Multi-channel self-distillation seizure detection model based on "squeeze-activate" mechanism

Abstract

The application discloses a multi-channel self-distillation epilepsy detection model based on an "extrusion-stimulation" mechanism, which comprises a multi-channel convolutional neural network module, an attention module, a self-distillation module and a classifier module. The model takes the original EEG signal as the input of the model, extracts the deep features containing different information by using the multi-channel convolutional neural network module, adaptively fuses the deep features from two channels by the attention module, forms a space with better separability in the feature space, and improves the accuracy of model reasoning. In addition, the self-distillation mechanism transfers the knowledge learned by the deep layer of the model to the shallow layer, which accelerates the convergence of the model and improves the fitting ability. The application also introduces response-based knowledge distillation for knowledge transmission. At the same time, in order to better transmit the knowledge, mutual information is introduced to reduce the gap between the front layer network and the deep layer network.

Description

[Technical Field]

[0001] This invention relates to the field of biomedical signal processing technology, and in particular to the field of a multi-channel self-distillation epilepsy detection model based on a "squeeze-excitation" mechanism. [Background Technology]

[0002] In recent years, deep learning has received widespread attention as a feature extraction method in image processing, natural language processing, and intelligent medicine. Neural networks adjust the weights of each layer to obtain more suitable deep features, thus deep learning algorithms can adapt well to the nonlinearity and complexity of signals. Compared with traditional feature extraction methods, deep learning methods can extract deeper and more discriminative information from signals for classification tasks. Meanwhile, deep learning technology is also widely used in the medical field. Generally, the aforementioned convolutional neural networks (CNNs) have good performance in extracting deep features and are effective in epilepsy detection. However, most works neglect the information contained in different perspectives from electroencephalogram (EEG) signals.

[0003] To fully utilize raw EEG signals, epileptic seizures require multi-view learning for detection. Multi-view models possess excellent feature extraction capabilities, enabling the extraction and fusion of the most relevant information for classification tasks. However, the relationship between multi-view features and the classification task is often complex and unpredictable. As the number of views increases, the number of parameters also increases significantly, leading to increased model size and training time. Knowledge distillation (KD) is a model compression method that compresses the model into a more compact one, resulting in lower memory requirements and computation time. KD is implemented by training a student model under the supervision of a teacher model. Application of knowledge distillation demonstrates that SeizureNet can effectively improve the performance of a compact network. In other words, multi-view learning allows the model to acquire richer multi-view features, while knowledge distillation reduces the number of trainable parameters in the feature extraction network, thus lowering model training costs.

[0004] However, models combining multi-view learning and knowledge distillation still have certain drawbacks. Specifically, while multi-view learning can extract richer features, it suffers from complex data preprocessing, increased number of channels in deep feature extraction networks, and higher model training costs compared to single-view models. Knowledge distillation, while reducing the number of training parameters and improving model performance, is not stable due to its influence from model structure, dataset, and application scenario. Furthermore, although fine-tuning can reduce the training cost of student models in real-world medical scenarios, training teacher models also requires additional costs, and the choice of teacher model structure also affects the effectiveness of knowledge distillation. [Summary of the Invention]

[0005] To address the issues of high training costs, large parameter counts, and unstable improvement effects of knowledge distillation in existing models, this invention designs a multi-channel self-distillation epilepsy detection model based on a "squeeze-excitation" mechanism, namely, a multi-channel epilepsy detection model based on attention mechanisms and self-distillation. In summary, this multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism fuses multi-channel deep features through an attention mechanism. It not only utilizes self-distillation to transfer knowledge learned by the deep network to the shallow network while training the model, but also introduces response-based knowledge distillation and mutual information to ensure more effective knowledge transfer. This allows for better transfer of knowledge learned by the deep network to the shallow network, ultimately achieving more efficient multi-channel epilepsy detection.

[0006] To achieve the above objectives, this invention proposes a multi-channel self-distillation epilepsy detection model based on a "squeeze-excitation" mechanism, including a multi-channel convolutional neural network module, an attention module, a self-distillation module, and a classifier module;

[0007] The deep features extracted by the multi-channel convolutional neural network module are represented as follows:

[0008] X deep =δ(ω*x+b) (1);

[0009] Where δ represents the activation function, ω represents the convolution kernel, x is the model input data, b represents the bias, and ω and b are set to random initial values;

[0010] The attention module is used to capture the spatial correlation in multi-channel deep features and divides the information processing into a global information embedding process and an adaptive calibration process, which are respectively represented as follows:

[0011]

[0012] s = F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)) (3);

[0013] Among them, z c For deep features that embed global information, s is the deep feature obtained after adaptive adjustment, σ represents the Sigmoid activation function, δ represents the ReLU activation function, and...

