A method for epileptic EEG recognition based on enhanced deep TSK fuzzy classifier

Abstract

The present invention proposes an epileptic EEG recognition method based on an enhanced deep TSK fuzzy classifier. In the TSK fuzzy classifier (ED‑TSK‑FC), the original EEG signal is used as input data, and deep feature learning of the epileptic EEG signal is performed through a one-dimensional convolutional neural network. The deep features of the hidden layer and the potential category information of the output layer can be obtained through layer-by-layer neural expression, and the deep features and the potential category information are merged into enhanced deep features. In addition, the enhanced deep features are used as training variables for the antecedent and consequent parts of the ED‑TSK‑FC fuzzy rules. In the antecedent part, a fuzzy clustering algorithm is used to fuzzily divide the enhanced deep feature space and obtain the corresponding antecedent parameters, and the ridge regression extreme learning algorithm is introduced in the consequent part to learn the parameters. The enhanced deep features make fuller use of the useful information of 1D‑CNN and improve the classification performance of ED‑TSK‑FC. Based on the high interpretability of fuzzy rules, ED‑TSK‑FC makes a good explanation of the enhanced deep features.

Description

[Technical field]

[0001] The invention relates to the technical field of epileptic EEG signal recognition, and in particular to an epileptic EEG recognition method based on an enhanced deep TSK fuzzy classifier. [Background technology]

[0002] Epilepsy is a common neurological disease with the characteristics of recurrent seizures and difficulty in cure, requiring patients to undergo long-term physical sign monitoring and drug treatment. Since most epilepsy patients usually have sudden attacks without obvious symptoms, this often causes unexpected harm to patients. Therefore, proposing an intelligent epilepsy identification method is an important means to effectively improve the life safety of patients.

[0003] EEG signals record the voltage changes in different parts of the cerebral cortex, thereby revealing the bioelectric activity of brain neurons. In the past few years, epilepsy EEG signal recognition technology using EEG signal preprocessing, feature extraction and classification has made significant progress at home and abroad. Since EEG signals are easily contaminated by noise such as ECG signals, eye movements and muscle activities, methods such as fast Fourier transform, wavelet packet decomposition and independent component analysis are used to extract important information from EEG signals. However, how to obtain effective features from raw EEG signals and use machine learning methods such as decision trees and support vector machines to achieve intelligent recognition is still a challenging task.

[0004] As an emerging type of multi-layer neural network model, deep learning can extract abstract deep features from raw inputs through its layer-by-layer neural expression method. It has developed rapidly in the fields of image recognition, natural language processing, and medical diagnosis. In the task of epileptic EEG signal recognition, convolutional neural networks have demonstrated good feature extraction capabilities and have shown accuracy rates that exceed those of clinicians. Existing literature records the application of convolutional neural networks to epileptic seizure detection tasks based on EEG signals, without the need for manual feature extraction and feature selection; in addition, a general architecture for epileptic EEG signal classification and analysis based on convolutional neural networks is also recorded, and related brain mapping images are generated for further research by medical experts. Due to the inability to truly understand the meaning of each parameter in different network layers, deep learning has become a black box model. As the architecture of deep learning becomes more complex, it is difficult to explain what useful knowledge deep learning has learned from the raw input to produce the final recognition results.

[0005] The TSK fuzzy classifier is an interpretable universal approximator that can interpret the learned knowledge in a semantic way. It has attracted wide attention in fields such as automated control and medical diagnosis. In the TSK fuzzy classifier, determining the antecedent fuzzy rules and consequent parameters are two key steps. Usually, they can be assigned manually by referring to the experience of experts or automatically obtained by using data-driven modeling methods. In recent years, deep fuzzy models that integrate deep learning and fuzzy classifiers are a feasible direction to further improve the approximation accuracy and provide interpretability.

[0006] Although deep convolutional neural networks have achieved great success in the accuracy of epileptic seizure detection, there is still much room for improvement in interpretability. In contrast, fuzzy rules have the ability to semantically interpret the learned deep knowledge. In view of this, a TSK fuzzy classifier (ED-TSK-FC) based on enhanced deep features is constructed in the present invention and applied to the task of epileptic EEG recognition. [Summary of the invention]

[0007] The purpose of the present invention is to solve the problem of poor interpretability of deep neural networks in the process of epilepsy recognition in the prior art, and to propose an epilepsy EEG recognition method based on an enhanced deep TSK fuzzy classifier, which can effectively characterize the feature information of the original input.

[0008] To achieve the above object, the present invention proposes an epileptic EEG recognition method based on an enhanced deep TSK fuzzy classifier, comprising the following steps:

[0009] S1. Construct a TSK fuzzy classifier based on enhanced deep features, which includes the following stages:

[0010] S11. Using the original EEG signal as input data, deep feature learning of the epileptic EEG signal is performed through a one-dimensional convolutional neural network, wherein the one-dimensional convolutional neural network is denoted as 1D-CNN, and the deep features of the hidden layer and the latent category information of the output layer are obtained through layer-by-layer neural expression, and the deep features and the latent category information are combined into enhanced deep features;

[0011] S12. The TSK fuzzy classifier based on enhanced depth features is denoted as ED-TSK-FC, and the enhanced depth features are used as training variables for the antecedent and consequent parts of the ED-TSK-FC fuzzy rule. In the antecedent part, a fuzzy clustering algorithm is used to fuzzily divide the enhanced depth feature space and obtain the corresponding antecedent parameters. In the consequent part, a ridge regression extreme learning algorithm is introduced to learn the parameters. The fuzzy clustering algorithm is denoted as FCM algorithm.

