Hand motion fNIRS signal classification method based on t distribution optimization feature extraction

Abstract

The invention discloses a hand motion FNIRS signal classification method based on t distribution optimization feature extraction, and the method specifically comprises the steps: employing a portable functional near-infrared collection device to collect an optical density signal of a forehead cortex brain region during the motion execution period during the operation of the device; after pretreatment, the concentration of oxyhemoglobin HBO and the concentration of deoxidized hemoglobin HBR are calculated according to the corrected Beer-Lambert law; according to motion execution and resting state segmentation data, time domain features and time-frequency domain features of oxyhemoglobin concentration data signals of each channel of each experiment test are extracted, and a first feature matrix is formed; performing kernel density estimation and kurtosis and skewness test on the first feature matrix to obtain a second feature matrix, performing t distribution on each corresponding feature satisfying normal distribution in the second feature matrix of the motion execution and resting states, and optimizing feature selection on the basis to obtain an optimized third feature matrix; the optimized third feature matrix is used as classifier input, a brain-computer interface system model training result is obtained, and effective feature selection and extraction are achieved; according to the method, the brain-computer interface system feature extraction of the hand motion FNIRS signals can be optimized, so that the classification performance is improved.

Description

Technical Field

[0001] This invention belongs to the technical field of fNIRS brain-computer interface systems for motion pattern recognition, specifically involving a method for classifying hand motion fNIRS signals by t-distribution optimized feature extraction. Background Technology

[0002] In the field of near-infrared spectroscopy (fNIRS) brain-computer interface systems for motion pattern recognition, conventional feature extraction methods, under conditions of data reconstruction, can generally complete the recognition of motor execution and motor imagery, and the classification of different motor execution and motor image content. Brain activation during motor execution tasks can serve as a reference activation pattern for motor imagery and motor observation. In existing fNIRS applications, input signals are generally acquired from multiple cortical regions. It is generally believed that the hemodynamic response of the primary motor cortex is more significant in motor execution and motor imagery, but the participation of the prefrontal cortex in the task process can also be observed. Compared to signals from other cortical regions, brain activation signals generated by the prefrontal cortex can now be acquired using portable devices. However, methods that directly use all time-domain or frequency-domain features as input have some problems:

[0003] (1) The classification accuracy is not high in small and medium-sized sample sets;

[0004] (2) The prefrontal cortex is involved in higher cognitive activities and its response is affected by motor proficiency, which may lead to instability of effective features between different sample sets during the measurement process.

[0005] Many published patents have proposed various methods for classifying hand motor imagery based on electroencephalogram (EEG) signals. However, due to the high motor sensitivity of EEG signals, practical applications face numerous inconveniences. A brain-computer interface (BCI) study on motor execution and motor imagery reported successfully classifying various motor execution and motor imagery patterns by training a BCI system model using temporal features of hemodynamic signals from the primary motor cortex and prefrontal cortex. However, this study simultaneously collected signals from both the prefrontal cortex and the primary motor cortex, resulting in a larger total data volume and longer computation time compared to the prefrontal cortex alone.

[0006] Chinese patent application CN116269366A discloses a method for optimizing and screening brain region channels in fNIRS analysis. The method, based on the temporal characteristics of preprocessed blood oxygen concentration signals, first uses the least squares method for feature fusion, then uses a permutation and combination method to construct a first set of ROI regions, containing all channel selection combinations for each ROI region. A preorder traversal algorithm is then used to construct a second set of ROI regions. A local-global optimization method is employed to find the maximum values ​​of local evaluation indicators and global brain evaluation indicators for each ROI region, thus finding the optimal channel combination for each brain ROI region and optimizing channel screening. However, this method still suffers from cumbersome computation, and the local evaluation indicators only reflect the degree of activation and not the specific activation pattern. Since it only screens channels, it may ignore many effective features while inputting invalid features when applied to brain-computer interface systems. These problems lead to a decrease in the classification accuracy and stability of brain-computer interface systems.

