Motion intention recognition method and system for brain-computer interface

Abstract

The invention discloses a motion intention recognition method and system for a brain-computer interface, and relates to the technical field of signal processing and recognition, original light intensity data is obtained by collecting functional near infrared spectrum signals of a frontal lobe brain region of a user under a normal form task, the light intensity data is preprocessed, and a motion intention recognition result is obtained. The preprocessed data are converted into hemodynamic response signals; extracting a plurality of features from the hemodynamic response signals, and combining the extracted features into a feature matrix; performing statistical feature optimization on the feature matrix, generating a feature selection mask, and determining an optimal feature set based on the feature selection mask; and adopting a support vector machine as a classifier, and inputting the optimized feature set as a training set into the classifier for model training. Redundant feature removal and high-contribution feature recognition are realized based on feature statistics optimization, the information mining depth and the operation efficiency of the blood oxygen concentration signal are considered, and the accuracy and the reliability of the blood oxygen concentration signal recognition are improved. And the classification result is presented in real time.

Description

Technical Field

[0001] The present disclosure relates to the technical field of signal processing and recognition, and more particularly, to a method and system for recognizing movement intentions for a brain-computer interface. Background Art

[0002] Functional Near-Infrared Spectroscopy (hereinafter referred to as fNIRS) is a functional detection method for calculating brain activities by measuring the degree of scattering and absorption of the light actively emitted by an optical system when passing through brain tissues. Compared with other brain imaging technologies, fNIRS has significant advantages such as high ecological validity, environmental robustness, and low deployment cost. In recent years, the brain-computer interface system (Brain-Computer Interface, hereinafter referred to as BCI) based on fNIRS has shown great practical potential in obtaining and recognizing brain nerve signals in daily scenarios and has received extensive attention from researchers.

[0003] The fNIRS signal analysis technology converts the light intensity signal into a blood oxygen concentration signal based on the modified Lambert-Beer law. When the subject uses a specific brain region, the metabolic pressure will cause the capillaries in the corresponding area to dilate, resulting in a higher hemoglobin concentration. The light scattering and absorption effects brought about by the above changes will be directly reflected in the measured light intensity data to match the activation pattern of the subject's brain region. After obtaining the hemodynamic activation data, the analysis method used determines the recognition accuracy of the stimulus.

[0004] However, in the related technologies, there are problems such as it being difficult to balance the depth of information mining of the blood oxygen concentration signal and the operation efficiency, the feature selection being greatly affected by the specificity of the brain activation time, and poor real-time performance. Summary of the Invention

[0005] In view of the many deficiencies of the prior art, the present disclosure provides a method and system for recognizing movement intentions for a brain-computer interface.

[0006] In a first aspect of the present disclosure, a method for recognizing movement intention for a brain-computer interface is provided, including the following steps: acquiring original light intensity data by collecting functional near-infrared spectroscopy signals of the frontal brain region of a user under a paradigm task, preprocessing the light intensity data, and converting the preprocessed data into a hemodynamic response signal; extracting multiple features from the hemodynamic response signal respectively, and combining the extracted multiple features into a feature matrix, wherein feature extraction includes time-domain feature extraction, frequency-domain feature extraction, network feature extraction, and curve fitting parameter extraction; performing statistical feature optimization on the feature matrix to generate a feature selection mask, determining a preferred feature set based on the feature selection mask, and statistical feature optimization includes dataset stratification, normality test, significance test, and feature reproducibility test; using a support vector machine as a classifier, inputting the preferred feature set as a training set into the classifier for model training to obtain a trained model; based on the trained model, classifying the preferred feature set to be recognized to obtain a classification result, and outputting the movement intention corresponding to the classification result to a user interface.

[0007] Further, in some embodiments, before collecting the functional near-infrared spectroscopy signals of the frontal brain region of the user under the paradigm task, the method further includes: according to the experimental paradigm design, setting parameters for the paradigm prompt on a real-time classification interface to guide the user to execute the paradigm task, wherein parameter setting includes classification task setting, paradigm trial setting, baseline time setting, task time setting, and rest time setting.

[0008] In some embodiments, preprocessing the light intensity data and converting the preprocessed data into a hemodynamic response signal includes: extracting the light intensity data before the end moment of each trial to obtain partial light intensity data; filtering the partial light intensity data to remove physiological noise, baseline drift, and spike noise in the partial light intensity data to obtain filtered data; converting the filtered data into a blood oxygen concentration signal by using the modified Lambert-Beer law, wherein the blood oxygen concentration signal includes oxyhemoglobin concentration data and deoxyhemoglobin concentration data.

[0009] In some embodiments, multiple feature extractions are performed on the hemodynamic response signal respectively, including: performing time-domain feature extraction on the hemodynamic response signal to obtain the time information of the hemodynamic response signal, where the time information includes mean, variance, slope, peak value, peak time, kurtosis, quadratic coefficient, and approximate entropy; performing frequency-domain feature extraction on the hemodynamic response signal to obtain the frequency information of the hemodynamic response signal, where the frequency information includes center frequency, mean square frequency, and frequency variance; performing network feature extraction on the hemodynamic response signal to obtain the spatial information of the hemodynamic response, and the spatial information includes node importance, average path length, clustering coefficient, node clustering coefficient, average degree, and degree distribution; using a deep learning network to extract the curve fitting parameters of the hemodynamic response signal.

[0010] In some embodiments, the deep learning network includes a multi-layer perceptron and a long short-term memory network. Using the deep learning network to extract the curve fitting parameters of the hemodynamic response signal includes: simultaneously using the multi-layer perceptron and the long short-term memory network to fit the response curve of each channel for the hemodynamic response signal, and using the trained network parameters as features to form a feature matrix. Among them, the network of the multi-layer perceptron fitting the response curve includes an input layer, a hidden layer with 10 neurons, and an output layer. The network of the long short-term memory network fitting the response curve includes a sequence input layer, a long short-term memory layer, a rectified linear unit activation layer, a fully connected layer, and a regression layer.

[0011] In some embodiments, statistically generating a feature selection mask for the feature matrix preferably includes: using the feature matrix obtained from valid trials as samples, performing stratified sampling on the data set including all samples according to a set rule to obtain a stratified data subset, and combining each data subset with the same samples according to the labels and classification tasks of the samples to obtain a sample set; performing a normality test on the sample set to determine whether the channel feature distribution of the sample set conforms to a normal distribution, and generating a first feature selection mask to indicate the result of the normality test; performing a significance test on the sample set to determine whether the channel feature distributions under different labels in the sample set conform to significance, and generating a second feature selection mask to indicate the result of the significance test; performing a feature reproducibility test on the sample set, counting the reproducibility of the channel features, and generating a third feature selection mask according to the reproducibility and the configured reproducibility threshold to indicate the result of the reproducibility test.

[0012] In some embodiments, a normality test is performed on the sample set to determine whether the channel feature distribution of the sample set conforms to a normal distribution, and a first feature selection mask is generated to indicate the normality test result, including: for each stratified sample set, the sample set is grouped according to different labels; the Jarque–Bera test method is independently used for each group of samples to calculate the p-value; if the p-value meets the configured first floating threshold, it is determined that the channel feature distribution within the current sample group conforms to a normal distribution, and the first feature selection mask is 1, or, if the p-value does not meet the configured first floating threshold, it is determined that the channel feature distribution within the current sample group does not conform to a normal distribution, and the first feature selection mask is 0.

