Electroencephalogram signal category identification method, device and equipment and storage medium

Abstract

The invention provides an electroencephalogram signal category identification method and device, equipment and a storage medium. The method comprises the following steps: acquiring electroencephalogram signal data to be identified; performing feature extraction processing on the to-be-recognized electroencephalogram signal data based on each feature extraction model in the plurality of feature extraction models to obtain electroencephalogram signal features output by each feature extraction model; determining a plurality of fuzzy rules based on the to-be-identified electroencephalogram signal data and the electroencephalogram signal features; and determining fuzzy output based on the fuzzy rule, and determining category information of the to-be-identified electroencephalogram signal data based on the fuzzy output. According to the method, more and richer electroencephalogram signal features are combined with the fuzzy rule to perform category identification on the electroencephalogram signal data, so that the accuracy of category identification of the electroencephalogram signals is improved, and the performance of the category identification method of the electroencephalogram signals is optimized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of signal processing, and particularly relates to a brain electrical signal category identification method and device, equipment and a storage medium. BACKGROUND

[0002] With the development and application of artificial intelligence models, the artificial intelligence models can be applied to the processing of signals, wherein the artificial intelligence models can be used to identify the category of brain electrical signals.

[0003] In the prior art, after obtaining brain electrical signal data, the brain electrical signal data can be input into a trained fuzzy classifier for classification processing to obtain a classification result of the brain electrical signal data. For example, the classification result of the brain electrical signal data can be an epileptic seizure period brain electrical signal, or can also be an epileptic non-seizure period brain electrical signal.

[0004] However, in the above manner, when the brain electrical signal data is classified based on the fuzzy classifier, only the shallow features of the brain electrical signal data can be learned by the fuzzy classifier to obtain the classification result of the brain electrical signal data, which leads to poor classification performance of the fuzzy classifier and low recognition accuracy. SUMMARY

[0005] The present application provides a brain electrical signal category identification method, device, equipment and storage medium to solve the problem of low accuracy of identifying the category of brain electrical signal data.

[0006] In a first aspect, the present application provides a brain electrical signal category identification method, comprising: obtaining brain electrical signal data to be identified; performing feature extraction processing on the brain electrical signal data to be identified based on each feature extraction model in a plurality of feature extraction models to obtain brain electrical signal features output by each feature extraction model; the brain electrical signal features are any one of the following: spatial and temporal features of brain electrical signal data, deep features of brain electrical signal data; determining a plurality of fuzzy rules based on the brain electrical signal data to be identified and the brain electrical signal features; wherein the fuzzy rules indicate the category information of the brain electrical signal data to be identified under the corresponding fuzzy subset; the fuzzy subset indicates at least part of the data in the brain electrical signal data to be identified that meets the similarity requirement; determining a fuzzy output based on the fuzzy rules, and determining the category information of the brain electrical signal data to be identified based on the fuzzy output.

[0007] In a possible implementation, the fuzzy rules include fuzzy antecedent rules and fuzzy consequent rules; the fuzzy antecedent rules indicate fuzzy subsets corresponding to the brain electrical signal data to be identified; the fuzzy consequent rules indicate category information of the brain electrical signal data to be identified under the fuzzy subsets; the determining the plurality of fuzzy rules based on the brain electrical signal data to be identified and the brain electrical signal features includes: dividing the brain electrical signal data to be identified into a plurality of fuzzy subsets based on a preset number of fuzzy rules; determining membership information of the brain electrical signal data to be identified under each fuzzy subset, and determining the fuzzy antecedent rules based on the membership information; the membership information indicates a degree of association between the brain electrical signal data to be identified and the fuzzy subsets; and setting a corresponding feature weight for the brain electrical signal features under each data dimension, and obtaining the fuzzy consequent rules by weighted summation of the brain electrical signal features under each data dimension and the corresponding feature weight.

[0008] In a possible implementation, the fuzzy subsets are obtained by clustering each one-dimensional brain electrical signal data corresponding to the brain electrical signal data to be identified according to the preset number of fuzzy rules; the fuzzy subsets include corresponding fuzzy subset centers and fuzzy subset width information; the fuzzy subset centers indicate center points corresponding to each one-dimensional brain electrical signal data satisfying a similarity requirement; the fuzzy subset width information indicates a distribution range corresponding to each one-dimensional brain electrical signal data satisfying the similarity requirement; and the determining the membership information of the brain electrical signal data to be identified under each fuzzy subset includes: determining a Gaussian membership between each one-dimensional brain electrical signal data and each fuzzy subset center based on the fuzzy subset center and the fuzzy subset width information; the Gaussian membership indicates a degree of association between the one-dimensional brain electrical signal data and the fuzzy subset center; performing amplification and activation processing on the Gaussian membership based on a hyperbolic tangent (Tanh) function, to obtain an amplified Gaussian membership; and performing a multiplication operation on each amplified Gaussian membership corresponding to a same fuzzy subset center, to obtain the membership information.

[0009] In a possible implementation, the determining the fuzzy output based on the fuzzy rules, and determining the category information of the EEG signal data to be identified based on the fuzzy output comprises: performing product operation on the membership information indicated by the fuzzy antecedent rules corresponding to each fuzzy rule after being multiplied, and the EEG signal features after being weighted and summed indicated by the fuzzy consequent rules, and then performing summation to determine the fuzzy output; constructing a target optimization function containing the fuzzy output and the category information of the EEG signal data to be identified based on a ridge regression algorithm, and solving the target optimization function to obtain the category information of the EEG signal data to be identified.

[0010] In a possible implementation, the feature extraction model is a spatial-temporal feature extraction model; the spatial-temporal feature extraction model is a model constructed based on a convolutional neural network structure and a recurrent neural network structure; and the feature extraction processing of the EEG signal data to be identified based on the feature extraction model to obtain the EEG signal features output by the feature extraction model comprises: inputting the EEG signal data to be identified into the spatial-temporal feature extraction model for processing to obtain the spatial-temporal features of the EEG signal data to be identified.

[0011] In a possible implementation, the feature extraction model is a multi-layer perceptron; and the feature extraction processing of the EEG signal data to be identified based on the feature extraction model to obtain the EEG signal features output by the feature extraction model comprises: inputting the EEG signal data to be identified into the multi-layer perceptron for processing, and determining the features extracted by a perceptron located at a layer before an output layer in the multi-layer perceptron as the deep-level features of the EEG signal data to be identified.

[0012] In a possible implementation, the feature extraction model is obtained by training according to the following steps: obtaining a training sample data set of EEG signal data and a classification label corresponding to each training sample in the training sample data set; inputting the training sample data set into a feature extraction model to be trained for processing to obtain a processing result; calculating a cross-entropy loss value based on the processing result and the classification label; and iteratively training the feature extraction model to be trained based on the cross-entropy loss value, to obtain the feature extraction model.

[0013] In a second aspect, the application provides a brain electrical signal category recognition device, comprising: an acquisition module configured to acquire brain electrical signal data to be recognized; a feature extraction module configured to perform feature extraction processing on the brain electrical signal data to be recognized based on each feature extraction model in a plurality of feature extraction models, to obtain brain electrical signal features output by the each feature extraction model; the brain electrical signal features are any one of the following: spatial and temporal features of the brain electrical signal data, deep features of the brain electrical signal data; a determination module configured to determine a plurality of fuzzy rules based on the brain electrical signal data to be recognized and the brain electrical signal features; wherein the fuzzy rules indicate category information of the brain electrical signal data to be recognized under corresponding fuzzy subsets; the fuzzy subsets indicate at least part of the brain electrical signal data to be recognized that meets a similarity requirement; and a classification module configured to determine fuzzy output based on the fuzzy rules, and determine category information of the brain electrical signal data to be recognized based on the fuzzy output.

[0014] In a third aspect, the application provides a computer device, comprising: a processor, and a memory connected to the processor in communication, the memory storing computer execution instructions, and the processor executes the computer execution instructions stored in the memory to implement the steps of the first aspect or any one of the implementation manners of the first aspect.

[0015] In a fourth aspect, the application provides a computer readable storage medium, the computer readable storage medium storing computer execution instructions, and the computer execution instructions are executed by the processor to implement the steps of the first aspect or any one of the implementation manners of the first aspect.

[0016] In a fifth aspect, the application provides a computer program product, the computer program product carrying program code, and the instructions included in the program code are used to execute the steps of the first aspect or any one of the implementation manners of the first aspect.

