An imagined movement decoding method based on tactile sensation calibration

Abstract

The present invention discloses a method for decoding imagined movements based on tactile sensation calibration, comprising: (1) setting up a tactile stimulation device and applying tactile stimulation to the dorsal side of the left or right wrist of the subject; (2) setting up an EEG device to record EEG signals, with the original signals being digitally sampled at a frequency of 500 Hz and using a filter to filter the original signals; (3) using the tactile stimulation device and the EEG device to collect the pure imagined movement data of the subject, as well as the imagined movement data assisted by tactile sensation; (4) using the common spatial pattern algorithm to extract the EEG data features and training a classifier for classifying the EEG data features; wherein, the imagined movement data assisted by tactile sensation is used as the training set; (5) applying the trained classifier to pure imagined movement decoding. By using the present invention, EEG data with a higher signal-to-noise ratio can be obtained, and then a more robust and accurate model can be trained for movement decoding.

Description

Technical Field

[0001] The present invention relates to the field of decoding of imagined movement brain-computer interfaces, and more particularly to an imagined movement decoding method based on tactile sensation calibration. Background Art

[0002] Brain-computer interfaces directly decode motor intentions from electroencephalogram (EEG) signals for external device control. Brain-computer interfaces can help stroke patients communicate with the outside world and control prosthetics, and are used for neurorehabilitation training using neuroplasticity. Since the concept of brain-computer interfaces was proposed in the 1970s, various brain-computer interface systems have been developed. Non-invasive brain-computer interfaces based on EEG electroencephalogram have received extensive attention due to their high safety. Brain-computer interfaces based on EEG electroencephalogram are mainly divided into three types: brain-computer interfaces based on visually evoked P300 signals, brain-computer interfaces based on steady-state visually evoked potentials, and brain-computer interfaces based on motor imagery rhythms. The first two types of brain-computer interfaces both require external visual stimuli, occupying the visual channel. While the brain-computer interface based on imagined movement is completely generated by the user's inner spontaneous activity and does not affect visual interaction, which is very important for people who have lost their visual ability.

[0003] However, the brain-computer interface based on motor imagery still has problems such as low decoding accuracy and long calibration process time. In recent years, many calibration methods have been developed to improve the decoding performance of imagined movement. Such as common spatial pattern algorithms and popular deep learning algorithms. However, due to the non-stationarity and low signal-to-noise ratio of EEG data, the performance improvement of these algorithms is limited. Another problem that hinders the wide application of the brain-computer interface based on imagined movement is that the calibration stage of the brain-computer interface requires a long time. EEG signals are non-stationary, and users need a certain amount of time to re-collect training data to train the model, otherwise the decoding accuracy will decrease over time.

[0004] In order to make brain-computer interfaces widely used, how to decode more accurately and calibrate more quickly are two urgent problems to be solved. The brain activation patterns induced by active movement, passive movement, functional electrical stimulation, and imagined movement are similar, and based on this, new strategies for calibrating imagined movement brain-computer interfaces have been developed.

[0005] Kaiser's research shows that robot-assisted passive movement can be used to decode imagined movement, which is very meaningful for brain-computer interfaces in stroke rehabilitation. In Vidaurre's work, the brain patterns induced by neuromuscular electrical stimulation were used to decode imagined movement, and offline analysis showed that this method can support the calibration of brain-computer interface systems.

[0006] Both of the above two methods generate proprioceptive information transmitted to the brain after stimulating the corresponding nerve fibers, thereby generating a recognized brain pattern to detect the movement intention. However, passive movement and electrical stimulation are completely different tasks from real imagined movement, and a large difference between the training set and the test set will lead to a decrease in the decoding accuracy.

[0007] Therefore, there is an urgent need to design an auxiliary training and decoding method that can maintain the consistency with pure imagined movement to the greatest extent. Summary of the Invention

[0008] The present invention provides an imagined movement decoding method based on tactile sensation calibration. By applying tactile stimulation to assist imagined movement in the imagined movement task during the calibration stage to enhance brain movement-related activation, EEG data with a higher signal-to-noise ratio can be obtained, and then a more robust and accurate model can be trained for movement decoding.

[0009] An imagined movement decoding method based on tactile sensation calibration includes the following steps:

[0010] (1) Set up a tactile stimulation device, apply tactile stimulation to the dorsal side of the left or right wrist of the subject, and use a linear resonant actuator to generate tactile stimulation with a frequency of 27 Hz.