[0013] In some embodiments, a significance test is performed on the sample set to determine whether the channel feature distributions under different labels in the sample set conform to significance, and a second feature selection mask is generated to indicate the significance result, including: for each stratified sample set, the sample set is grouped according to different classification tasks; the t-test method and the analysis of variance method are respectively used for each group of samples to calculate the p-value; if the p-value meets the configured second floating threshold, it is determined that the channel feature distribution within the current sample group conforms to significance, and the second feature selection mask is 1, or, if the p-value does not meet the configured second floating threshold, it is determined that the channel feature distribution within the current sample group does not conform to significance, and the second feature selection mask is 0.

[0014] In some embodiments, a feature reproducibility test is performed on the sample set, the reproducibility rate of the channel features is statistically analyzed, and a third feature selection mask is generated according to the reproducibility rate and the configured reproducibility rate threshold to indicate the reproducibility test result, including: statistically analyzing the number of channel features that simultaneously conform to a normal distribution and significance in all stratifications, and combining the number of all stratifications to obtain the reproducibility rate; if the reproducibility rate meets the configured reproducibility rate threshold, it is determined that the channel feature passes the reproducibility test, and the third feature selection mask is 1, or, if the reproducibility rate does not meet the configured reproducibility rate threshold, it is determined that the channel feature fails the reproducibility test, and the third feature selection mask is 0.

[0015] The second aspect of the present disclosure provides a motion intention recognition device for a brain-computer interface, including: a data acquisition module, which is used to obtain original light intensity data by collecting the functional near-infrared spectroscopy signals of the frontal brain region of the user under a paradigm task, preprocess the light intensity data, and convert the preprocessed data into hemodynamic response signals; a feature extraction module, which is used to perform various feature extractions on the hemodynamic response signals respectively, and combine the extracted various features into a feature matrix, where the feature extraction includes time-domain feature extraction, frequency-domain feature extraction, network feature extraction, and curve fitting parameter extraction; a feature statistics and optimization module, which is used to generate a feature selection mask by performing statistical feature optimization on the feature matrix, determine the optimized feature set through the feature selection mask, and the statistical feature optimization includes data set stratification, normality test, significance test, and feature reproducibility test; a training module, which is used to use a support vector machine as a classifier, input the optimized feature set as a training set into the classifier for model training, and obtain a trained model; a classification module, which is used to perform task classification on the optimized feature set to be recognized based on the trained model to obtain a classification result, and output the motion intention corresponding to the classification result to the user interface.

[0016] The third aspect of the present disclosure provides a real-time classification interface for motion intention recognition of a brain-computer interface, including a paradigm setting module, a data reading module, an identification module, and a classification result presentation module. Among them, the paradigm setting module is used to set the relevant parameters for paradigm prompt presentation, the data reading module is used to read the signals collected in real time by the fNIRS device in real time to obtain the original data, the identification module is used to preprocess, extract features, and optimize features from the collected original data and then input them into the classifier to obtain the classification result, and the classification result presentation module is used to present the classification result in a visual form in real time.

[0017] The fourth aspect of the present disclosure provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the foregoing first aspect.

[0018] The fifth aspect of the present disclosure provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the foregoing first aspect.

[0019] The sixth aspect of the present disclosure provides a computer program product, including a computer program, and the computer program realizes the foregoing first aspect when executed by a processor.

[0020] Due to the above technical solution, the beneficial effect of the present disclosure is that the present disclosure calculates the hemodynamic activation signal under the paradigm task based on the fNIRS data of the frontal brain region, combines multiple feature extraction methods and combines deep learning to obtain a feature matrix to fully express the data information, and realizes the removal of redundant features and the identification of high-contribution features through statistical optimization methods, taking into account both the depth of information mining of the blood oxygen concentration signal and the operation efficiency, and completes model training and deployment on the real-time classification interface based on the above methods, taking into account both the depth of information mining and the operation efficiency, solving the problems that feature selection is greatly affected by the time specificity of brain activation and poor real-time performance, and realizing high-precision motion decision recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Through the following description of the embodiments of the present disclosure with reference to the accompanying drawings, the above and other objects, features and advantages of the present disclosure will become clearer. In the drawings: Figure 1 Schematically shows a flowchart of a method for recognizing a motion intention for a brain-computer interface according to an embodiment of the present disclosure; Figure 2 Schematically shows a schematic diagram of a paradigm experiment design of a method for recognizing a motion intention for a brain-computer interface according to an embodiment of the present disclosure; Figure 3 Schematically shows a flowchart of a preprocessing method of a method for recognizing a motion intention for a brain-computer interface according to an embodiment of the present disclosure; Figure 4 Schematically shows a flowchart of a feature extraction method of a method for recognizing a motion intention for a brain-computer interface according to an embodiment of the present disclosure; Figure 5 Schematically shows a flowchart of a feature statistical optimization method of a method for recognizing a motion intention for a brain-computer interface according to an embodiment of the present disclosure; Figure 6 Schematically shows a schematic diagram of a system for recognizing a motion intention for a brain-computer interface according to an embodiment of the present disclosure; Figure 7 Schematically shows a schematic diagram of an electronic device according to an embodiment of the present disclosure. DETAILED IMPLEMENTATION MANNER

[0022] The following describes exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, the description below omits the description of well-known functions and structures.

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the related technologies of the technical solutions of the present disclosure will be described first.

[0024] In the research on fNIRS-BCI motion intention recognition in the related technologies, it is found that the feature extraction methods based on deep learning in the related technologies generate a larger number of features, resulting in longer operation time and a decrease in recognition accuracy. In particular, it is difficult to achieve high-precision recognition of complex motion intentions for the measurement of the frontal brain region, etc. It is not suitable for application in daily scenarios.

[0025] For fNIRS signals, a reliable decoding algorithm is required to extract effective information to achieve the recognition of different motion patterns. The quality of the measured data first depends on the performance of the acquisition device and the acquisition skills. Under the condition of minimizing the noise generated in the measurement stage, a suitable filtering algorithm needs to be selected to address the disturbances accompanying the measurement. At the same time, in the BCI system, when all features are selected as inputs, there will be more redundancies. Non-task-related features will lead to longer operation time and a decrease in the training effect of the classification model. When a neural network is used for feature extraction, a large number of features will be extracted. At this time, the adverse factors of retaining redundant features will be more significant.

[0026] In view of the problems of the existing fNIRS-BCI system, such as being difficult to balance the depth of information mining and the operation efficiency, the feature selection being greatly affected by the specificity of brain activation time, and the poor real-time performance, etc., there are limitations in that the classification accuracy and the complexity of the classification task cannot meet the requirements of brain-computer interface applications in daily scenarios.

[0027] Therefore, to solve the problems in the related technologies, the embodiments of the present invention provide a motion intention recognition method for a brain-computer interface, which is used to optimize the effective feature extraction based on the hemodynamic response signals measured in the frontal brain region of fNIRS signals. A variety of feature extraction methods are combined and deep learning is used to obtain a feature matrix to fully express the data information. Through statistical optimization methods, redundant feature removal and the recognition of high-contribution features are achieved, and model training and deployment on a real-time classification interface are completed, balancing the depth of information mining and the operation efficiency, and solving the problems such as the feature selection being greatly affected by the specificity of brain activation time and the poor real-time performance.

