Single-feature brain fatigue recognition method based on multi-kernel learning

Abstract

The invention provides a single-feature brain fatigue recognition method based on multi-kernel learning. The method comprises the steps that firstly, L different kernel functions are subjected to weighted combination to form a combined kernel function of an MK-SVM; secondly, optimizing the weight of a combined kernel function through kernel matrix combination and by using a kernel function standardization and cross validation method, and forming an improved MK-SVM multi-classification model; and finally, the electroencephalogram signals are classified by the MK-SVM multi-classification model through weighted combination of a plurality of kernel functions, behaviors of the kernel functions are adjusted according to specific tasks, and an optimal decision boundary is searched in different kernel spaces. According to the multi-kernel SVM, a plurality of basic kernel functions such as a linear kernel, a Gaussian kernel and a polynomial kernel are combined to form a composite kernel, the weight of each kernel is optimized, and the flexibility and robustness of the model are improved so as to adapt to electroencephalogram feature complex data distribution and improve electroencephalogram signal classification indexes.

Description

Technical Field

[0001] This invention relates to a single-feature brain fatigue recognition method based on multi-core learning, belonging to the field of brain fatigue recognition technology. Background Technology

[0002] Fatigue (drowsiness), a common state of brain activity, is ubiquitous in human life. However, when it occurs in actual production activities, it often causes significant harm. The reduced brain activity, decreased reaction time, and other changes brought about by fatigue prevent operators from making timely and accurate judgments and actions as they would under normal conditions. Specifically, in driving, production operations, research, and learning scenarios, fatigue can lead to decreased efficiency, misjudgments, and even serious accidents—drowsy driving has been a major cause of traffic accidents for many years. Therefore, it is necessary to find an effective method to detect fatigue. Thus, the detection of human brain fatigue is of great significance.

[0003] Support Vector Machine (SVM) is a machine learning method based on statistical learning theory. It has been successfully applied in fields such as speech recognition, character recognition, and EEG signal recognition. The basic principle of SVM is to construct a hyperplane in the feature space such that this hyperplane can separate samples with the largest possible classification margin. This hyperplane is called the discriminative hyperplane, and the hyperplane with the largest classification margin is called the optimal discriminative hyperplane. The two boundaries are composed of different training samples, which are called support vectors. The hyperplane formed by these boundaries is called the support hyperplane. The goal of SVM is to maximize the distance between the discriminative hyperplane and the support hyperplane. Therefore, the SVM classification method has significant advantages in the classification of brain fatigue states. SVM can be understood as a constraint optimization problem. For its specific algorithm, refer to "Application of Support Vector Machine in EEG Signal Classification" by Li Gang, Wang Wei, Zhang Sheng, et al. [J]. Computer Applications, 2006, (06): 1431-1433+1436.

[0004] Currently, Gaussian kernel SVM and polynomial kernel SVM are commonly used for EEG signal feature classification. Gaussian kernel SVM classifies EEG signal features by mapping data to an infinite-dimensional feature space through nonlinear mapping, making linearly inseparable data linearly separable in the high-dimensional space. The expression and decision function of the Gaussian kernel function are referenced in Wang Xingfu and Chen Jiawei's "Parameter Selection of Gaussian Kernel-Based SVM" [J]. Computer Systems Applications, 2014, 23(07):242-245. Polynomial kernel SVM classifies EEG signal features by mapping data to a high-dimensional polynomial space, capturing the interactions between features, such as quadratic and cubic relationships, and is suitable for nonlinear classification problems of moderate complexity. The expression and decision function of the polynomial kernel function are referenced in Zuo Sen, Guo Xiaosong, Wan Jing, et al.'s "Fast Classification Algorithm of Polynomial Kernel SVM" [J]. Computer Engineering, 2007, (06):27-29+32. Gaussian kernel SVM and multinomial kernel SVM are used for EEG feature classification. However, because they are single kernel functions, they are difficult to adapt to the complex data distribution of EEG signal features and have a high risk of overfitting. Gaussian kernel SVM is more suitable for datasets with smaller data size and moderate dimensionality, while multinomial kernel SVM is more suitable for feature data with clear multinomial relationships between features.

[0005] A search revealed that Chinese patent CN108520239B discloses a method and system for classifying electroencephalogram (EEG) signals, which uses multi-kernel learning support vector machine (SVM) to classify EEG signals after feature extraction; Chinese patent CN108960299B discloses a method for recognizing multiple types of motor imagery EEG signals, which also uses multi-kernel learning SVM. The multi-kernel combination methods used in these two patents rely on manually set weight coefficients, lack a dynamic adjustment mechanism, and cannot automatically optimize the kernel function weight parameters, resulting in suboptimal model performance. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a single-feature brain fatigue recognition method based on multi-kernel learning, which improves the classification performance of EEG signals by weighted combination of multiple kernel functions.