[0011] (2) Set up an EEG device to record EEG signals. During the entire recording process, the electrode impedance is lower than 10 KΩ, the original signal is digitally sampled at a frequency of 500 Hz, and the original signal is filtered using a bandwidth filter between 0.5 - 70 Hz and a notch filter of 48 - 52 Hz.

[0012] (3) Use the tactile stimulation device and the EEG device to collect the pure imagined movement data of the subject and the imagined movement data assisted by tactile sensation.

[0013] (4) Use the common spatial pattern algorithm to extract the EEG data features and train a classifier for classifying the EEG data features; among them, the imagined movement data assisted by tactile sensation is used as the training set.

[0014] (5) Use the trained classifier for pure imagined movement decoding.

[0015] The present invention proposes to apply tactile sensation calibration in a brain-computer interface based on imagined movement, use an imagined movement task with tactile sensation to train a model to decode a pure imagined movement task. The tactile sensation increases the brain movement-related activation of the imagined movement task, and EEG data with a higher signal-to-noise ratio is obtained, and then a more robust and accurate model is trained to decode the pure imagined movement task.

[0016] In step (1), a linear resonant actuator (10 mm, C10-100, Precision Microdrives td., with a normalised amplitude of 1.4 G typically) is used to generate tactile stimuli. The frequency of the stimuli is 27 Hz, which is generated by a 27 Hz sine wave with a 175 Hz sine wave as the carrier. Such stimuli can generate the sense of touch because they activate both Pacinian and Meissner corpuscles simultaneously.

[0017] Preferably, the linear resonant actuator uses a piezoelectric ceramic stimulator.

[0018] In step (2), a 64-channel wireless NeuSenW64 EEG system (Neuracle, Changzhou, China) is used to record EEG signals; an EEG cap with 64 channels is used. The electrodes are arranged according to the 10-20 international system; the reference electrode is CPz and the ground electrode is AFz.

[0019] In step (3), when collecting the pure imagined movement data of the subject, the imagined movement paradigm is designed as follows:

[0020] Throughout the process, the subject is required to sit quietly on a chair facing a computer screen; the subject places their hands on the armrests while keeping their body still, without generating facial and hand muscle movements, and not blinking during each task;

[0021] The tactile stimulator is wrapped around the subject's wrist and placed on the dorsal outer side of the wrist; the screen background is black, and at the start of each trial, t = 0 s, a white cross appears in the centre of the screen and the subject needs to fixate on the cross; at t = 2 s, a tactile stimulus with a duration of 200 milliseconds acts on both hands simultaneously to remind the subject that the task is about to start; at t = 3 s, a red cue bar is superimposed on the white cross, which randomly appears on the left or right; when the red cue bar appears on the left, the subject needs to perform the left hand imagined movement task; when the red cue bar appears on the right, the subject needs to perform the right hand imagined movement task; the duration of the red cue bar is 1.5 s, i.e., the red cue bar disappears at t = 4.5 s; from the appearance of the red cue bar, the subject continuously performs the imagined movement task for 5 s until the white cross disappears at t = 8 s; finally, there is a random period of time between 1.5 and 3.5 s for the subject to rest and to limit the subject's adaptation effect.

[0022] When collecting the imagined movement data of the subject assisted by tactile sensation, the tactile sensation paradigm is designed as follows:

[0023] The tactile stimulator is wrapped around the subject's wrist and placed on the dorsal lateral side of the wrist; the screen background is black. At the start of each trial, t = 0 s, a white cross appears in the center of the screen, and the subject is required to fixate on the cross. At t = 2 s, a tactile stimulus with a duration of 200 milliseconds acts on both hands simultaneously to remind the subject that the task is about to start. At t = 3 s, a red cue bar is superimposed on the white cross, and it randomly appears on the left or right side. When the red cue bar appears on the left side, the subject performs a left-hand imagination movement task assisted by tactile stimulation, where the subject imagines the movement of the left hand, focuses on the left hand, and then feels the tactile stimulation on the left hand. When the red cue bar appears on the right side, the subject needs to perform a right-hand imagination movement task assisted by stimulation, where the subject imagines the movement of the right hand, focuses on the right hand, and feels the tactile stimulation on the right hand. The duration of the red cue bar is 1.5 s, that is, the red cue bar disappears at t = 4.5 s.