[0028] Next, a motion intention recognition method, system, real-time classification interface, and electronic device for a brain-computer interface according to the embodiments of the present disclosure will be described with reference to the accompanying drawings.

[0029] Figure 1 Schematically shows a flowchart of a motion intention recognition method for a brain-computer interface according to an embodiment of the present disclosure.

[0030] As Figure 1 shown, the method includes the following steps 101-105.

[0031] Step 101: By collecting the functional near-infrared spectroscopy signals of the frontal brain region of the user under the paradigm task, obtaining the original light intensity data, preprocessing the light intensity data, and converting the preprocessed data into hemodynamic response signals.

[0032] In some embodiments, before collecting the functional near-infrared spectroscopy signals of the frontal brain region of the user under the paradigm task, the method further includes: According to the experimental paradigm design, setting parameters for the paradigm prompt on the real-time classification interface to guide the user to execute the paradigm task, where the parameter settings include classification task settings, paradigm trial settings, baseline time settings, task time settings, and rest time settings.

[0033] In some embodiments, the classification task settings are used to deploy the algorithm model according to the selected classification task and change the picture presentation of the classification result presentation module to adapt to the selected task. The paradigm trial settings are used to configure the number of cycles of the task prompt and rest prompt on the paradigm prompt screen. The baseline time settings are used to configure the duration of the baseline rest state before the start of the paradigm. The task time settings are used to set the duration of the task prompt presentation on the paradigm prompt screen. The rest time settings are used to set the duration of the rest prompt presentation on the paradigm prompt screen.

[0034] In some embodiments, the paradigm is designed to perform five types of movement execution tasks: rest, left hand grasping, right hand grasping, left finger tapping, and right finger tapping.

[0035] In one embodiment of the present disclosure, as Figure 2 shown, a single paradigm experiment starts with a 30s baseline Baseline time, and then consists of a total of 20 task Blocks in five cycles in the order of rest - left hand grasping - rest - right hand grasping (Trial). Among them, the duration of each rest and task Block can be 5s, 10s, or 15s, and the total experimental time is 230s. When the task duration is 15s, the paradigm experiment consists of a total of 12 task Blocks in three cycles in the order of rest - left hand grasping - rest - right hand grasping.

[0036] Further, the movement execution in the paradigm experiment can be a tapping action. A single paradigm experiment starts with a 30s baseline Block, and then consists of a total of 20 task Blocks in five cycles in the order of rest - left hand tapping - rest - right hand tapping. Among them, the duration of each rest and task Block can be 5s, 10s, or 15s, and the total experimental time is 230s. When the task duration is 15s, the paradigm experiment consists of a total of 12 task Blocks in three cycles in the order of rest - left hand tapping - rest - right hand tapping.

[0037] It should be noted that to ensure the standardization of paradigm execution, the subjects are seated to maintain a comfortable body posture and a comfortable screen distance, and complete the tasks according to the prompts presented on the screen. A complete grasping motion is set as follows: starting from the natural state, each finger first extends outward, pauses slightly after reaching the maximum abduction state, then grasps inward forcefully, and also maintains for a short period of time in the fist state, and finally returns to the natural state; the finger tapping motion is executed at a frequency of 3 Hz.

[0038] In an embodiment of the present disclosure, the single-cycle data sample collection amounts for a total of three subjects are 500, including 100 groups each of rest, left hand grasping, right hand grasping, left finger tapping, and right finger tapping.

[0039] In some embodiments, an fNIRS signal acquisition device realizes parallel acquisition of all measurement channels based on a high-precision digital phase-locked algorithm. Two near-infrared light wavelengths of 775 nm and 855 nm are equipped, 10 light sources and 4 detectors are arranged, and the sampling frequency is set to 2 Hz.

[0040] In some embodiments of the present disclosure, in the embodiments of the present disclosure, according to the paradigm task, all the data before the end moment of each order is extracted, and this part of the data is filtered. Before feature extraction, the light intensity signal is first converted into a blood oxygen concentration signal based on the modified Lambert-Beer law, and then according to the label, only the data during the task period is retained.

[0041] As Figure 3 shown, the light intensity data is preprocessed, and the preprocessed data is converted into a hemodynamic response signal, including the following steps 201-203.

[0042] Step 201, extract the light intensity data before the end moment of each trial to obtain partial light intensity data.

[0043] Step 202, filter the partial light intensity data to remove physiological noise, baseline drift, and spike noise in the partial light intensity data to obtain the filtered data.

[0044] In some embodiments, the filtering method uses an 8th-order 0.45 Hz Butterworth low-pass filter to remove physiological noises such as heartbeat and respiration, uses a moving standard deviation with a window of 5 and a Sobel filter to identify the segment to be processed for each channel of data, and uses cubic spline interpolation and a Savitzky-Golay filter to remove baseline drift and spike noise for the segment to be processed respectively.

[0045] Step 203, convert the filtered data into blood oxygen concentration signals using the modified Lambert-Beer law, where the blood oxygen concentration signals include oxygenated hemoglobin (HbO) concentration data and deoxygenated hemoglobin (HbR) concentration data.

[0046] In some embodiments, the operation formula of the modified Lambert-Beer law is as follows: ; ; where L is the total average path length, is the light intensity attenuation at 775 nm wavelength, is the light intensity attenuation at 855 nm wavelength, is the extinction coefficient of deoxygenated hemoglobin for near-infrared light at 775 nm wavelength, is the extinction coefficient of deoxygenated hemoglobin for near-infrared light at 855 nm wavelength, is the extinction coefficient of oxygenated hemoglobin for near-infrared light at 775 nm wavelength, is the extinction coefficient of oxygenated hemoglobin for near-infrared light at 855 nm wavelength, is the concentration of oxygenated hemoglobin, is the concentration of deoxygenated hemoglobin.

[0047] In some embodiments, input the oxygenated hemoglobin concentration data into Step 102 for feature extraction.

[0048] Step 102, perform various feature extractions on the hemodynamic response signals respectively, and combine the extracted various features into a feature matrix, where the feature extraction includes time-domain feature extraction, frequency-domain feature extraction, network feature extraction, and curve fitting parameter extraction.

[0049] In some embodiments of the present disclosure, convert the preprocessed data into hemodynamic response signals and input them into the feature extraction module, where the feature extraction module includes a time-domain feature extraction module, a frequency-domain feature extraction module, a network feature extraction module, and a deep learning module; among them, the time-domain feature extraction module is used to extract the time information of the hemodynamic response, the frequency-domain feature extraction module is used to extract the frequency information of the hemodynamic response, the network feature extraction module is used to extract the spatial information of the hemodynamic response, and the deep learning module is used to extract the curve fitting parameters of the hemodynamic response.

[0050] In some embodiments, as Figure 4 shown, performing various feature extractions on the hemodynamic response signals respectively includes the following steps 301-304.

[0051] Step 301: Extract the time-domain features of the hemodynamic response signal to obtain the time information of the hemodynamic response signal, where the time information includes the mean, variance, slope, peak value, peak time, kurtosis, quadratic term coefficient, and approximate entropy.

[0052] In some embodiments, the mean value of the hemodynamic response signal within the task response window is calculated by the formula where represents the mean value, N represents the number of data points within the window, and represents the HbO / HbR concentration value at the q-th time point within the window.