[0017] The EEG signal category recognition method, apparatus, device, and storage medium provided in this application can extract multiple EEG signal features from the EEG signal data to be recognized using multiple feature extraction models after acquiring the EEG signal data to be recognized. This results in a larger number and richer variety of EEG signal features, which helps in the accurate identification of EEG signal categories. After obtaining the EEG signal features, multiple fuzzy rules can be determined based on the EEG signal data to be recognized and the EEG signal features, and fuzzy outputs can be determined based on these multiple fuzzy rules. This allows the EEG signal category recognition method to be interpretable through these multiple fuzzy rules, providing a foundation for subsequent EEG signal category recognition and facilitating further research. Simultaneously, the method of determining fuzzy outputs through fuzzy rules handles and expresses the uncertainty and fuzziness of the EEG signal data to be recognized, making EEG signal category recognition more flexible and further improving the performance of the EEG signal category recognition method. Attached Figure Description

[0018] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0019] Figure 1 A flowchart illustrating a method for classifying electroencephalogram (EEG) signals provided in an embodiment of this application;

[0020] Figure 2 A flowchart illustrating another method for classifying electroencephalogram (EEG) signals provided in this application embodiment;

[0021] Figure 3 A schematic diagram of an intelligent model corresponding to a method for classifying electroencephalogram (EEG) signals provided in an embodiment of this application;

[0022] Figure 4 A schematic diagram of a Frideman Ranking analysis result provided in an embodiment of this application;

[0023] Figure 5 A graph illustrating the accuracy of two sets of controlled experiments on the CHB-MIT dataset, provided as an embodiment of this application;

[0024] Figure 6 A schematic diagram of the F-Measure statistic for two groups of controlled experiments on the CHB-MIT dataset, provided as an embodiment of this application;

[0025] Figure 7 A schematic diagram showing the comparison of the accuracy of a log-TSK algorithm and an MV-Tanh-TSK-FC algorithm provided in an embodiment of this application;

[0026] Figure 8 A comparison result diagram of statistical quantity F-Measure of a log-TSK algorithm and an MV-Tanh-TSK-FC algorithm provided by an embodiment of the present application;

[0027] Figure 9 A structure diagram of a brain electrical signal category recognition device provided by an embodiment of the present application;

[0028] Figure 10 A structure diagram of another brain electrical signal category recognition device provided by an embodiment of the present application;

[0029] Figure 11 A structure diagram of a computer device provided by an embodiment of the present application.

[0030] The above drawings have shown the specific embodiments of the present application, which will be described in more detail hereinafter. These drawings and the written description are not intended to limit the scope of the present application in any way, but to illustrate the concept of the present application for those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0031] The exemplary embodiments will be described in detail herein with reference to the accompanying drawings. The following description is presented with reference to the accompanying drawings, in which the same numbers represent the same or similar elements throughout the several drawings. The embodiments described in the following exemplary embodiments are not meant to represent all implementations consistent with the present application. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present application as detailed in the appended claims.

[0032] The term "and / or", used herein, merely describes an associated relationship, which means that there can be three relationships, for example, A and / or B, which can mean that A exists alone, A and B exist together, and B exists alone. In addition, the term "at least one" herein means any one of a plurality or any combination of at least two of a plurality, for example, including at least one of A, B, and C, which means including any one or more elements selected from the set consisting of A, B, and C.

[0033] With the development and application of artificial intelligence models, artificial intelligence models can be applied to the processing of signals, for example, the above-mentioned signals can be brain electrical signals, or electrocardiogram signals, etc.

[0034] When an artificial intelligence model is used to identify the category of a brain electrical signal, after obtaining brain electrical signal data, the brain electrical signal data can be input into a trained artificial intelligence model for processing, and the category of the brain electrical signal data is output after processing.

[0035] For example, the artificial intelligence model can be a fuzzy classifier. After the electroencephalogram signal data is input into the fuzzy classifier, the shallow features of the electroencephalogram signal data can be learned, and a classification result of the electroencephalogram signal data can be obtained. For example, it can be determined that the electroencephalogram signal data is an electroencephalogram signal in a seizure period, or an electroencephalogram signal in a non-seizure period.

[0036] However, in the above manner, when the electroencephalogram signal data is classified based on the fuzzy classifier, only the shallow features of the electroencephalogram signal data are learned, which leads to poor classification performance of the fuzzy classifier and low recognition accuracy

[0037] The present application provides a category recognition method, device and equipment of electroencephalogram signal and storage medium, aiming at solving the above technical problems of the prior art.

[0038] The technical solutions of the present application and how the technical solutions of the present application solve the above technical problems will be described in detail below with specific examples. The following specific examples can be combined with each other, and the same or similar concepts or processes may not be described again in some examples. The embodiments of the present application will be described below with reference to the drawings.

[0039] Figure 1 A flow chart of a category recognition method of electroencephalogram signal provided by an embodiment of the present application is shown in Figure 1 The method comprises:

[0040] S101, obtaining electroencephalogram signal data to be recognized.

[0041] In the embodiment of the present application, in order to recognize the category of the electroencephalogram signal, the electroencephalogram signal data to be recognized can be obtained, and the category recognition of the electroencephalogram signal data to be recognized can be performed to obtain the category information of the electroencephalogram signal data. The category information of the electroencephalogram signal can be an electroencephalogram signal in a seizure period, or an electroencephalogram signal in a non-seizure period, etc.

[0042] In the embodiment of the present application, the electroencephalogram signal data can be collected by electrodes, and the collected electroencephalogram signal data can be preprocessed to obtain the above-mentioned electroencephalogram signal data to be recognized.

[0043] For example, the preprocessing of the electroencephalogram signal data collected by the electrodes can include the following methods: low-pass filter filtering, segmenting, etc. The noise with a frequency higher than 64 Hz can be filtered out by a low-pass filter, and the collected electroencephalogram signal data can be segmented into segments with a time length meeting the processing requirements of the category recognition method of the electroencephalogram signal provided by the embodiment, for example, the segmented segments can be electroencephalogram signal data with a time length of 2 seconds, or electroencephalogram signal data with a time length of 5 seconds, etc.

[0044] S102, perform feature extraction processing on the brain electrical signal data to be identified based on each feature extraction model in the plurality of feature extraction models, to obtain brain electrical signal features output by each feature extraction model.

[0045] The brain electrical signal features are any one of the following: spatial and temporal features of the brain electrical signal data, deep-level features of the brain electrical signal data.

[0046] For example, the number of feature extraction models in the embodiments of the present application can be 2, 3, or 5, and the number of feature extraction models is not limited in the embodiments of the present application.

[0047] In the embodiments of the present application, each feature extraction model can perform feature extraction processing on the brain electrical signal data to be identified, and obtain brain electrical signal features of the brain electrical signal data to be identified after processing.

[0048] In the embodiments of the present application, the same network structure can be included in different feature extraction models, but the network parameters corresponding to the network structure included in each feature extraction model are different. The network parameters corresponding to the network structure included in different feature extraction models are different parameters obtained after training according to different training parameters.

[0049] For example, assuming that the number of feature extraction models is 2, the two feature extraction models are feature extraction model 1 and feature extraction model 2, and both of the two feature extraction models include network structure 1. At this time, the network structure 1 in the feature extraction model 1 and the feature extraction model 2 can include different network parameters.

[0050] In the embodiments of the present application, after the brain electrical signal data to be identified is processed by the feature extraction model, brain electrical signal features of the brain electrical signal data can be obtained.

[0051] S103, determine a plurality of fuzzy rules based on the brain electrical signal data to be identified and the brain electrical signal features.

[0052] The fuzzy rules indicate the class information of the brain electrical signal data to be identified under the corresponding fuzzy subset, and the fuzzy subset indicates at least part of the data in the brain electrical signal data to be identified that meets the similarity requirement.

[0053] Due to the complexity and nonlinearity of the brain electrical signal, the class information of the brain electrical signal data cannot be accurately determined, so in the embodiments of the present application, the class information of the brain electrical signal data is more effectively and accurately determined through fuzzy reasoning.

[0054] Exemplarily, in the embodiment of the present application, a plurality of fuzzy rules can be determined based on the brain electrical signal data to be identified and the brain electrical signal features, and the acquired brain electrical signal data can be converted into an output based on the plurality of fuzzy rules.

[0055] In S104, a fuzzy output is determined based on the fuzzy rules, and the category information of the brain electrical signal data to be identified is determined based on the fuzzy output.

[0056] Exemplarily, in the process of determining the fuzzy output based on the fuzzy rules, the fuzzy output corresponding to the acquired brain electrical signal data can be obtained by fuzzy reasoning on the fuzzy rules, and at this time, the category information of the brain electrical signal data can be determined based on the fuzzy output.

[0057] Exemplarily, the fuzzy output can be 0 or 1, when the fuzzy output is 1, it can represent that the category information of the brain electrical signal data is the brain electrical signal data in the seizure period, and when the fuzzy output is 0, it can represent that the category information of the brain electrical signal data is the brain electrical signal data in the non-seizure period.

[0058] In another example, the fuzzy output can also be any value between 0 and 1, at this time, a fuzzy output threshold can be set, and in the case where the fuzzy output is greater than the set fuzzy output threshold, it is determined that the category information of the brain electrical signal data is the brain electrical signal data in the seizure period, and in the case where the fuzzy output is less than or equal to the set fuzzy output threshold, it is determined that the category information of the brain electrical signal data is the brain electrical signal data in the non-seizure period.

