained.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. The present invention extracts detection signal fragments by diffusing from the initial feature extraction point to both sides through the feature extraction kernel, matches the detection signal fragments of different standard emotional EEG data with each other to generate emotional signal fragments, which helps to explore the potential emotional features in EEG signals and improves the accuracy and effectiveness of feature extraction.

[0047] 2. By obtaining the spatiotemporal correlation weights of feature fragments and establishing a multi-dimensional feature spatiotemporal matrix, it is possible to comprehensively analyze emotional information in the time and space dimensions while considering the spatiotemporal characteristics of EEG signals, more accurately describe the electrical activity patterns of the brain under emotional states, and improve the accuracy of emotion recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention.

[0049] Figure 1 This is the flowchart of the method of the present invention. Specific embodiments

[0050] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be described in detail below. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope protected by the present invention.

[0051] As Figure 1 shown, the method for optimizing the extraction of electroencephalogram features based on a deep learning model includes the following steps:

[0052] Step S1: Set up electroencephalogram acquisition channels to collect several original emotional electroencephalogram signals under various emotional stimuli, and after standardizing the original emotional electroencephalogram signals, obtain standard emotional electroencephalogram data;

[0053] Step S2: Set up a feature extraction network, the feature extraction network is provided with multiple feature extraction kernels, and then randomly select multiple initial feature extraction points in each standard emotional electroencephalogram data. Then, the feature extraction kernels spread to both sides from the initial feature extraction points to extract detection signal segments, and match the detection signal segments of different standard emotional electroencephalogram data with each other to generate emotional signal segments;

[0054] Step S3: Denote each emotional signal segment as a feature emotional signal segment, traverse several feature segments in each standard emotional electroencephalogram data through the feature emotional signal segment, obtain the spatio-temporal correlation weight of each feature segment relative to the standard emotional electroencephalogram data where it is located, and then establish a multi-dimensional feature spatio-temporal matrix under various emotional stimuli;

[0055] Step S4: Establish an emotional correlation tree corresponding to the multi-dimensional feature spatio-temporal matrix, and then match and connect the emotional correlation trees corresponding to different emotional stimuli in sequence to obtain a full-emotion feature forest. Convert the newly generated standard emotional electroencephalogram data into a detection feature sub-matrix sequence and input it into the full-emotion feature forest, and then output the emotional proportion associated with the corresponding standard emotional electroencephalogram data.

[0056] Further, the step S1 is implemented through the following process:

[0057] Install a 32-channel electroencephalogram acquisition device on the heads of multiple subjects, and the coverage range of the 32-channel electroencephalogram acquisition device includes the prefrontal lobe, parietal lobe, temporal lobe and occipital lobe regions of the subjects' heads;

[0058] Set the signal acquisition frequency for the 32-channel EEG acquisition device, with a bandwidth of 0.5 - 45 Hz. Then, use the 32-channel EEG acquisition device to collect the original emotional EEG signals of the subjects under different emotional stimuli, and each emotional stimulus lasts for about 10 minutes;

[0059] Set a signal normalization mechanism at the signal output position of each channel of the 32-channel EEG acquisition device to convert the original emotional EEG signals into corresponding standard emotional EEG data. The specific process includes:

[0060] Set m time nodes, obtain the mean and standard deviation of each original emotional EEG signal, and then perform standard processing on each original emotional EEG signal through the Z-score normalization method. The processing formula: Where Z i 、X i represents the normalized value and the original value of the original emotional EEG signal at the i-th time node, μ is the mean, σ is the standard deviation, and m is a natural number greater than 100;

[0061] Connect the normalized values at each time node in sequence according to the time order to obtain the standard emotional EEG data corresponding to the original emotional EEG signal.

[0062] Furthermore, the step S2 is implemented through the following process:

[0063] The feature extraction network is set with n feature extraction kernels. Input each standard emotional EEG data into the feature extraction network at the same time. Then, the feature extraction network randomly selects n initial feature extraction points from the standard emotional EEG data, where n is a natural number greater than 50;

[0064] Set the unit diffusion length, and let each feature extraction kernel attach to the initial feature extraction point in turn. Then, starting from both ends of each initial feature extraction point at the same time, each feature extraction kernel intercepts the detection signal segment of the unit diffusion length centered on the initial feature extraction point to which it attaches;

[0065] Match the detection signal segments of the standard emotional EEG data under different emotional stimuli. If the similarity between two detection signal segments is above 90%, then eliminate the initial feature extraction points corresponding to the two detection signal segments, and randomly set a new initial feature extraction point on the corresponding standard emotional EEG data;

[0066] If the similarity between two detection signal segments is 90% or less, then retain the two detection signal segments and continue the next detection signal segment matching;

[0067] After all the detected signal segments are completely matched with each other, the detected signal segments corresponding to the same emotional stimulus but from different standard emotional EEG data in the retained detected signal segments are matched with each other, and a similarity threshold is set.