[0053] In some embodiments, the variance of the hemodynamic response signal within the task response window is calculated by the formula where represents the variance.

[0054] In some embodiments, the peak value of the hemodynamic response signal within the task response window is calculated by the formula where E[Z 4 is the expectation (mean) operator, EX represents the mean μ of the sample point X, DX represents the variance of X , represents the standard deviation of the sample point X, and Z = represents calculating the standardized value for each sample point.

[0055] In some embodiments, for each channel, a quadratic polynomial is fitted, and the polynomial fitting is solved by QR decomposition to extract the slope and the quadratic term coefficient.

[0056] Specifically, assuming there are m channels to be fitted, the signal of each channel can be expressed as an n-th order polynomial of time x: ; where is the coefficient of the -th order term of the polynomial, is the coefficient of the highest order term, and y is the observed signal value of each channel.

[0057] Further, the following system of equations is constructed: ; where is the matrix composed of vectors of each order variable; , where n is the highest order of the variable; , where m is the number of channels to be fitted; Then, the matrix A is QR decomposed according to the following formula to solve the vector

[0058] ; wherein, is a coefficient vector, and is a constant, the value represents the slope to be obtained, represents the quadratic term coefficient to be obtained, is an orthogonal matrix, represents the transpose matrix of the orthogonal matrix Q, is a non - singular upper triangular matrix.

[0059] Step 302: Extract the frequency domain features of the hemodynamic response signal to obtain the frequency information of the hemodynamic response signal. Among them, the frequency information includes the center frequency, mean square frequency, and frequency variance.

[0060] In some embodiments, the center frequency is obtained through the following operation expression: ; wherein, represents the center frequency, represents the power spectral density (PSD), and the PSD can be calculated by the Welch method.

[0061] In some embodiments, the mean square frequency is obtained through the following operation expression: ; In some embodiments, the frequency variance is obtained based on the center frequency and PSD. Specifically, the frequency variance can be obtained through the following operation expression: ; Step 303: Extract the network features of the hemodynamic response signal to obtain the spatial information of the hemodynamic response. The spatial information includes node importance, average path length, clustering coefficient, node clustering coefficient, average degree, and degree distribution.

[0062] In some embodiments, first calculate the Pearson correlation coefficient between each channel to construct an adjacency matrix, and construct the connection of the edges between nodes with 0.3, 0.4, 0.5, 0.6, 0.7, and 0.8 as thresholds respectively. The number of rows of the network feature matrix represents the selected threshold, and the number of columns represents the extracted network features. Calculate the node importance, average path length, clustering coefficient, node clustering coefficient, average degree, and degree distribution features for the adjacency matrix under each threshold.

[0063] In some embodiments, the node importance includes degree centrality, betweenness centrality, and eigenvector centrality. Degree centrality refers to the sum of the weights of the edges connected to the node. Betweenness centrality refers to the frequency of the node appearing in all the shortest paths. Eigenvector centrality refers to the weighted sum of the importance of the node's neighbors.

[0064] In some embodiments, the average path length is obtained through the following operation expression: ; wherein, represents the average path length, is the shortest path length from node z to j, represents the maximum number of edges.

[0065] In some embodiments, the clustering coefficient is obtained through the following operation expression: ; wherein, represents a node, and j and k are nodes belonging to the set of adjacent nodes of node ; and , is the adjacent node of node , and is the actual number of edges existing between them, is the number of adjacent nodes of node .

[0066] Step 304: Use a deep learning network to extract the curve fitting parameters of the hemodynamic response signal.

[0067] In the embodiments of the present disclosure, the core objective of the Fitting Response Curves is to establish a mathematical relationship between brain activity and hemodynamic signals (such as changes in HbO / HbR concentration).

[0068] In some embodiments of the present disclosure, step 304 includes: simultaneously using a Multilayer Perceptron (MLP) and a Long Short-Term Memory (LSTM) to fit the response curve of each channel for the hemodynamic response signal, and using the trained network parameters as features to form a feature matrix.

[0069] Among them, the network for the MLP to fit the response curve includes an input layer, a hidden layer with 10 neurons, and an output layer, and the network for the LSTM to fit the response curve includes a sequence input layer, an LSTM layer, a rectified linear unit activation layer, a fully connected layer, and a regression layer.

[0070] Step 103: Perform statistical feature optimization on the feature matrix to generate a feature selection mask, determine the preferred feature set based on the feature selection mask, and the statistical feature optimization includes data set stratification, normality test, significance test, and feature reproducibility test.

[0071] In some embodiments of the present disclosure, the features extracted by each feature extraction module are combined into a feature matrix and input into a statistical feature optimization module. The statistical feature optimization module includes a data set stratification module, a normality test module, a difference significance test module, and a feature recurrence rate test module. Among them, the data set stratification module is used to divide the sample set for statistical test calculation and label the input samples in each group. The normality test module is used to test the normality level of the sample distribution in the sample set. The difference significance test module is used to test the difference significance of the sample distributions under different labels in the sample set. The feature recurrence rate test module is used to test the number of times the same feature is repeatedly selected under different stratifications.

[0072] In some embodiments, as Figure 5 shown, performing statistical feature optimization on the feature matrix to generate a feature selection mask includes the following steps 401-404.

[0073] Step 401: Using the feature matrix obtained from valid trials as samples, perform stratified sampling on the data set including all samples according to a set rule to obtain stratified data subsets. According to the labels and classification tasks of the samples, combine each data subset with the same samples to obtain a sample set.

[0074] In the embodiments of the present disclosure, the data set stratification module first determines that every 5 samples are used for one stratified sampling according to the number of samples in the input entire data set, obtains the stratified data set, and combines each data set with the same sample data according to the labels and classification tasks and inputs them into the subsequent normality test module, difference significance test module, and feature recurrence rate test module respectively.

[0075] In some embodiments, calculate the degree of difference between the corresponding elements of two groups of samples. The features with significant differences retained by this method are used as the primary selection feature matrix 1. After increasing the number of sample extractions, form the primary selection feature matrix 2 by the same method, and keep this operation process after more increases in the number of sample extractions to obtain the feature matrix n. After obtaining all the primary selection feature matrices, finally, judge whether to input as an optimized feature based on the recurrence times of each feature in the primary selection feature matrix.

[0076] In some embodiments, the size of the time-domain feature matrix is (7, 10), that is, 7 time-domain features are extracted for each of the 10 channels; the size of the frequency-domain feature matrix is (4, 10), that is, 4 frequency-domain features are extracted for each of the 10 channels; the size of the network feature matrix is (6, 24), that is, 24 network features are extracted for each of the 6 threshold selections when establishing the network for the adjacency matrix, the size of the MLP network parameter feature matrix is (31, 10), that is, 31 features are extracted for each of the 10 channels; the size of the LSTM network parameter feature matrix is (571, 10), that is, 571 features are extracted for each of the 10 channels.

[0077] Step 402, perform a normality test on the sample set to determine whether the channel feature distribution of the sample set conforms to a normal distribution, and generate a first feature selection mask to indicate the result of the normality test.