[0012] The entire learning algorithm of S2.ED-TSK-FC includes the following stages:

[0013] S21. In the training phase, the weights W in 1D-CNN are updated iteratively l and bias b l ,After the iteration is completed, the enhanced deep features are extracted, the fuzzy division is performed using the FCM algorithm in the fuzzy rules, and the ,consequent parameters are solved using the ridge regression extreme learning machine;

[0014] S22. In the testing phase, the trained ED-TSK-FC is used to predict the test samples, and the enhanced deep features of the test samples are extracted through 1D-CNN. The membership of the test samples under different rules is calculated, and finally the fuzzy rule consequent output is obtained.

[0015] Preferably, step S11 specifically includes the following steps:

[0016] a1.1D-CNN extracts deep features layer by layer through multiple convolutional layers and maximum pooling layers. After several convolutional and maximum pooling operations, the extracted deep features are input into the first fully connected layer, and then classified by several fully connected layers and output layers. The convolutional layer and the maximum pooling layer are respectively denoted as C layer and MP layer, and the fully connected layer is denoted as FC layer;

[0017] a2.1 Weights W of D-CNN l and bias b l Through the error back propagation algorithm, it is iterated and optimized. Once the loss function converges, W is determined. l and b l The optimal parameters of are used to extract the first fully connected layer deep features and the output layer features as enhanced deep features.

[0018] Preferably, the algorithm of step a1 includes the following stages:

[0019] a11. Assume that the training data set is and label sets Where d represents the sample dimension and N represents the number of samples. In 1D-CNN, the output result of the lth layer is marked as H l , l∈{0,1,...,L}, where H0 represents the original input x i , the output result refers to the feature map. In addition, the weights and biases of the lth layer are randomly initialized as W l and b l ;

[0020] a12. In layer C, convolution operation is performed using local connection and the bias b is l Add it to the feature map, and then use the activation function f(·) for nonlinear transformation. The calculation process is as follows:

[0021] H l =f(Wl *H l-1 +b l ) (1)

[0022] In formula (1), * represents the convolution operation, and f(·) uses the PReLU activation function to improve the computational efficiency and fitting ability of the deep model. The PReLU activation function is expressed as follows:

[0023]

[0024] Where a represents the input variable and ε represents the coefficient of the slope;

[0025] a13. In the MP layer, the maximum pooling strategy is used to select the local significant features of each feature map to improve the robustness of the features. The pooling process is expressed as follows:

[0026] H l =Pool(H l-1 ) (3)

[0027] Where Pool(·) represents the maximum pooling operation;

[0028] a14. After several convolution and maximum pooling operations, the extracted deep features are input into the first FC layer; in the FC layer, the weight W l and bias b l The input is calculated and then the activation function is used for nonlinear transformation; the calculation process is as follows:

[0029] H l =f(W l ·H l-1 +b l ) (4)

[0030] Among them, f(·) still uses the PReLU activation function; the output layer is the Softmax layer, and its output H l =(h1,h2,...,h s ) T Converted to the probability result E of the label i =(e1,e2,...,e s ) T , where s represents the total number of categories, and is calculated as follows:

[0031]

[0032] During the training process, the cross entropy loss function is used to measure the difference between the output of 1D-CNN and the true label. The calculation formula is as follows:

[0033]

[0034] The output of the 1D-CNN is the potential category information.

[0035] Preferably, the process of iteration and optimization in step a2 specifically includes the following stages:

[0036] a21. Back-propagate the loss value of formula (6) from the last layer to the first layer, and update the parameters according to the error of each layer; assuming W l and b l The derivative of is ΔW l and Δb l , parameter update formula:

[0037]

[0038]

[0039] Where θ represents the number of training iterations and η represents the initial learning rate;

[0040] a22. Once the loss function converges, W can be determined l and b l The optimal parameter for a test sample x test Use this network to extract the deep features H of any hidden layer l ; The first fully connected layer deep features and output layer features are extracted as enhanced deep features;

[0041] a23. For easy distinction, the following definition is made: The first fully connected layer depth feature is marked as H σ =(h1,h2,...,h r ) T , the output layer result is marked as E ω =(e1,e2,...,e s ) T , the calculation method is as follows:

[0042]

[0043]

[0044] Preferably, in step a2, during the iteration process, the Adam algorithm is used to calculate the adaptive learning rates of different parameters.