[0007] Therefore, existing technologies need a feature extraction method that can select effective features to improve the classification accuracy and stability of brain-computer interface systems while ensuring data acquisition efficiency, and at the same time, effectively classify the hemodynamic signal response of prefrontal lobe hand movement. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a feature extraction method for a brain-computer interface system based on FNIRS that can select effective features, in order to overcome the limitation of traditional brain-computer interface systems in classifying hand movements based on FNIRS signals from the prefrontal cortex, where input redundancy leads to a decrease in classification accuracy and stability.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0010] A hand motion FNIRS signal classification method based on t-distribution optimized feature extraction includes the following steps:

[0011] Step 1: Use a portable functional near-infrared acquisition device to acquire light intensity signals of the prefrontal cortex brain region during motor execution when the device is operating. After preprocessing, calculate the oxyhemoglobin concentration (HBO) and deoxyhemoglobin concentration (HBR) according to the modified Beer-Lambert law.

[0012] Step 2: Based on the segmented data of exercise execution and resting state, extract the time-domain and frequency-domain features of the oxyhemoglobin concentration signal of each channel in each experimental trial to form the first feature matrix;

[0013] Step 3: Perform kernel density estimation and kurtosis and skewness tests on the first feature matrix to obtain the second feature matrix. For each feature in the second feature matrix that satisfies the normal distribution, take the t-distribution. Based on this, optimize the feature selection to obtain the optimized third feature matrix.

[0014] Step 4: Use the optimized third feature matrix as input to the classifier to obtain the training results of the brain-computer interface system model, thereby achieving effective feature selection and extraction.

[0015] A further improvement of this invention lies in the following steps: In step 1, the acquisition of light intensity signals from the prefrontal cortex brain region during motor execution while the device is operating, followed by preprocessing and calculation of oxyhemoglobin concentration (HBO) and deoxyhemoglobin concentration (HBR) based on the modified Beer-Lambert law, specifically includes the following steps:

[0016] Step 11: Set the acquisition device parameters, using near-infrared light with wavelengths of 775 and 855 nm, and set the differential path factor values ​​to 6.35 and 5.88 respectively, with an integration time of 500 ms;

[0017] Step 12: Collect training data and perform tasks on two types of movements: grasping and finger tapping. Each trial includes 10 sets of movement executions and 10 sets of resting states, with each set lasting 10 seconds.

[0018] Step 13: Remove physiological noise and baseline shift. Use an 8th-order 0.45Hz Butterworth low-pass filter to remove physiological noise caused by breathing and heartbeat, as well as low-frequency baseline drift from the data.

[0019] Step 14: Remove spike noise. Combine moving window standard deviation with Sobel filter to detect spike noise, and use SG filtering method to remove spike noise.

[0020] Step 15: Calculate concentration data. Using the first 30 seconds of each test as the baseline time, calculate the oxyhemoglobin concentration (HBO) and deoxyhemoglobin concentration (HBR) according to the modified Beer-Lambert law.

[0021] A further improvement of the present invention lies in the following steps: In step 2, the extraction of time-domain and frequency-domain features of the oxyhemoglobin concentration signal of each channel in each experimental trial, based on the segmented data of motion execution and resting state, to form a first feature matrix, specifically includes the following steps:

[0022] Step 21: Segment the data. Segment the data according to the motion execution and resting state to obtain 20 sampling points for each motion execution and resting state group. In each experimental run, a set of 460 sampling points from 10 channels can be obtained. in This represents the set of sampling points on the nth channel of the i-th resting state group in an experiment. This represents the set of sampling points of the nth channel in the i-th occurrence of a motion execution group in an experiment.

[0023] Step 22: Extract time-domain features. The time-domain features used in this invention include mean, variance, kurtosis, skewness, peak value, peak time, coefficient of the highest-order term in the polynomial fitting, and approximate entropy, constructing a time-domain feature vector for each experimental trial. This represents the eight temporal features extracted from the i-th resting state group in the n-th channel. The eight temporal features extracted from the i-th occurrence of the motion in the n-th channel are represented by a1, a2, a3, kurtosis, a4, skewness, a5, peak value, a6, peak time, a7, coefficient of the highest-order term of the polynomial fitting, and a8, approximate entropy.

[0024] Step 23: Extract frequency domain features. The frequency domain features used in this invention include the mean of the time series corresponding to each frequency band after wavelet transform and the wavelet coherence coefficients between channels, constructing a time-frequency domain feature vector for each experimental trial. This represents the (k+100) frequency domain features extracted from the i-th occurrence of the resting state group in the n-th channel. This represents the (k+100) frequency domain features extracted from the motion execution group when the nth channel appears for the ith time, where k is the number of frequency bands obtained after wavelet transformation, and the 100 features after the kth feature are the wavelet coherence coefficients between each pair of the 10 channels.