[0014] The ReLU activation function is preceded and followed by two fully connected layers to form a gating mechanism, resulting in the following final output of the attention module:

[0015] X fuse =F scale (u c,s c )=u c ·s c (4);

[0016] Among them, F scale Representing the feature map u c and s c Perform channel-by-channel multiplication, X fuse This represents the adaptive fusion depth feature output by the attention module.

[0017] The mathematical expression for the objective function of the self-distillation module is as follows:

[0018] L=αL hard (P,y)+(1-α)L soft (E,E i )+I(E,E i (5);

[0019] Where P is the model's final prediction, y is the true label of the dataset, and L... hard It is the hard loss using the real labels, L soft It is the knowledge distillation loss using the output of each channel subclassifier logic unit, and α is a hyperparameter that balances soft loss and hard loss;

[0020] The self-distillation module uses the hard loss L of a real label. hard And the knowledge distillation loss L using the output of each channel subclassifier logic unit soft The mathematical expressions for them are as follows:

[0021]

[0022]

[0023]

[0024]

[0025] Where c represents the number of channels, M is the total number of channels, and v i It is the final logical unit output of the deepest part of the model, z i This is the output of the logical unit of the sub-classifier in the multi-channel feature extraction network of the model's front layer. T is the temperature coefficient, and N is the total number of true labels. This is the model's predicted value for the i-th class under a temperature coefficient of T. It is the prediction value of the subclassifier of the front-layer deep extraction network for the label of class i under the condition of temperature coefficient T;

[0026] The classifier module consists of four fully connected layers and finally passes through a Softmax layer to obtain the final prediction result.

[0027] Preferably, in the attention module, the second fully connected layer has a larger dimension than the first fully connected layer.

[0028] As a preferred approach, the model uses mutual information to train the feature extractor.

[0029] Furthermore, the mathematical expression for the mutual information loss term is as follows:

[0030]

[0031] Where I(·) is the mutual information between the two random variables, and E is the final logic unit output of the model. i It is the logical unit output of the feature extraction network subclassifier for the i-th channel.

[0032] Preferably, in the classifier module, the sizes of the four fully connected layers are 256, 64, and 32, respectively.

[0033] The beneficial effects of this invention are:

[0034] The multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism consists of four parts: a multi-channel deep feature extraction network module, an attention module, a multi-channel self-distillation framework, and a classifier module. This model uses the original EEG signal as input, extracts deep features containing different information using a multi-channel convolutional neural network module, and adaptively fuses the deep features from two channels through the attention module, forming a more separable feature space and improving the accuracy of model inference. Furthermore, the self-distillation mechanism transfers knowledge learned in the deep layers to the shallow layers, accelerating model convergence and improving fitting ability. This invention also introduces response-based knowledge distillation for knowledge transfer. Simultaneously, to better transfer knowledge, mutual information is introduced to narrow the gap between the front-end and deep networks.

[0035] Compared to knowledge distillation in the teacher-student model, self-distillation effectively transfers knowledge from deep networks to shallow networks without requiring a separate teacher model, thus improving model performance. Furthermore, it is unaffected by the performance of the teacher model, ensuring the stability of knowledge transfer and enhancing the model's generalization ability.

[0036] In summary, the multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism proposed in this invention has the following characteristics:

[0037] (1) The model proposed in this invention can extract depth features by performing convolution operations on the original EEG signal through a multi-channel convolutional neural network, and then adaptively fuse the depth features through an attention module to form more distinguishable depth features in a new feature space, which effectively improves the accuracy of automatic epilepsy detection.

[0038] (2) Compared with the traditional teacher-student model of knowledge distillation, the self-distillation proposed in this invention, which combines multi-channel convolutional neural networks and mutual information, does not require the introduction of a separate teacher model. It can transfer the knowledge learned by the deep network to the shallow network through the output of the logic unit of the intermediate sub-classifier, effectively improving the accuracy of the model in automatically detecting epilepsy. At the same time, the introduction of response-based knowledge distillation and mutual information effectively transfers the knowledge learned by the deep network to the shallow network.

