Deep TSK fuzzy classifier based on multi-level feature fusion

Abstract

The present invention proposes a deep TSK fuzzy classifier based on multi-level feature fusion, including a feature learning module and a knowledge reasoning module; the feature learning module is a feature learning module based on a convolutional neural network, which takes the original data as the input of the deep TSK fuzzy classifier based on multi-level feature fusion, and the feature learning module extracts local information in a local connection manner (i.e., convolution kernel), and obtains the deep features of the hidden layer through layer-by-layer neural expression to automatically extract deep features from the original data; the knowledge reasoning module is a knowledge reasoning module based on the TSK fuzzy classifier, which takes the deep features as the training parameters of the fuzzy rules, adopts the FCM clustering algorithm to generate the antecedent parameters of the fuzzy rules, and uses the original data to train the consequent parameters of the fuzzy rules. By fusing the feature extraction capability of the convolutional neural network and the uncertainty processing capability of the fuzzy representation, a deep TSK fuzzy classifier that is easier to understand is formed.

Description

Technical field

[0001] The present invention relates to the technical field of classifiers, and in particular to a deep TSK fuzzy classifier based on multi-level feature fusion. [Background Technology]

[0002] The TSK fuzzy classifier based on enhanced deep features (ED-TSK-FC) can construct interpretable fuzzy rules for the enhanced deep features composed of the deep features of convolutional neural networks and latent class information, and provide a semantic explanation of the impact of the enhanced deep features in the classification process.

[0003] Compared to the TSK fuzzy classifier based on raw data, ED-TSK-FC achieves better classification performance. However, ED-TSK-FC still has some issues that need to be addressed. For example, fuzzy rules are only interpretable for enhanced deep features, which limits the model's ability to interpret raw data. Furthermore, in some scenarios, the model must be able to directly identify and interpret anomalies in the raw data. [Summary of the invention]

[0004] The purpose of this invention is to solve the problem of insufficient interpretability of ED-TSK-FC in the feature space of original data, and propose a deep TSK fuzzy classifier based on multi-level feature fusion. By integrating the feature extraction capability of convolutional neural networks and the uncertainty processing capability of fuzzy representation, a deeper TSK fuzzy classifier that is easier to understand is formed.

[0005] To achieve the above objectives, the present invention proposes a deep TSK fuzzy classifier based on multi-level feature fusion, including a feature learning module and a knowledge reasoning module. The feature learning module is a feature learning module based on a convolutional neural network, which takes the original data as the input of the deep TSK fuzzy classifier based on multi-level feature fusion. The feature learning module extracts local information by adopting a local connection method (i.e., a convolution kernel), and automatically obtains the deep features of the hidden layer from the original data through layer-by-layer neural expression; the knowledge reasoning module is a knowledge reasoning module based on the TSK fuzzy classifier, which takes the deep features as the antecedent training data of the fuzzy rules, adopts the FCM (Fuzzy C-Means Algorithm) clustering algorithm to generate the antecedent parameters of the fuzzy rules, and uses the original data to train the consequent parameters of the fuzzy rules.

[0006] Preferably, the feature learning module is composed of an input layer, a convolutional layer, a pooling layer, a fully connected layer and an output layer, wherein the convolutional layer, the pooling layer and the fully connected layer are respectively referred to as Conv layer, Pool layer and FC layer; first, multiple convolutional layers and pooling layers are stacked layer by layer to extract deep features, and the convolutional layers and pooling layers are alternately set, and then classification is performed through several fully connected layers and output layers;

[0007] Assume that the training data set is X={x i ,x i =(x1,x2,...,x n ) T ,x i ∈R n ,i=1,2,...,N} and label set Where n represents the sample dimension, N represents the number of samples, and K represents the category of the sample; in the convolutional neural network, the output result (feature map) of the tth layer is marked as Z t , where Z0 represents the original data x i ;

[0008] In the Conv layer, a local connection method is used (convolution kernel W t ) performs convolution operations, which can effectively reduce the number of parameters of the entire model, and then biases b t Add it to the feature map, and then use the activation function f(·) for nonlinear transformation. The calculation process of the Conv layer is as follows:

[0009] Z t =f(W t *Z t-1 +b t ) (1)

[0010] After the convolution calculation is completed, the nonlinear mapping of the Softplus activation function can further improve the feature representation ability of the model; the mathematical expression of Softplus is as follows:

[0011] f(a)=log(1+e a ) (2)

[0012] Where a represents the input variable;

[0013] In the Pool layer, the maximum pooling operation is selected by selecting the maximum value in the pooling window as the result, thereby retaining more texture features; the pooling process can be expressed as follows:

[0014] Z t =Pool(Z t-1 ) (3)

[0015] After several convolution and maximum pooling operations, the extracted deep features are input into the first FC layer. In the FC layer, all neurons between layers are connected, and the deep features are further mapped to the new feature space to complete the classification task. Specifically, through the weight W t and bias b t Perform calculations and use the activation function f(·) for nonlinear transformation; the calculation process is as follows:

[0016] Z t =f(W t ·Z t-1 +b t ) (4)

[0017] The output layer of the convolutional neural network uses the Softmax activation function, and its output Z t =(z1,z2,...,z K ) T is converted into the corresponding probability result E i =(e1,e2,...,e K ) T , where K represents the total number of categories, is calculated as follows:

[0018]

[0019] During the training process, the cross entropy loss function is used to measure the difference between the output of the convolutional neural network and the true label. Its calculation formula is as follows:

[0020]

[0021] The weight W of the convolutional neural network t and bias b t 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; assuming W t and b t The derivative of ΔW t and Δb t , the parameter update formula is expressed as follows:

[0022]

[0023] Where l represents the number of training iterations and θ represents the learning rate;

[0024] Once the loss function converges, W can be determined t and b t The optimal parameters; for a test sample x test The network can be used to extract the deep features H of any hidden layert′ ; Extract the first FC layer deep feature and mark it as H σ =(h1,h2,...,h r ) T , the calculation method is as follows:

[0025]

[0026] Where σ represents the network layer index value of the first FC layer.

[0027] As a preference, the deep TSK fuzzy classifier based on multi-level feature fusion is denoted as CNN-TSK-FC, and the deep feature H σ =(h1,h2,...,h r ) T and the original data x i =(x1,x2,...,x n ) T As the input data of the antecedent and consequent of the fuzzy classifier in the knowledge reasoning module, the knowledge reasoning module includes the following algorithms:

[0028]

[0029] in, Represents the variable x n In the fuzzy subset of the mth rule, ∧ represents the fuzzy combination operator, Represents the variable x n The consequent parameter in the mth rule, y m is the output of the mth fuzzy rule, M represents the total number of fuzzy rules; the deep feature H σ =(h1,h2,...,h r ) T and the original data x i =(x1,x2,...,x n ) T are the input data of the antecedent and consequent of the fuzzy rule respectively;

[0030] In CNN-TSK-FC, the antecedent part is divided using the FCM clustering algorithm, and the cluster center of each rule is estimated in the deep features. and core width The calculation process is as follows:

[0031]

[0032] Where i and d represent the i-th sample and the d-th component respectively, represents the membership degree under the mth rule, and λ is the scale factor. The normalized membership degree of each rule based on the Gaussian membership function is calculated as follows:

[0033]

[0034] make:

[0035]

[0036] P=((p 1 ) T ,(p 2 ) T ,...,(p m ) T ) T (17)

[0037]

[0038] Therefore, the output of CNN-TSK-FC It can be expressed as:

[0039]

[0040] The output of CNN-TSK-FC is considered as the following linear regression problem:

[0041]

[0042] Preferably, the fuzzy rule representation of the CNN-TSK-FC further includes the following algorithm:

[0043] The subsequent parameter learning problem of the CNN-TSK-FC fuzzy classifier is regarded as a typical linear regression problem. The optimization objective function of the CNN-TSK-FC fuzzy classifier based on the ε-insensitive loss function is expressed as follows:

[0044]

[0045] By introducing the penalty term of L2 norm and As well as the structural risk term, the optimization objective function is modified as follows:

[0046]

[0047] According to the optimization learning theory, the dual problem can be expressed as the following quadratic programming problem:

[0048]

[0049] make:

[0050]

[0051] Λ=(τy T,-τy T ) T ,y=(Y1,...,Y N ) T (27)

[0052] The dual problem can be reduced to a standard quadratic programming form:

[0053]

[0054] Where A=[A ij ] 2N×2N ,

[0055] When getting λ + ,λ - At the same time as the optimal solution of , P is expressed as follows:

[0056]

[0057] As an example, the deep TSK fuzzy classifier based on multi-level feature fusion is denoted as CNN-TSK-FC; the input of the learning algorithm of CNN-TSK-FC includes: the training data set is X={x i ,x i =(x1,x2,...,x n ) T ,x i ∈R n ,i=1,2,...,N} and Maximum number of iterations for convolutional neural networks Initial learning rate θ, number of rules M, scale factor λ and regularization factor α, test sample x test ; Output of the CNN-TSK-FC learning algorithm: Output of the trained CNN-TSK-FC test sample The learning algorithm of the CNN-TSK-FC consists of the following two stages:

[0058] S1. Training phase:

[0059] S10. Initialization: Use random initialization W t and b t , at the maximum number of iterations Steps S11 to S13 are performed sequentially within the range; S11. Input training sample X = {x i ,x i =(x1,x2,...,x n ) T ,x i ∈R n,i=1,2,...,N}, the training samples are convolved and pooled through the Conv layer and the Pool layer to extract deep features;

[0060] S12. In the FC layer, the probability of the sample belonging to each label is calculated;

[0061] S13. Perform error backpropagation on the convolutional neural network based on the cross entropy loss function;

[0062] S14. After the iteration is completed, extract the first fully connected layer depth feature H σ =(h1,h2,...,h r ) T ;