[0068] If the similarity between two detected signal segments is greater than or equal to the similarity threshold, the same feature annotation is set for the two detected signal segments; otherwise, no operation is performed.

[0069] The positions of the initial feature extraction points corresponding to each detected signal segment without feature annotation are removed, and a new feature extraction point is randomly selected again. For the detected signal segments with feature annotation, the positions of the initial feature extraction points are removed, and the detected signal segments with a double unit diffusion length are extracted centered on the initial feature extraction points.

[0070] First, for different standard emotional EEG data under different emotional stimuli, if there is a detected signal segment shorter than the detected signal segment with a double unit diffusion length or more, and the similarity of the shorter detected signal segment in the secondary matching result is above 90%, the initial feature extraction point corresponding to the shorter detected signal segment is removed, and a new initial feature extraction point is set.

[0071] If the similarity with a detected signal segment of equal length or longer in the secondary matching result is above 90%, the two corresponding detected signal segments are reverted to a detected signal segment with a single unit diffusion length, and the reverted detected signal segment is recorded as an emotional signal segment.

[0072] The matching process of the previous detected signal segments is repeated for the detected signal segments with a single unit diffusion length. For the detected signal segments with three, four,..., i unit diffusion lengths subsequently, the matching process of the detected signal segments with a double unit diffusion length is adopted, where i is a natural number greater than 3.

[0073] It should be noted that whenever a new initial feature extraction point falls within the emotional signal segment that has been recorded, and the detected signal segment with a single unit diffusion length centered on the initial feature extraction point is completely within the emotional signal segment, one initial feature extraction point is permanently reduced on the corresponding standard emotional EEG data.

[0074] Repeat the above matching process of the detected signal segments until there are no initial feature extraction points in the standard emotional EEG data.

[0075] Furthermore, the step S3 is implemented through the following process:

[0076] After the matching among the standard EEG data of various emotions is completed in step S2, the emotional signal segments in the standard EEG data of various emotions are recorded as characteristic emotional signal segments, and the characteristic emotional signal segments are matched with each other in the standard EEG data under the same emotional stimulus, and the characteristic emotional signal segments with a similarity greater than or equal to the similarity threshold are spliced and fused;

[0077] Furthermore, traverse each standard EEG data through the characteristic emotional signal segments, and mark a number of characteristic segments in each standard EEG data according to the traversal results;

[0078] Obtain the query matrix, key matrix and value matrix of each characteristic segment through the multi-head self-attention mechanism method, record the entire standard EEG data as the brain region, and then obtain the spatio-temporal correlation weight between the corresponding characteristic segment and the brain region according to the query matrix, key matrix and value matrix;

[0079] The calculation formula of the spatio-temporal correlation weight is:

[0080]

[0081] where A, B, and C respectively represent the query matrix, key matrix, and value matrix of the characteristic segment, represents the correction parameter, and t represents the occurrence time sequence of the characteristic segment in the entire standard EEG data;

[0082] Integrate the spatio-temporal correlation weights of the same kind of characteristic segments at each occurrence time sequence to generate the corresponding time sequence weight matrix, and at the same time select the occurrence time sequence when the spatio-temporal correlation weight of each characteristic segment is the largest, and overlap each characteristic segment in turn and convert it into the corresponding multi-dimensional characteristic spatio-temporal matrix.

[0083] Further, step S4 is implemented through the following process:

[0084] Divide the multi-dimensional characteristic spatio-temporal matrix into several characteristic sub-matrices, set a characteristic tree node for each characteristic sub-matrix, and input each characteristic segment and the corresponding time sequence weight matrix into the characteristic tree node;

[0085] At the same time, according to the spatio-temporal positions of each characteristic sub-matrix in the multi-dimensional characteristic spatio-temporal matrix, connect each characteristic tree node in turn to obtain the emotion correlation tree under the corresponding type of emotion stimulus;

[0086] Match the characteristic tree nodes in the emotion correlation trees under different types of emotion stimuli, set the correlation similarity threshold, it should be noted that the correlation similarity threshold is less than 0.9;

[0087] If the similarity between the feature segments in two feature tree nodes is greater than or equal to the associated similarity threshold, then connect the corresponding two feature tree nodes to each other; otherwise, do nothing.