[0078] In some embodiments, performing a normality test on the sample set to determine whether the channel feature distribution of the sample set conforms to a normal distribution, and generating a first feature selection mask to indicate the result of the normality test includes: for each stratified sample set, grouping the sample set according to different labels; independently using the Jarque–Bera test method to calculate the p-value for each group of samples; if the p-value satisfies the configured first floating threshold, determining that the channel feature distribution within the current sample group conforms to a normal distribution, and the first feature selection mask is 1, or, if the p-value does not satisfy the configured first floating threshold, determining that the channel feature distribution within the current sample group does not conform to a normal distribution, and the first feature selection mask is 0.

[0079] Specifically, first calculate the normality of the feature sample distribution according to different labels, use the Jarque–Bera (JB) test to calculate the p-value, and then configure whether the feature distribution on the corresponding channel conforms to a normal distribution according to the floating set threshold. The calculation formula of the JB test is as follows: ; where N represents the sample size, that is, the number of data points within the window, S represents the skewness value, K represents the kurtosis value, and the calculation formulas of the skewness value S and the kurtosis value K are as follows: ; ; where is the sample mean.

[0080] Furthermore, generate a first feature selection mask indicating the result of the normality test, which consists of 0 and 1. 1 indicates selecting the feature at the corresponding position, and 0 indicates removing the feature at the corresponding position.

[0081] Step 403: Perform a test for significant differences on the sample set to determine whether the channel feature distributions under different labels in the sample set meet the significance of differences, and generate a second feature selection mask to indicate the results of the significance of differences.

[0082] In some embodiments, performing a test for significant differences on the sample set to determine whether the channel feature distributions under different labels in the sample set meet the significance of differences, and generating a second feature selection mask to indicate the results of the significance of differences includes: for each stratified sample set, grouping the sample set according to different classification tasks; calculating p-values for each group of samples using the t-test method and the analysis of variance method respectively; if the p-value meets the configured second floating threshold, determining that the channel feature distribution within the current sample group meets the significance of differences, and the second feature selection mask is 1, or, if the p-value does not meet the configured second floating threshold, determining that the channel feature distribution within the current sample group does not meet the significance of differences, and the second feature selection mask is 0.

[0083] Specifically, calculate the p-value using the statistical test methods of t-test and analysis of variance according to different classification tasks, and then configure the significance of differences of the features on the corresponding channels according to the floating set threshold. For the independent samples t-test method, compare whether there is a significant difference in the means of two independent samples X and Y. The calculation formula of the statistic is as follows: ; where represents the mean of sample X, represents the mean of sample Y, σ(X) is the standard deviation of sample X, σ(Y) is the standard deviation of sample Y, n1 is the sample size of X, and n2 is the sample size of Y.

[0084] The samples are divided into 1 - j groups according to different classification tasks. The calculation formula of the statistic for the analysis of variance method is as follows: First, calculate the sample mean and sample variance: ; ; The between-group variance is calculated by the following formula: ; where represents the sample mean of the jth classification task, represents the sample size of the jth classification task, represents the value of the ith sample point in the jth classification task, is the between-group variance, represents the sum of squares between groups (Sum of Squares Between Groups, abbreviated as SSB), Represents the total sample mean of all groups, Is the degree of freedom of SSB, the total sample mean The calculation formula is as follows: ; Among them, Represents the sum of each sample size. At this time, the within-group variance MSE can be calculated as: ; Among them, Is the sum of squared errors (Sum of Squared Errors, abbreviated as SSE), Is the degree of freedom of SSE.

[0085] Furthermore, generate a second feature selection mask Indicates the result of the significance test of differences, consisting of 0 and 1. 1 indicates selecting the feature at the corresponding position, and 0 indicates removing the feature at the corresponding position.

[0086] Step 404, perform a feature reproducibility test on the sample set, count the reproducibility of channel features, and generate a third feature selection mask according to the reproducibility and the configured reproducibility threshold to indicate the result of the reproducibility test.

[0087] In the embodiments of the present disclosure, the number of features that simultaneously meet the requirements of normal distribution and significant differences in channel features under each stratified sampling is sequentially added to obtain the number of channel features that meet the configured normality requirements and significant differences under different stratified conditions, and then output as a feature selection mask by the configured reproducibility threshold.

[0088] In some embodiments, performing a feature reproducibility test on the sample set, counting the reproducibility of channel features, and generating a third feature selection mask according to the reproducibility and the configured reproducibility threshold to indicate the result of the reproducibility test includes: counting the number of each channel feature that simultaneously meets the requirements of normal distribution and significant differences in all stratifications, and combining the number of all stratifications to obtain the reproducibility; if the reproducibility meets the configured reproducibility threshold, determine that the channel feature passes the reproducibility test, and the third feature selection mask is 1, or, if the reproducibility does not meet the configured reproducibility threshold, determine that the channel feature fails the reproducibility test, and the third feature selection mask is 0.

[0089] In the embodiments of the present disclosure, the output of feature preference statistics is: ; Among them, Is the input feature matrix of the i-th sample, and isnan means removing the 0 elements in the matrix, Is all the input features of the i-th sample, 、 、 They are respectively the results of normality test, significance test of differences, and replication rate test, which are composed of 0 and 1. 1 indicates selecting the feature at the corresponding position, while 0 indicates removing the feature at the corresponding position.

[0090] Step 104: Use a support vector machine as the classifier, input the set of preferred features as the training set into the classifier for model training, and obtain a trained model.

[0091] According to the embodiments of the present disclosure, use a support vector machine as the classifier, input the set of preferably selected features as the training set into the classifier to train the model parameters, use the newly input data as the test set, and verify the model recognition accuracy. Take 70% of the entire data set as the training set and 30% as the validation set. Input the training set into the model to train the model parameters, input the validation set data into the trained model, and use precision recall score and accuracy ; ; ; ; where TP (True Positive) is the number of samples that the model correctly predicts as the positive class, FP (False Positive) is the number of samples that the model incorrectly predicts as the positive class but is actually the negative class, FN (false negative) is the number of positive class samples missed by the model, and TN (true negative) is the number of negative class samples correctly excluded by the model. In the formula for calculating represents accuracy and R represents recall .

[0092] According to the embodiments of the present disclosure, the maximum classification accuracy for the left - hand grip - right - hand grip binary classification task is 91%, the maximum classification accuracy for the left - hand tap - right - hand tap binary classification task is 91%, the maximum classification accuracy for the rest - grip - tap ternary classification task is 93%, and the maximum classification accuracy for the rest - left - hand grip - right - hand grip - left - hand tap - right - hand tap quinary classification task is 65%.

[0093] Step 105: Based on the trained model, perform task classification on the set of preferred features to be recognized to obtain a classification result, and output the corresponding motion intention of the classification result to the user interface.

[0094] Figure 6Schematically shown is a motion intention recognition system 500 for a brain-computer interface according to an embodiment of the present disclosure, as Figure 6 shown. The system includes: A data acquisition module 510, which is used to acquire the functional near-infrared spectroscopy signals of the frontal brain region of the user under a paradigm task, obtain the original light intensity data, preprocess the light intensity data, and convert the preprocessed data into a hemodynamic response signal; A feature extraction module 520, which is used to perform various feature extractions on the hemodynamic response signal respectively, and combine the extracted various features into a feature matrix. Among them, feature extraction includes time-domain feature extraction, frequency-domain feature extraction, network feature extraction, and curve fitting parameter extraction; A feature statistics and optimization module 530, which is used to perform statistical feature optimization on the feature matrix to generate a feature selection mask, and determine the optimized feature set through the feature selection mask. Statistical feature optimization includes data set stratification, normality test, significance test, and feature reproducibility test; A training module 540, which is used to use a support vector machine as a classifier, input the optimized feature set as a training set into the classifier for model training, and obtain a trained model; A classification module 550, which is used to perform task classification on the optimized feature set to be recognized based on the trained model to obtain a classification result, and output the motion intention corresponding to the classification result to the user interface.