[0063] S15. Run the FCM algorithm to generate a fuzzy membership matrix;

[0064] S16. Estimating cluster centers of fuzzy rules and core width

[0065] S17. Calculate the membership of samples in different rules

[0066] S18. Calculate and generate Q i ,The output of CNN-TSK-FC can be expressed as:

[0067] S19. Calculate the analytical solution of P:

[0068]

[0069] S2. Testing phase:

[0070] S21. Input test sample x test ;

[0071] S22. Extract the first fully connected layer depth feature H σ =(h1,h2,...,h r ) T ;

[0072] S23. Calculate the membership of the test sample in different rules

[0073] S24. According to the output Output test samples

[0074] The beneficial effects of the present invention are as follows: in the training process of fuzzy rules, the performance of the model can be improved by selecting features at different levels for training the antecedent parameters and the consequent parameters respectively. Using original data containing more noise as training data for the consequent parameters can improve the generalization ability of CNN-TSK-FC, and has good accuracy, sensitivity, specificity and interpretability.

[0075] The features and advantages of the present invention will be described in detail through embodiments with reference to the accompanying drawings.

Brief Description of the Drawings

[0076] Figure 1 is a schematic diagram of the convolutional neural network structure of the present invention;

[0077] Figure 2 It is a schematic diagram of the internationally standardized 10-20 lead system;

[0078] Figure 3 This is a performance ranking comparison chart of Shallow-TSK-FC, Deep-TSK-FC, and CNN-TSK-FC;

[0079] Figure 4 This is a performance ranking comparison chart of the base classifier and CNN-TSK-FC;

[0080] Figure 5 It is a schematic diagram of the average membership of all training samples under different rules;

[0081] Figure 6 Schematic diagram of applying CNN-TSK-FC to diagnose epilepsy signals. [Specific implementation method]

[0082] 1. CNN-TSK-FC model architecture

[0083] See Figure 1 , this paper proposes a deep TSK fuzzy classifier based on multi-level feature fusion (CNN-TSK-FC); the original data x i =(x1,x2,...,x n ) T As input, the feature learning module adopts a neural network structure, uses local connection (i.e. convolution kernel) to extract local information, and obtains the deep feature H of the hidden layer through layer-by-layer neural expression. σ =(h1,h2,...,h r ) T In the knowledge reasoning module, the deep feature H σ =(h1,h2,...,h r ) T As the training data of the fuzzy rule antecedent, the original data xi =(x1,x2,...,x n ) T It is used to train the parameters of the consequent part of the fuzzy rule. The implementation process of CNN-TSK-FC will be described in detail below.

[0084] 2. CNN-TSK-FC Implementation Process

[0085] First, multiple convolutional layers (Conv layers) and pooling layers (Pool layers) are stacked layer by layer to extract deep features. The convolutional layers and pooling layers are set alternately, and then classification is performed through several fully connected layers (FC layers) and output layers.

[0086] Assume that the training data set is X={x i ,x i =(x1,x2,...,x n ) T ,x i ∈R n ,i=1,2,...,N} and label set Where n represents the sample dimension, N represents the number of samples, and K represents the category of the sample; in the convolutional neural network, the output result (feature map) of the tth layer is marked as Z t , where Z0 represents the original data x i ;

[0087] In the Conv layer, a local connection method is used (convolution kernel W t ) performs convolution operations, which can effectively reduce the number of parameters of the entire model, and then biases b t Add it to the feature map, and then use the activation function f(·) for nonlinear transformation. The calculation process of the Conv layer is as follows:

[0088] Z t =f(W t *Z t-1 +b t ) (1)

[0089] After the convolution calculation is completed, the nonlinear mapping of the Softplus activation function can further improve the feature representation ability of the model; the mathematical expression of Softplus is as follows:

[0090] f(a)=log(1+e a ) (2)

[0091] Where a represents the input variable;

[0092] In the Pool layer, the maximum pooling operation is selected by selecting the maximum value in the pooling window as the result, thereby retaining more texture features; the pooling process can be expressed as follows:

[0093] Z t =Pool(Z t-1 ) (3)

[0094] After several convolution and maximum pooling operations, the extracted deep features are input into the first FC layer. In the FC layer, all neurons between layers are connected, and the deep features are further mapped to the new feature space to complete the classification task. Specifically, through the weight W t and bias b t Perform calculations and use the activation function f(·) for nonlinear transformation; the calculation process is as follows:

[0095] Z t =f(W t ·Z t-1 +b t ) (4)

[0096] The output layer of the convolutional neural network uses the Softmax activation function, and its output Z t =(z1,z2,...,z K ) T is converted into the corresponding probability result E i =(e1,e2,...,e K ) T , where K represents the total number of categories, is calculated as follows:

[0097]

[0098] During the training process, the cross entropy loss function is used to measure the difference between the output of the convolutional neural network and the true label. Its calculation formula is as follows:

[0099]

[0100] The weight W of the convolutional neural network t and bias b t 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; assuming W t and b t The derivative of ΔW t and Δb t , the parameter update formula is expressed as follows:

[0101]

[0102] Where l represents the number of training iterations and θ represents the learning rate;

[0103] Once the loss function converges, W can be determined t and b t The optimal parameter; for a test sample x test The network can be used to extract the deep features H of any hidden layer t′ ; Extract the first FC layer deep feature and mark it as H σ =(h1,h2,...,h r ) T , the calculation method is as follows:

[0104]

[0105] Where σ represents the network layer index value of the first FC layer.