[0059] Here, the value of the fuzzy output and the category information indicated by the value of the fuzzy output are not specifically limited in the embodiment.

[0060] In the embodiment, after acquiring the brain electrical signal data to be identified, a plurality of brain electrical signal features of the brain electrical signal data to be identified can be extracted through a plurality of feature extraction models, so that a larger number and more types of brain electrical signal features can be obtained, thereby facilitating accurate identification of the category of the brain electrical signal. After obtaining the brain electrical signal features, a plurality of fuzzy rules can be determined based on the brain electrical signal data to be identified and the brain electrical signal features, and a fuzzy output can be determined based on the plurality of fuzzy rules, so that the category identification method of the brain electrical signal can have explainability through the plurality of fuzzy rules, which provides a certain basis for subsequent category identification of the brain electrical signal and facilitates further research. At the same time, the uncertainty and fuzziness of the brain electrical signal data to be identified are processed and expressed by the fuzzy rules to determine the fuzzy output, so that the category identification of the brain electrical signal can be more flexible, and the performance of the category identification method of the brain electrical signal is further improved.

[0061] Figure 2A flowchart illustrating another method for classifying electroencephalogram (EEG) signals provided in this application embodiment is shown below. Figure 2 As shown, the method includes:

[0062] S201. Acquire the EEG signal data to be identified.

[0063] The acquired electroencephalogram (EEG) signal data can be denoted as X = (x1, x2, ..., x...). d ) T Where x1, x2, ..., x d This represents the one-dimensional EEG signal data corresponding to the EEG signal data X to be identified. When multiple EEG signal data points are acquired, they can be denoted as X. i =(x i1 ,x i2 ,…,x id ) T , where i represents the i-th EEG signal data to be identified, and the value of i is a positive integer greater than 1.

[0064] For example, this step can be referred to Figure 1 Step S101 will not be repeated here.

[0065] S202. Based on each of the multiple feature extraction models, feature extraction processing is performed on the EEG signal data to be identified to obtain the EEG signal features output by each feature extraction model.

[0066] Among them, the characteristics of EEG signals are any one of the following: spatial and temporal characteristics of EEG signal data, and deep-level characteristics of EEG signal data.

[0067] In one example, step S202 may include the following two steps:

[0068] In the first step of step S202, when the feature extraction model is a spatial-temporal feature extraction model, the EEG signal data to be identified can be input into the spatial-temporal feature extraction model for processing to obtain the spatial-temporal features of the EEG signal data to be identified.

[0069] Among them, the spatial temporal feature extraction model is a model built based on convolutional neural network structure and recurrent neural network structure.

[0070] In one example, a convolutional neural network structure can be added to the front end of a recurrent neural network structure; for instance, multiple convolutional pooling layers can be added to the front end of a recurrent neural network.

[0071] In this way, multiple convolutional layers and pooling layers can be stacked at the front end of the recurrent neural network to improve the ability of the spatial and temporal feature extraction model to extract local spatial features of the EEG signal data to be recognized.

[0072] Here, the number of convolutional pooling layers can be 3, 5, etc., and the embodiments of the present application do not make specific limitations thereto.

[0073] Here, the convolutional layer in the convolutional pooling layer can be a one-dimensional convolutional neural network structure, thereby preserving the translation invariance of the convolutional neural network structure through a one-dimensional network.

[0074] In the embodiments of the present application, the convolution kernels corresponding to each convolutional pooling layer in the multiple convolutional pooling layers are different from each other, thereby extracting more abundant local spatial features of the EEG signal data to be recognized based on convolution kernels of different sizes.

[0075] Since the EEG signal data is composed of time series data, the embodiments of the present application consider the relationship between the past EEG signal data state and the current EEG signal data state, and therefore use a recurrent neural network to consider the influence of past EEG signal data on subsequent EEG signal data and the characteristics of dynamic time series information, to improve the accuracy of the EEG signal classification method.

[0076] At this time, the long short-term memory network LSTM (Long Short Term Memory) is a special recurrent neural network structure, which can effectively capture the semantic association between long sequences through the gate structure control, and can alleviate the gradient disappearance or explosion phenomenon, thereby better processing time series tasks.

[0077] Therefore, in the embodiments of the present application, the short-term memory network LSTM can be used as the above-mentioned recurrent neural network structure, at this time, the first EEG signal feature can be input into the long short-term memory network LSTM for feature extraction, to obtain the spatial and temporal features of the EEG signal data to be recognized.

[0078] The long short-term memory network LSTM used in the embodiments of the present application is composed of a forward long short-term memory network LSTM and a reverse long short-term memory network LSTM (at this time, the forward long short-term memory network LSTM and the reverse long short-term memory network LSTM are combined to obtain a bidirectional long short-term memory network BiLSTM (Bi-directional Long Short-Term Memory)). This implementation can simultaneously consider the history and future information of the electroencephalogram signal data, so as to selectively forget unimportant information, add new information of the current input, and finally integrate to obtain the spatial and temporal features of the electroencephalogram signal data to be recognized.

[0079] In another possible implementation, the above recurrent neural network structure can also be a gate recurrent unit GRU (Gate Recurrent Unit). At this time, the gate recurrent unit GRU reduces the number of gate structures to reduce the calculation amount of the network structure, saves the training time, and improves the performance of the network structure. Therefore, the gate recurrent unit GRU can also be used as the above recurrent neural network to extract features of the electroencephalogram signal data to be recognized, to obtain the spatial and temporal features of the electroencephalogram signal data to be recognized.

[0080] This implementation can combine the recurrent neural network structure with the convolutional pooling layer, so as to extract more rich spatial and temporal features of the electroencephalogram signal data, and further help the modeling of the electroencephalogram signal data composed of time series data.

[0081] In the second step of step S202, in the case that the feature extraction model is a multilayer perceptron, the electroencephalogram signal data to be recognized is input into the multilayer perceptron for processing, and the features extracted by the perceptron located in the layer before the output layer in the multilayer perceptron are determined as the deep features of the electroencephalogram signal data to be recognized.

[0082] In the embodiments of the present application, the number of perceptrons included in the multilayer perceptron can be 3, 4, 5, etc., and the number of perceptrons included in the multilayer perceptron is not limited in the embodiments of the present application.

[0083] In the embodiments of the present application, a specific number of neurons can be set for each layer of the perception machine, and the number of neurons contained in each perception machine in the multi-layer perception machine is in a decreasing state from the first layer perception machine, and the number of neurons contained in adjacent perception machines can be the same or different. For example, in the case where the multi-layer perception machine contains 3 perception machines, the number of neurons contained in the first layer perception machine can be set to 5, the number of neurons contained in the second layer perception machine can be set to 3, and the number of neurons contained in the third layer perception machine can be set to 1. Alternatively, the number of neurons contained in the first layer perception machine can be set to 5, the number of neurons contained in the second layer perception machine can be set to 5, and the number of neurons contained in the third layer perception machine can be set to 3, and so on. Here, the number of neurons set for each layer of the perception machine in the multi-layer perception machine and the decreasing rule are not specifically limited to be able to be implemented.

[0084] In the embodiments of the present application, after the brain electrical signal data to be identified is processed by the first layer perception machine, the processing result and the weight information determined by the first layer perception machine can be assigned to the next layer perception machine through the neurons between the layers, until the last layer perception machine.

[0085] At this time, the output result of the perception machine in the multi-layer perception machine located in the layer before the output layer can be determined as the deep-level feature of the brain electrical signal data to be identified.

[0086] In the embodiments of the present application, the deep-level feature of the brain electrical signal data to be identified can be quickly extracted through the data transmission of the neurons between the multi-layer perception machines, and the performance of the brain electrical signal classification method is improved.

[0087] For example, the brain electrical signal feature of the brain electrical signal data to be identified determined by each feature extraction model can be denoted as wherein k represents the kth feature extraction model, f represents the brain electrical signal feature in the fth data dimension corresponding to the brain electrical signal feature, at this time, the multiple brain electrical signal features corresponding to the multiple feature extraction models are denoted as H=(h 1 ,h 2 ,…,h k ), in the case where the number of brain electrical signal data to be processed is multiple, the brain electrical signal feature extracted by each feature extraction model can be denoted as The multiple brain electrical signal features corresponding to the multiple feature extraction models are denoted as

[0088] S203, divide the brain electrical signal data to be identified into multiple fuzzy subsets based on the number of fuzzy rules set in advance.

[0089] The fuzzy rule indicates category information of the electroencephalogram signal data to be identified under a corresponding fuzzy subset.

[0090] The fuzzy rule includes a fuzzy antecedent rule and a fuzzy consequent rule. The fuzzy antecedent rule indicates a fuzzy subset corresponding to the electroencephalogram signal data to be identified. The fuzzy consequent rule indicates category information of the electroencephalogram signal data to be identified under the fuzzy subset.

[0091] The number of fuzzy rules can be a preset value M, where M is greater than or equal to 1 and is a positive integer.