[0024] From the appearance of the red cue bar, the actuators of both the left and right hands vibrate, and the tactile stimulation lasts for 5 s. During this process, the subject continuously performs the imagination movement task for 5 s until the white cross disappears at t = 8 s; finally, there is a random period of time between 1.5 and 3.5 s for the subject to rest and limit the subject's adaptation effect.

[0025] In step (4), the specific process of extracting the EEG data features using the common spatial pattern algorithm is as follows:

[0026] Find the two most separable linear projection directions in a data-driven manner, in these directions, the difference information between the two classes is maximized and the common information is minimized. The difference information and common information between the two classes are respectively described as the difference S of the covariance d and S c :

[0027] S d = Σ (l) - Σ (r)

[0028] S c = Σ (l) + Σ (r)

[0029] Σ (l) represents the covariance of the left-hand imagination movement task, Σ (r) represents the covariance of the right-hand imagination movement task. In addition, the objective function is expressed as:

[0030]

[0031]

[0032] Through the Lagrange multiplier method, the objective function is transformed into a generalized eigenvalue problem:

[0033] S d w = λ d w

[0034] For the eigenvector w corresponding to the eigenvalue λ d it satisfies:

[0035] wΣ (l) w + wΣ (r) w = 1

[0036] In the direction of w, the sum of the variances of the two classes is 1; therefore, when the projection direction maximizes the variance of one class, it will simultaneously minimize the variance of the other class; after solving for w, the original data is transformed into:

[0037]

[0038] where, X ∈ R C×T , X CSP ∈ R C×T , W ∈ R C×C , the column vectors of W are generalized eigenvectors; X CPS is the projection of the original signal in the direction of the generalized eigenvector, X is the original signal, t represents the current time, C is the number of channels, and T is the total time length. Select the eigenvectors corresponding to the three largest eigenvalues and the three smallest eigenvalues as filters; in the first three filters, the variance of the left hand imagined movement task is larger, while the variance of the right hand imagined movement task is smaller, and in the last three filters, it is the opposite; finally, apply the logarithm to the variance of the filtered data to obtain six-dimensional features.

[0039] Optionally, the classifier uses the linear discriminant analysis algorithm. When the number of classes is 2, its basic idea is to find a projection direction that maximizes the distance between the two classes while minimizing the distance within each class; the distances between and within the two classes are defined as the between-class scatter matrix S B and the within-class scatter matrix S W :

[0040] S B = (μ1 - μ2)(μ1 - μ2) T

[0041] S W = (S1 + S2)

[0042] where, μ1 and μ2 represent the centers of the two classes respectively, and S1 and S2 represent the variances of the two classes respectively. The objective function of the linear discriminant analysis algorithm is expressed as:

[0043]

[0044]

[0045] By the Lagrange multiplier method, the objective function is transformed into the following generalized eigenvalue problem:

[0046] S B ω = λS W ω

[0047] Since S B ω = (μ1 - μ2)(μ1 - μ2) T ω, so S B ω and μ1 - μ2 have the same direction, that is: S B ω = k(μ1 - μ2) = λS W ω, where λ is a constant scaling factor. Therefore, the direction of ω is expressed as:

[0048]

[0049] The offset b of the classifier is:

[0050]

[0051] Finally, the separating surface between the two classes is determined by the following equation:

[0052] y = ω T x + b

[0053] Optionally, the classifier uses a support vector machine, and the optimization objective is as follows:

[0054]

[0055] s.t.y i (w T x i + b) ≥ 1 - ξ i

[0056]

[0057] where w is the normal vector of the decision boundary. In the case of linearly separable, the classification margin of the support vector machine is 2 / w T w, C is the penalty coefficient, ξ is the scaling factor, and b is the deviation of the decision boundary.

[0058] Compared with the prior art, the present invention has the following beneficial effects:

[0059] The present invention uses a data training model based on tactile sensations for decoding an imagined movement brain-computer interface; through the above calibration strategy, compared with the traditional imagined movement calibration strategy, it can significantly improve the decoding accuracy of the imagined movement brain-computer interface, and to a certain extent solve the problems of low accuracy and brain-computer interface blindness in the brain-computer interface system. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a schematic diagram comparing the decoding paradigms of tactile sensation calibration and traditional imagined movement calibration in an embodiment of the present invention;

[0061] Figure 2 It is a comparison diagram of the decoding effects of tactile sensation calibration and traditional imagined movement calibration in an embodiment of the present invention;

[0062] Figure 3 It is a comparison diagram of the decoding effects of tactile sensation calibration and traditional imagined movement calibration with the number of trials in an embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0063] The following further describes the present invention in detail with reference to the drawings and embodiments. It should be noted that the following embodiments are intended to facilitate the understanding of the present invention, but do not limit it in any way.