[0106] Deep feature H σ =(h1,h2,...,h r ) T and the original data x i =(x1,x2,...,x n ) T They are used as the input data of the antecedent and consequent of the fuzzy classifier in the knowledge reasoning module. Therefore, the proposed CNN-TSK-FC fuzzy rules can be expressed as follows:

[0107]

[0108] in, Represents the variable x n In the fuzzy subset of the mth rule, ∧ represents the fuzzy combination operator, Represents the variable x n The consequent parameter in the mth rule, y m is the output of the mth fuzzy rule, M represents the total number of fuzzy rules; the deep feature H σ =(h1,h2,...,h r ) T and the original data x i =(x1,x2,...,x n ) T are the input data of the antecedent and consequent of the fuzzy rule respectively;

[0109] In CNN-TSK-FC, the antecedent part is still divided using the FCM clustering algorithm to estimate the cluster center of each rule in the deep features. and core width The calculation process is as follows:

[0110]

[0111] Where i and d represent the i-th sample and the d-th component respectively, represents the membership degree under the mth rule, and λ is the scale factor. The normalized membership degree of each rule based on the Gaussian membership function is calculated as follows:

[0112]

[0113] make:

[0114]

[0115] P=((p 1 ) T ,(p 2 ) T ,...,(p m ) T ) T (17)

[0116]

[0117] Therefore, the output of CNN-TSK-FC It can be expressed as:

[0118]

[0119] The output of CNN-TSK-FC can be viewed as the following linear regression problem:

[0120]

[0121] The subsequent parameter learning problem of the CNN-TSK-FC fuzzy classifier can be regarded as a typical linear regression problem. The optimization objective function of the CNN-TSK-FC fuzzy classifier based on the ε-insensitive loss function is expressed as follows:

[0122]

[0123] By introducing the penalty term of the L2 paradigm and As well as the structural risk term, the optimization objective function is further modified as follows:

[0124]

[0125] According to the optimization learning theory, the dual problem can be expressed as the following quadratic programming problem:

[0126]

[0127] make:

[0128]

[0129] Λ=(τy T ,-τy T ) T ,y=(Y1,...,Y N ) T (27)

[0130] The dual problem can be reduced to a standard quadratic programming form:

[0131]

[0132] Where A=[A ij ] 2N×2N ,

[0133] When getting λ + ,λ - At the same time as the optimal solution of , P in the original problem can be expressed as follows:

[0134]

[0135] 3. CNN-TSK-FC Algorithm Process

[0136] The entire learning algorithm of CNN-TSK-FC is described as Algorithm 1

[0137]

[0138]

[0139] 4. Time complexity analysis

[0140] The computational complexity of the feature extraction module in CNN-TSK-FC is approximately in Represents the maximum number of iterations of the convolutional neural network, C t represents the number of feature maps in layer t, w t represents the convolution kernel size, χ t Indicates the dimension of the feature map. In fuzzy rules, the complexity of the FCM algorithm is O(l2NM 2 r), where l2 represents the maximum number of iterations of FCM.

[0141] Due to the introduction of the ε-insensitive loss function after adding the regularization term, the optimization problem of the consequent parameter is equivalent to the corresponding center-constrained minimum containing ball problem. Therefore, given a fixed minimum containing ball approximation parameter Θ, the time complexity of the consequent parameter learning algorithm is bounded by O(N / Θ 2 +1 / Θ 4 ). Therefore, the total time complexity of CNN-TSK-FC is bounded by

[0142] 5. Experiment and Analysis

[0143] The CHB-MIT dataset was used to verify the effectiveness of CNN-TSK-FC. The experimental environment was: AMD Ryzen 5 3600 3.6 GHz and 32 GB RAM, GeForce RTX 2070 SUPER with 8 GB RAM, the 64-bit Microsoft Windows 10; the programming environment was: Python 3.5.2 with Torch 1.2.0.

[0144] 5.1 Dataset

[0145] The CHB-MIT dataset provided by Boston Children's Hospital was used. This dataset contains 24 sets of scalp EEG data from 23 epilepsy patients, including 5 males and 18 females, aged 2 to 22 years old. The chb01 and chb21 groups were obtained from the same patient at different times. Figure 2 As shown, the internationally standardized 10-20 lead system is used to describe the location of scalp electrodes. Almost all files contain 23-channel EEG signals recorded at 256Hz with 16-bit resolution. In addition, each set of scalp EEG data consists of multiple files, and the duration of the electrical signals contained in the files varies, as does the duration of the epileptic seizures.