[0088] Repeat the above process of matching and connecting feature tree nodes to obtain the full emotion feature forest.

[0089] Furthermore, obtain the standard emotion EEG data to be analyzed, traverse the feature segments under various emotion stimuli against the standard emotion EEG data, and obtain the feature segment sequences under multiple emotions according to the traversal results.

[0090] Convert each feature segment sequence into a sequence of detection feature sub-matrices, mark the corresponding occurrence time sequence for the detection feature sub-matrices, and input the sequence of detection feature sub-matrices into the full emotion feature forest.

[0091] Starting from the first detection feature sub-matrix according to the occurrence time sequence, synchronously compare with each emotion association tree in the full emotion feature forest in turn, and then statistically retrieve the corresponding time sequence weight matrix according to the occurrence time sequence of each detection feature sub-matrix to obtain the time sequence weight values of each detection feature sub-matrix in different emotion association trees.

[0092] Statistically calculate the total time sequence weight values of the sequence of detection feature sub-matrices in different emotion association trees, accumulate the total time sequence weight values of different emotion association trees, and obtain the proportion of each emotion.

[0093] The above embodiments are only used to illustrate the technical method of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical method of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical method of the present invention.

Claims

1. An optimized extraction method for electroencephalogram features based on a deep learning model, characterized in that Including the following steps: Step S1: Set up EEG acquisition channels to collect a number of original emotional EEG signals under various emotional stimuli, and after performing standardization processing on the original emotional EEG signals, obtain standard emotional EEG data; Step S2: Set up a feature extraction network, the feature extraction network is provided with multiple feature extraction kernels, randomly select multiple initial feature extraction points in each standard emotional EEG data, and then the feature extraction kernels spread from the initial feature extraction points to both sides to extract detection signal segments, and match the detection signal segments of different standard emotional EEG data with each other, and then generate emotional signal segments; Step S3: Denote each emotional signal segment as a characteristic emotional signal segment, traverse a number of characteristic segments in each standard emotional EEG data through the characteristic emotional signal segment, obtain the spatio-temporal correlation weight of each characteristic segment relative to the standard emotional EEG data where it is located, and then establish a multi-dimensional characteristic spatio-temporal matrix under various emotional stimuli; Step S4: Establish an emotional correlation tree corresponding to the multi-dimensional characteristic spatio-temporal matrix, and then match and connect the emotional correlation trees corresponding to different emotional stimuli in sequence to obtain a full-emotion characteristic forest, convert the newly generated standard emotional EEG data into a sequence of detection characteristic sub-matrices and input them into the full-emotion characteristic forest, and then output the emotional proportion associated with the corresponding standard emotional EEG data.

2. The method for optimizing the extraction of electroencephalogram features based on a deep learning model according to claim 1, wherein, The acquisition process of the original emotional EEG signal includes: Install 32-channel EEG acquisition devices on the heads of multiple subjects, and then collect the original emotional EEG signals of the subjects under different emotional stimuli through the 32-channel EEG acquisition devices.

3. The method for optimizing the extraction of electroencephalogram features based on a deep learning model according to claim 2, wherein The process of performing standardization processing on the original emotional EEG signal includes: Set a number of time nodes, obtain the mean and standard deviation of each original emotional EEG signal, and then perform standard processing on each original emotional EEG signal through the Z-score standardization method, and connect the standardized values at each time node in sequence to obtain the standard emotional EEG data.

4. The method for optimizing the extraction of electroencephalogram features based on a deep learning model according to claim 3, wherein The extraction process of the detection signal segment includes: The feature extraction network is provided with n feature extraction kernels, input each standard emotional EEG data into the feature extraction network at the same time, and then the feature extraction network randomly selects n initial feature extraction points in the standard emotional EEG data, where n is a natural number greater than 50; Set the unit diffusion length, make each feature extraction kernel attach to the initial feature extraction point in turn, and each feature extraction kernel intercepts the detection signal segment of the unit diffusion length from both ends with the initial feature extraction point to which it attaches as the center.