[0025] Step 24: Combine the feature matrix by combining all time-domain and frequency-domain features of all trials into a single feature vector. This combination yields the resting state group vector for the i-th occurrence of the n-th channel. The motion execution group that appears for the i-th time in channel n. This leads to the resting state feature matrix, composed of the feature vectors of the nth channel in the tth experimental trial. and motion execution feature matrix

[0026]

[0027] The combination of T trials can be used to obtain the first characteristic matrix. and

[0028] A further improvement of the present invention lies in the following steps: In step 3, the kernel density estimation and kurtosis and skewness tests are performed on the first feature matrix to obtain the second feature matrix. For each feature in the second feature matrix corresponding to the motion execution and resting state that satisfies the normal distribution, a t-distribution is taken. Based on this, feature selection is optimized to obtain the optimized third feature matrix. Specifically, this includes the following steps:

[0029] Step 31: Perform kernel density estimation on the first feature matrix, and perform kernel density estimation and fit the probability density curve for each feature of each group under each channel. The kernel density estimation formula is as follows:

[0030]

[0031] Where K is the kernel function, K h Here, h is the scaling kernel function, and h is the bandwidth smoothing parameter.

[0032] Step 32: Examine the distribution type of the feature samples. Kurtosis and skewness tests are performed to check whether the feature samples meet the normal distribution. Using the standard normal distribution as a reference, kurtosis and skewness tests are performed on the fitted probability density curve to screen features that meet the normal distribution. The formulas for kurtosis and skewness tests are as follows:

[0033] Skewness:

[0034] Kuroshi:

[0035] Skewness test:

[0036] Kurtosis test:

[0037] Z s and Z k A value greater than 1.96 indicates that the sample distribution is not normally distributed, where n represents the sample size. The standard error coefficient;

[0038] With CRT n and CMT n The corresponding column vectors are used as a set of samples, and the second feature matrix that conforms to a normal distribution is selected after calculation as CRT1. n and CMT1 n , where n represents the nth channel, and the matrix has the following relationship:

[0039] CRT1 1 ∈CRT 1 CRT1 2 ∈CRT 2 ,…,CRT1 10 ∈CRT 10

[0040] CMT1 1 ∈CMT 1 CMT1 2 ∈CMT 2 CMT1 10 ∈CMT 10

[0041] Step 33: Calculate the t-value. Calculate the t-distribution for the feature samples of each group of sub-task execution and resting states. The calculation formula is as follows:

[0042]

[0043] Where n1 is the number of samples X and n2 is the number of samples Y;

[0044] Step 34: Set the feature selection threshold, verify the significance of the difference with P < 0.01, and reorganize the features that meet the requirements into a new feature matrix, thus obtaining the third feature matrix CRT2. n and CMT2 n .

[0045] A further improvement of this invention lies in step 4, where the optimized third feature matrix is ​​used as the input to the classifier to obtain the brain-computer interface system model training result, thereby achieving effective feature selection and extraction. This specifically includes the following steps:

[0046] Step 41, Classifier selection, using the third feature matrix CRT2 n and CMT2 n The inputs are fed into a decision tree, a K-nearest neighbor algorithm, and a support vector machine for training, respectively.

[0047] Step 42: Verify model performance. Use cross-validation to verify model performance and select the model with the best classification performance among each classification method.

[0048] Step 43: Test model performance. Collect test data using the same process as collecting training data. Construct the feature matrix of the test data based on the optimized feature selection method to test model performance.

[0049] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:

[0050] (1) This invention provides a feature extraction method for a brain-computer interface system based on FNIRS that can select effective features, in order to solve the limitation of traditional brain-computer interface systems in classifying hand movements based on FNIRS signals from the prefrontal cortex, where input redundancy leads to a decrease in classification accuracy and stability.