[0146] Because epileptic seizures last so briefly, even if all epileptic segments are extracted and merged, there is still a significant imbalance in the amount of data compared to the normal EEG signals in the case. Therefore, the EEG signals can be segmented into multiple 2-second time segments for study. To increase the number of epileptic segments, a resampling method was used for sample extraction, with the overlapping sampling duration set to 0.5 seconds. For the normal EEG signal files, a portion of the EEG signals from non-epileptic seizures was randomly discarded. The ratio of epileptic samples to normal samples in each case group is approximately 1:2. The specific sample numbers are shown in Table 1 below.

[0147] Studies have shown that the frequency domain characteristics of epileptic seizures mainly appear between 4Hz and 30Hz. The frequency domain characteristics between 4Hz and 30Hz were extracted using fast Fourier transform, and the sampling interval was set to 0.5Hz. Finally, the epileptic samples were marked as 1 and the normal samples were marked as 0 to form the experimental data set.

[0148] Table 1 CHB-MIT epilepsy dataset

[0149]

[0150] 6. Experimental Setup

[0151] The optimal structure of the feature learning module of CNN-TSK-FC is shown in Table 2. At the same time, the optimization range of fuzzy rules is set to M∈{5,10,20,30,40,50}, the optimization range of the scale parameter of the FCM algorithm is set to λ∈{0,0.1,1,10,100}, the optimization range of the regularization factor is set to ζ∈{0,0.1,1,10,100}, and other parameters are set to the default values.

[0152] Table 2 Feature learning module structure of CNN-TSK-FC on CHB-MIT dataset

[0153]

[0154] The knowledge reasoning module of CNN-TSK-FC is a novel TSK fuzzy classifier that integrates deep and shallow features. In experiments, two different versions of first-order TSK fuzzy classifiers (Shallow-TSK-FC and Deep-TSK-FC) were compared. Shallow-TSK-FC uses only raw data as training variables for the antecedent and consequent to determine optimal parameters, while Deep-TSK-FC uses deep features based on convolutional neural networks as training variables for the antecedent and consequent to obtain optimal parameters. The optimization range for all fuzzy rules used in fuzzy classification is set to M∈{5,10,20,30,40,50}, the optimization range for the scaling factor of the FCM algorithm is set to λ∈{0,0.1,1,10,100}, and the optimization range for the regularization factor is set to ζ∈{0,0.1,1,10,100}. All other parameters are set to their default values.

[0155] In addition, the base classifiers compared in this experiment include decision tree (DT), K-nearest neighbor (KNN), linear discriminant analysis (LDA), support vector machine (SVM) and deep belief network (DBN). The parameter settings of these five algorithms are as follows: the maximum depth optimization space of decision tree is {5, 6, ..., 10}, the search range of K-nearest neighbor is {3, 5, ..., 11}, the optimization space of penalty factor of SVM is {0.1, 0.2, ..., 2}, the number of hidden layers of DBN is set to 3, and other parameters are set to default values.

[0156] The evaluation indicators Accuracy, Sensitivity and Specificity were selected to reasonably evaluate and analyze the experimental results. The specific formula is as follows:

[0157] Accuracy=(TP+TN) / (TN+FP+TP+FN)

[0158] Sensitivity = TP / (TP+FN)

[0159] Specificity = TN / (TN + FP)

[0160] Where 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.

[0161] 7. Experimental Results

[0162] Table 3 reports the CNN classification results on the CHB-MIT dataset. In the patient-specific epilepsy identification task, the average accuracy was 96.14 ± 0.94%, the sensitivity was 97.17 ± 0.94%, and the specificity was 94.09 ± 2.11%. The accuracy, sensitivity, and specificity of the chb01, chb02, chb05, chb11, chb22, and chb23 groups all exceeded 98%, while the accuracy, sensitivity, and specificity of the chb13 and chb24 groups were all below the average.

[0163] Table 3 Classification performance of CNN on CHB-MIT dataset

[0164]

[0165]

[0166] Table 4 reports the detailed classification performance of CNN-TSK-FC. It can be seen that the average accuracy, sensitivity, and specificity all outperformed those of most groups listed in Table 3, with the average specificity exceeding 99%. In particular, in group 13, where the CNN model performed poorly, CNN-TSK-FC's classification performance significantly improved. This is because the fuzzy representation of CNN-TSK-FC, based on multi-feature fusion, has stronger uncertainty handling capabilities than the single neural representation of CNN in the epilepsy classification task.