[0095] In some embodiments, the system further includes a parameter setting module, which is used to set the parameters of the paradigm prompt on the real-time classification interface according to the experimental paradigm design to guide the user to perform the paradigm task. Among them, parameter setting includes classification task setting, paradigm trial setting, baseline time setting, task time setting, and rest time setting.

[0096] In some embodiments, the data acquisition module 510 is specifically used to: extract the light intensity data before the end moment of each trial to obtain partial light intensity data; perform filtering processing on the partial light intensity data to remove physiological noise, baseline drift, and spike noise in the partial light intensity data to obtain filtered data; use the modified Lambert-Beer law to convert the filtered data into a blood oxygen concentration signal, where the blood oxygen concentration signal includes oxyhemoglobin concentration data and deoxyhemoglobin concentration data.

[0097] In some embodiments, the feature extraction module 520 is specifically configured to: extract time-domain features from the hemodynamic response signal to obtain the time information of the hemodynamic response signal, where the time information includes mean, variance, slope, peak value, peak time, kurtosis, quadratic coefficient, and approximate entropy; extract frequency-domain features from the hemodynamic response signal to obtain the frequency information of the hemodynamic response signal, where the frequency information includes center frequency, mean square frequency, and frequency variance; extract network features from the hemodynamic response signal to obtain the spatial information of the hemodynamic response, and the spatial information includes node importance, average path length, clustering coefficient, node clustering coefficient, average degree, and degree distribution; and extract the curve fitting parameters of the hemodynamic response signal using a deep learning network.

[0098] In some embodiments, the deep learning network includes a multi-layer perceptron and a long short-term memory network. Extracting the curve fitting parameters of the hemodynamic response signal using a deep learning network includes: simultaneously using the multi-layer perceptron and the long short-term memory network to fit the response curve of each channel for the hemodynamic response signal, and using the trained network parameters as features to form a feature matrix. Among them, the network of the multi-layer perceptron fitting the response curve includes an input layer, a hidden layer with 10 neurons, and an output layer, and the network of the long short-term memory network fitting the response curve includes a sequence input layer, a long short-term memory layer, a rectified linear unit activation layer, a fully connected layer, and a regression layer.

[0099] In some embodiments, the feature statistics optimization module 530 is specifically configured to: use the feature matrix obtained from valid trials as samples, perform stratified sampling on the data set including all samples according to a set rule to obtain a stratified data subset, and combine each data subset with the same samples according to the labels and classification tasks of the samples to obtain a sample set; perform a normality test on the sample set to determine whether the channel feature distribution of the sample set conforms to a normal distribution, and generate a first feature selection mask to indicate the result of the normality test; perform a significance test on the sample set to determine whether the channel feature distributions under different labels in the sample set conform to significance, and generate a second feature selection mask to indicate the result of the significance test; perform a feature reproducibility test on the sample set, count the reproducibility of the channel features, and generate a third feature selection mask according to the reproducibility and the configured reproducibility threshold to indicate the result of the reproducibility test.

[0100] In some embodiments, a normality test is performed on the sample set to determine whether the channel feature distribution of the sample set conforms to a normal distribution, and a first feature selection mask is generated to indicate the normality test result, including: for each stratified sample set, the sample set is grouped according to different labels; the Jarque–Bera test method is independently used for each group of samples to calculate the p-value; if the p-value meets the configured first floating threshold, it is determined that the channel feature distribution within the current sample group conforms to a normal distribution, and the first feature selection mask is 1, or, if the p-value does not meet the configured first floating threshold, it is determined that the channel feature distribution within the current sample group does not conform to a normal distribution, and the first feature selection mask is 0.

[0101] In some embodiments, a significance test is performed on the sample set to determine whether the channel feature distributions under different labels in the sample set conform to significance, and a second feature selection mask is generated to indicate the significance result, including: for each stratified sample set, the sample set is grouped according to different classification tasks; the t-test method and the analysis of variance method are respectively used for each group of samples to calculate the p-value; if the p-value meets the configured second floating threshold, it is determined that the channel feature distribution within the current sample group conforms to significance, and the second feature selection mask is 1, or, if the p-value does not meet the configured second floating threshold, it is determined that the channel feature distribution within the current sample group does not conform to significance, and the second feature selection mask is 0.

[0102] In some embodiments, a feature reproducibility test is performed on the sample set to count the reproducibility of the channel features, and a third feature selection mask is generated according to the reproducibility and the configured reproducibility threshold to indicate the reproducibility test result, including: counting the number of channel features that simultaneously conform to a normal distribution and significance in all stratifications, and combining the number of all stratifications to obtain the reproducibility; if the reproducibility meets the configured reproducibility threshold, it is determined that the channel feature passes the reproducibility test, and the third feature selection mask is 1, or, if the reproducibility does not meet the configured reproducibility threshold, it is determined that the channel feature fails the reproducibility test, and the third feature selection mask is 0.

[0103] According to an embodiment of the present disclosure, a real-time class interface for motion intention is provided, including a paradigm setting module, a data reading module, an identification module, and a classification result presentation module, wherein the paradigm setting module is used to set relevant parameters for paradigm prompt presentation, the data reading module is used to read the signals collected by the fNIRS device in real time, the identification module is used to input the collected raw data into a classifier after preprocessing, feature extraction, and feature optimization to obtain a classification result, and the classification result presentation module is used to present the classification result in a visual form in real time.

[0104] According to an embodiment of the present disclosure, the paradigm setting module includes classification task setting, paradigm trial setting, baseline time setting, task time setting and rest time setting, wherein the classification task setting is used to deploy the algorithm model according to the selected classification task and change the image presentation of the classification result presentation module to adapt to the selected task, the paradigm trial setting is used to configure the number of cycles of task prompts and rest prompts on the paradigm prompt screen, the baseline time setting is used to configure the duration of the baseline resting state before the start of the paradigm, the task time setting is used to set the duration of the task prompt presentation on the paradigm prompt screen, and the rest time setting is used to set the duration of the rest prompt presentation on the paradigm prompt screen.

[0105] According to an embodiment of the present disclosure, the data reading module reads data in real time according to the storage path when the fNIRS device collects data through the classification interface, and inputs the data into the recognition module.

[0106] According to an embodiment of the present disclosure, the recognition module first performs data segmentation on the input reading data according to the number of completed paradigm task presentations, retains the data before the end of the last rest presentation and inputs it into the preprocessing module, and after feature extraction and feature selection, inputs it into the trained support vector machine model to obtain the output classification result and input it into the classification result presentation module.

[0107] According to the embodiments of the present disclosure, the classification result presentation module reads the classification results according to the set classification tasks, selects the corresponding hand movement pictures that meet the classification results and presents them in the interface in real time, and finally saves the data and classification results generated during the operation through MATLAB.