[0092] In one example, the fuzzy subset is obtained by clustering each one-dimensional electroencephalogram signal data corresponding to the electroencephalogram signal data to be identified according to the preset number of fuzzy rules. For example, a clustering algorithm can be used to cluster the electroencephalogram signal data to be identified based on the preset number of fuzzy rules, and a number of fuzzy subsets equal to the number of fuzzy rules can be obtained. In an example, the clustering algorithm can be a fuzzy c-means (FCM) algorithm, or can be another type of clustering algorithm.

[0093] The fuzzy subset can include a fuzzy subset center and fuzzy subset width information. The fuzzy subset center indicates a center point of each one-dimensional electroencephalogram signal data that meets the similarity requirement. The fuzzy subset width information indicates a distribution range of each one-dimensional electroencephalogram signal data that meets the similarity requirement.

[0094] S204, determine membership information of the electroencephalogram signal data to be identified under each fuzzy subset.

[0095] The membership information indicates an association degree between the electroencephalogram signal data to be identified and the fuzzy subset.

[0096] In one example, the fuzzy subset is obtained by clustering each one-dimensional electroencephalogram signal data corresponding to the electroencephalogram signal data to be identified according to the preset number of fuzzy rules. The fuzzy subset includes a fuzzy subset center and fuzzy subset width information. The fuzzy subset center indicates a center point of each one-dimensional electroencephalogram signal data that meets the similarity requirement. The fuzzy subset width information indicates a distribution range of each one-dimensional electroencephalogram signal data that meets the similarity requirement. Based on this, step S204 can include the following process:

[0097] The first step of step S204 is determining the Gaussian membership degree between each one-dimensional electroencephalogram signal data and each fuzzy subset center based on the fuzzy subset center and fuzzy subset width information; the Gaussian membership degree indicates the correlation degree between the one-dimensional electroencephalogram signal data and the fuzzy subset center.

[0098] The Gaussian membership degree between each one-dimensional electroencephalogram signal data and the mth fuzzy subset center (corresponding to the mth fuzzy rule) can be calculated according to the following formula (1).

[0099]

[0100]

[0101]

[0102] In the above formula (1), the represents the one-dimensional electroencephalogram signal data x ij and A mj corresponding to the Gaussian membership degree between the fuzzy subset, x ij represents the jth feature of the electroencephalogram signal data X i to be identified (i.e., the jth one-dimensional electroencephalogram signal data of the electroencephalogram signal data X i ), A mj is the fuzzy subset of the jth feature of the input data in the mth fuzzy rule, c m represents the fuzzy subset center, σ m in formula (2) represents the fuzzy subset width information, and ε represents a preset scale parameter.

[0103] The second step of step S204 is performing amplification and activation processing on the Gaussian membership degree based on the hyperbolic tangent Tanh function, and obtaining the amplified Gaussian membership degree after processing.

[0104] Here, the specific calculation formula of the hyperbolic tangent Tanh function for amplifying and activating the Gaussian membership degree can be represented as shown in the following formula (3).

[0105]

[0106] The third step of step S204 is performing a multiplication operation on each amplified Gaussian membership degree corresponding to the same fuzzy subset center, and obtaining the membership degree information after multiplication.

[0107] Here, the process of performing a multiplication operation on each amplified Gaussian membership degree corresponding to the same fuzzy subset center is shown in the following formula (4).

[0108]

[0109] ​The multiplication manner of the magnified and activated Gaussian membership by the hyperbolic tangent function Tanh can solve the problem that the membership influence is small or even invalid due to the multiplication of the membership in the related art when the value of the Gaussian membership is much smaller than 1, thereby further improving the performance of the electroencephalogram signal classification method.

[0110] In S205, the fuzzy antecedent rule is determined based on the membership information.

[0111] Exemplarily, after obtaining the multiplication operation result of each magnified Gaussian membership corresponding to the same fuzzy subset center, the obtained multiplication result can be normalized to obtain membership information meeting the value range requirement. Then, the membership information after the normalization can be determined as the fuzzy antecedent rule.

[0112] In S206, the corresponding feature weight is set for the electroencephalogram signal feature in each data dimension, and the electroencephalogram signal features in each data dimension are weighted and summed with the corresponding feature weight to obtain the fuzzy consequent rule.

[0113] Exemplarily, the obtained mth fuzzy consequent rule can be expressed as shown in the following formula (5).

[0114]

[0115] wherein, p mf represents the fth feature weight under the current fuzzy rule, represents the electroencephalogram signal feature in the fth data dimension of the electroencephalogram signal feature corresponding to the mth fuzzy rule.

[0116] In S207, the multiplication operation is performed on the multiplication membership information indicated by the fuzzy antecedent rule corresponding to each fuzzy rule and the electroencephalogram signal feature after the weighted summation indicated by the fuzzy consequent rule, and then the summation is determined to obtain the fuzzy output.

[0117] Exemplarily, the calculation formula of the fuzzy output can be shown in the following formula (6).

[0118]

[0119] Then, the fuzzy output shown in the formula (6) can be converted into a fuzzy output in a matrix form to facilitate the calculation of the fuzzy output. The conversion process of the fuzzy output shown in the formula (6) into the fuzzy output in the matrix form can be shown in the following formula (7) to formula (12).

[0120] p m =(p m0 ,p m1 ,p m2..., p mf ) T (7)

[0121] P = ((p1) T ..., (p m ) T ..., (p M ) T ) T (8)

[0122]

[0123]

[0124]

[0125]

[0126] In the embodiment of the present application, the feature weight corresponding to each electroencephalogram signal feature in the fuzzy output can be determined through the conversion process shown in the above formulas (7) to (12), that is, the fuzzy output formula after conversion can be represented as shown in the following formula (13). Wherein, the feature weight is determined based on the label information of the electroencephalogram signal data in the model training process, and the specific training process is described in the following embodiment, which will not be described in detail.

[0127]

[0128] S208, based on the ridge regression algorithm, a target optimization function containing fuzzy output and class information of the electroencephalogram signal data to be identified is constructed, and the class information of the electroencephalogram signal data to be identified is obtained after the target optimization function is solved.

[0129] Exemplarily, based on the ridge regression learning algorithm, a target optimization function containing the fuzzy output as shown in formula (13) and the class information (that is, Z in the following formula (14)) corresponding to the electroencephalogram signal data to be identified can be constructed, and the constructed target optimization function can be shown in the following formula (14).

[0130]

[0131] Wherein, λ is a pre-set ridge regression parameter, and Z indicates the class information of the electroencephalogram signal data to be identified.

[0132] In the embodiment of the present application, the value of Z can be obtained by solving the target optimization function as shown in formula (14), so that the class information of the electroencephalogram signal data to be identified can be determined according to the value of Z.

[0133] As can be seen from the above description, the EEG signal category recognition method provided in this application mainly consists of two parts: one part is to extract features from the EEG signal data to be recognized to obtain the EEG signal features of the EEG signal data to be recognized, and the other part is to determine fuzzy rules and fuzzy output based on the EEG signal data to be recognized and the EEG signal features.

[0134] Based on this Figure 3 This is a schematic diagram of the intelligent model corresponding to the brainwave signal category recognition method provided in this application embodiment. The following will be combined with... Figure 3 This application provides a detailed description of the implementation process of the electroencephalogram (EEG) signal category recognition method provided in its embodiments.

[0135] Depend on Figure 3 It can be seen that after acquiring the EEG signal data to be identified, data preprocessing can be performed on the EEG signal data to be identified. That is, the acquired EEG signal data to be identified can be transformed to obtain one-dimensional EEG signal data to be identified (wherein, the one-dimensional EEG signal data to be identified is also each one-dimensional EEG signal data). At this time, the EEG signal data to be identified can be denoted as X = (x1, x2, ..., x...). d ) T .

[0136] Subsequently, feature extraction can be performed based on the EEG signal data to be identified. That is, the preprocessed EEG signal data to be identified can be input into, for example,... Figure 3 In the three feature extraction models shown, the EEG signal features output by each feature extraction model are obtained. At this point, by... Figure 3 As can be seen, the first feature extraction model includes a Convolutional Neural Network (CNN) structure (i.e., the convolutional pooling layer mentioned above) and a Recurrent Neural Network (BiLSTM) structure (i.e., Figure 3 As shown in the CNN+BiLSTM example, the EEG signal features output by this feature extraction model are the spatial-temporal features of the EEG signal data to be processed. Figure 3 It can be seen that the second feature extraction model includes a convolutional neural network structure CNN (i.e., the convolutional pooling layer mentioned above) and a recurrent neural network structure GRU (i.e., Figure 3 As shown in the CNN+GRU example, the EEG signal features output by this feature extraction model are the spatial-temporal features of the EEG signal data to be processed. Figure 3 As can be seen, the third feature extraction model is a multilayer perceptron (MLP). In this case, the EEG signal features output by this feature extraction model are the deep features of the EEG signal data to be processed.