[0051] (2) Compared with the training results of typical classifiers, the accuracy of the prefrontal cortex brain activation decoding task can be improved by 7.3% ± 0.97%, the classification performance is significantly improved, the task-related features can be extracted from small and medium-sized samples, the effective model can be trained, and a practical solution for brain-computer interface technology in daily scenarios is provided. Attached Figure Description

[0052] Figure 1 This is a flowchart provided for an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of the preprocessing method and concentration calculation process provided in the embodiments of the present invention;

[0054] Figure 3 This is a schematic diagram of the device acquisition channel location provided in an embodiment of the present invention;

[0055] Figure 4 A schematic diagram illustrating the data segmentation and construction of the first feature matrix provided in this embodiment of the invention;

[0056] Figure 5 This is a schematic diagram illustrating the calculation process of the second and third feature matrices provided in an embodiment of the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0058] In view of the limitations of traditional brain-computer interface systems, such as input redundancy leading to decreased classification accuracy and stability, this invention provides a hand movement fNIRS signal classification method based on t-distribution optimized feature extraction. This method optimizes the effective features extracted by brain-computer interface systems during prefrontal cortex motor execution in fNIRS signal detection.

[0059] The following detailed description of the hand motion FNIRS signal classification method based on t-distribution optimized feature extraction provided by the present invention, in conjunction with the accompanying drawings and specific embodiments, is described in detail.

[0060] To verify the hand motion FNIRS signal classification method based on t-distribution optimized feature extraction described in this invention, an within-group design paradigm was adopted. Hemodynamic signals of grasping movements, finger tapping, and resting states were continuously collected from healthy subjects. Each experimental trial included 10 sets of movement executions and 10 sets of resting states, alternating between the two. The training dataset included a total of 240 sets of movement executions and 160 sets of resting states.

[0061] like Figure 1As shown in the figure, the hand motion FNIRS signal classification method based on t-distribution optimized feature extraction provided by the present invention may include the following steps:

[0062] Step 100: Use a portable functional near-infrared spectroscopy (FIR) acquisition device to acquire light intensity signals from the prefrontal cortex brain region during motor execution while the device is operating. After preprocessing, calculate the oxyhemoglobin concentration (HBO) and deoxyhemoglobin concentration (HBR) according to the modified Beer-Lambert law. Specific steps are as follows: Figure 2 As shown:

[0063] Step 101: Set the acquisition device parameters. Use near-infrared light with wavelengths of 775nm and 855nm, with differential path factors of 6.35 and 5.88 respectively, an integration time of 500ms, and the device acquisition channel position as shown below. Figure 3 As shown.

[0064] Step 102: Collect training data and perform tasks on two types of movements: grasping and finger tapping. Each trial includes 10 sets of movement executions and 10 sets of resting states, with each set lasting 10 seconds.

[0065] Step 103: Remove physiological noise and baseline shift. Use an 8th-order 0.45Hz Butterworth low-pass filter to filter out physiological noise caused by breathing and heartbeat, as well as low-frequency baseline drift in the data.

[0066] Step 104: Remove spike noise. Combine moving window standard deviation with Sobel filter to detect spike noise, and use SG filtering method to remove spike noise.

[0067] Step 105: Calculate concentration data. Using the first 30 seconds of each test as the baseline time, calculate the oxyhemoglobin concentration (HBO) and deoxyhemoglobin concentration (HBR) according to the modified Beer-Lambert law.

[0068] Step 200: Based on the segmented data of exercise execution and resting state, extract the time-domain and frequency-domain features of the oxyhemoglobin concentration signal for each channel in each trial, and form the first feature matrix. Specific steps are as follows: Figure 4 As shown:

[0069] Step 201: Segment the data. Segment the data according to the motion execution and resting state to obtain 20 sampling points for each motion execution and resting state group. In each experimental run, a set of 460 sampling points from 10 channels can be obtained. in This represents the set of sampling points on the nth channel of the i-th resting state group in an experiment. Let represent the set of sampling points of the nth channel of the i-th motion execution group in an experimental trial.

[0070] Step 202: Extract time-domain features. The time-domain features used in this invention include mean, variance, kurtosis, skewness, peak value, peak time, coefficient of the highest-order term in the polynomial fitting, and approximate entropy, constructing a time-domain feature vector for each experimental trial.

[0071] This represents the eight temporal features extracted from the i-th resting state group in the n-th channel. The expression represents the eight temporal features extracted from the i-th occurrence of the motion execution group in the n-th channel, where a1 is the mean, a2 is the variance, a3 is the kurtosis, a4 is the skewness, a5 is the peak value, a6 is the peak time, a7 is the coefficient of the highest-order term of the polynomial fitting, and a8 is the approximate entropy.