[0108] Due to the adoption of the above technical solution, the technical progress achieved by the present invention is: The present disclosure provides a method, device, and real-time classification interface for identifying movement intentions in a brain-computer interface, and provides an fNIRS-BCI movement intention recognition scheme based on feature statistical optimization that can improve the performance of hand movement intention recognition, so as to solve the contradictory limitations of the related art in that the accuracy of fNIRS recognition of complex movement classification tasks in the frontal lobe region is low and the adequacy of feature extraction and the difficulty in taking into account both the efficiency of operation. The method has high reliability, can meet the optimization of the classification effects of multiple individuals, and at the same time achieves generalization performance, and can train an overall model suitable for all multiple individuals. A comprehensive feature extraction method and a feature statistical optimization method with high reliability and validity are provided, which meet the optimization of the classifier training effect after feature optimization while maintaining a high interpretability, and support a more intuitive understanding of the relationship between statistical parameters and the classifier training effect. The designed real-time classification interface can display the subject's action execution intention in real time using the trained classification model, and provide important tools and method references for practical brain-computer interface applications for movement decision recognition.

[0109] First, use the fNIRS acquisition device to collect brain activation data in the frontal lobe area of the subject according to the paradigm requirements. After preprocessing to identify and filter out non-task-related noise, use the modified Lambert-Beer law to achieve the conversion between the light intensity signal and the blood oxygen concentration signal.

[0110] Then, the network parameters obtained through the time feature extraction module, frequency feature extraction module, network feature extraction module, and deep learning methods such as MLP and LSTM network fitting training are combined into a feature matrix. Each element of the combined feature matrix is respectively subjected to a normality test, a significance test of differences, and a reproducibility test. By adjusting three statistical parameters, namely the significance level of the normality test, the significance level of the significance test of differences, and the number of feature reproduction times, the construction of the feature selection mask matrix is completed. According to the mask matrix, the features to be retained and the redundant features to be removed are selected, and the classifier training is completed by the support vector machine model.

[0111] Finally, deploy the trained model to the real-time classification interface. After configuring the paradigm parameters, the subject starts to complete the paradigm task according to the prompts presented on the screen. The interface reads the real-time collected data and, after the same preprocessing and feature extraction, outputs the classification result and displays it on the interface.

[0112] Figure 7 FIG. shows a schematic block diagram of an example electronic device 600 that can be used to implement embodiments of the present disclosure. The electronic device is intended to represent various forms of digital computers, such as, for example, a laptop computer, a desktop computer, a workbench, a personal digital assistant, a server, a blade server, a mainframe computer, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as, for example, a personal digital processor, a cellular phone, a smart phone, a wearable device, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely exemplary and are not intended to limit the implementations of the present disclosure described and / or claimed herein.

[0113] As Figure 7 shown, the device 600 includes a computing unit 601 that can perform various appropriate actions and processes according to a computer program stored in a ROM (Read-Only Memory) 602 or a computer program loaded from a storage unit 608 into a RAM (Random Access Memory) 603. In the RAM 603, various programs and data required for the operation of the device 600 can also be stored. The computing unit 601, the ROM 602, and the RAM 603 are connected to each other via a bus 604. An I / O (Input / Output) interface 605 is also connected to the bus 604.

[0114] Multiple components in device 600 are connected to I / O interface 605, including: an input unit 606, such as a keyboard, a mouse, etc.; an output unit 607, such as various types of displays, speakers, etc.; a storage unit 608, such as a disk, an optical disc, etc.; and a communication unit 609, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 609 allows device 600 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0115] The computing unit 601 can be various general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include but are not limited to a CPU (Central Processing Unit), a GPU (Graphic Processing Units), various dedicated AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, a DSP (Digital Signal Processor), and any suitable processor, controller, microcontroller, etc. The computing unit 601 executes the various methods and processes described above, such as the motion intention recognition method for a brain-computer interface. For example, in some embodiments, the motion intention recognition method for a brain-computer interface can be implemented as a computer software program that is tangibly contained in a machine-readable medium, such as the storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed onto device 600 via the ROM 602 and / or the communication unit 609. When the computer program is loaded into the RAM 603 and executed by the computing unit 601, one or more steps of the methods described above can be executed. Alternatively, in other embodiments, the computing unit 601 can be configured to execute the aforementioned motion intention recognition method for a brain-computer interface in any other suitable manner (e.g., by means of firmware).

[0116] The various embodiments of the systems and techniques described above in this document may be implemented in digital electronic circuitry, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application Specific Standard Products), SOCs (System On Chip), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include: being implemented in one or more computer programs that may be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a special-purpose or general-purpose programmable processor that receives data and instructions from a storage system, at least one input device, and at least one output device, and transmits the data and instructions to the storage system, the at least one input device, and the at least one output device.

[0117] The program code for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when the program codes are executed by the processor or controller, the functions / operations specified in the flowchart and / or block diagram are implemented. The program code may be executed entirely on the machine, partially on the machine, as a stand-alone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0118] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in connection with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of a machine-readable storage medium would include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a RAM, a ROM, an EPROM (Electrically Programmable Read-Only-Memory), or a flash memory, an optical fiber, a CD-ROM (Compact Disc Read-Only Memory), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0119] In order to provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (Cathode-Ray Tube) or an LCD (Liquid Crystal Display) monitor); and a keyboard and a pointing device (e.g., a mouse or a trackball), by which the user can provide input to the computer. Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, speech input, or tactile input).

[0120] The systems and techniques described herein can be implemented in a computing system including backend components (e.g., as a data server), or a computing system including middleware components (e.g., an application server), or a computing system including frontend components (e.g., a user computer having a graphical user interface or a web browser through which the user can interact with an implementation of the systems and techniques described herein), or a computing system including any combination of such backend components, middleware components, or frontend components. The components of the system can be interconnected to each other by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a LAN (Local Area Network), a WAN (Wide Area Network), the Internet, and a blockchain network.

[0121] A computer system may include a client and a server. The client and the server are generally far from each other and usually interact through a communication network. The client-server relationship is generated by computer programs running on respective computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system, solving the defects of difficult management and weak business scalability existing in traditional physical hosts and VPS services ("Virtual Private Server", or simply "VPS"). The server may also be a server of a distributed system or a server combined with a blockchain.

[0122] Among them, it should be noted that artificial intelligence is a discipline that studies how to make a computer simulate certain thinking processes and intelligent behaviors of humans (such as learning, reasoning, thinking, planning, etc.), and it has both hardware-level technologies and software-level technologies. Artificial intelligence hardware technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, and big data processing; artificial intelligence software technologies mainly include several major directions such as computer vision technology, natural language processing technology, and machine learning / deep learning, big data processing technology, and knowledge graph technology.

[0123] It should be understood that various forms of processes shown above can be used, steps can be reordered, added, or deleted. For example, the steps recorded in the present disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved, and no limitations are imposed herein.

[0124] The above specific embodiments do not constitute a limitation on the protection scope of the present disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present disclosure shall be included within the protection scope of the present disclosure.