[0039] (3) Experimental results of the proposed model on the Bonn epilepsy dataset show that the model outperforms state-of-the-art methods on different evaluation criteria.

[0040] The features and advantages of the present invention will be described in detail through embodiments and in conjunction with the accompanying drawings. [Attached Image Description]

[0041] Figure 1 This is a network structure diagram of a multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism;

[0042] Figure 2 It is mutual information;

[0043] Figure 3 It is the accuracy curve during the training process;

[0044] Figure 4 It is the loss curve during training.

Detailed Implementation Methods

[0045] Example 1: Model architecture, implementation process, and algorithm flow of a multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism:

[0046] 1. Model architecture of a multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism:

[0047] This invention focuses on shallow networks where the dark knowledge contained in the output of logical units in deep networks is passed to two channels respectively. It proposes an adaptive multi-channel self-distillation epilepsy detection module based on a "squeeze-excitation" mechanism, including a multi-channel convolutional neural network module, an attention module, a self-distillation module, and a classifier module. For example... Figure 1 As shown, this invention will process the original EEG data x i =(x1,x2,…,x n )T As input to the model, multi-channel deep features are obtained through layer-by-layer convolution and pooling operations. The introduction of sub-classifiers enables intermediate layers to obtain the logical unit output E through fully connected layers. i =(e1,e2) T A temperature coefficient T is introduced to transfer dark knowledge contained in deep network logic units through response-based knowledge distillation. The introduction of an attention module enables the multi-channel deep features obtained through convolution and pooling operations to generate more representative deep features in a new feature space.

[0048] 2. Implementation process of a multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism:

[0049] Convolutional neural networks are widely used to extract depth features from input signals. This invention uses a one-dimensional convolutional neural network as the multi-channel depth feature extractor of the model and represents the extracted depth features as follows:

[0050] X deep =δ(ω*x+b) (1);

[0051] Where δ represents the activation function, ω represents the convolution kernel, x is the model input data, b represents the bias, and ω and b are set to random initial values.

[0052] To fully utilize the deep features obtained from two-channel convolutional neural networks, improve the model's sensitivity to information, and reduce information redundancy, this invention proposes an attention module based on a "squeeze-excitation" mechanism to capture spatial correlations in multi-channel deep features. If the convolutional neural network uses a local receptive field, the obtained features cannot utilize the contextual information of that receptive field. To alleviate the channel dependency problem, this invention divides the information processing process into embedding global information (i.e., the "squeeze" process) and adaptive calibration (i.e., the "excitation" process).

[0053] To fully utilize the information contained in the depth features of each channel and alleviate the channel dependency problem, global average pooling is used to compress global spatial information. Assume the input depth features are u. c The feature map has a size of H×W, and the process of embedding global information can be represented as follows:

[0054]

[0055] Among them, z c This is a deep feature that embeds global information.

[0056] The depth features obtained after the "squeezing" operation embed rich spatial information, which can effectively alleviate the channel dependency problem. To further utilize this aggregated information, capture the spatial dependencies between channels, and adaptively calibrate the depth features, this invention applies the Sigmoid activation function to improve the nonlinear representation capability of the attention module, and the "activation" process is represented as follows:

[0057] s = F ex (z,W)=σ(g(z,W))=σ(W2δ(W1z)) (3);

[0058] Where s represents the depth feature obtained after adaptive adjustment, σ represents the Sigmoid activation function, and δ represents the ReLU activation function.

[0059] To reduce the complexity of the attention module and improve the model's generalization ability, this invention sets two fully connected layers before and after the ReLU activation function, with the second fully connected layer having a larger dimension than the first, thus forming a gating mechanism. The final output of the attention module can be represented as follows:

[0060] X fuse =F scale (u c ,s c )=u c ·s c (4);

[0061] Among them, F scale Representing the feature map u c and s c Perform channel-by-channel multiplication, X fuse This represents the adaptive fusion deep features that the attention module ultimately outputs.

[0062] Knowledge distillation typically improves model performance by reducing the cross-entropy loss between the model and the target distribution, thereby enhancing generalization and mitigating overfitting. This invention combines knowledge distillation and mutual information to design a new objective function, mathematically expressed as follows:

[0063] L=αL hard (P,y+(1-α)L soft (E,E i )+I(E,E i (5);

[0064] Where P is the model's final prediction, y is the true label of the dataset, and L... hard It is the hard loss using the real labels, L softIt is the knowledge distillation loss using the output of each channel subclassifier logic unit, and α is a hyperparameter that balances soft loss and hard loss.