[0072] Step 203: Extract frequency domain features. The frequency domain features used in this invention include the mean of the time series corresponding to each frequency band after wavelet transform and the wavelet coherence coefficients between channels, constructing a time-frequency domain feature vector for each experimental trial. This represents the (k+100) frequency domain features extracted from the i-th occurrence of the resting state group in the n-th channel. This represents the (k+100) time-frequency domain features extracted from the i-th occurrence of the motion execution group in the n-th channel, where k is the number of frequency bands obtained after wavelet transformation, and the 100 features after the k-th feature are the wavelet coherence coefficients between each pair of the 10 channels.

[0073] Step 204: Combine the feature matrix by combining all time-domain and frequency-domain features of all trials into a single feature vector. After combination, the resting state group vector of the i-th occurrence of the n-th channel can be obtained. The motion execution group that appears for the i-th time in channel n. This leads to the resting state feature matrix, composed of the feature vectors of the nth channel in the tth experimental trial. and motion execution feature matrix

[0074]

[0075] The combination of T trials can be used to obtain the first characteristic matrix. and

[0076]

[0077] Step 300: Perform kernel density estimation and kurtosis / skewness tests on the first feature matrix to obtain the second feature matrix. For each feature in the second feature matrix corresponding to the motion execution and resting state that satisfies a normal distribution, take a t-distribution. Based on this, optimize feature selection to obtain the optimized third feature matrix. The specific steps are as follows: Figure 5As shown:

[0078] Step 301: Perform kernel density estimation on the first feature matrix, and perform kernel density estimation and fit the probability density curve for each feature of each group under each channel. The kernel density estimation formula is as follows:

[0079]

[0080] Where K is the kernel function, K h Here, h is the scaling kernel function, and h is the bandwidth smoothing parameter.

[0081] Step 302: Examine the distribution type of the feature samples. Kurtosis and skewness tests are performed to check whether the feature samples meet the normal distribution. Using the standard normal distribution as a reference, kurtosis and skewness tests are performed on the fitted probability density curve to screen features that meet the normal distribution. The formulas for kurtosis and skewness tests are as follows:

[0082] Skewness:

[0083] Kuroshi:

[0084] Skewness test:

[0085] Kurtosis test:

[0086] Z s and Z k A value greater than 1.96 indicates that the sample distribution is not normally distributed, where n represents the sample size. The standard error coefficient is denoted as .

[0087] With CRT n and CMT n The corresponding column vectors are used as a set of samples, and the second feature matrix that conforms to a normal distribution is selected after calculation as CRT1. n and CMT1 n , where n represents the nth channel, and the matrix has the following relationship:

[0088] CRT1 1 ∈CRT 1 CRT1 2 ∈CRT 2 ,…,CRT1 10 ∈CRT 10

[0089] CMT1 1 ∈CMT 1 CMT1 2 ∈CMT 2 CMT1 10 ∈CMT 10

[0090] Step 303: Calculate the t-value. For each group of feature samples in both the task execution and resting states, calculate the t-distribution. The calculation formula is as follows:

[0091]

[0092] Where n1 is the number of samples X and n2 is the number of samples Y.

[0093] Step 304: Set the feature selection threshold, verify the significance of the difference with P < 0.01, and reorganize the features that meet the requirements into a new feature matrix, thus obtaining the third feature matrix CRT2. n and CMT2 n .

[0094] Step 400: Use the optimized third feature matrix as input to the classifier to obtain the brain-computer interface system model training results, achieving effective feature selection and extraction. The specific steps are as follows:

[0095] Step 401, Classifier selection, using the third feature matrix CRT2 n and CMT2 n The data are fed into a decision tree, a K-nearest neighbor algorithm, and a support vector machine for training.

[0096] Step 402: Verify model performance. Use cross-validation to verify model performance and select the model with the best classification performance among each classification method.

[0097] Step 403: Test model performance. Collect test data using the same process as collecting training data. Construct the feature matrix of the test data based on the optimized feature selection method to test model performance.

[0098] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows:

[0099] This invention provides a feature extraction method for brain-computer interface systems based on FNIRS that can extract effective features, thereby overcoming the limitations of traditional brain-computer interface systems where input redundancy leads to a decrease in classification accuracy and stability. At the same time, the method is simple and effective, can achieve ideal classification results in dynamic small and medium-sized sample sets, and provides a solution for classifying motor execution using portable devices.