[0064] In an embodiment of the present invention, 14 subjects (8 males, 6 females, average age 23.6 ± 1.3 years) were recruited by posting invitations on the official forum of Zhejiang University. They were all exposed to the brain-computer interface experiment for the first time, right-handed, with normal or corrected vision, and in good physical and mental health. All subjects signed an informed consent form before participating. This study was approved by the Ethics Committee of Zhejiang University. All experimental procedures in this study were approved by the Ethics Committee of Zhejiang University, and each person participating in the experiment signed an informed consent form.

[0065] An imagined movement decoding method based on tactile sensation calibration includes the following steps:

[0066] (1) Acquisition of EEG brain electrical signals

[0067] EEG data was acquired from a 59-channel wireless NeuSenW64 EEG system (Neuracle, Changzhou, China) and a 64-channel EEG cap. The EEG cap was arranged according to the international extended 10 / 20 system, with the reference at CPz and the ground at AFz. Subjects were required to wash their hair within 12 hours before the experiment to keep the scalp clean. During the entire recording process, the impedance was kept below 10 KΩ. The original signal was digitally sampled at a frequency of 500 Hz.

[0068] (2) Tactile stimulation device

[0069] Tactile sensation was generated by continuous tactile stimulation produced by linear resonant actuators (10 mm, C10 - 100, Precision Microdrives Ltd., typical standardized amplitude 1.4G) placed on the dorsal sides of the left and right wrists of the subjects. The continuous tactile stimulation was a 27 Hz sine wave modulated by a 175 Hz sine carrier. It is known that such stimulation can activate Pacinian corpuscles sensitive to vibration stimulation above 100 Hz and Meissner corpuscles sensitive to vibration stimulation in the range of 20 to 50 Hz.

[0070] (3) Imaginary movement paradigm

[0071] As Figure 1 shown, throughout the experiment, the subjects were required to sit as quietly as possible in a chair facing a computer screen. The subjects needed to place their hands on the armrests while keeping their bodies still, especially not generating facial and hand muscle movements, and try not to blink during each task. The tactile stimulator was wrapped around the subjects' wrists and placed on the dorsal outer side of the wrists. The screen background was black. At the start of each trial (t = 0 s), a white cross appeared in the center of the screen, and the subjects needed to fixate on the cross. At the 2nd second (t = 2 s), a tactile stimulation with a duration of 200 milliseconds was applied to both hands simultaneously to remind the subjects that the task was about to start. At the 3rd second (t = 3 s), a red cue bar was superimposed on the white cross, which randomly appeared on the left or right side. When the red cue bar appeared on the left side, the subjects needed to perform the left - hand imaginary movement task. When the red cue bar appeared on the right side, the subjects needed to perform the right - hand imaginary movement task. The duration of the red cue bar was 1.5 seconds, that is, the red cue bar disappeared at t = 4.5 s. From the appearance of the red cue bar, the subjects continuously performed the imaginary movement task for 5 seconds until the white cross disappeared at t = 8 s. Finally, there was a random time between 1.5 and 3.5 seconds for the subjects to rest and to limit the subjects' adaptation.

[0072] (4) Tactile - assisted imaginary movement paradigm

[0073] In this module, except for applying tactile stimulation to the subjects when performing the imaginary movement task, other settings were the same as those in the imaginary movement calibration module. When the red cue bar appeared, the actuators of both the left and right hands vibrated, and the tactile stimulation lasted for 5 s until the white cross disappeared. When the red cue bar appeared on the left side, the subjects needed to perform the left - hand imaginary movement task assisted by tactile stimulation, where the subjects imagined the movement of the left hand, focused on the left hand, and then felt the tactile stimulation of the left hand. When the red cue bar appeared on the right side, the subjects needed to perform the right - hand imaginary movement task assisted by stimulation, where the subjects imagined the movement of the right hand, concentrated on the right hand, and felt the tactile stimulation of the right hand. This module had 3 rounds and 120 trials, and each run contained 40 trials.