[0065] This invention uses the cross-entropy loss function as the hard loss, and its mathematical expression is as follows:

[0066]

[0067] Where c represents the number of channels and M is the total number of channels.

[0068] This invention employs response-based knowledge distillation to pass information contained in the final logical unit output of the deep network to the shallow logical units, thereby improving model performance. Simultaneously, it obtains the logical unit outputs of the deep feature extraction networks for each channel through the sub-classifiers of the shallow network. The mathematical expression is as follows:

[0069]

[0070]

[0071]

[0072] Among them, v i It is the final logical unit output of the deepest part of the model, z i This is the output of the logical unit of the sub-classifier in the multi-channel feature extraction network of the model's front layer. T is the temperature coefficient, and N is the total number of true labels. This is the model's predicted value for the i-th class under a temperature coefficient of T. It is the prediction value of the subclassifier of the front-layer deep extraction network for the label of class i under the condition of temperature coefficient T.

[0073] Many recent unsupervised and self-supervised representation learning methods utilize mutual information for training feature extractors. To ensure effective knowledge transfer, this application introduces mutual information alongside knowledge distillation. The aim is to improve the model's knowledge learning efficiency and performance by maximizing the mutual information between the outputs of the logical units in deep networks and the logical unit outputs of the sub-classifiers in shallow networks. Figure 2 As shown, mutual information can be used to measure the dependence between two variables, and the mathematical expression for the loss term is as follows:

[0074]

[0075] Where I(·) is the mutual information between the two random variables, and E is the final logic unit output of the model. i It is the logical unit output of the feature extraction network subclassifier for the i-th channel.

[0076] The model's classifier consists of four fully connected layers with sizes of 256, 64, and 32. The final prediction result is obtained after passing through a Softmax layer.

[0077] 3. Algorithm flow of a multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism:

[0078] The algorithm flow of the multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism is shown in Table 1 below:

[0079]

[0080]

[0081] Table 1. Algorithm flow of the multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism.

[0082] Example 2: Experiment and analysis of a multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism:

[0083] 1. Dataset:

[0084] The EEG signal experimental data used in this embodiment came from the database of the Epilepsy Research Center at the University of Bonn, Germany, and was publicly released in the UCI database after preprocessing. As shown in Table 2, the dataset is divided into five groups: A, B, C, D, and E. Each group contains 100 single-channel EEG signals with a sampling rate of 173.6 Hz, a sampling time of 23.6 s, and 4097 sampling points. Groups D and E collected EEG signals from 5 healthy volunteers, while groups A, B, and C collected EEG signals from epilepsy patients. Group A collected EEG data from the epileptogenic focus during the seizures of 5 epilepsy patients, group B collected EEG data from the contralateral region of the epileptogenic focus during the interictal periods of 5 epilepsy patients, and group C collected EEG data from the epileptogenic focus during the seizure-free periods of 5 epilepsy patients. Group D collected signals with the volunteers' eyes closed, and group E collected signals with the volunteers' eyes open.

[0085]

[0086] Table 2 Description of Bonn Epilepsy Dataset

[0087] To verify the model's performance under different experimental conditions, the Bonn dataset was further divided into four experimental datasets. As shown in Table 3, Dataset1-Dataset3 contain EEG signals of epilepsy patients in different states, while Dataset4 is divided according to the EEG signals of epilepsy patients and healthy volunteers at different seizure stages.

[0088]

[0089] Table 3 Description of Experimental Datasets

[0090] 2. Experimental setup:

[0091] The experimental environment for the algorithm was: GeForce RTX 2070 SUPER with 8GB RAM, AMD Ryzen 5 3600 3.6GHz and 32GB RAM, and 64-bit Microsoft Windows 10. The programming environment was: Python 3.6.1 and PyTorch v1.6.

[0092] In this embodiment, we evaluate the model's performance using accuracy, sensitivity, specificity, and F1-score. These three metrics can be expressed as follows:

[0093]

[0094]

[0095]

[0096]

[0097] Among them, TP is a true positive, indicating the number of epileptic segments correctly detected; FN is a false negative, indicating the number of non-epilepsy segments incorrectly detected; TN is a true negative, indicating the number of non-epilepsy segments correctly detected; and FP is a false positive, indicating the number of epileptic segments incorrectly detected.