[0100] In the embodiments provided by the present invention, the initial feature selection includes a total of 80 time-domain features and hundreds of frequency-domain features, with each of the 10 channels having 8 time-domain features.

[0101] First, the first feature matrix is ​​extracted from the preprocessed data. Then, the kernel density of the sample distribution of one feature in a single channel is estimated to fit the probability density curve. The curve is then tested for kurtosis and skewness to select the features whose sample distribution satisfies the normal distribution, thus obtaining the second feature matrix.

[0102] Then, the t-distribution is calculated for the resting state feature distribution and motion execution feature distribution of the features retained in the second feature matrix. The difference is verified with P < 0.01, and the part with significant difference in distribution is selected as the third feature matrix.

[0103] Finally, the third feature matrix is ​​used as the training set for the classifier and input into the decision tree, K-nearest neighbor model and support vector machine respectively, and the model with the best performance is selected.

[0104] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Any modifications or improvements made without departing from the spirit of this application shall fall within the scope of protection claimed in this application.

Claims

1. A hand motion fNIRS signal classification method based on t-distribution optimized feature extraction, characterized in that: Includes the following steps: Step 1: Use a portable functional near-infrared acquisition device to acquire light intensity signals of the prefrontal cortex brain region during motor execution when the device is operating. After preprocessing, calculate the oxyhemoglobin concentration (HBO) and deoxyhemoglobin concentration (HBR) according to the modified Beer-Lambert law. Step 2: Based on the segmented data of exercise execution and resting state, extract the time-domain and frequency-domain features of the oxyhemoglobin concentration signal of each channel in each experimental trial to form the first feature matrix; Step 3: Perform kernel density estimation and kurtosis and skewness tests on the first feature matrix to obtain the second feature matrix. For each feature in the second feature matrix that satisfies the normal distribution, take the t-distribution. Based on this, optimize the feature selection to obtain the optimized third feature matrix. Step 4: Use the optimized third feature matrix as the input signal for the classifier to obtain the training results of the brain-computer interface system model, thereby achieving effective feature selection and extraction.

2. A hand motion fNIRS signal classification method based on t-distribution optimized feature extraction, characterized in that: In step 1, a portable functional near-infrared spectroscopy (FIR) acquisition device is used to acquire light intensity signals from the prefrontal cortex brain region during motor execution while the device is operating. After preprocessing, the oxyhemoglobin concentration (HBO) and deoxyhemoglobin concentration (HBR) are calculated according to the modified Beer-Lambert law. The specific steps include: Step 11: Set the acquisition device parameters, using near-infrared light with wavelengths of 775 and 855 nm, and set the differential path factor values ​​to 6.35 and 5.88 respectively, with an integration time of 500 ms; Step 12: Collect training data and perform tasks on two types of movements: grasping and finger tapping. Each trial includes 10 sets of movement executions and 10 sets of resting states, with each set lasting 10 seconds. Step 13: Remove physiological noise and baseline shift. Use an 8th-order 0.45Hz Butterworth low-pass filter to filter out physiological noise caused by breathing and heartbeat, as well as low-frequency baseline drift in the data. Step 14: Remove spike noise. Combine moving window standard deviation with Sobel filter to detect spike noise, and use SG filtering method to remove spike noise. Step 15: Calculate concentration data. Using the first 30 seconds of each test as the baseline time, calculate the oxyhemoglobin concentration (HBO) and deoxyhemoglobin concentration (HBR) according to the modified Beer-Lambert law.