5. The method for optimally extracting electroencephalogram features based on a deep learning model according to claim 4, wherein The process of matching the detection signal segments of different standard emotional EEG data with each other includes: Match the detection signal segments of the standard emotional EEG data under different emotional stimuli with each other, remove the initial feature extraction points corresponding to the two detection signal segments according to the matching result, and randomly set a new initial feature extraction point on the corresponding standard emotional EEG data, or keep the two detection signal segments and continue the next matching of the detection signal segments; When all the detection signal segments have completed mutual matching, match the detection signal segments corresponding to the same emotional stimulus but from different standard emotional EEG data among the remaining detection signal segments, and set a similarity threshold. If the similarity between two detected signal segments is greater than or equal to the similarity threshold, the same feature annotation is set for the two detected signal segments; otherwise, no operation is performed. The initial feature extraction point positions corresponding to the detected signal segments without feature annotations are removed, and a new feature extraction point is randomly selected. For the detected signal segments with feature annotations, the initial feature extraction point positions are removed, and a detected signal segment with a two-fold unit diffusion length is extracted centered on the initial feature extraction point. If the similarity between the detected signal segment and an equal-length or longer detected signal segment in the secondary matching result is above 90%, the corresponding two detected signal segments are rolled back to a detected signal segment with a unit diffusion length, and the rolled-back detected signal segment is recorded as an emotional signal segment.

6. The method for optimizing the extraction of electroencephalogram features based on a deep learning model according to claim 5, wherein Whenever a new initial feature extraction point falls within the emotional signal segment that has been recorded, and the detected signal segment with a unit diffusion length centered on the initial feature extraction point is completely within the emotional signal segment, one initial feature extraction point is permanently reduced on the corresponding standard emotional EEG data. The above-mentioned detected signal segment matching process is repeated until there are no initial feature extraction points in the standard emotional EEG data.

7. The method for optimizing the extraction of electroencephalogram features based on a deep learning model according to claim 6, wherein The process of establishing a multi-dimensional feature spatio-temporal matrix includes: The emotional signal segments in each standard emotional EEG data are recorded as feature emotional signal segments, and each standard emotional EEG data is traversed through the feature emotional signal segments. Based on the traversal results, several feature segments are marked on each standard emotional EEG data. The query matrix, key matrix, and value matrix of each feature segment are obtained through the multi-head self-attention mechanism method. The entire standard emotional EEG data is recorded as a brain region, and then the spatio-temporal correlation weights between the corresponding feature segments and the brain region are obtained based on the query matrix, key matrix, and value matrix. The spatio-temporal correlation weights of the same type of feature segments at each appearance time sequence are integrated to generate the corresponding time sequence weight matrix. At the same time, the appearance time sequence when the spatio-temporal correlation weight of each feature segment is the largest is selected, and each feature segment is overlapped in turn and transformed into the corresponding multi-dimensional feature spatio-temporal matrix.

8. The method for optimally extracting electroencephalogram features based on a deep learning model according to claim 7, wherein The process of establishing a full-emotion feature forest: The multi-dimensional feature spatio-temporal matrix is divided into several feature sub-matrices, and a feature tree node is set for each feature sub-matrix. At the same time, according to the spatio-temporal positions of the feature sub-matrices in the multi-dimensional feature spatio-temporal matrix, the feature tree nodes are connected in sequence to obtain an emotion correlation tree corresponding to the emotional stimuli of the corresponding type. The feature tree nodes in the emotion correlation trees under different types of emotional stimuli are matched with each other. According to the matching results, the feature tree nodes are connected to each other, and then a full-emotion feature forest is obtained.

9. The method for optimizing the extraction of electroencephalogram features based on a deep learning model according to claim 8, wherein, The process of obtaining the emotion proportion of the standard emotional EEG data to be analyzed includes: The standard emotional EEG data to be analyzed is obtained, and the feature segments under various emotional stimuli are used to traverse the standard emotional EEG data. According to the traversal results, the feature segment sequences under multiple emotions are obtained. Each feature segment sequence is transformed into a sequence of detection feature sub-matrices, and the corresponding appearance time sequences are marked for the detection feature sub-matrices. The sequence of detection feature sub-matrices is input into the full-emotion feature forest. Starting from the first detected feature sub-matrix according to the occurrence time sequence, sequentially perform synchronous comparison with each emotion association tree in the full emotion feature forest, and then statistically retrieve the corresponding time sequence weight matrix according to the occurrence time sequence of each detected feature sub-matrix, and obtain the time sequence weight values of each detected feature sub-matrix in different emotion association trees; Statistically calculate the total time sequence weight values of the detected feature sub-matrix sequence in different emotion association trees, accumulate the total time sequence weight values of different emotion association trees, and obtain the proportion of each emotion.

Images