[0098] In this embodiment, the ten-fold cross-validation method is used to find the optimal parameters. The dataset is divided into ten subsets, nine of which are used to train the model and the remaining subset is used to test the model performance. This process is repeated ten times.

[0099] 3. Experimental Results:

[0100] This embodiment trains the model using ten-fold cross-validation. The changes in accuracy and loss during training are as follows: Figure 3 and Figure 4 As shown. From Figure 3 and Figure 4As can be seen, the model initially exhibits high loss and low accuracy during training. However, with increasing training iterations, the loss decreases and the model accuracy improves. This embodiment compares the experimental results with recent machine learning and deep learning methods applied to the Bonn dataset. As shown in Table 4, this embodiment summarizes the experimental results of the proposed method and other methods, with the best results marked in bold. The average accuracy, sensitivity, specificity, and F1-score for all data subsets are 99.35±0.76%, 99.25±0.73%, 99.35±1.03%, and 99.35±0.76%, respectively (in the table, NA indicates "not applicable," meaning that the data was not used in this study). Furthermore, in Table 4, group 1 is the method proposed in "Epileptic seizure detection using multiwavelet transform based approximate entropy and artificial neural networks", group 2 is the method proposed in "Online seizure prediction using an adaptive learning approach", group 3 is the method proposed in "Epileptic seizure classification of EEG time-series using rational discrete short-time Fourier transform", group 4 is the method proposed in "Automatic pileptic seizure detection based on empirical mode decomposition and deep neural network", and group 5 is the detection method of the present invention.

[0101]

[0102] Table 4 compares the classification performance of existing seizure detection methods on the Bonn dataset.

[0103]

[0104] Table 5. Comparison of classification performance of different feature fusion methods

[0105] The attention module based on the "squeeze-excitation" mechanism can adaptively fuse deep features. To verify the effectiveness of the attention module, this embodiment adds and concatenates the obtained multi-channel deep features and compares the results. The experimental results are shown in Table 6. In the method proposed in this embodiment, the performance of addition is better than that of concatenation. Furthermore, due to the combination of addition and concatenation, the attention module achieves the best performance in accuracy, sensitivity, specificity, and F1-score. The experimental results demonstrate that the attention module based on the "squeeze-excitation" mechanism can adaptively fuse multi-channel deep features, forming deep features in the new feature space that are more conducive to the model's classification and recognition.

[0106]

[0107] Table 6. Classification performance of different feature fusion methods on the Bonn dataset.

[0108] To verify that the introduction of self-distillation can effectively transfer knowledge learned by the deep network to the shallow network, this embodiment conducts experiments on four partitioned datasets, comparing the model performance with and without knowledge distillation. As shown in Table 7, on datasets 1 and 3, the model's accuracy, sensitivity, specificity, and F1-score are all better with knowledge distillation than without it; on datasets 2 and 4, except for sensitivity, accuracy, specificity, and F1-score are all better with knowledge distillation. The experimental results show that the application of self-distillation can effectively transfer knowledge learned by the deep network to the shallow network through the logic unit output, effectively improving the model's performance.

[0109]

[0110] Table 7. The impact of knowledge distillation on model classification performance.

[0111] To further verify the reliable classification ability of the obtained deep features, this embodiment compares its method with five different machine learning classifiers: Support Vector Machine (SVM), Random Forest (FR), Decision Tree (DT), K-Nearest Neighbors (KNN), and Naive Bayes (NB). The experimental results are shown in Tables 8 to 10. In the four datasets, the method proposed in this embodiment outperforms the other five classifiers in accuracy, sensitivity, specificity, and F1-score. Therefore, the model proposed in this invention can effectively extract information from deep features and has reliable classification ability.

[0112]

[0113]

[0114] Table 8 Classification performance of different classifiers on Dataset1

[0115]

[0116] Table 9 Classification performance of different classifiers on Dataset2

[0117]

[0118] Table 10 Classification performance of different classifiers on Dataset3

[0119] This embodiment employs a response-based knowledge distillation method, whereby knowledge learned by the deep network is passed to the shallow network through the output of the logic unit. The output of the logic unit in the multi-channel shallow network is obtained through two sub-classifiers, and their performance is shown in Tables 11 to 15. Experimental data show that the model's final prediction outperforms the two sub-classifiers in accuracy, sensitivity, specificity, and F1-score. Furthermore, the introduction of knowledge distillation can improve the performance of the intermediate layer sub-classifiers while simultaneously enhancing the performance of the model's final prediction results, and the two-channel sub-classifiers (C1, C2) exhibit relatively similar performance.