[0007] This invention provides a single-feature brain fatigue recognition method based on multi-core learning, comprising the following steps:

[0008] Step 1: Weighted combination of L different kernel functions to form the combined kernel function of MK-SVM;

[0009] Step 2: Optimize the weights of the combined kernel functions by combining kernel matrices and using kernel function standardization and cross-validation methods to form an improved MK-SVM multi-classification model;

[0010] Step 3: The MK-SVM multi-classification model classifies EEG signals by weighted combination of multiple kernel functions, adjusts the behavior of the kernel functions according to the specific task, and searches for the optimal decision boundary in different kernel spaces.

[0011] The Multi-Kernel Support Vector Machine (MK-SVM) provided in this invention is an extended algorithm based on the Support Vector Machine (SVM) framework. Its core idea is to improve the model's classification performance on complex data by combining multiple kernel functions. This invention uses an optimized and improved MK-SVM multi-class classification model for EEG signal feature classification, achieving better performance on complex datasets. For multi-class problems, the MK-SVM model employs a one-to-many strategy, i.e., for each category, a classifier is trained to distinguish that category from all other categories. The MK-SVM of this invention improves classification performance by weightedly combining multiple kernel functions.

[0012] The following is a further optimized technical solution of the present invention:

[0013] In step 1, L different kernel functions k1(x,x) are given. i ),k2(x,x i ),…,k L (x,x i The L kernel functions are weighted and combined to form the combined kernel function K(x,x) of MK-SVM. i ),

[0014]

[0015] In the formula, K(x,x) i ) represents the combined composite kernel function, x represents the feature vector of the new sample to be classified, i represents the sample index subscript, l represents the kernel function index subscript, and x represents the composite kernel function. i Let β be the feature vector of the i-th sample in the input space. l Let k be the weight coefficient of the l-th kernel function. l (x,x i ) is the l-th basic kernel function.

[0016] In step 2, for a multi-class classification problem, assume there are N samples and M classes, and the sample features x i ∈R d The tag is y i ∈{1,2,…,M},

[0017] The optimization objective function of MK-SVM is expressed as:

[0018]

[0019] In the formula, w cLet b be the weight of category c. c For the bias term of category c, Let β be the slack variable for the i-th sample in class c. l Here, w represents the weights for each kernel function, b is the bias term, c is the class index, and y represents the weights. i Let be the true class label of the i-th sample, where i is the sample index, l is the kernel function index, and C is the regularization parameter;

[0020] For each sample i, the label y i The following constraints should be satisfied:

[0021]

[0022] In the formula, These are the sample points after mapping using a kernel function;

[0023] After optimization, the final classification decision function formula obtained by MK-SVM is as follows:

[0024] f(x)=∑ i∈SV α i y i K(x,x i )+b,

[0025] In the formula, x is the feature vector of the new sample to be classified, f(x) is the decision function value of sample x, SV is the set of support vectors, and x i Let y be the i-th support vector. i Let α be the class label of the i-th support vector. i Let K(x,x) be the Lagrange multiplier for the i-th support vector. i ) is a composite kernel function, i is the sample index subscript, and b is the bias term.

[0026] In step 3, the MK-SVM multi-classification model finds the optimal classification decision function by minimizing an objective function.

[0027] This invention applies the one-to-many classification principle to the MK-SVM model. By combining kernel matrices and optimizing weights using kernel function standardization and cross-validation, an improved MK-SVM multi-classification model is formed. This model performs classification through a weighted combination of multiple kernel functions, and can adjust the behavior of the kernel functions according to the specific task, thereby finding the optimal decision boundary in different kernel spaces to adapt to more complex linear data structures.

[0028] In step 1, the weighted combination of linear kernels, Gaussian kernels, and polynomial kernels forms the combined kernel function of MK-SVM, which is derived from the linear kernel function. Gaussian kernel function k2(x,x)i )=exp(-γ||xx i || 2 ), γ = 0.5, polynomial kernel function The composite kernel function with d=3 is expressed as follows:

[0029] K(x,x i )=β1k1(x,x i )+β2k2(x,x i )+β3k3(x,x i )

[0030] In the formula, K is the composite kernel function after combination, and x is the feature vector of the new sample to be classified. i Let be the feature vector of the i-th sample in the input space, β1 be the weight coefficient of the linear kernel function, β2 be the weight coefficient of the Gaussian kernel function, β3 be the weight coefficient of the polynomial kernel function, k1 be the linear kernel function, k2 be the Gaussian kernel function, k3 be the polynomial kernel function, γ be the Gaussian kernel parameter, d be the polynomial order, and i be the sample index.