3. A hand motion fNIRS signal classification method based on t-distribution optimized feature extraction, characterized in that: In step 2, based on the segmentation data of exercise execution and resting state, the time-domain and time-frequency domain features of the oxyhemoglobin concentration data of each channel in each experimental trial are extracted to form the first feature matrix. This specifically includes the following steps: Step 21: Segment the data. Segment the data according to the motion execution and resting state to obtain 20 sampling points for each motion execution and resting state group. In each experimental run, a set of 460 sampling points from 10 channels can be obtained. in This represents the set of sampling points on the nth channel of the i-th resting state group in an experiment. This represents the set of sampling points of the nth channel in the i-th occurrence of a motion execution group in an experiment. Step 22: Extract time-domain features. The time-domain features used in this invention include mean, variance, kurtosis, skewness, peak value, peak time, coefficient of the highest-order term in the polynomial fitting, and approximate entropy, constructing a time-domain feature vector for each experimental trial. This represents the eight temporal features extracted from the i-th resting state group in the n-th channel. The eight temporal features extracted from the i-th occurrence of the motion in the n-th channel are represented by a1, a2, a3, kurtosis, a4, skewness, a5, peak value, a6, peak time, a7, coefficient of the highest-order term of the polynomial fitting, and a8, approximate entropy. Step 23: Extract time-frequency domain features. The time-frequency domain features used in this invention include the mean of the time series corresponding to each frequency band after wavelet transform and the wavelet coherence coefficients between channels, constructing a time-frequency domain feature vector for each experimental trial. This represents the (k+100) time-frequency domain features extracted from the i-th occurrence of the resting state group in the n-th channel. This represents the (k+100) time-frequency domain features extracted from the i-th occurrence of the motion execution group in the n-th channel, where k is the number of frequency bands obtained after wavelet transformation, and the 100 features after the k-th feature are the wavelet coherence coefficients between each pair of the 10 channels. Step 24: Combine the feature matrix by combining all time-domain and time-frequency domain features of all trials into a single feature vector. After combination, the resting state group vector of the i-th occurrence of the n-th channel can be obtained. The motion execution group that appears for the i-th time in channel n. This leads to the resting state feature matrix, composed of the feature vectors of the nth channel in the tth experimental trial. and motion execution feature matrix The combination of T trials can be used to obtain the first characteristic matrix. and 4. A hand motion FNIRS signal classification method based on t-distribution optimized feature extraction, characterized in that: In step 3, kernel density estimation and kurtosis and skewness tests are performed on the first feature matrix to obtain the second feature matrix. For each feature in the second feature matrix that satisfies a normal distribution, a t-distribution is taken. Based on this, feature selection is optimized to obtain the optimized third feature matrix. The specific steps include: Step 31: Perform kernel density estimation on the first feature matrix, and perform kernel density estimation and fit the probability density curve for each feature of each group under each channel. The kernel density estimation formula is as follows: Where K is the kernel function, K h Here, h is the scaling kernel function, and h is the bandwidth smoothing parameter. Step 32: Examine the distribution type of the feature samples. Kurtosis and skewness tests are performed to check whether the feature samples meet the normal distribution. Using the standard normal distribution as a reference, kurtosis and skewness tests are performed on the fitted probability density curve to screen features that meet the normal distribution. The formulas for kurtosis and skewness tests are as follows: Skewness: Kuroshi: Skewness test: Kurtosis test: Z s and Z k A value greater than 1.96 indicates that the sample distribution is not normally distributed, where n represents the sample size. The standard error coefficient; With CRT n and CMT n The corresponding column vectors are used as a set of samples, and the second feature matrix that conforms to a normal distribution is selected after calculation as CRT1. n and CMT1 n , where n represents the nth channel, and the matrix has the following relationship: CRT1 1 ∈CRT 1 ,CRT1 2 ∈CRT 2 ,…,CRT1 10 ∈CRT 10 CMT1 1 ∈CMT 1 ,CMT1 2 ∈CMT 2 ,…,CMT1 10 ∈CMT 10 Step 33: Calculate the t-value. Calculate the t-distribution for the feature samples of each group of sub-task execution and resting states. The calculation formula is as follows: Where n1 is the number of samples X and n2 is the number of samples Y; Step 34: Set the feature selection threshold, verify the significance of the difference with P < 0.01, and reorganize the features that meet the requirements into a new feature matrix, thus obtaining the third feature matrix CRT2. n and CMT2 n .

5. A hand motion FNIRS signal classification method based on t-distribution optimized feature extraction, characterized in that: In step 4, the optimized third feature matrix is ​​used as input to the classifier to obtain the training results of the brain-computer interface system model, achieving effective feature selection and extraction. This specifically includes the following steps: Step 41, Classifier selection, using the third feature matrix CRT2 n and CMT2 n The inputs are fed into a decision tree, a K-nearest neighbor algorithm, and a support vector machine for training, respectively. Step 42: Verify model performance. Use cross-validation to verify model performance and select the model with the best classification performance among each classification method. Step 43: Test model performance. Collect test data using the same process as collecting training data. Construct the feature matrix of the test data based on the optimized feature selection method to test model performance.

Images