[0045] Preferably, step S12 specifically includes the following steps:

[0046] b1. Output the potential category information E ω =(e1,e2,...,e s ) T Expand to the deep feature space Hσ =(h1,h2,...,h r ) T Generating enhanced deep features The proposed ED-TSK-FC adopts the following fuzzy rules:

[0047]

[0048] Where M represents the total number of fuzzy rules, m represents the current fuzzy rule, and denote the fuzzy subset and consequent parameter of the rth variable in the mth rule, respectively. m represents the fuzzy output of the mth rule, ^ represents the fuzzy sum operator;

[0049] b2. The antecedent part uses the FCM clustering algorithm to perform fuzzy space division and estimate the cluster center of each rule and core width

[0050]

[0051]

[0052] where h it represents the tth component of the ith sample, represents the membership of the component under the mth rule, and λ is the scale factor; in addition, the membership of each rule is estimated using the Gaussian membership function:

[0053]

[0054]

[0055] Normalize the membership of each rule:

[0056]

[0057] Therefore, the output of ED-TSK-FC It can be expressed as:

[0058]

[0059] make:

[0060]

[0061] P=((p 1 ) T ,(p 2 ) T ,...,(p M )T ) T (19)

[0062]

[0063]

[0064] Formula (17) can be regarded as the following linear regression problem:

[0065]

[0066] b3. In ED-TSK-FC, the consequent parameter P is learned by the ridge regression extreme learning algorithm, as follows: for a given training data set and label sets The target formula for solving the consequent parameters is as follows:

[0067]

[0068] Where Q=(Q1,Q2,...,Q N ) T ,Z=(z1,z2,...,z N ) T , α is the regularization factor, and then, based on formula (23), the consequent parameter is obtained:

[0069] P=(Q Τ Q+αI) -1 Q Τ Z (24)

[0070] Where I represents the identity matrix.

[0071] Preferably, in step S11, learning is performed only on the last layer, the parameters of the hidden layer can be assigned arbitrarily, and a cheap algorithm is used for training. The specific method includes the following steps:

[0072] c1. Terminate learning early before the loss of 1D-CNN converges;

[0073] c2. Extract enhanced deep features in the 1D-CNN with early termination learning and use them as training variables for the antecedent and consequent parts of fuzzy rules.

[0074] Beneficial effects of the present invention:

[0075] 1) This paper proposes a TSK fuzzy classifier based on enhanced deep features (ED-TSK-FC). 1D-CNN is used to learn deep features of epileptic EEG signals. The deep features of its hidden layer often have lower data dimensions and higher generalization ability. The output layer also contains corresponding latent category information, which can effectively characterize the feature information of the original input.

[0076] 2) In the process of constructing enhanced deep features, the latent category information is expanded to the deep feature space, which further opens up the manifold structure of the feature space, improves the identifiability of the features, and uses interpretable fuzzy rules to make a good explanation for the enhanced deep features.

[0077] 3) Considering the cumbersome training process of 1D-CNN, the present invention provides a cheap training method for ED-TSK-FC. First, the learning is terminated before the loss function of 1D-CNN converges, and the enhanced deep features of the training samples are extracted; then, the parameters of the ED-TSK-FC fuzzy rules are quickly solved. This cheap training method can shorten the training time of the entire model without significantly reducing the classification accuracy.

[0078] The features and advantages of the present invention will be described in detail through embodiments in conjunction with the accompanying drawings.

Brief Description of the Drawings

[0079] Figure 1 This is the ED-TSK-FC framework diagram;

[0080] Figure 2 It is a schematic diagram of 5 types of typical EEG signals;

[0081] Figure 3 This is a schematic diagram of the performance difference between EP-TSK-FC's cheap training method and 1D-CNN under four data sets and different numbers of iterations;

[0082] Figure 4 It is a schematic diagram of the impact of fuzzy rules on the performance of ED-TSK-FC. [Specific implementation method]

[0083] The present invention extracts deep features and potential category information from deep convolutional neural networks and then constructs interpretable fuzzy rules, deriving an interpretable fuzzy classifier to solve the problem of poor interpretability of deep neural networks in the process of epilepsy identification. In the TSK fuzzy classifier (ED-TSK-FC), the original EEG signal is used as input data, and deep feature learning of epileptic EEG signals is performed through a one-dimensional convolutional neural network (1D-CNN). The deep features of the hidden layer and the potential category information of the output layer can be obtained through layer-by-layer neural expression, and the deep features and potential category information are merged into enhanced deep features. In addition, the enhanced deep features are used as training variables for the antecedent and consequent parts of the ED-TSK-FC fuzzy rule. In the antecedent part, the fuzzy clustering (FCM) algorithm is used to fuzzily divide the enhanced deep feature space and obtain the corresponding antecedent parameters. The ridge regression extreme learning algorithm is introduced in the consequent part to learn the parameters. The specific implementation method of the present invention is described in detail below.

[0084] 1 ED-TSK-FC

[0085] 1.1 ED-TSK-FC framework

[0086] Figure 1 The overall framework of ED-TSK-FC is described, which converts the original EEG signal x i =(x1,x2,...,x d ) T As input, 1D-CNN is used to learn deep features of epileptic EEG signals, and the deep features H of the hidden layer are obtained through layer-by-layer neural expression. σ =(h1,h2,...,h r ) T and the potential class information E of the output layer ω =(e1,e2,...,e s ) T Then, the deep features and latent class information are combined into enhanced deep features And it is used as the training variables of the antecedent and consequent parts of the fuzzy rules in ED-TSK-FC to learn the parameters of the fuzzy rules. The specific workflow of ED-TSK-FC is as follows:

[0087] 1.1.1 1D-CNN

[0088] like Figure 1 As shown in Figure 1, 1D-CNN extracts deep features layer by layer by multiple convolutional layers and maximum pooling layers (represented as C layers and MP layers, respectively), and then classifies them by several fully connected layers (represented as FC layers) and output layers. Assume that the training dataset is and label sets Where d represents the sample dimension and N represents the number of samples.