[0074] (4) Electroencephalogram data feature extraction:

[0075] The Common Spatial Pattern (CSP) algorithm is a classic feature extraction algorithm based on the imagined movement brain-computer interface. It finds the two most separable linear projection directions in a data-driven manner. In these directions, the discriminative information between the two classes is maximized and the common information is minimized. Mathematically, the discriminative information and the common information between the two classes can be described as the difference S d and sum S c :

[0076] S d = Σ (l) - Σ (r)

[0077] S c = Σ (l) + Σ (r)

[0078] Σ (l) represents the covariance of the left hand imagined movement task, and Σ (r) represents the covariance of the right hand imagined movement task. In addition, the objective function can be expressed as:

[0079]

[0080]

[0081] Through the Lagrange multiplier method, the objective function can be transformed into a generalized eigenvalue problem:

[0082] S d w = λ d w

[0083] For the eigenvector w corresponding to the eigenvalue λ d it satisfies:

[0084] wΣ (l) w + wΣ (r) w = 1

[0085] In other words, in the w direction, the sum of the variances of the two classes is 1. Therefore, when the projection direction maximizes the variance of one class, it will simultaneously minimize the variance of the other class. After solving for w, we can transform the original data to:

[0086]

[0087] where, X ∈ R C×T , X CSP ∈ R C×T , W ∈ R C×C, the column vectors of W are generalized eigenvectors, usually called filters. X CSP The projection of the original signal in the direction of the generalized eigenvector, where X is the original signal, t represents the current time, C is the number of channels, and T is the total time length. We select the eigenvectors corresponding to the largest three eigenvalues and the smallest three eigenvalues as filters. Among the first three filters, the variance of the left-hand imagined movement task is large, while that of the right-hand imagined movement task is small, and vice versa for the last three filters. Finally, the logarithm is applied to the variance of the filtered data, and finally, we obtain six-dimensional features.

[0088] (5) Classifier

[0089] After extracting the CSP features, a classifier is trained to classify the EEG data features. Specifically, the imagined movement data assisted by tactile sensations is used as the training set during training.

[0090] The classifier can use the Linear Discriminant Analysis (LDA) algorithm. LDA is one of the classic algorithms in the BCI field because it is simple and effective. When the number of classes is 2, its basic idea is to find a projection direction that maximizes the distance between the two classes and minimizes the distance within each class. Mathematically, the distances between and within the two classes are defined as the between-class scatter matrix and the within-class scatter matrix:

[0091] S B =(μ1 - μ2)(μ1 - μ2) T

[0092] S W =(S1 + S2)

[0093] where μ1 and μ2 represent the centers of the two classes respectively, and S1 and S2 represent the variances of the two classes respectively. The objective function of LDA can be expressed as:

[0094]

[0095]

[0096] Through the method of Lagrange multipliers, the objective function can be transformed into the following generalized eigenvalue problem:

[0097] S B ω = λS W ω

[0098] Since S B ω=(μ1 - μ2)(μ1 - μ2) T ω, so S B ω and μ1 - μ2 have the same direction, that is: S B ω = k(μ1 - μ2)=λS Wω and λ are constant scaling factors. Therefore, the direction of ω can be expressed as:

[0099]

[0100] The offset b of the classifier is:

[0101]

[0102] Finally, the decision boundary between the two classes is determined by the following equation:

[0103] y = ω T x + b

[0104] In addition, the support vector machine can also be used as a classifier because it is more robust to outlier data:

[0105] The support vector machine is a classic algorithm for two-class tasks. Based on statistical ideas, it seeks an optimal separation plane to maximize the distance between the two classes of data and the decision boundary. We only use the linear support vector machine, and the optimization objective is as follows:

[0106]

[0107] s.t. y i (w T x i + b) ≥ 1 - ξ i

[0108]

[0109] This is a quadratic programming problem. Here, w is the normal vector of the decision boundary. In the linearly separable case, the classification margin of the SVM is 2 / w T w. C is the penalty coefficient, ξ is the scaling factor that allows some samples to be misclassified by the decision function, making the model more robust, b is the bias of the decision boundary. The decision boundary of the support vector machine is mainly determined by the support vectors at the connection of the two classes, so it is more robust to outliers. We determine the value of C through 5x5 cross-validation.

[0110] (6) Decoding imagined movement

[0111] Use pure imagined movement data as the test set to simulate the real brain-computer interface application scenario. Use the model trained with tactile sensory-assisted imagined movement data in step (2) to decode pure imagined movement.