[0137] Then, as shown in Figure 3 , the fuzzy consequent rules can be determined based on the EEG signal features extracted by each feature extraction model. Specifically, a corresponding feature weight can be set for the EEG signal features in each data dimension, and the EEG signal features in each data dimension and the corresponding feature weight are weighted and summed to obtain the fuzzy consequent rules.

[0138] Meanwhile, fuzzy logic reasoning can also be performed based on the EEG signal data to be identified, that is, the fuzzy antecedent rules are determined based on the EEG signal data to be identified. Specifically, as shown in Figure 3 , each one-dimensional EEG signal data can be used as an antecedent variable, and the Gaussian membership degrees of the one-dimensional EEG signal data (i.e., the antecedent variable) in each fuzzy subset are first determined. Then, the Gaussian membership degrees are amplified and activated based on the hyperbolic tangent function Tanh, and the amplified Gaussian membership degrees are obtained after processing. Finally, the amplified Gaussian membership degrees corresponding to the same fuzzy subset center are multiplied to obtain the membership information. At this time, the fuzzy antecedent rules can be determined based on the membership information.

[0139] Next, the fuzzy antecedent rules and the fuzzy consequent rules corresponding to each fuzzy rule are multiplied and then summed to obtain the fuzzy output. After converting the fuzzy output into a matrix form, the fuzzy output expression can be determined as shown in the above formula (13).

[0140] Then, the fuzzy output can be solved by a ridge regression learning algorithm, and the class information of the EEG signal data to be identified is obtained after solving.

[0141] As described above, the EEG signal class identification method provided in the present application can extract multiple EEG signal features of the EEG signal data to be identified through multiple feature extraction models after obtaining the EEG signal data to be identified, so that more numerous and more diverse EEG signal features can be obtained, thereby facilitating accurate identification of the EEG signal class. After obtaining the EEG signal features, multiple fuzzy rules can be determined based on the EEG signal data to be identified and the EEG signal features, and a fuzzy output can be determined based on the multiple fuzzy rules, so that the EEG signal class identification method can be enhanced in terms of explainability through the multiple fuzzy rules, providing a certain basis for subsequent identification of the EEG signal class, and facilitating further research. Meanwhile, the uncertainty and fuzziness of the EEG signal data to be identified are processed and expressed through the fuzzy rules to determine the fuzzy output, so that the EEG signal class identification is more flexible, and the performance of the EEG signal class identification method is further improved.

[0142] In the embodiments of the present application, as shown in Figure 3Before the intelligent model shown in the figure identifies the electroencephalogram signal, the intelligent model needs to be trained, and after the training, the intelligent model is used to identify the category of the electroencephalogram signal. Among them, the hardware environment of the training is AMD Ryzen 53400 3.7GHz 32G RAM 64-bit Microsoft Windows 10, and the programming environment is Python 3.8 with Torch 1.6.0+CPU. At this time, the training of the intelligent model is mainly for the training of the feature extraction model, and the specific training process can be described as the following process.

[0143] In an example, the feature extraction model can be trained. Specifically, the training sample data set of the electroencephalogram signal data and the classification label corresponding to each training sample in the training sample data set can be obtained first.

[0144] Among them, the training sample data set of the electroencephalogram signal data is the CHB-MIT data set.

[0145] The inventors found in experiments that in the case of identifying whether the electroencephalogram signal is the electroencephalogram signal of the seizure period based on the category identification method of the electroencephalogram signal provided in the embodiments of the application, because the seizure time is short and the electroencephalogram signal data collection time is long, there is a serious data sample imbalance problem. Based on this, in order to ensure the completeness of the seizure data and balance the samples to the maximum extent, in the experiments, the electroencephalogram signal data during the seizure period is adopted in the way of overlapping sampling and data segmentation. The data is segmented into 2-second time periods, and the overlapping sampling time is 0.5 seconds. In addition, the random discarding method is also adopted to reduce the amount of electroencephalogram signal data samples of normal non-seizure epilepsy.

[0146] In the embodiments of the application, after obtaining the training sample data set, the training sample data set is preprocessed, that is, the low-pass filter is used to filter out the noise irrelevant to the seizure event in the training sample data set with a frequency higher than 64Hz, so as to further ensure the accuracy of the training sample data set.

[0147] In the embodiments of the application, the obtained training sample data set can be as shown in Table 1.

[0148] Table 1 CHB-MIT data set

[0149]

[0150] In the embodiments of the application, after obtaining the training sample of the electroencephalogram signal data and the classification label corresponding to each training sample in the training sample data set, the training sample data set can be input to the feature extraction model as shown in the figure. Figure 3The processing result can be understood as the class probability of the training sample obtained after full connection calculation of the EEG signal feature output by each feature extraction model based on the Softmax function.

[0151] Then, the cross-entropy loss value is calculated based on the processing result and the classification label.

[0152] Finally, the feature extraction model to be trained is iteratively trained based on the cross-entropy loss value, and the feature extraction model is obtained after training.

[0153] The number of iterations for iteratively training each feature extraction model is a pre-set value, for example, the number of iterations can be set to 30, or can be set to 50, etc. The number of iterations is not limited in the embodiments of the present application, and can be determined according to actual needs.

[0154] In the embodiments of the present application, after the feature extraction model is trained to obtain a trained feature extraction model, the feature weight corresponding to the EEG signal feature can be solved in the pre-set scale parameter range and the pre-set ridge regression parameter range, so as to obtain the fuzzy output.

[0155] Specifically, the fuzzy output containing variable feature weight can be constructed according to the training sample data, and then the target optimization function can be constructed based on the training sample data set and the classification label set corresponding to the training sample data set. At this time, the target optimization function is shown in the above formula (14), at this time, P in the target optimization function is an unknown variable, and Z is a classification label set, at this time, P obtained by solving can be expressed as shown in the following formula (15).

[0156] P=(K T K+λI) -1 K T Z (15)

[0157] In the embodiments of the present application, the training sample data set of the acquired EEG signal data can be divided into a training set and a test set, at this time, after the intelligent model shown in the above formula (1) is trained based on the training set, the test set can be used to test the intelligent model. Figure 3

[0158] The process of training the intelligent model shown in the above formula (1) provided by the embodiments of the present application will be introduced in detail below with reference to the pseudo code. Figure 3

[0159] Input: the training sample data set corresponding to the EEG signal data classification label set ​​The training set and the test set are divided, the number of fuzzy rules is M, the maximum number of iterations corresponding to each feature extraction model is γ, the scale parameter is ε, and the ridge regression parameter is λ;

[0160] Output: the class information of each test sample data in the test set;

[0161] Initialization: randomly initialize weights and biases;

[0162] forγ’<maxγdo:

[0163] Input the training set data to perform 1-dimensional convolutional neural network CNN convolution and pooling operation, bidirectional long short-term memory network BiLSTM or gated recurrent unit GRU calculation, and get the sample class probability through the Softmax function after the full connection layer. Cross-entropy loss function error back propagation;

[0164] Input the training set data to calculate the PReLU activation function nonlinear conversion by multilayer perceptron MLP, calculate the class probability by Softmax function, and perform error back propagation by cross-entropy loss function;

[0165] end

[0166] Respectively extract the feature vectors output by the bidirectional long short-term memory network BiLSTM in CNN-BiLSTM, the feature vectors output by the gated recurrent unit GRU in CNN-GRU, and the deep features output by MLP.

[0167] forεin 2 -6 ,2 -5 ,…,2 2 do:

[0168] forλin 10 -3 ,10 -2 ,…,10 2 do:

[0169] Training set data is calculated by fuzzy C clustering FCM algorithm; the fuzzy subset width information is calculated by formula (2);

[0170] Gaussian membership is calculated by formula (1), and Gaussian membership is converted by Tanh activation function in formula (4);

[0171] Each electroencephalogram feature is taken as a fuzzy consequent rule parameter;

[0172] According to formula (11), construct

[0173] Solve the consequent variable P in formula (15);

[0174] Extract each electroencephalogram feature of the test set;

[0175] calculating the fuzzy subset center and the fuzzy subset width information of the test set;

[0176] calculating the Gaussian membership of the test set by formula (2), and converting the Gaussian membership by a Tanh activation function according to formula (3);

[0177] calculating the electroencephalogram feature H i of the test set according to formula (10);

[0178] obtaining the prediction output according to formula (13);

[0179] end

[0180] end

[0181] It should be noted that the number M of the fuzzy rules can be a pre-set fixed value, for example, the value of M can be set to 3, or can be set to 5, etc. Alternatively, the number of fuzzy rules can also be a pre-set value range, so that the intelligent model as shown in FIG. 1 is trained in the value range, wherein the pre-set value range of the number M of fuzzy rules can be M = {3, 4, …, 10}. Figure 4 In the embodiments of the present application, after training the intelligent model as shown in FIG. 1, the effectiveness of the intelligent model can be analyzed by using the pointer accuracy Accuracy (ACC) and the statistical quantity F-Measure (F1). Wherein the calculation formula of the accuracy ACC is as shown in the following formula (16), and the statistical quantity F1 is as shown in the following formula (17).