Claims

1. A method for recognizing movement intention for a brain-computer interface, characterized in that, The method includes: Acquiring original light intensity data by collecting functional near-infrared spectroscopy signals of the frontal brain region of a user under a paradigm task, preprocessing the light intensity data, and converting the preprocessed data into a hemodynamic response signal; Performing multiple feature extractions on the hemodynamic response signal respectively, and combining the multiple extracted features into a feature matrix, wherein the feature extraction includes time-domain feature extraction, frequency-domain feature extraction, network feature extraction, and curve fitting parameter extraction; Performing statistical feature optimization on the feature matrix to generate a feature selection mask, and determining a preferred feature set based on the feature selection mask, wherein the statistical feature optimization includes data set stratification, normality test, significance test, and feature reproducibility test; Using a support vector machine as a classifier, inputting the preferred feature set as a training set into the classifier for model training to obtain a trained model; Based on the trained model, performing task classification on the preferred feature set to be recognized to obtain a classification result, and outputting the motion intention corresponding to the classification result to a real-time classification interface.

2. The motion intention recognition method for a brain-computer interface according to claim 1, characterized in that Wherein, Before collecting the functional near-infrared spectroscopy signals of the frontal brain region of the user under the paradigm task, the method further includes: According to the experimental paradigm design, setting parameters for the paradigm prompt on the real-time classification interface to guide the user to execute the paradigm task, wherein the parameter setting includes classification task setting, paradigm trial setting, baseline time setting, task time setting, and rest time setting.

3. The method for recognizing movement intention for a brain-computer interface according to claim 2, wherein, Wherein, Preprocessing the light intensity data and converting the preprocessed data into a hemodynamic response signal includes: Selecting the light intensity data before the end moment of each trial to obtain partial light intensity data; Performing filtering processing on the partial light intensity data to remove physiological noise, baseline drift, and spike noise in the partial light intensity data to obtain filtered data; Converting the filtered data into a blood oxygen concentration signal by using a modified Lambert-Beer law, wherein the blood oxygen concentration signal includes oxyhemoglobin concentration data and deoxyhemoglobin concentration data.

4. The method for recognizing movement intention for a brain-computer interface according to claim 1, wherein, Wherein, Performing multiple feature extractions on the hemodynamic response signal respectively includes: Performing time-domain feature extraction on the hemodynamic response signal to obtain the time information of the hemodynamic response signal, wherein the time information includes mean, variance, slope, peak value, peak time, kurtosis, quadratic term coefficient, and approximate entropy; Performing frequency-domain feature extraction on the hemodynamic response signal to obtain the frequency information of the hemodynamic response signal, wherein the frequency information includes center frequency, mean square frequency, and frequency variance; Performing network feature extraction on the hemodynamic response signal to obtain the spatial information of the hemodynamic response, and the spatial information includes node importance, average path length, clustering coefficient, node clustering coefficient, average degree, and degree distribution; Extracting the curve fitting parameters of the hemodynamic response signal by using a deep learning network.

5. The method for recognizing movement intention for a brain-computer interface according to claim 4, characterized in that, Wherein, The deep learning network includes a multi-layer perceptron and a long short-term memory network. The curve fitting parameters of the hemodynamic response signal extracted by using the deep learning network include: Using both the multi-layer perceptron and the long short-term memory network to fit the response curves of each channel for the hemodynamic response signal, and taking the trained network parameters as features to form a feature matrix. Among them, the network of the multi-layer perceptron for fitting the response curve includes an input layer, a hidden layer with 10 neurons, and an output layer. The network of the long short-term memory network for fitting the response curve includes a sequence input layer, a long short-term memory layer, a rectified linear unit activation layer, a fully connected layer, and a regression layer.

6. The method for recognizing movement intention for a brain-computer interface according to claim 1, characterized in that, Among them, Statistically optimizing the features of the feature matrix to generate a feature selection mask includes: Taking the feature matrix obtained from valid trials as samples, performing stratified sampling on the data set including all samples according to a set rule to obtain a stratified data subset, and combining each data subset with the same samples according to the labels and classification tasks of the samples to obtain a sample set. Performing a normality test on the sample set to determine whether the channel feature distribution of the sample set conforms to a normal distribution, and generating a first feature selection mask to indicate the result of the normality test. Performing a significance test on the sample set to determine whether the channel feature distributions under different labels in the sample set conform to significance, and generating a second feature selection mask to indicate the result of the significance test. Performing a feature reproducibility test on the sample set, counting the reproducibility of channel features, and generating a third feature selection mask according to the reproducibility and the configured reproducibility threshold to indicate the result of the reproducibility test.

7. The method for recognizing movement intention for a brain-computer interface according to claim 6, wherein, Among them, Performing a normality test on the sample set to determine whether the channel feature distribution of the sample set conforms to a normal distribution, and generating a first feature selection mask to indicate the result of the normality test includes: For each stratified sample set, grouping the sample set according to different labels. Independently using the Jarque–Bera test method to calculate the p-value for each group of samples. If the p-value meets the configured first floating threshold, it is determined that the channel feature distribution within the current sample group conforms to a normal distribution, and the first feature selection mask is 1, or If the p-value does not meet the configured first floating threshold, it is determined that the channel feature distribution within the current sample group does not conform to a normal distribution, and the first feature selection mask is 0.

8. The motion intention recognition method for a brain-computer interface according to claim 6, characterized in that, Among them, Performing a significance test on the sample set to determine whether the channel feature distributions under different labels in the sample set conform to significance, and generating a second feature selection mask to indicate the result of the significance test includes: For each stratified sample set, grouping the sample set according to different classification tasks. Respectively using the t-test method and the analysis of variance method to calculate the p-value for each group of samples. If the p-value meets the configured second floating threshold, it is determined that the channel feature distribution within the current sample group conforms to significance, and the second feature selection mask is 1, or If the p-value does not meet the configured second floating threshold, it is determined that the channel feature distribution within the current sample group does not conform to significance, and the second feature selection mask is 0.

9. The method for recognizing movement intention for a brain-computer interface according to claim 6, characterized in that, Among them, Perform a feature reproduction rate test on the sample set, count the reproduction rate of channel features, and generate a third feature selection mask based on the reproduction rate and the configured reproduction rate threshold to indicate that the reproduction rate test result includes: Count the number of channel features that simultaneously meet the normal distribution and significance of differences in each layer, and combine the number of all layers to obtain the reproduction rate; If the reproduction rate meets the configured reproduction rate threshold, determine that the channel feature passes the reproduction rate test, and the third feature selection mask is 1, or, If the reproduction rate does not meet the configured reproduction rate threshold, determine that the channel feature fails the reproduction rate test, and the third feature selection mask is 0.

10. A motion intention recognition system for a brain-computer interface, characterized in that, Including: A data acquisition module, which is used to obtain original light intensity data by collecting the functional near-infrared spectroscopy signals of the frontal brain region of the user under the paradigm task, preprocess the light intensity data, and convert the preprocessed data into hemodynamic response signals; A feature extraction module, which is used to perform various feature extractions on the hemodynamic response signals respectively, and combine the extracted various features into a feature matrix, where the feature extraction includes time-domain feature extraction, frequency-domain feature extraction, network feature extraction, and curve fitting parameter extraction; A feature statistics and optimization module, which is used to perform statistical feature optimization on the feature matrix to generate a feature selection mask, and determine the optimized feature set through the feature selection mask. The statistical feature optimization includes data set stratification, normality test, significance test, and feature reproduction rate test; A training module, which is used to use a support vector machine as a classifier, input the optimized feature set as a training set into the classifier for model training, and obtain a trained model; A classification module, which is used to perform task classification on the optimized feature set to be recognized based on the trained model to obtain a classification result, and output the corresponding motion intention of the classification result to the user interface.

Images