[0089] In 1D-CNN, the output result (feature map) of the lth layer is marked as H l , l∈{0,1,...,L}, where H0 represents the original input x i In addition, the weights and biases of the lth layer are randomly initialized as W l and b l .

[0090] In the C layer, a local connection method (i.e., convolution kernel W l ) performs the convolution operation and sets the bias b l Add it to the feature map, and then use the activation function f(·) for nonlinear transformation. The calculation process is as follows:

[0091] H l =f(W l *H l-1+b l ) (1)

[0092] In formula (1), * represents the convolution operation, and f(·) uses the PReLU activation function to improve the computational efficiency and fitting ability of the deep model. The PReLU activation function is expressed as follows:

[0093]

[0094] Where a represents the input variable and ε represents the coefficient of the slope.

[0095] In the MP layer, the maximum pooling strategy is used to select the local significant features of each feature map to improve the robustness of the features. The pooling process can be expressed as follows:

[0096] H l =Pool(H l-1 ) (3)

[0097] Where Pool(·) represents the maximum pooling operation.

[0098] After several convolution and maximum pooling operations, the extracted deep features are input into the first FC layer. In the FC layer, the weights W l and bias b l The input is calculated and then the activation function is used for nonlinear transformation. The calculation process is as follows:

[0099] H l =f(W l ·H l-1 +b l ) (4)

[0100] Among them, f(·) still uses the PReLU activation function. In particular, the output layer is a Softmax layer, whose output H l =(h1,h2,...,h s ) T Converted to the probability result E of the label i =(e1,e2,...,e s ) T , where s represents the total number of categories, and is calculated as follows:

[0101]

[0102] During the training process, the cross entropy loss function is used to measure the difference between the output (latent category information) of 1D-CNN and the true label. The calculation formula is as follows:

[0103]

[0104] 1D-CNN weight Wl and bias b l Iterate and optimize through the error back propagation algorithm. Specifically, the loss value of formula (6) is back propagated from the last layer to the first layer, and the parameters are updated according to the error of each layer. Assume W l and b l The derivative of is ΔW l and Δb l , parameter update formula:

[0105]

[0106]

[0107] Where θ represents the number of training iterations and η represents the initial learning rate. During the iteration process, the Adam algorithm is used to calculate the adaptive learning rates of different parameters.

[0108] Once the loss function converges, we can determine W l and b l The optimal parameters. For a test sample x test The network can be used to extract the deep features H of any hidden layer l In the present invention, the first fully connected layer depth feature and the output layer feature are extracted as enhanced depth features. For the convenience of distinction, the following definition is made: the first fully connected layer depth feature is marked as H σ =(h1,h2,...,h r ) T , the output layer result is marked as E ω =(e1,e2,...,e s ) T , the calculation method is as follows:

[0109]

[0110]

[0111] 1.1.2 Fuzzy rules

[0112] The output layer potential category information E ω =(e1,e2,...,e s ) T Expand to the deep feature space H σ =(h1,h2,...,h r ) T Generating enhanced deep features Therefore, the proposed ED-TSK-FC adopts the following fuzzy rules:

[0113]

[0114] Where M represents the total number of fuzzy rules, m represents the current fuzzy rule, and denote the fuzzy subset and consequent parameter of the rth variable in the mth rule, respectively. m represents the fuzzy output of the mth rule, and ^ represents the fuzzy sum operator.

[0115] The antecedent part can use FCM clustering algorithm to perform fuzzy space partition and estimate the cluster center of each rule and core width

[0116]

[0117]

[0118] where h it represents the tth component of the ith sample, represents the membership of the component under the mth rule, and λ is the scale factor. In addition, the membership of each rule is estimated using the Gaussian membership function:

[0119]

[0120]

[0121] Normalize the membership of each rule:

[0122]

[0123] Therefore, the output of ED-TSK-FC It can be expressed as:

[0124]

[0125] make:

[0126]

[0127] P=((p 1 ) T ,(p 2 ) T ,...,(p M ) T ) T (19)

[0128]

[0129]

[0130] Formula (17) can be regarded as the following linear regression problem:

[0131]

[0132] In ED-TSK-FC, the consequent parameter P is introduced into the ridge regression extreme learning algorithm for learning. Specifically, for a given training data set and label sets The objective formula for solving the consequent parameters can be expressed as follows:

[0133]

[0134] Where Q=(Q1,Q2,...,Q N ) T ,Z=(z1,z2,...,z N ) T , α is the regularization factor, and then, based on formula (23), the consequent parameter is obtained:

[0135] P=(Q Τ Q+αI) -1 Q Τ Z (24)

[0136] Where I represents the identity matrix.