[0167] Table 4 Classification performance of CNN-TSK-FC on the CHB-MIT dataset

[0168]

[0169]

[0170] Table 5 compares the classification results of Shallow-TSK-FC, Deep-TSK-FC and the CNN-TSK-FC proposed in this invention. Table 6 gives the values ​​of the scale parameter, regularization factor and number of fuzzy rules used by these models. Figure 3The average rankings of Shallow-TSK-FC, Deep-TSK-FC and CNN-TSK-FC under different performance indicators are listed. Each histogram represents the result of a performance indicator, and the horizontal axis represents the corresponding classifier. First, Deep-TSK-FC using deep features has better performance than Shallow-TSK-FC using original data directly, indicating that the use of deep features can improve the separability of data. Secondly, CNN-TSKFC has the highest average ranking in terms of accuracy, sensitivity and specificity, which indeed shows that CNN-TSK-FC has better classification performance than Shallow-TSK-FC and Deep-TSK-FC. This directly illustrates the advantages of the model proposed in the present invention: in the training process of fuzzy rules, selecting features at different levels for training of antecedent parameters and consequent parameters can improve the performance of the model, and using original data containing more noise as training data for consequent parameters can improve the generalization ability of CNN-TSK-FC.

[0171] Table 5 Classification performance of Shallow-TSK-FC, Deep-TSK-FC and CNN-TSK-FC on the CHB-MIT dataset

[0172]

[0173]

[0174] Table 6 Parameter settings of Shallow-TSK-FC, Deep-TSK-FC and CNN-TSK-FC

[0175]

[0176] The performance comparison results of the base classifier and CNN-TSK-FC are shown in Table 7 and Figure 4 As shown in the figure, CNN-TSK-FC outperformed the other models in average accuracy, sensitivity, and specificity, with the exception of DBN, which had a higher sensitivity than CNN-TSK-FC. CNN-TSK-FC achieved a significant improvement in specificity, reaching 1.69%, 0.60%, 1.95%, 0.61%, 3.36%, and 5.07%, respectively. Furthermore, CNN-TSK-FC had lower standard deviations in accuracy, sensitivity, and specificity than the other models, indicating that CNN-TSK-FC provided more stable classification performance.

[0177] Table 7 Classification performance of base classifier and CNN-TSK-FC on CHB-MIT dataset

[0178]

[0179]

[0180] Table 8 compares other methods with CNN-TSK-FC (Note: NA in the table indicates not applicable). It shows that the proposed CNN-TSK-FC has better accuracy, sensitivity, and specificity. However, the method proposed by Ke et al. achieves the best performance, achieving a sensitivity of 98.85%. Furthermore, CNN-TSK-FC achieves higher sensitivity than the method reported by Tian et al., while maintaining comparable accuracy and specificity, which is of great significance for epilepsy detection.

[0181] Table 8 Comparison of previous research results and CNN-TSK-FC on the CHB-MIT dataset

[0182]

[0183] In addition, Table 9 reports the average classification results of CNN-TSK-FC using different fuzzy rules on the CHB-MIT dataset. It can be seen that the average performance of CNN-TSK-FC does not significantly improve with the increase in the number of fuzzy rules. Specifically, when the number of fuzzy rules is set to 30, CNN-TSK-FC achieves the best performance. However, the interpretability of CNN-TSK-FC is affected. Generally speaking, fuzzy rules require a balance between classification performance and interpretability. Therefore, setting the number of fuzzy rules to 10 achieves performance close to the average, and CNN-TSK-FC also provides concise interpretability.

[0184] at the same time, Figure 5 The figure shows the average fuzzy membership of all training samples in the CHB01 group of the CHB-MIT dataset. Epileptic samples exhibit high membership in Rules 3 and 10, while normal samples have an average membership of approximately 0.1 under all rules, except Rule 10. Preliminary analysis suggests that epileptic seizures directly influence certain specific fuzzy rules in CNN-TSK-FC. For example, in the current system, if a test sample has high membership in Rules 3 and 10, it is likely an epileptic sample.

[0185] Table 9 Effect of the number of fuzzy rules on the classification performance of CNN-TSK-FC

[0186]

[0187] The following example illustrates the interpretability of CNN-TSK-FC in the chb01 group and explains the IF-THEN rule in detail. For example, for a given test data X test =(x1,x2...,x n ) T , the deep feature can be expressed as Htest =(h1,h2...,h r ) T In the first rule of CNN-TSK-FC, the parameter of the IF part (c 1 is the cluster center vector of the FCM algorithm, σ 1 is the kernel width vector), where the cluster center is c 1 =(0.600,0.150,0.180,0.057,3.990×10 -5 ,3.450×10 -11 ,8.660×10 -11 ,2.890×10 -8 ) T , the kernel width is σ 1 =(30.570,8.300,7.590,1.500,1.290×10 -6 ,1.610×10 -18 ,1.470×10 -17 ,1.060×10 -12 ) T , the consequent parameter is P = (4.850 × 10 -5 ,7.560×10 -5 ,9.160×10 -5 ,...,-3.040×10 -6 ) T Therefore, the first fuzzy rule of CNN-TSK-FC in group chb01 can be expressed in language as:

[0188] The first fuzzy rule of CNN-TSK-FC:

[0189] IF:

[0190] h1 is exp(-(h1-0.600) 2 ) / (2×30.570)∧

[0191] h2 is exp(-(h2-0.150) 2 ) / (2×8.300)∧

[0192] h3 is exp(-(h3-0.180) 2 ) / (2×7.590)∧

[0193] h4 is exp(-(h4-0.057) 2 ) / (2×1.500)∧

[0194] h5 is exp(-(h5-3.990×10 -5 ) 2 ) / (2×1.290×10 -6 )∧

[0195] h6 is exp(-(h6-3.450×10 -11 ) 2 ) / (2×1.610×10 -18 )∧

[0196] h7 is exp(-(h7-8.660×10 -11 ) 2 ) / (2×1.470×10 -17 )∧

[0197] h8 is exp(-(h8-2.890×10 -8 ) 2 ) / (2×1.060×10 -12 )

[0198] THEN:

[0199] y 1 =4.850×10 -5 +7.560×10 -5 ×x1+9.160×10 -5 ×

[0200] x²+,...,-3.040×10 -6 ×x n

[0201] After training, CNN-TSK-FC can be used to diagnose epilepsy. After training, CNN-TSK-FC uses the deep features of neural representation to establish the antecedent of fuzzy rules, and the original data is used to optimize the consequent parameters of fuzzy rules. The output of these 10 rules, such as Figure 6 As shown in Figure 2, the method has high interpretability due to the rule-based reasoning method. From the results, we can see that Rules 3 and 10 play an important role in diagnosis, and the final fuzzy output can correctly identify epilepsy.

[0202] The CNN-TSK-FC of the present invention is mainly composed of a feature learning module based on a convolutional neural network and a knowledge reasoning module based on a TSK fuzzy classifier. Taking the original data as the input of CNN-TSK-FC, the feature learning module can automatically extract deep features from the original data. Since the original data does not have good separability, the deep features after being represented by the convolutional neural network form a more separable area in the derived feature space. Therefore, in the knowledge reasoning module, the deep features are used as the training parameters of the antecedent part, and the FCM clustering algorithm is used to generate the antecedent parameters of the fuzzy rules. Subsequently, the original data is used to train the consequent parameters of the fuzzy rules, ensuring that the original data can appear directly in each rule, thereby improving the interpretability of the model to the original data. In addition, since the original data is usually interfered with by random noise, the generalization ability of the model can be further improved. The above experiments show that the method of the present invention has good accuracy, sensitivity, specificity and interpretability on the CHB-MIT dataset.

[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 falls within the protection scope of the present invention.

Claims

1. Deep TSK fuzzy classifier based on multi-level feature fusion, characterized by: It includes a feature learning module and a knowledge reasoning module. The feature learning module is a feature learning module based on a convolutional neural network. It takes the original data as the input of a deep TSK fuzzy classifier based on multi-level feature fusion. The feature learning module extracts local information by local connection and automatically obtains the deep features of the hidden layer from the original data through layer-by-layer neural expression. The knowledge reasoning module is a knowledge reasoning module based on the TSK fuzzy classifier. It takes the deep features as the antecedent training data of the fuzzy rules, adopts the FCM clustering algorithm to generate the antecedent parameters of the fuzzy rules, and uses the original data to train the consequent parameters of the fuzzy rules. The CHB-MIT dataset is used to verify the effectiveness of the deep TSK fuzzy classifier based on multi-level feature fusion.

2. The deep TSK fuzzy classifier based on multi-level feature fusion according to claim 1, characterized in that: The feature learning module consists of an input layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. The convolutional layer, the pooling layer, and the fully connected layer are respectively referred to as Conv layer, Pool layer, and FC layer. First, multiple convolutional layers and pooling layers are stacked layer by layer to extract deep features. The convolutional layers and pooling layers are alternately set, and then classification is performed through several fully connected layers and the output layer. Assume that the training data set is X={x i ,x i =(x1,x2,...,x n ) T ,x i ∈R n ,i=1,2,...,N} and label set Where n represents the sample dimension, N represents the number of samples, and K represents the category of the sample; in the convolutional neural network, the output result of the tth layer is marked as Z t , where Z0 represents the original data x i ; In the Conv layer, the convolution operation is performed using a local connection method, and then the bias b t Add it to the feature map, and then use the activation function f(·) for nonlinear transformation. The calculation process of the Conv layer is as follows: Z t =f(W t *Z t-1 +b t ) (1) After the convolution calculation is completed, it undergoes nonlinear mapping through the Softplus activation function; the mathematical expression of Softplus is as follows: f(a)jlog(1+e a ) (2) Where a represents the input variable; In the Pool layer, the maximum pooling operation is selected by taking the maximum value within the pooling window as the result; the pooling process can be expressed as follows: WITH t =Pool(Z t-1 ) (3) After several convolution and maximum pooling operations, the extracted deep features are input into the first FC layer. In the FC layer, all neurons between layers are connected, and the deep features are further mapped to the new feature space to complete the classification task. t and bias b t Perform calculations and use the activation function f(·) for nonlinear transformation; The calculation process is as follows: Z t =f(W t ·Z t-1 +b t ) (4) The output layer of the convolutional neural network uses the Softmax activation function, and its output Z t =(z1,z2,...,z K ) T is converted into the corresponding probability result E i =(e1,e2,...,e K ) T , where K represents the total number of categories, is calculated as follows: During the training process, the cross entropy loss function is used to measure the difference between the output of the convolutional neural network and the true label. Its calculation formula is as follows: The weight W of the convolutional neural network t and bias b t Iterate and optimize through the error back propagation algorithm; 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; assume that W t and b t The derivative of ΔW t and Δb t , the parameter update formula is expressed as follows: Where l represents the number of training iterations and θ represents the learning rate; Once the loss function converges, W can be determined t and b t The optimal parameter; for a test sample x test The network can be used to extract the deep features H of any hidden layer t′ ; Extract the first FC layer deep feature and mark it as H σ =(h1,h2,...,h r ) T , the calculation method is as follows: Where σ represents the network layer index value of the first FC layer.