[0182] Figure 4

[0183]

[0184]

[0185] Wherein TP represents True Positive (TP), which means that the sample of the epilepsy seizure period is correctly detected, TN represents True Negative (TN), which means that the healthy sample of the epilepsy non-seizure period is correctly detected, FP represents False Positive (FP), which means that the sample of the epilepsy non-seizure period is misdetected as the sample of the epilepsy seizure period, and FN represents False Negative (FN), which means that the sample of the epilepsy seizure period is misdetected as the healthy sample of the epilepsy non-seizure period.

[0186] ​​In the embodiments of the present application, the inventors compared the category recognition method of the electroencephalogram signal provided in the embodiments of the present application with other category recognition methods. The other category recognition methods are specifically divided into the following three cases.

[0187] Case one: a single deep learning network

[0188] The deep learning network compared in the embodiments of the present application includes: a one-dimensional convolutional neural network CNN combined with a bidirectional long short-term memory network BiLSTM (that is, CNN-BiLSTM), a convolutional neural network CNN combined with a gated recurrent unit GRU (that is, CNN-GRU), and a fully connected neural network (that is, a multilayer perceptron MLP).

[0189] Case two: a TSK fuzzy classification algorithm

[0190] The TSK fuzzy classification algorithm compared in the embodiments of the present application includes: a traditional TSK fuzzy classification algorithm (denoted as TSK), a TSK fuzzy classification algorithm with a log function (denoted as log-TSK), and a TSK fuzzy classification algorithm with a Tanh activation function (denoted as Tanh-TSK).

[0191] Case three: a TSK fuzzy classification algorithm taking electroencephalogram signal features as fuzzy consequent rules

[0192] The TSK fuzzy classification algorithm taking electroencephalogram signal features as fuzzy consequent rules compared in the embodiments of the present application includes: a TSK fuzzy classification algorithm taking electroencephalogram signal features extracted by CNN-BiLSTM, CNN-GRU, and MLP as fuzzy consequent rules (denoted as MV-TSK), which is constructed by fuzzy antecedent rules from original data.

[0193] The category recognition method of the electroencephalogram signal provided in the embodiments of the present application is a TSK fuzzy classification algorithm taking electroencephalogram signal features as fuzzy consequent rules with a Tanh activation function, which can be denoted as MV-Tanh-TSK (at this time, the MV-Tanh-TSK is also denoted as MV-Tanh-TSK-FC).

[0194] In the embodiments of the present application, in the process of verifying the category recognition method MV-Tanh-TSK-FC of the electroencephalogram provided by the embodiments of the present application, the preset value of the number of fuzzy rules M is M={3, 4, …, 10}, at this time, in the case where it is determined that the value of the number of fuzzy rules M is greater than the number of electroencephalogram features extracted based on the plurality of feature extraction models, the plurality of electroencephalogram features can be recycled to determine the fuzzy consequent rules corresponding to the fuzzy rules. When it is determined that the value of the number of fuzzy rules M is less than the number of electroencephalogram features, M electroencephalogram features can be selected from the plurality of electroencephalogram features to determine the fuzzy consequent rules.

[0195] In the embodiments of the present application, in the process of verifying the category recognition method MV-Tanh-TSK-FC of the electroencephalogram provided by the embodiments of the present application, the preset value of the scale parameter ε is ε={2 -6 ,2 -5 ,…,2 2}, and the preset ridge regression parameter is λ={10 -3 ,10 -2 ,…,10 2}.

[0196] At this time, the TSK, log-TSK, Tanh-TSK, and MV-TSK are all first order, and the adjustable parameter settings in each algorithm are the same as MV-Tanh-TSK-FC.

[0197] At this time, each of the above algorithms can be tested on the CHB-MIT dataset, and the statistical results of the accuracy Accuracy and the statistical quantity F-Measure of each algorithm are obtained after the test, and the results are shown in Tables 2 and 3.

[0198] Table 2 Accuracy Accuracy values of each algorithm on the CHB-MIT dataset

[0199]

[0200]

[0201] From Table 2, it can be seen that the average accuracy of the MV-Tanh-TSK-FC algorithm provided by the embodiments of the present application can reach 98.19%, which has better classification performance compared to other algorithms. In the classification task of the electroencephalogram of the epilepsy patient, MV-Tanh-TSK-FC has obvious advantages.

[0202] Table 3 Statistical quantity F-Measure values of each algorithm on the CHB-MIT dataset

[0203]

[0204] From the above Table 3, it can be seen that the average statistical index of the MV-Tanh-TSK-FC algorithm provided in the embodiment of the present application can reach 96.71%, which has certain improvement compared with other algorithms, and therefore, the MV-Tanh-TSK-FC algorithm provided in the embodiment of the present application has obvious advantages.

[0205] In the embodiment of the present application, in order to observe whether the MV-Tanh-TSK algorithm provided in the embodiment of the present application has significant difference, the MV-Tanh-TSK algorithm provided in the embodiment of the present application and the above-mentioned other classification recognition methods are subjected to non-parametric statistical analysis. Specifically, Frideman Ranking and Holm post-hoc test can be performed for non-parametric statistical analysis.

[0206] Figure 5 A schematic diagram of a Frideman Ranking analysis result provided in the embodiment of the present application is shown in FIG. 3. Figure 6 It can be seen that the MV-Tanh-TSK-FC algorithm provided in the embodiment of the present application achieves the optimal level, which indicates that the MV-Tanh-TSK-FC algorithm provided in the embodiment of the present application has significant difference with the other seven comparative algorithms.

[0207] In the embodiment of the present application, the Holm post-hoc result after the non-parametric statistical analysis can be shown in Table 4. From Table 4, it can be seen that under the condition of confidence of 0.05, the Holm post-hoc test result shows that the MV-Tanh-TSK is rejected (Rejected) to its comparative algorithm, which indicates that in the comparative experiment, the deep fuzzy classification algorithm MV-Tanh-TSK-FC proposed in the present application is the optimal one and has significant difference.

[0208] Table 4: Holm post-hoc result of each algorithm

[0209]

[0210] Meanwhile, in the embodiment of the present application, the effectiveness of converting Gaussian membership degree by using Tanh activation function can also be verified, at this time, two sets of control experiments can be used to verify the effectiveness of converting Gaussian membership degree by using Tanh activation function. The first kind: TSK algorithm and Tanh-TSK algorithm based on the same data; the second kind: MV-TSK algorithm and MV-Tanh-TSK-FC algorithm, at this time, both of the two algorithms use original data to construct fuzzy antecedent rules, and the EEG signal feature information extracted by CNN-BiLSTM, CNN-GRU and MLP is used as fuzzy consequent rules.

[0211] As shown in Table 2, the accuracy of the TSK algorithm is 96.51%, and the accuracy of the Tanh-TSK algorithm is 96.87%, and the accuracy is increased by 0.36% after converting Gaussian membership degree by using Tanh activation function. As shown in Table 2, the accuracy of the MV-TSK algorithm is 97.80%, and the accuracy of the MV-Tanh-TSK-FC algorithm is 98.19%, and the accuracy is increased by 0.39% after converting Gaussian membership degree by using Tanh activation function. The accuracy of the MV-TSK algorithm is 97.80%, and the accuracy of the MV-Tanh-TSK-FC algorithm is 98.19%, and the accuracy is increased by 0.39% after converting Gaussian membership degree by using Tanh activation function.

[0212] Meanwhile, as shown in Table 3, the statistical quantity F-Measure value of the TSK algorithm is 93.62%, and the statistical quantity F-Measure value of the Tanh-TSK algorithm is 94.12%, and the statistical quantity F-Measure value is increased by 0.5% after converting Gaussian membership degree by using Tanh activation function. The statistical quantity F-Measure value of the MV-TSK algorithm is 96.04%, and the statistical quantity F-Measure value of the MV-Tanh-TSK-FC algorithm is 96.71%, and the statistical quantity F-Measure value is increased by 0.67% after converting Gaussian membership degree by using Tanh activation function.

[0213] Therefore, the TSK fuzzy classification algorithm with Tanh activation function is more stable in performance when classifying EEG signal data.

[0214] In the embodiment of the present application, the indexes of each algorithm on the CHB-MIT data set can be drawn based on the statistics in Table 2 and Table 3. Therefore, the effectiveness of converting Gaussian membership degree by using Tanh activation function can be more clearly and intuitively determined based on the line graph.

[0215] Figure 5 The accuracy Accuracy line graph of two sets of control experiments on the CHB-MIT data set provided in the embodiment of the present application,Figure 6 A statistical quantity F-Measure fold line diagram of two groups of control experiments provided by an embodiment of the present application on the CHB-MIT data set. Here, Figure 5 and Figure 6 The numerical values corresponding to the fold line graphs in Table 2 and Table 3, so in Figure 5 and Figure 6 Detailed annotations are not made.