[0137] At this point, by integrating the 1D-CNN training method in Section 1.1.1 and the antecedent and consequent parameter learning algorithm of fuzzy rules in Section 1.1.2, the ED-TSK-FC model can be obtained.

[0138] 1.2 ED-TSK-FC learning algorithm and its cheap learning algorithm

[0139] The present invention describes the entire learning algorithm of ED-TSK-FC as Algorithm 1, which mainly includes two stages:

[0140] 1) In the training phase, the parameters W in 1D-CNN are updated iteratively l and b l After the iteration, the enhanced deep features The extraction is performed, the FCM algorithm is used to perform fuzzy partitioning in the fuzzy rules, and the ridge regression extreme learning machine is used to solve the consequent parameter P.

[0141] 2) In the testing phase, the trained ED-TSK-FC is used to test the sample x test Specifically, the enhanced deep features of the test samples are extracted through 1D-CNN, and their membership under different rules is calculated, and finally the fuzzy rule consequence output is obtained.

[0142]

[0143]

[0144]

[0145] Next, we will analyze the time complexity of the ED-TSK-FC algorithm. The computational complexity of 1D-CNN is approximately in represents the maximum number of iterations of 1D-CNN, C l represents the number of feature maps in layer l, w l represents the convolution kernel size, χ l Indicates the dimension of the feature graph. In fuzzy rules, the complexity of the FCM algorithm is O(θ2NM 2 (r+s)), where θ2 represents the maximum number of iterations. The time complexity of the ridge regression extreme learning algorithm is O(N(r+s)) 2 +(r+s) 3 ). Therefore, the total time complexity of ED-TSK-FC is

[0146] Considering the high complexity of the training process of 1D-CNN in ED-TSK-FC, the present invention further proposes a cheap learning algorithm. Since the batch learning process of the neural network actually expresses a kernel learning machine, when the accuracy requirement is not high, the learning of the network only needs to be implemented in the last layer, and the parameters in the hidden layer can be assigned arbitrarily. Therefore, the present invention proposes the following cheap algorithm method:

[0147] 1) Before the loss of 1D-CNN converges, terminate the learning early, that is

[0148] 2) Extract enhanced deep features from the 1D-CNN with early termination learning and use them as training variables for the antecedent and consequent parts of fuzzy rules.

[0149] This cheap training method quickly learns the parameters of 1D-CNN and introduces the ridge regression extreme learning algorithm in the consequent learning of ED-TSK-FC fuzzy rules to ensure the convergence of ED-TSK-FC. The cheap training method of ED-TSK-FC can shorten the training time of the model without significantly reducing the classification accuracy.

[0150] 2 Experimental studies

[0151] The proposed ED-TSK-FC was verified and evaluated experimentally based on the classification task of epileptic EEG signals. The hardware environment for this experiment is: AMD Ryzen 5 3600 3.6GHz and 32G RAM, GeForce RTX 2070SUPER with 8G RAM, the 64-bit Microsoft Windows 10; the programming environment is: Python 3.5.2with torch 1.2.0.

[0152] 2.1 Dataset and Preprocessing

[0153] The epilepsy data used in this experiment comes from the University of Bonn in Germany and is publicly released in the UCI database after preprocessing. Table 1 gives the detailed information of the dataset, including the type of EEG signal, the test state and number of samples corresponding to each class. The dimension of each sample is 176. There are 5 types of typical EEG signals such as Figure 2 shown.

[0154] The Bonn dataset is divided into four experimental datasets, as shown in Table 2, to verify the stability of ED-TSK-FC under different experimental conditions. Among them, Dataset1-Dataset3 contain EEG signals of epilepsy patients in different states, and Dataset4 is divided according to the EEG signals of healthy volunteers and epilepsy patients during the attack period.

[0155] Table 1 Bonn epilepsy dataset

[0156]

[0157] Table 2 Experimental dataset

[0158] Dataset epilepsy normal Dataset1 A B Dataset2 A C Dataset3 A B / C Dataset4 A D / E

[0159] 2.2 Experimental Setup and Performance Indicators

[0160] The hidden layer structure, kernel size, step size and other parameters of 1D-CNN are determined by trial and error. The specific structure of 1D-CNN is reported in Table 3. In addition, the initial learning rate η is set to 0.001, and the remaining parameters are set to the default values. To ensure the interpretability of the fuzzy system, the fuzzy rule optimization range of ED-TSK-FC is set to M = {5, 8, ..., 32}, and the optimization range of the scale factor in formula (13) is λ = {2 -6 ,2 -5 ,...,2 2}, the optimization range of the penalty factor in formula (24) is α = {0.001, 0.01, ..., 100}.

[0161] Table 3 Structure of 1D-CNN

[0162]

[0163] In order to evaluate the effectiveness of the model, the ED-TSK-FC proposed in this paper is compared with three first-order TSK fuzzy classifiers, two classifiers based on deep features and 1D-CNN. Specifically, the three TSK fuzzy classifiers are TSK fuzzy classifier based on original data (Shallow-TSK-FC), TSK fuzzy classifier based on PCA algorithm (PCA-TSK-FC) and TSK fuzzy classifier based on 1D-CNN deep features (Deep-TSK-FC). The optimization range of fuzzy rules is set to M = {5, 8, ..., 32}, and the optimization range of scale factor in formula (13) is λ = {2 -6 ,2 -5 ,...,2 2}, the optimization range of the penalty factor in formula (24) is α = {0.001, 0.01, ..., 100}. The two classifiers based on 1D-CNN deep features are support vector machine (Deep-SVM) and decision tree (Deep-DT). The optimization range of the penalty factor of Deep-SVM is {0.01, 0.1, ..., 100}, and the optimization range of the maximum depth of Deep-DT is {4, 5, ..., 10}. The other parameters are set to the default values.