[0031] Specifically, linear kernel function Gaussian kernel function k2(x,x) i )=exp(-γ||xx i || 2 ), γ = 0.5, polynomial kernel function After optimization, the composite kernel function consisting of d=3 provides a better understanding of the label y for each sample i. i The following constraints should be satisfied:

[0032]

[0033] Compared with existing technologies, the multi-kernel SVM of this invention forms a composite kernel by combining multiple basic kernel functions such as linear kernel, Gaussian kernel and polynomial kernel, and optimizes the weights of each kernel, thereby improving the flexibility and robustness of the model to adapt to the complex data distribution of EEG features and improve the classification index of EEG signals. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the present invention. Detailed Implementation

[0035] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings: This embodiment is implemented under the premise of the technical solution of the present invention, and provides detailed implementation methods and specific operation processes, but the protection scope of the present invention is not limited to the following embodiments.

[0036] This invention provides a single-feature brain fatigue recognition method based on multi-core learning, such as... Figure 1 As shown, it includes the following steps:

[0037] Step 1: The weighted linear kernel function, Gaussian kernel function, and polynomial kernel function are combined to form the combined kernel function of MK-SVM.

[0038] By linear kernel function Gaussian kernel function k2(x,x) i )=exp(-γ||xx i || 2 ), γ = 0.5, polynomial kernel function d=3

[0039] Combinatorial kernel functions:

[0040] K(x,x i )=β1k1(x,x i )+β2k2(x,x i )+β3k3(x,x i )

[0041] In the formula, K is the composite kernel function after combination, and x is the feature vector of the new sample to be classified. i Let be the feature vector of the i-th sample in the input space, β1 be the weight coefficient of the linear kernel function, β2 be the weight coefficient of the Gaussian kernel function, β3 be the weight coefficient of the polynomial kernel function, k1 be the linear kernel function, k2 be the Gaussian kernel function, k3 be the polynomial kernel function, γ be the Gaussian kernel parameter, and d be the polynomial order.

[0042] Step 2: By combining kernel matrices and using kernel function standardization and cross-validation methods to optimize the weights of the combined kernel functions, an improved MK-SVM multi-classification model is formed.

[0043] For a multi-class classification problem, suppose there are N samples and M classes, and sample features x i ∈R d The tag is y i ∈{1,2,…,M},

[0044] The optimization objective function of MK-SVM is expressed as:

[0045]

[0046] In the formula, w c Let b be the weight of category c. c For the bias term of category c, ξ i c Let β be the slack variable for the i-th sample in class c. l Here, w represents the weights for each kernel function, b is the bias term, c is the class index, and y represents the weights. iLet be the true class label of the i-th sample, where i is the sample index, l is the kernel function index, and C is the regularization parameter;

[0047] For each sample i, the label y i The following constraints should be satisfied:

[0048]

[0049] In the formula, These are the sample points after mapping using a kernel function;

[0050] After optimization, the final classification decision function formula obtained by MK-SVM is as follows:

[0051] f(x)=∑ i∈SV α i y i K(x,x i )+b,

[0052] In the formula, x is the feature vector of the new sample to be classified, f(x) is the decision function value of sample x, SV is the set of support vectors, and x i Let y be the i-th support vector. i Let α be the class label of the i-th support vector. i Let K(x,x) be the Lagrange multiplier for the i-th support vector. i ) is a composite kernel function, i is the sample index subscript, and b is the bias term.

[0053] Step 3: The MK-SVM multi-classification model classifies EEG signals by weighted combination of multiple kernel functions, adjusts the behavior of the kernel functions according to the specific task, and searches for the optimal decision boundary in different kernel spaces.

[0054] The MK-SVM multi-class classification model finds the optimal classification decision function by minimizing an objective function.

[0055] Example 1

[0056] Using an eight-channel OpenBCI portable EEG acquisition device with a sampling frequency of 250 Hz, subjects were asked to perform three-digit arithmetic calculations in a quiet room. 24 sets of EEG data were collected at rest, when the subjects were awake in the morning, 20 minutes after fatigue, and 40 minutes after fatigue.

[0057] The acquired EEG data was imported into the eeglab EEG preprocessing toolbox in MATLAB for preprocessing. The preprocessing process was as follows: first, the acquired EEG data was imported into electrode localization; second, a 0.5 Hz high-pass filter and a 40 Hz low-pass filter were performed to remove high-frequency and low-frequency noise; finally, the filtered data was processed by independent component analysis (ICA) to decompose it into 8 signal sources, and signal sources with EEG components below 60% were removed.