[0216] From Figure 7 and Figure 8 It can be seen that the TSK fuzzy classification algorithm with Tanh activation function is more stable in performance and has higher accuracy in the classification of electroencephalogram data of epilepsy patients.

[0217] The inventors found that the TSK fuzzy classification algorithm with log function uses a logarithmic function to smooth the Gaussian membership degree, thereby improving the prediction performance of the algorithm. Therefore, in the embodiments of the present application, the log-TSK algorithm is used as a comparative algorithm to compare the MV-Tanh-TSK-FC algorithm with the added Tanh activation function of the present application.

[0218] Figure 7 A comparison result diagram of the accuracy Accuracy of a log-TSK algorithm and an MV-Tanh-TSK-FC algorithm provided by an embodiment of the present application, Figure 8 A comparison result diagram of the statistical quantity F-Measure of a log-TSK algorithm and an MV-Tanh-TSK-FC algorithm provided by an embodiment of the present application. Among them, Figure 7 indicates the comparison result of the accuracy Accuracy of the log-TSK algorithm and the MV-Tanh-TSK (i.e. MV-Tanh-TSK-FC) algorithm, Figure 8 indicates the comparison result of the statistical quantity F-Measure of the log-TSK algorithm and the MV-Tanh-TSK (i.e. MV-Tanh-TSK-FC) algorithm. Here, Figure 7 and Figure 8 The numerical values in Table 2 and Table 3, so in Figure 7 and Figure 8 The specific numerical values are not marked.

[0219] From Figure 9 and Figure 9 It can be seen that the MV-Tanh-TSK-FC algorithm, after amplifying the Gaussian membership degree by the Tanh function, alleviates the problem of membership degree multiplication close to 0 caused by too small membership degree value, thereby making its classification performance better.

[0220] In the embodiments of the present application, the MV-Tanh-TSK (i.e., MV-Tanh-TSK-FC) algorithm provided in the embodiments of the present application can also be compared with related technologies, and the comparison results are shown in Table 5.

[0221] Table 5 Comparison of related technologies and MV-Tanh-TSK on CHB-MIT dataset

[0222]

[0223] As shown in Table 5, the accuracy of the MV-Tanh-TSK algorithm provided in the embodiments of the present application is higher. The classification accuracy of the method proposed by He et al. based on graph attention network (GAT) to extract spatial features and based on Bi-LSTM to learn the spatial and temporal features of electroencephalogram data is higher than that of the present application. However, this method is a black box classification algorithm, which leads to poor interpretability of the model.

[0224] Those skilled in the art can understand that, in the above method of the specific implementation, the writing order of each step does not mean a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0225] Based on the same inventive concept, the embodiments of the present application also provide an electroencephalogram category identification device corresponding to the electroencephalogram category identification method. Since the principle of solving problems in the device of the embodiments of the present application is similar to the above-mentioned electroencephalogram category identification method of the embodiments of the present application, the implementation of the device can be referred to the implementation of the method, and the repeated parts will not be described here.

[0226] Figure 10 A structural schematic diagram of an electroencephalogram category identification device provided in the embodiments of the present application is shown in FIG. 9. Figure 10 As shown in FIG. 9, the electroencephalogram category identification device 900 includes an acquisition unit 901, a feature extraction unit 902, a determination unit 903, and a classification unit 904.

[0227] The first acquisition unit 901 is configured to acquire electroencephalogram data to be identified.

[0228] The feature extraction unit 902 is configured to perform feature extraction processing on the electroencephalogram data to be identified based on each feature extraction model in a plurality of feature extraction models, to obtain electroencephalogram features output by each feature extraction model. The electroencephalogram features are any one of the following: spatial and temporal features of the electroencephalogram data, deep features of the electroencephalogram data.

[0229] The determining unit 903 is configured to determine a plurality of fuzzy rules based on the electroencephalogram data to be identified and the electroencephalogram features; wherein the fuzzy rules indicate the category information of the electroencephalogram data to be identified under corresponding fuzzy subsets; and the fuzzy subsets indicate at least part of the electroencephalogram data to be identified that meets the similarity requirement.

[0230] The classifying unit 904 is configured to determine a fuzzy output based on the fuzzy rules, and determine the category information of the electroencephalogram data to be identified based on the fuzzy output.

[0231] By way of example, the embodiments can refer to the above-described method embodiments, which have similar principles and technical effects, and thus will not be described herein.

[0232] Figure 11 Another structure diagram of an electroencephalogram category identification device provided by the embodiments of the present application is shown in FIG. 10. Figure 11 As shown in FIG. 10, the electroencephalogram category identification device 1000 includes an obtaining unit 1001, a feature extraction unit 1002, a determining unit 1003, and a classifying unit 1004.

[0233] The first obtaining unit 1001 is configured to obtain electroencephalogram data to be identified.

[0234] The feature extraction unit 1002 is configured to perform feature extraction processing on the electroencephalogram data to be identified based on each feature extraction model in a plurality of feature extraction models, to obtain electroencephalogram features output by each feature extraction model; and the electroencephalogram features are any one of the following: spatial and temporal features of the electroencephalogram data, deep features of the electroencephalogram data.

[0235] The determining unit 1003 is configured to determine a plurality of fuzzy rules based on the electroencephalogram data to be identified and the electroencephalogram features; wherein the fuzzy rules indicate the category information of the electroencephalogram data to be identified under corresponding fuzzy subsets; and the fuzzy subsets indicate at least part of the electroencephalogram data to be identified that meets the similarity requirement.

[0236] The classifying unit 1004 is configured to determine a fuzzy output based on the fuzzy rules, and determine the category information of the electroencephalogram data to be identified based on the fuzzy output.

[0237] In one example, the fuzzy rules include fuzzy antecedent rules and fuzzy consequent rules; the fuzzy antecedent rules indicate the fuzzy subsets corresponding to the electroencephalogram data to be identified; and the fuzzy consequent rules indicate the category information of the electroencephalogram data to be identified under the fuzzy subsets.

[0238] The determining unit 1003 includes:

[0239] The first determination module 10031 is configured to divide the brain electrical signal data to be recognized into a plurality of fuzzy subsets based on a preset number of fuzzy rules.

[0240] The second determination module 10032 is configured to determine membership information of the brain electrical signal data to be recognized in each fuzzy subset, and determine fuzzy antecedent rules based on the membership information. The membership information indicates a degree of association between the brain electrical signal data to be recognized and the fuzzy subset.

[0241] The third determination module 10033 is configured to set a corresponding feature weight for the brain electrical signal feature in each data dimension, and obtain a fuzzy consequent rule by weighted summation of the brain electrical signal feature in each data dimension and the corresponding feature weight.

[0242] In one example, the fuzzy subset is obtained by clustering each one-dimensional brain electrical signal data corresponding to the brain electrical signal data to be recognized according to the preset number of fuzzy rules; the fuzzy subset includes corresponding fuzzy subset center and fuzzy subset width information; the fuzzy subset center indicates a center point corresponding to each one-dimensional brain electrical signal data satisfying a similarity requirement; and the fuzzy subset width information indicates a distribution range corresponding to each one-dimensional brain electrical signal data satisfying the similarity requirement.

[0243] The second determination module 10032 includes:

[0244] The first determination submodule 100321 is configured to determine a Gaussian membership between each one-dimensional brain electrical signal data and each fuzzy subset center based on the fuzzy subset center and the fuzzy subset width information; the Gaussian membership indicates a degree of association between the one-dimensional brain electrical signal data and the fuzzy subset center.

[0245] The second determination submodule 100322 is configured to perform amplification and activation processing on the Gaussian membership based on a hyperbolic tangent Tanh function, and obtain an amplified Gaussian membership after processing.

[0246] The third determination submodule 100323 is configured to perform a multiplication operation on each amplified Gaussian membership corresponding to the same fuzzy subset center, and obtain the membership information after multiplication.

[0247] In one example, the classification unit 1004 includes:

[0248] The fourth determination module 10041 is configured to perform a product operation on the multiplied membership information indicated by the fuzzy antecedent rule corresponding to each fuzzy rule and the brain electrical signal feature after weighted summation indicated by the fuzzy consequent rule, and then perform summation to determine a fuzzy output.

[0249] The fifth determining module 10042 is configured to construct a target optimization function containing the fuzzy output and the category information of the brain electrical signal data to be identified based on a ridge regression algorithm, and obtain the category information of the brain electrical signal data to be identified by solving the target optimization function.

[0250] In one example, the feature extraction model is a spatial-temporal feature extraction model; and the spatial-temporal feature extraction model is a model constructed based on a convolutional neural network structure and a recurrent neural network structure.

[0251] The feature extraction unit 1002 comprises:

[0252] The first processing module 10021 is configured to input the brain electrical signal data to be identified into the spatial-temporal feature extraction model for processing, to obtain spatial-temporal features of the brain electrical signal data to be identified.