[0164] In order to reasonably evaluate and analyze the experimental results of the present invention, the evaluation indicators Accuracy and F-Measure are selected. The specific formulas are as follows:

[0165] Accuracy=(TP+TN) / (TP+FP+FN+TN) (25a)

[0166] P=TP / (TP+FP) (25b)

[0167] R=TP / (TP+FN) (25c)

[0168] F-Measure=(2×P×R) / (P+R) (25d)

[0169] Among them, TP (true positive) represents the number of samples correctly identified as epilepsy, TN (true negative) represents the number of samples correctly identified as normal, FP (false positive) represents the number of samples incorrectly identified as epilepsy, and FN (false negative) represents the number of samples incorrectly identified as normal.

[0170] In this experiment, a five-fold cross validation method was used to determine the optimal parameters of each model. Specifically, the data set was divided into five subsets, four of which were used for training, and the remaining subset was used to verify the classification performance. The program was repeated five times, and the mean and standard deviation of the evaluation index were expressed as mean ± standard deviation in all relevant tables.

[0171] 2.3 Results and Analysis

[0172] Tables 4 and 5 show the experimental results of ED-TSK-FC and six comparison classifiers on four datasets. The experiments compare the average test accuracy and F-Measure and its standard deviation, and the best results are in bold. The following results are summarized:

[0173] 1) From the experimental results in Table 4, in terms of classification accuracy, the classifiers Deep-SVM, Deep-DT and Deep-TSK-FC based on 1D-CNN deep features are better than Shallow-TSK-FC and PCA-TSK-FC, and in most cases higher than 1D-CNN, which shows that the deep features extracted by 1D-CNN have more reliable classification capabilities. In addition, ED-TSK-FC performs better than the six comparison classifiers in all four data sets. The reason is that the enhanced deep features expand the latent category information of the output layer to the deep feature space, further opening up the feature space manifold structure of the original input, and achieving more robust classification of epileptic EEG signals.

[0174] Table 4 Accuracy comparison results of ED-TSK-FC and six comparison classifiers on four datasets

[0175]

[0176]

[0177] 2) From the experimental results in Table 5, ED-TSK-FC shows a higher F-Measure index than the other six classifiers, especially in Dataset 4, where the F-Measure of ED-TSK-FC is more obvious than that of other classifiers. This shows that ED-TSK-FC can effectively capture the characteristics of epileptic EEG signals and complete epilepsy recognition in a dataset of mixed EEG signals.

[0178] Table 5 F-Measure comparison results of ED-TSK-FC and six comparison classifiers on four datasets

[0179]

[0180] Figure 3The performance differences of the proposed ED-TSK-FC cheap training method and 1D-CNN at different iteration numbers on four datasets are fully compared. Specifically, after each 1D-CNN iteration, its enhanced deep features are extracted and the parameters of the fuzzy rules in ED-TSK-FC are trained, where the 0th iteration means that 1D-CNN is only initialized. Figure 3 It can be seen that when the number of training iterations of 1D-CNN is low, 1D-CNN does not have good classification performance, while the derived ED-TSK-FC can achieve higher classification performance. As the number of iterations increases, the classification ability of 1D-CNN continues to improve, and the classification performance of ED-TSK-FC is also improved accordingly. This shows that before the loss function of 1D-CNN converges, good classification performance cannot be obtained. Introducing the ridge regression extreme learning algorithm in the subsequent learning of ED-TSK-FC can ensure that ED-TSK-FC has good convergence.

[0181] In addition, Table 6 further shows Figure 3 The total training time and accuracy of the five typical cheap versions of ED-TSK-FC (i.e., ED-TSK-FC-*, * represents the number of iterations of 1D-CNN). Taking Dataset1 as an example, when 1D-CNN is initialized, the accuracy of the derived ED-TSK-FC-0 is 91.11%, which has good classification ability and the training time is 0.3907. As the number of iterations of 1D-CNN increases, the accuracy of the derived ED-TSK-FC-80 is 96.64%, which exceeds the accuracy of 1D-CNN, and the training time is shortened by 2.3020 seconds; although there is a certain gap in accuracy between the cheap version of ED-TSK-FC and ED-TSK-FC, the cheap training method of ED-TSK-FC can significantly shorten the training time of the model and quickly realize the classification ability. At the same time, in the high-accuracy classification scenario, the number of iterations of 1D-CNN can be increased, that is, the complete ED-TSK-FC training process can be realized, thereby improving the classification ability.