[0120]

[0121] Table 11 Classification performance of different classifiers on Dataset4

[0122]

[0123] Table 12 shows the classification performance of the subclassifiers on Dataset1.

[0124]

[0125]

[0126] Table 13 shows the classification performance of the subclassifiers on Dataset2.

[0127]

[0128] Table 14 shows the classification performance of the subclassifiers on Dataset3.

[0129]

[0130] Table 15 shows the classification performance of the subclassifiers on Dataset4.

[0131] The above embodiments are illustrative of the present invention and are not intended to limit the present invention. Any simple modifications to the present invention are within the scope of protection of the present invention.

Claims

1. A multi-channel autodistillation method for epilepsy detection based on a "squeeze-excitation" mechanism, characterized in that: This is achieved through a multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism. The multi-channel self-distillation epilepsy detection model based on the "squeeze-excitation" mechanism includes a multi-channel convolutional neural network module, an attention module, a self-distillation module, and a classifier module. The model input is the raw EEG signal, and the model output is the epilepsy prediction result. The deep features extracted by the multi-channel convolutional neural network module are represented as follows: （1）； in, This represents the activation function. Represents the convolution kernel. Input data to the model, Indicates bias. and Set to a random initial value; The attention module is used to capture the spatial correlation in multi-channel deep features and divides the information processing into a global information embedding process and an adaptive calibration process, which are respectively represented as follows: （2）； （3）； in, Deep features for embedding global information The depth features are obtained after adaptive adjustment. express Activation function express Activation function, and , , and It is the size of the feature map. Represents the scaling factor; The Two fully connected layers are set before and after the activation function to form a gating mechanism, and the final output of the attention module is represented as follows: （4）； in, Representation of feature map and Perform channel-by-channel multiplication. This represents the adaptive fusion depth feature output by the attention module. The mathematical expression for the objective function of the self-distillation module is as follows: （5）； in, It is the model's final prediction. These are the true labels of the dataset. It is the hard loss using the real label. It is the knowledge distillation loss using the output of each channel subclassifier logic unit. It is a hyperparameter that balances soft and hard losses. It is the mutual information of two random variables. It is the final logical unit output of the model. It is the first The logical unit output of the feature extraction network subclassifier for each channel; The self-distillation module uses hard loss with real labels. And the knowledge distillation loss using the output of each channel subclassifier logic unit. The mathematical expressions for them are as follows: （6）； （7）； （8）； （9）； Where c represents the number of channels, and M is the total number of channels. It is the final logical unit output of the deepest part of the model. It is the output of the logical unit of the sub-classifier of the multi-channel feature extraction network in the front layer of the model. It is the temperature coefficient. It is the total number of real tags. The model has a temperature coefficient of In the case of label number Predicted values ​​for the class It is the sub-classifier of the front-layer deep extraction network with a temperature coefficient of In the case of label number The predicted value of the class; The classifier module consists of four fully connected layers and is finally processed by… The final prediction result is obtained from the layer.

2. The multi-channel autodistillation epilepsy detection method based on the "squeeze-excitation" mechanism as described in claim 1, characterized in that: In the attention module, the second fully connected layer has a larger dimension than the first fully connected layer.

3. The multi-channel self-distillation epilepsy detection method based on the "squeeze-excitation" mechanism as described in claim 1, characterized in that: The model uses mutual information to train the feature extractor.

4. The multi-channel autodistillation epilepsy detection method based on the "squeeze-excitation" mechanism as described in claim 3, characterized in that: The mathematical expression for the mutual information loss term is as follows: （10）； in, It is the mutual information of two random variables. It is the final logical unit output of the model. It is the first The output of the logical unit of the feature extraction network subclassifier for each channel.

5. The multi-channel self-distillation epilepsy detection method based on the "squeeze-excitation" mechanism as described in claim 1, characterized in that: In the classifier module, the sizes of the four fully connected layers are 256, 64, and 32, respectively.

Images