[0058] After labeling the data, feature extraction was performed using four types of features: eight-channel differential entropy, sample entropy, approximate entropy, and root mean square value. The extracted feature data was saved to an Excel spreadsheet, and the four extracted features were named Feature 1, Feature 2, Feature 3, and Feature 4.

[0059] Model training and automatic optimization process for kernel function weight parameters:

[0060] 1. Read data: Read features and labels from an Excel file.

[0061] 2. Label Binarization: Binarize the multi-class labels (one binary output for each category).

[0062] 3. Define a composite kernel function: using a combination of a linear kernel, a Gaussian kernel (gamma = 0.5), and a polynomial kernel (degree = 3). The steps to construct a composite kernel function are as follows:

[0063] The combined kernel functions of linear kernels, Gaussian kernels, and polynomial kernels are formed by weighted combinations of these kernel functions to create the combined kernel function of MK-SVM.

[0064] By linear kernel function Gaussian kernel function k2(x,x) i )=exp(-γ||xx i || 2 ), γ = 0.5, polynomial kernel function d=3

[0065] Combinatorial kernel functions:

[0066] K(x,x i )=β1k1(x,x i )+β2k2(x,x i )+β3k3(x,x i )

[0067] In the formula, K is the composite kernel function after combination, and x is the feature vector of the new sample to be classified. iLet be the feature vector of the i-th sample in the input space, β1 be the weight coefficient of the linear kernel function, β2 be the weight coefficient of the Gaussian kernel function, β3 be the weight coefficient of the polynomial kernel function, k1 be the linear kernel function, k2 be the Gaussian kernel function, k3 be the polynomial kernel function, γ be the Gaussian kernel parameter, and d be the polynomial order.

[0068] 4. Weight Search Space: Generate weight combinations. Take 5 equally spaced points for weights β1 and β2 in the interval [0.1, 0.9], then calculate β3 = 1 - β1 - β2, and ensure that β3 is also in (0, 1).

[0069] 5. Cross-validation: Use 10-fold hierarchical cross-validation to evaluate each weight combination.

[0070] 6. Normalization: Normalize the training and test data in each fold.

[0071] 7. Calculate kernel matrices: Calculate the training kernel matrix (between training sets) and the test kernel matrix (between test sets) using the current weight combination.

[0072] 8. Training the model: Wrap MK-SVM with OneVsRestClassifier, use pre-computed kernels, set the regularization parameter to C=100, output probabilities, and balance the classes.

[0073] For a multi-class classification problem, suppose there are N samples and M classes, and sample features x i ∈R d The tag is y i ∈{1,2,…,M},

[0074] The optimization objective function of MK-SVM is expressed as:

[0075]

[0076] In the formula, w c Let b be the weight of category c. c For the bias term of category c, ξ i c Let β be the slack variable for the i-th sample in class c. l Here, w represents the weights for each kernel function, b is the bias term, c is the class index, and y represents the weights. i Let be the true class label of the i-th sample, where i is the sample index, l is the kernel function index, and C is the regularization parameter;

[0077] For each sample i, the label y i The following constraints should be satisfied:

[0078]

[0079] In the formula, These are the sample points after mapping using a kernel function;

[0080] After optimization, the final classification decision function formula obtained by MK-SVM is as follows:

[0081] f(x)=∑ i∈SV α i y i K(x,x i )+b,

[0082] In the formula, x is the feature vector of the new sample to be classified, f(x) is the decision function value of sample x, SV is the set of support vectors, and x i Let y be the i-th support vector. i Let α be the class label of the i-th support vector. i Let K(x,x) be the Lagrange multiplier for the i-th support vector. i ) is a composite kernel function, i is the sample index subscript, and b is the bias term.

[0083] 9. Evaluate performance: Calculate the accuracy of each fold in the 10-fold cross-validation, and record the ROC curve and AUC value for each class, as well as the classification report.

[0084] 10. Select the best weights: Select the kernel function weight combination that has the highest average accuracy in 10-fold cross-validation.

[0085] 11. Save the best model and summarize the results: Calculate the average ROC curve and AUC for each category to obtain the MK-SVM model with the optimal kernel function weight combination. Summarize the classification report and calculate the average values ​​of each metric for each category. The report includes the best MK-SVM multi-class classification model and its evaluation results.