[0182] Table 6 Total training time and accuracy of EP-TSK-FC and its cheap version and 1D-CNN on four datasets

[0183]

[0184] In ED-TSK-FC, fuzzy rules calculate the membership of samples under different rules through Gaussian membership. The antecedent part of each fuzzy rule and the corresponding consequent output provide interpretability for the enhanced deep features of 1D-CNN. In order to explain in more detail how ED-TSK-FC interprets enhanced deep features, Dataset1 is used as an example.

[0185] Figure 4 It shows that ED-TSK-FC has good performance at λ=2 -6 ,When α=0.01, the classification performance of ED-TSK-FC changes with the increase of the total number of fuzzy rules.,It can be seen that adding fuzzy rules does not significantly improve the,classification performance of ED-TSK-FC, and when there are 8 and 20 fuzzy rules,,ED-TSK-FC achieves a higher average classification performance.

[0186] Considering the interpretability and classification performance of ED-TSK-FC, 8 fuzzy rules are used to interpret the enhanced deep features. Table 7 shows the cluster center and kernel width of each rule. test , whose enhanced depth feature is Among them, h1, h2, ..., h6 represent deep features, and h7 and h8 represent potential category information. Here we explain the first fuzzy rule corresponding to the first row:

[0187] In rule 1:

[0188] IF h1 is exp(-(h1-57.5348) 2 / 2×27.8546)^

[0189] h2 is exp(-(h2-98.2213) 2 / 2×53.9427)^

[0190] h3 is exp(-(h3-98.8597) 2 / 2×57.2677)∧

[0191] h4 is exp(-(h4-99.5494) 2 / 2×56.5404)∧

[0192] h5 is exp(-(h5-102.6806) 2 -2×54.5038)∧

[0193] h6 is exp(-(h6-67.4723) 2 / 2×38.1890)∧

[0194] h7 is exp(-(h7-0.0196) 2 / 2×0.0003)∧

[0195] h8 is exp(-(h8-0.9804) 2 -2×0.0003)

[0196] THEN y 1 =2.1900×10 -6 +9.1670×10 -5 ×h1+1.4805×10 -4 ×h2

[0197] +2.8400×10 -6 ×10 -4 ×h3-3.0844×10 -3 ×h4+1.3000×10 -7 ×h5

[0198] +2.0980×10 -5 ×h6+1.4663×10 -2 ×h7+1.2900×10 -6 ×h8

[0199] Table 7 Cluster centers and core widths of each rule of ED-TSK-FC

[0200]

[0201]

[0202] The present invention proposes a TSK fuzzy classifier ED-TSK-FC based on enhanced deep features and applies it to epileptic EEG signal recognition. The enhanced deep features make fuller use of the useful information of 1D-CNN and improve the classification performance of ED-TSK-FC. Based on the high interpretability of fuzzy rules, ED-TSK-FC provides a good explanation for the enhanced deep features. In addition, the present invention provides a cheap training method for ED-TSK-FC. Experiments show that this method can shorten the training time of the entire model without significantly reducing the classification accuracy.

[0203] The above embodiments are intended to illustrate the present invention, not to limit the present invention. Any solution that is a simple transformation of the present invention belongs to the protection scope of the present invention.

Claims

1. A method for epileptic EEG recognition based on enhanced deep TSK fuzzy classifier, characterized by: The following steps are involved: S1. Construct a TSK fuzzy classifier based on enhanced deep features, which includes the following stages: S11. Using the original EEG signal as input data, deep feature learning of the epileptic EEG signal is performed through a one-dimensional convolutional neural network, wherein the one-dimensional convolutional neural network is denoted as 1D-CNN, and the deep features of the hidden layer and the latent category information of the output layer are obtained through layer-by-layer neural expression, and the deep features and the latent category information are combined into enhanced deep features; S12. The TSK fuzzy classifier based on enhanced depth features is denoted as ED-TSK-FC, and the enhanced depth features are used as training variables for the antecedent and consequent parts of the ED-TSK-FC fuzzy rule. In the antecedent part, a fuzzy clustering algorithm is used to fuzzily divide the enhanced depth feature space and obtain the corresponding antecedent parameters. In the consequent part, a ridge regression extreme learning algorithm is introduced to learn the parameters. The fuzzy clustering algorithm is denoted as FCM algorithm. The learning algorithm of S2.ED-TSK-FC includes the following stages: S21. In the training phase, the weights W in 1D-CNN are updated iteratively l and bias b l ,After the iteration is completed, the enhanced deep features are extracted, the fuzzy division is performed using the FCM algorithm in the fuzzy rules, and the ,consequent parameters are solved using the ridge regression extreme learning machine; S22. In the testing phase, the trained ED-TSK-FC is used to predict the test samples, and the enhanced deep features of the test samples are extracted through 1D-CNN, and their membership under different rules is calculated, and finally the fuzzy rule consequent output is obtained; Step S11 specifically includes the following steps: a1.1D-CNN extracts deep features layer by layer through multiple convolutional layers and maximum pooling layers. After several convolutional and maximum pooling operations, the extracted deep features are input into the first fully connected layer, and then classified by several fully connected layers and output layers. The convolutional layer and the maximum pooling layer are respectively denoted as C layer and MP layer, and the fully connected layer is denoted as FC layer; a2.1 Weights W of D-CNN l and bias b l Through the error back propagation algorithm, it is iterated and optimized. Once the loss function converges, W is determined. l and b l The optimal parameters of the first fully connected layer depth feature and the output layer feature are extracted as enhanced depth features; In step S11, learning is performed only on the last layer, and the parameters of the hidden layer can be assigned arbitrarily. A cheap algorithm is used for training. The specific method includes the following steps: c1. Terminate learning early before the loss of 1D-CNN converges; c2. Extract enhanced deep features in the 1D-CNN with early termination learning and use them as training variables for the antecedent and consequent parts of fuzzy rules.