3. The deep TSK fuzzy classifier based on multi-level feature fusion according to claim 2, characterized in that: The deep TSK fuzzy classifier based on multi-level feature fusion is denoted as CNN-TSK-FC, and the deep feature H σ =(h1,h2,...,h r ) T and the original data x i =(x1,x2,...,x n ) T As the input data of the antecedent and consequent of the fuzzy classifier in the knowledge reasoning module, the knowledge reasoning module includes the following algorithms: in, Represents the variable x n In the fuzzy subset of the mth rule, ∧ represents the fuzzy combination operator, Represents the variable x n The consequent parameter in the mth rule, y m is the output of the mth fuzzy rule, M represents the total number of fuzzy rules; the deep feature H σ =(h1,h2,...,h r ) T and the original data x i =(x1,x2,...,x n ) T are the input data of the antecedent and consequent of the fuzzy rule respectively; In CNN-TSK-FC, the antecedent part is divided using the FCM clustering algorithm, and the cluster center of each rule is estimated in the deep features. and core width The calculation process is as follows: Where i and d represent the i-th sample and the d-th component respectively, represents the membership degree under the mth rule, and λ is the scale factor. The normalized membership degree of each rule based on the Gaussian membership function is calculated as follows: make: Therefore, the output of CNN-TSK-FC It can be expressed as: The output of CNN-TSK-FC is considered as the following linear regression problem:

4. The deep TSK fuzzy classifier based on multi-level feature fusion according to claim 3, characterized in that: The fuzzy rule representation of the CNN-TSK-FC also includes the following algorithms: The subsequent parameter learning problem of the CNN-TSK-FC fuzzy classifier is regarded as a typical linear regression problem. The optimization objective function of the CNN-TSK-FC fuzzy classifier based on the ε-insensitive loss function is expressed as follows: By introducing the penalty term of the L2 paradigm and As well as the structural risk term, the optimization objective function is modified as follows: According to the optimization learning theory, the dual problem can be expressed as the following quadratic programming problem: make: L=(τy T ,-the T ) T ,y=(Y1,...,Y N ) T (27) The dual problem can be reduced to a standard quadratic programming form: Where A=[A ij ] 2N×2N , When obtaining λ + ,λ - At the same time as the optimal solution of , P is expressed as follows:

5. The deep TSK fuzzy classifier based on multi-level feature fusion according to any one of claims 1 to 4, characterized in that: The deep TSK fuzzy classifier based on multi-level feature fusion is denoted as CNN-TSK-FC; The input of the CNN-TSK-FC learning algorithm includes: the training data set is X={x i ,x i =(x1,x2,...,x n ) T ,x i ∈R n ,i=1,2,...,N} and convolution Maximum number of iterations of the neural network Initial learning rate θ, number of rules M, scale factor λ and regularization factor α, test sample x test ; Output of the CNN-TSK-FC learning algorithm: Output of the trained CNN-TSK-FC test sample The learning algorithm of the CNN-TSK-FC consists of the following two stages: S1. Training phase: S10. Initialization: Use random initialization W t and b t , at the maximum number of iterations Steps S11 to S13 are performed in sequence within the range; S11. Input training sample X={x i ,x i =(x1,x2,...,x n ) T ,x i ∈R n ,i=1,2,...,N}, the training samples are convolved and pooled through the Conv layer and the Pool layer to extract deep features; S12. In the FC layer, the probability of the sample belonging to each label is calculated; S13. Perform error backpropagation on the convolutional neural network based on the cross entropy loss function; S14. After the iteration is completed, extract the first fully connected layer depth feature H σ =(h1,h2,...,h r ) T ; S15. Run the FCM algorithm to generate a fuzzy membership matrix; S16. Estimating cluster centers of fuzzy rules and core width S17. Calculate the membership of samples in different rules S18. Calculate and generate Q i ,The output of CNN-TSK-FC can be expressed as: S19. Calculate the analytical solution of P: S2. Testing phase: S21. Input test sample x test ; S22. Extract the first fully connected layer depth feature H σ =(h1,h2,...,h r ) T ; S23. Calculate the membership of the test sample in different rules S24. According to the output Output test samples

Images