[0253] In one example, the feature extraction model comprises a multi-layer perceptron; and the number of neurons contained in the perceptrons at different levels of the multi-layer perceptron is different. The feature extraction unit 1002 can further comprise:

[0254] The second processing module 10022 is configured to input the brain electrical signal data to be identified into the multi-layer perceptron for processing, and determine the features extracted by the perceptron at a layer before the output layer of the multi-layer perceptron as deep-level features of the brain electrical signal data to be identified.

[0255] In one example, the apparatus provided by the embodiment can further comprise:

[0256] The second acquisition unit 1005 is configured to acquire a training sample data set of the brain electrical signal data and a classification label corresponding to each training sample in the training sample data set.

[0257] The processing unit 1006 is configured to input the training sample data set into the feature extraction model to be trained for processing, to obtain a processing result.

[0258] The calculation unit 1007 is configured to calculate a cross-entropy loss value based on the processing result and the classification label.

[0259] The training unit 1008 is configured to iteratively train the feature extraction model to be trained based on the cross-entropy loss value, to obtain the feature extraction model.

[0260] By way of example, the embodiment can refer to the method embodiments described above, and the principles and technical effects are similar, which will not be described herein.

[0261] ​ A structural schematic diagram of a computer device provided by the embodiment of the present application is shown in FIG. 1. ​As shown, the computer device 1100 includes a memory 1101 and a processor 1102.

[0262] The memory 1101 is configured to store instructions executable by the processor 1102.

[0263] The processor 1102 is configured to perform the method provided by the above-mentioned embodiments.

[0264] The computer device further includes a receiver 1103 and a transmitter 1104. The receiver 1103 is configured to receive instructions and data sent by external devices, and the transmitter 1104 is configured to send instructions and data to external devices.

[0265] Exemplarily, the embodiments can refer to the above-mentioned method embodiments, and the principles and technical effects are similar, which will not be described herein.

[0266] The embodiments of the present application further provide a computer readable storage medium, which stores a computer program. The computer program is run by a processor to perform the steps of the electroencephalogram signal category identification method in the above-mentioned method embodiments. The storage medium can be a volatile or non-volatile computer readable storage medium.

[0267] The embodiments of the present application further provide a computer program product, which carries a program code. The instructions included in the program code can be used to perform the steps of the electroencephalogram signal category identification method in the above-mentioned method embodiments. For details, please refer to the above-mentioned method embodiments, which will not be described herein.

[0268] Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. The specification and examples given are exemplary only and the true scope and spirit of the application is indicated by the following claims. It will be appreciated by persons skilled in the art that numerous variations and / or modifications can be made to the application as described above without departing from the scope or spirit of the application. It is intended that all such variations and / or modifications be included within the scope of the application. The specification and examples given are exemplary only and the true scope and spirit of the application is indicated by the following claims.

[0269] It should be understood that the application is not limited to the precise construction that has been described above and shown in the accompanying drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application is limited only by the appended claims.

Claims

1. A method for classifying electroencephalogram (EEG) signals, characterized in that, include: Acquire the EEG signal data to be identified; The EEG signal data to be identified is processed by feature extraction based on each of the multiple feature extraction models to obtain the EEG signal features output by each feature extraction model; the EEG signal features are any one of the following: spatial and temporal features of the EEG signal data, or deep features of the EEG signal data. Multiple fuzzy rules are determined based on the EEG signal data to be identified and the EEG signal features; wherein, the fuzzy rules indicate the category information of the EEG signal data to be identified under a corresponding fuzzy subset; the fuzzy subset indicates at least a portion of the EEG signal data to be identified that meets the similarity requirements; Based on the fuzzy rules, a fuzzy output is determined, and based on the fuzzy output, the category information of the EEG signal data to be identified is determined.

2. The method according to claim 1, characterized in that, The fuzzy rules include fuzzy antecedent rules and fuzzy consequent rules; the fuzzy antecedent rules indicate the fuzzy subset corresponding to the EEG signal data to be identified; The fuzzy consequent rule indicates the category information of the EEG signal data to be identified under the corresponding fuzzy subset; The determination of multiple fuzzy rules based on the EEG signal data to be identified and the EEG signal features includes: The EEG signal data to be identified is divided into multiple fuzzy subsets based on a pre-set number of fuzzy rules; The membership information of the EEG signal data to be identified in each of the fuzzy subsets is determined, and the fuzzy antecedent rule is determined based on the membership information; the membership information indicates the degree of association between the EEG signal data to be identified and the fuzzy subset. The EEG signal features under each data dimension are assigned corresponding feature weights. The EEG signal features under each data dimension are weighted and summed with their corresponding feature weights to obtain the fuzzy consequent rule.

3. The method according to claim 2, characterized in that, The fuzzy subset is obtained by clustering the one-dimensional EEG signal data corresponding to the EEG signal data to be identified according to the pre-set number of fuzzy rules; the fuzzy subset includes the corresponding fuzzy subset center and fuzzy subset width information. The fuzzy subset center indicates the center point corresponding to each of the one-dimensional EEG signal data that meets the similarity requirement; The fuzzy subset width information indicates the distribution range of each of the one-dimensional EEG signal data that meets the similarity requirements; Determining the membership information of the EEG signal data to be identified in each of the fuzzy subsets includes: Based on the fuzzy subset center and the fuzzy subset width information, the Gaussian membership degree between each of the one-dimensional EEG signal data and each of the fuzzy subset centers is determined; the Gaussian membership degree indicates the degree of association between the one-dimensional EEG signal data and the fuzzy subset center. The Gaussian membership degree is amplified and activated based on the hyperbolic tangent Tanh function, and the amplified Gaussian membership degree is obtained after processing. The amplified Gaussian membership degrees corresponding to the same fuzzy subset center are multiplied together to obtain the membership degree information.

4. The method according to claim 3, characterized in that, The step of determining the fuzzy output based on the fuzzy rules and determining the category information of the EEG signal data to be identified based on the fuzzy output includes: The fuzzy output is determined by multiplying the membership information indicated by the fuzzy antecedent rule corresponding to each fuzzy rule and the EEG signal features indicated by the weighted summation of the fuzzy consequent rule. A target optimization function containing the fuzzy output and the category information of the EEG signal data to be identified is constructed based on the ridge regression algorithm, and the category information of the EEG signal data to be identified is obtained by solving the target optimization function.

5. The method according to claim 1, characterized in that, The feature extraction model is a spatial-temporal feature extraction model; the spatial-temporal feature extraction model is a model constructed based on convolutional neural network structure and recurrent neural network structure; the feature extraction process is performed on the EEG signal data to be identified by each of the multiple feature extraction models to obtain the EEG signal features output by each feature extraction model, including: The EEG signal data to be identified is input into the spatial-temporal feature extraction model for processing to obtain the spatial-temporal features of the EEG signal data to be identified.

6. The method according to claim 1, characterized in that, The feature extraction model is a multilayer perceptron; the feature extraction process is performed on the EEG signal data to be identified based on each of the multiple feature extraction models to obtain the EEG signal features output by each feature extraction model, including: The EEG signal data to be identified is input into the multilayer perceptron for processing, and the features extracted by the perceptron located in the layer before the output layer of the multilayer perceptron are determined as the deep features of the EEG signal data to be identified.

7. The method according to any one of claims 1-6, characterized in that, The method further includes: Acquire a training sample dataset of EEG signal data and the classification label corresponding to each training sample in the training sample dataset; The training sample dataset is input into the feature extraction model to be trained for processing to obtain the processing result; Calculate the cross-entropy loss value based on the processing result and the classification label; The feature extraction model to be trained is iteratively trained based on the cross-entropy loss value, and the feature extraction model is obtained after training.

8. A device for classifying electroencephalogram (EEG) signals, characterized in that, include: The first acquisition unit is used to acquire the EEG signal data to be identified; The feature extraction unit is used to perform feature extraction processing on the EEG signal data to be identified based on each of the multiple feature extraction models to obtain the EEG signal features output by each feature extraction model; the EEG signal features are any one of the following: spatial temporal features of the EEG signal data, or deep features of the EEG signal data; A determining unit is configured to determine multiple fuzzy rules based on the EEG signal data to be identified and the EEG signal features; wherein the fuzzy rules indicate the category information of the EEG signal data to be identified under a corresponding fuzzy subset; the fuzzy subset indicates at least a portion of the EEG signal data to be identified that meets the similarity requirement; A classification unit is used to determine a fuzzy output based on the fuzzy rules, and to determine the category information of the EEG signal data to be identified based on the fuzzy output.

9. A computer device, characterized in that, include: A processor, and a memory communicatively connected to the processor, the memory storing computer-executable instructions, the processor executing the computer-executable instructions stored in the memory to implement the steps of the electroencephalogram (EEG) signal category recognition method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when executed by a processor, are used to implement the steps of the method for classifying EEG signals as described in any one of claims 1 to 7.

Images