2. The epileptic EEG recognition method based on enhanced deep TSK fuzzy classifier according to claim 1 is characterized in that: The algorithm of step a1 consists of the following stages: a11. Assume that the training data set is and label sets Where d represents the sample dimension and N represents the number of samples. In 1D-CNN, the output result of the lth layer is marked as H l , l∈{0,1,...,L}, where H0 represents the original input x i , the output result refers to the feature map. In addition, the weights and biases of the lth layer are randomly initialized as W l and b l ; a12. In layer C, convolution operation is performed using local connection and the bias b is l Add it to the feature map, and then use the activation function f(·) for nonlinear transformation. The calculation process is as follows: H l =f(W l *H l-1 +b l ) (1) In formula (1), * represents the convolution operation, and f(·) uses the PReLU activation function to improve the computational efficiency and fitting ability of the deep model. The PReLU activation function is expressed as follows: Where a represents the input variable and ε represents the coefficient of the slope; a13. In the MP layer, the maximum pooling strategy is used to select the local significant features of each feature map to improve the robustness of the features. The pooling process is expressed as follows: H l =Pool(H l-1 ) (3) Where Pool(·) represents the maximum pooling operation; a14. After several convolution and maximum pooling operations, the extracted deep features are input into the first FC layer; in the FC layer, the weight W l and bias b l Calculate its input and then use the activation function to perform nonlinear transformation; The calculation process is as follows: H l =f(W l ·H l-1 +b l ) (4) Among them, f(·) still uses the PReLU activation function; the output layer is the Softmax layer, and its output H l =(h1,h2,...,h s ) T Converted to the probability result E of the label i =(e1,e2,...,e s ) T , where s represents the total number of categories, and is calculated as follows: During the training process, a cross entropy loss function is used to measure the difference between the output of 1D-CNN and the true label. The output of 1D-CNN is the potential category information.

3. The epileptic EEG recognition method based on enhanced deep TSK fuzzy classifier as claimed in claim 2 is characterized by: The iteration and optimization process in step a2 specifically includes the following stages: a21. Back-propagate the cross entropy loss value from the last layer to the first layer, and update the parameters according to the error of each layer; assuming W l and b l The derivative of is ΔW l and Δb l , parameter update formula: Where θ represents the number of training iterations and η represents the initial learning rate; a22. Once the loss function converges, W can be determined l and b l The optimal parameter for a test sample x test Use this network to extract the deep features H of any hidden layer l ; The first fully connected layer deep features and output layer features are extracted as enhanced deep features; a23. For easy distinction, the following definition is made: The first fully connected layer depth feature is marked as H σ =(h1,h2,...,h r ) T , the output layer result is marked as E ω =(e1,e2,...,e s ) T , the calculation method is as follows: 。 4. The epileptic EEG recognition method based on enhanced deep TSK fuzzy classifier according to claim 1 or 3, characterized in that: In step a2, during the iteration process, the Adam algorithm is used to calculate the adaptive learning rates of different parameters.

5. The epileptic EEG recognition method based on enhanced deep TSK fuzzy classifier according to claim 1 or 3, characterized in that: Step S12 specifically includes the following steps: b1. Output the potential category information E ω =(e1,e2,...,e s ) T Expand to the deep feature space H σ =(h1,h2,...,h r ) T Generating enhanced deep features The proposed ED-TSK-FC adopts the following fuzzy rules: Where M represents the total number of fuzzy rules, m represents the current fuzzy rule, and denote the fuzzy subset and consequent parameter of the rth variable in the mth rule, respectively. m represents the fuzzy output of the mth rule, ∧ represents the fuzzy sum operator; b2. The antecedent part uses the FCM clustering algorithm to perform fuzzy space division and estimate the cluster center of each rule and core width where h it represents the tth component of the ith sample, represents the membership of the component under the mth rule, and λ is the scale factor; in addition, the membership of each rule is estimated using the Gaussian membership function: Normalize the membership of each rule: Therefore, the output of ED-TSK-FC It is expressed as: make: P=((p 1 ) T ,(p 2 ) T ,...,(p M ) T ) T (18) Formula (17) can be regarded as the following linear regression problem: b3. In ED-TSK-FC, the consequent parameter P is learned by the ridge regression extreme learning algorithm, as follows: for a given training data set and label sets The objective formula for solving the consequent parameters is expressed as follows: Where Q=(Q1,Q2,...,Q N ) T ,Z=(z1,z2,...,z N ) T , α is the regularization factor, and then, based on formula (22), the consequent parameter is obtained: P=(Q Τ Q+αI) -1 Q Τ Z (23) Where I represents the identity matrix.

Images