[0112] (7) Performance evaluation of tactile sensory calibration decoding

[0113] Compare the decoding performance of the two calibration schemes on the test data, namely the haptic perception calibration decoding test set data and the traditional imagined movement calibration decoding test set data.

[0114] To compare and illustrate the effectiveness of the calibration decoding method proposed by the present invention, a comparison is made between the results of haptic perception calibration and the traditional calibration method, and Figure 2 is obtained. It shows the BCI decoding accuracy in different calibration methods and algorithms. When using the CSP-LDA algorithm, all participants achieved the best performance in the haptic perception calibration method except for S4 and S11. The average accuracy of traditional imagined movement calibration was 71.31%, while the average accuracy of haptic perception calibration was 78.33%. When using the CSP-SVM algorithm, all participants could also obtain the best performance in haptic calibration except for S1 and S3 and the average accuracy. The traditional imagined movement calibration was 74.61%, and the haptic calibration was 79.23%. The paired t-test showed that the haptic-assisted calibration was significantly higher than the imagined movement calibration (CSP-LDA: p < 0.01, CSP-SVM: p < 0.01). These results indicate that haptic perception calibration can significantly improve the performance of the brain-computer interface based on imagined movement.

[0115] The calibration process in the brain-computer interface system is very time-consuming because sufficient data is required to obtain a high-performance classifier. We hope to minimize the number of trials as much as possible and shorten the calibration time on the premise of achieving the same excellent performance. We trained the classifier with different numbers of trial numbers and obtained the performance changes of haptic perception calibration and traditional imagined movement calibration with the number of trials. As Figure 3 shown, when the number of trials is between 20 and 120, the accuracy of haptic perception calibration is significantly higher than that of imagined movement calibration. When using haptic-assisted calibration, only 30 experimental trials are required to obtain a classification performance of over 70%, which is the generally recognized BCI acceptance level, while traditional imagined movement calibration requires 50 trials. The accuracy obtained with 120 trials in imagined movement calibration is the same as that obtained with 40 trials in haptic-assisted calibration. In haptic perception calibration, when the number of trials exceeds 80, the accuracy curve still has an upward trend, but in imagined movement calibration, the curve tends to level off. This shows that haptic perception calibration can achieve better performance with less time than traditional imagined movement calibration.

[0116] The above-described embodiments have detailed the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not used to limit the present invention. Any modifications, supplements, and equivalent replacements made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An imagination movement decoding method based on tactile feeling calibration, characterized in that, It includes the following steps: (1) Set up a tactile stimulation device to apply tactile stimulation to the dorsal side of the left or right wrist of the subject. Use a linear resonant actuator to generate tactile stimulation, and the frequency of the stimulation is 27 Hz; (2) Set up an EEG device to record EEG signals. During the entire recording process, the electrode impedance is lower than 10 KΩ. The original signal is digitally sampled at a frequency of 500 Hz, and the original signal is filtered using a bandwidth filter between 0.5 - 70 Hz and a notch filter of 48 - 52 Hz; (3) Utilize the tactile stimulation device and the EEG device to collect the pure imagined movement data of the subject, as well as the imagined movement data assisted by tactile sensation; (4) Use the common spatial pattern algorithm to extract the EEG data features and train a classifier for classifying the EEG data features; among them, the imagined movement data assisted by tactile sensation is used as the training set; the specific process of using the common spatial pattern algorithm to extract the EEG data features is as follows: Find two most separable linear projection directions in a data-driven manner, in which the difference information between the two classes is maximized and the common information is minimized. The difference information and the common information between the two classes are respectively described as the difference S of covariance d and sum S c : S d = Σ (l) - Σ (r) S c = Σ (l) + Σ (r) Σ (l) represents the covariance of the left - hand imaginary movement task, Σ (r) represents the covariance of the right - hand imaginary movement task. In addition, the objective function is expressed as: Through the Lagrange multiplier method, the objective function is transformed into a generalized eigenvalue problem: S d w = λ d w For the eigenvector w corresponding to the eigenvalue λ d it satisfies: wΣ (l) w + wΣ (r) w = 1 In the w direction, the sum of the variances of the two classes is 1; therefore, when the projection direction maximizes the variance of one class, it will simultaneously minimize the variance of the other class; after solving for w, the original data is transformed into: where X ∈ R C×T , X CSP ∈ R C×T , W ∈ R C×C , and the column vectors of W are generalized eigenvectors; X CSP is the projection of the original signal in the direction of the generalized eigenvector, X is the original signal, t represents the current time, C is the number of channels, and T is the total time length; Select the eigenvectors corresponding to the three largest eigenvalues and the three smallest eigenvalues as filters; among the first three filters, the variance of the left - hand imagined movement task is larger, while the variance of the right - hand imagined movement task is smaller, and in the last three filters, it is the opposite; finally, apply the logarithm to the variance of the filtered data to obtain six - dimensional features; The classifier mentioned above uses a support vector machine, and the optimization objective is as follows: s.t.y i (w T x i +b)≥1-ξ i Among them, w is the normal vector of the decision boundary. In the case of linear separability, the classification margin of the support vector machine is 2 / w T w, C is the penalty coefficient, ξ is the scaling factor, and b is the deviation of the decision boundary; (5) Use the trained classifier for pure imagined movement decoding.