[0086] An improved MK-SVM multi-classification model was used to classify EEG signals, and the classification results are shown in Table 2. The classification results of the multinomial kernel SVM model, Gaussian kernel SVM model, and linear kernel SVM model for EEG signals are shown in Tables 2, 3, and 4, respectively.

[0087] Table 1. Classification results of the MK-SVM model in this paper.

[0088] Table1 MK-SVM model classification results

[0089]

[0090] Table 2. Classification results of polynomial kernel SVM models

[0091] Table 2Polynomial kernel SVM model classification results

[0092]

[0093]

[0094] Table 3 Classification results of Gaussian kernel SVM model

[0095] Table3 Gaussian kernel SVM model classification results

[0096]

[0097] Table 4. Classification results of linear kernel SVM models

[0098] Table4 Linear kernel SVM model classification results

[0099]

[0100] As shown in the table above, the MK-SVM model significantly improves the classification results in all metrics compared to other single-core SVM models.

[0101] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any transformations or substitutions that can be conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0102] The program code for model training and automatic optimization of kernel function weight parameters is as follows:

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

Claims

1. A single-feature brain fatigue recognition method based on multi-core learning, characterized in that, Includes the following steps: Step 1: Weighted combination of L different kernel functions to form the combined kernel function of MK-SVM; Step 2: Optimize the weights of the combined kernel functions by combining kernel matrices and using kernel function standardization and cross-validation methods to form an improved MK-SVM multi-classification model; Step 3: The MK-SVM multi-classification model classifies EEG signals by weighted combination of multiple kernel functions, adjusts the behavior of the kernel functions according to the specific task, and searches for the optimal decision boundary in different kernel spaces.

2. The single-feature brain fatigue recognition method based on multi-core learning according to claim 1, characterized in that, In step 1, L different kernel functions k1(x,x) are given. i ),k2(x,x i ),…,k L (x,x i The L kernel functions are weighted and combined to form the combined kernel function K(x,x) of MK-SVM. i ), In the formula, K(x,x) i ) represents the combined kernel function, and x represents the feature vector of the new sample to be classified. i Let β be the feature vector of the i-th sample in the input space. l Let k be the weight coefficient of the l-th kernel function. l (x,x i ) is the l-th basic kernel function.

3. The single-feature brain fatigue recognition method based on multi-core learning according to claim 2, characterized in that, In step 2, for a multi-class classification problem, assume there are N samples and M classes, and the sample features x i ∈R d The tag is y i ∈{1,2,…,M}, The optimization objective function of MK-SVM is expressed as: In the formula, w c Let b be the weight of category c. c For the bias term of category c, ξ i c Let β be the slack variable for the i-th sample in class c. l Here, w represents the weights for each kernel function, b is the bias term, c is the class index, and y represents the weights for each kernel function. i Let C be the true class label of the i-th sample, and C be the regularization parameter. For each sample i, the label y i The following constraints should be satisfied: In the formula, These are the sample points after mapping using a kernel function; After optimization, the final classification decision function formula obtained by MK-SVM is as follows: f(x)=∑ i∈SV α i y i K(x,x i )+b, In the formula, x is the feature vector of the new sample to be classified, f(x) is the decision function value of sample x, SV is the set of support vectors, and x i Let y be the i-th support vector. i Let α be the class label of the i-th support vector. i Let K(x,x) be the Lagrange multiplier for the i-th support vector. i ) is the composite kernel function, and b is the bias term.

4. The single-feature brain fatigue recognition method based on multi-core learning according to claim 1, characterized in that, In step 3, the MK-SVM multi-classification model finds the optimal classification decision function by minimizing an objective function.

5. The single-feature brain fatigue recognition method based on multi-core learning according to claim 2, characterized in that, In step 1, the weighted combination of linear kernels, Gaussian kernels, and polynomial kernels forms the combined kernel function of MK-SVM, which is derived from the linear kernel function. Gaussian kernel function k2(x,x) i )=exp(-γ||xx i || 2 ), γ = 0.5, polynomial kernel function Combining composite kernel functions, the expression for a composite kernel function is shown below: K(x,x i )=β1k1(x,x i )+β2k2(x,x i )+β3k3(x,x i ) In the formula, K is the composite kernel function after combination, and x is the feature vector of the new sample to be classified. i Let be the feature vector of the i-th sample in the input space, β1 be the weight coefficient of the linear kernel function, β2 be the weight coefficient of the Gaussian kernel function, β3 be the weight coefficient of the polynomial kernel function, k1 be the linear kernel function, k2 be the Gaussian kernel function, k3 be the polynomial kernel function, γ be the Gaussian kernel parameter, and d be the polynomial order.

Images