2. The method for decoding imaginary movement based on haptic perception calibration according to claim 1, wherein In step (1), the linear resonant actuator uses a piezoelectric ceramic stimulator.

3. The method for decoding imagined movement calibrated based on tactile sensation according to claim 1, wherein In step (2), the EEG device uses a 64 - channel wireless NeuSenW64 EEG system, uses an EEG cap with 64 channels, the electrodes are arranged according to the 10 - 20 international system, the reference electrode is CPz, and the ground electrode is AFz.

4. The method for decoding imagined movement based on haptic perception calibration according to claim 1, wherein In step (3), when collecting the pure imagined movement data of the subject, the imagined movement paradigm is designed as follows: Throughout the process, the subject is required to sit quietly on a chair facing a computer screen; the subject places their hand on the armrest while keeping the body still, without generating facial and hand muscle movements, and not blinking during each task; The tactile stimulator is wrapped around the subject's wrist and placed on the dorsal lateral side of the wrist; the screen background is black. At the start of each trial, at t = 0 s, a white cross appears in the center of the screen, and the subject needs to fixate on the cross; at t = 2 s, a tactile stimulus with a duration of 200 ms acts on both hands simultaneously to remind the subject that the task is about to start; at t = 3 s, a red cue bar is superimposed on the white cross, and it randomly appears on the left or right side; when the red cue bar appears on the left side, the subject needs to perform a left-hand imagined movement task; when the red cue bar appears on the right side, the subject needs to perform a right-hand imagined movement task; the duration of the red cue bar is 1.5 s, that is, the red cue bar disappears at t = 4.5 s; from the appearance of the red cue bar, the subject continuously performs the imagined movement task for 5 s until the white cross disappears at t = 8 s; finally, there is a random period of time between 1.5 and 3.5 s for the subject to rest and to limit the subject's adaptation effect.

5. The method for decoding imagined movement calibrated based on tactile sensation according to claim 4, wherein In step (3), when collecting the imagined movement data assisted by the subject's tactile sensation, the tactile sensation paradigm is designed as follows: The tactile stimulator is wrapped around the subject's wrist and placed on the dorsal lateral side of the wrist; the screen background is black. At the start of each trial, at t = 0 s, a white cross appears in the center of the screen, and the subject needs to fixate on the cross; at t = 2 s, a tactile stimulus with a duration of 200 ms acts on both hands simultaneously to remind the subject that the task is about to start; at t = 3 s, a red cue bar is superimposed on the white cross, and it randomly appears on the left or right side; when the red cue bar appears on the left side, the subject performs a left-hand imagined movement task assisted by tactile stimulation, where the subject imagines the movement of the left hand, focuses on the left hand, and then feels the tactile stimulation of the left hand; when the red cue bar appears on the right side, the subject needs to perform a right-hand imagined movement task assisted by stimulation, where the subject imagines the movement of the right hand, focuses on the right hand, and feels the tactile stimulation of the right hand; the duration of the red cue bar is 1.5 s, that is, the red cue bar disappears at t = 4.5 s; From the appearance of the red cue bar, the actuators of both the left and right hands vibrate, and the tactile stimulation lasts for 5 s. During this process, the subject continuously performs the imagined movement task for 5 s until the white cross disappears at t = 8 s; finally, there is a random period of time between 1.5 and 3.5 s for the subject to rest and to limit the subject's adaptation effect.

Images