Electroencephalogram recognition system and method based on learnable entropy power transform and radial basis

Abstract

The present application relates to the technical field of electroencephalogram recognition, and provides an electroencephalogram recognition system and method based on a learnable entropy power transform and a radial basis function network. The electroencephalogram recognition method based on the learnable entropy power transform and the radial basis function network comprises an electroencephalogram acquisition module and a processor; the processor comprises a time-frequency window selection module; a local statistical feature extraction module calculates a local standard deviation corresponding to an electroencephalogram sample of each channel, frequency band and time window, so as to serve as a local statistical feature; a learnable entropy power transform module converts the local statistical feature into a task-adaptive entropy power transform feature according to a learnable scale power parameter; a radial basis function network module calculates a radial basis function network similarity according to a learnable network center and a learnable kernel width parameter, and then combines a linear output weight and a bias to output an electroencephalogram state probability; the method has the advantages of compact parameters, being suitable for small-sample electroencephalogram recognition, and being deployable in an embedded brain-computer interface or a medical monitoring device.

Description

Technical Field

[0001] This invention relates to the field of electroencephalogram (EEG) recognition technology, and in particular to an EEG recognition system and method based on learnable entropy power transformation and radial basis function networks. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Electroencephalogram (EEG) signals are among the most commonly used physiological signals in brain-computer interfaces and neurocognitive assessments. Existing EEG classification methods broadly include handcrafted features combined with traditional classifiers, spatial filtering methods, and deep neural network methods. Handcrafted feature methods, such as differential entropy, filter bank differential entropy, CSP, FBCSP, and wavelet transform, have advantages such as small parameter count and the ability to train on small sample data; however, their transformation forms are usually fixed in advance. For example, differential entropy or log-variance features use the same mathematical mapping for all frequency bands and all time windows, failing to adaptively enhance more discriminative time-frequency components based on the final recognition target, nor can they selectively compress components with high noise or weak contribution. Traditional classifiers such as SVM and KNN are usually separated from the feature extraction stage. Once the front-end features are fixed, the back-end classifier can only find the boundary within the existing feature space, making it difficult to influence the scale distribution of features. While deep convolutional networks, Transformers, and their variants can automatically learn features, they often face problems such as large parameter count, unstable training, susceptibility to individual noise, and overfitting in small EEG samples and cross-time period validation scenarios. Furthermore, existing technologies remain at the offline algorithm level and do not fully consider the system-level closed loop between electrode acquisition, amplification, analog-to-digital conversion, stimulus synchronization, embedded inference, and result output, making it difficult to directly form deployable EEG recognition devices. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides an EEG recognition system and method based on learnable entropy power transformation and radial basis function networks. This system can extract stable, low-dimensional statistical features containing discriminative information from the local time-frequency structure of multi-channel EEG data. It utilizes entropy power theory to perform learnable nonlinear scaling of the local amplitude scale and employs a nonlinear classifier with a small number of parameters to form a stable similarity mapping for high-dimensional EEG features. The system is implemented with a clear modular process on the acquisition hardware and embedded processor, making it suitable for scenarios such as brain-computer interfaces, neurocognitive assessment, and epilepsy detection.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides an EEG recognition system based on learnable entropy power transformation and radial basis function network.

[0006] In one or more embodiments, an EEG recognition system based on learnable entropy power transformation and radial basis function network is provided, including: an EEG acquisition module and a processor; The EEG acquisition module is used to acquire multi-channel EEG signals of a specified object and form an EEG sample; The processor includes: The time-frequency window selection module is used to divide EEG samples into multiple time windows and filter them into multiple frequency bands; The local statistical feature extraction module is used to calculate the local standard deviation corresponding to each channel, frequency band, and time window EEG sample, as a local statistical feature. The learnable entropy power transformation module is used to convert local statistical features into task-adaptive entropy power transformation features based on learnable scale power parameters. The normalization and vectorization module is used to expand the feature tensor composed of the task-adaptive entropy power transformation features into a one-dimensional vector in a set order, and perform normalization processing according to the learnable scaling parameter and the learnable translation parameter to obtain the feature vector. The radial basis function network module is used to calculate the radial basis function network similarity based on the learnable network center and learnable kernel width parameters, and then combine the linear output weights and the biased output EEG state probability. The result output module is used to output the cognitive state, pathological state, or brain event category based on the probability of the EEG state.

[0007] A second aspect of the present invention provides an EEG recognition method based on learnable entropy power transformation and radial basis function network.

[0008] In one or more embodiments, an EEG recognition method based on learnable entropy power transformation and radial basis function networks performs the following steps within a processor: Acquire multi-channel EEG signals from a specified object and form an EEG sample; The EEG samples were divided into multiple time windows and filtered into multiple frequency bands; Calculate the local standard deviation for each channel, frequency band, and time window EEG sample as a local statistical feature; Based on the learnable scale power parameter, local statistical features are transformed into task-adaptive entropy power transformation features. The feature tensor composed of the task-adaptive entropy power transformation features is expanded into a one-dimensional vector in a set order, and then normalized according to the learnable scaling parameter and the learnable translation parameter to obtain the feature vector. The radial basis function network similarity is calculated based on the learnable network center and learnable kernel width parameter, and then combined with the linear output weights and the biased output EEG state probability. Based on the probability of the EEG state, the cognitive state, pathological state, or EEG event category is output.

[0009] A third aspect of the present invention provides a computer-readable storage medium.

[0010] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the EEG recognition method based on learnable entropy power transformation and radial basis function network as described above.

[0011] A fourth aspect of the present invention provides an electronic device.

[0012] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the EEG recognition method based on learnable entropy power transformation and radial basis function network as described above.

[0013] Compared with the prior art, the beneficial effects of the present invention are: The present invention relates to an EEG recognition system based on learnable entropy power transformation and radial basis function networks. The system utilizes an EEG acquisition module to collect multi-channel EEG signals from a target object and form EEG samples. A time-frequency window selection module divides the EEG samples into multiple time windows and filters them into multiple frequency bands. A local statistical feature extraction module calculates the local standard deviation corresponding to each channel, frequency band, and time window EEG sample as a local statistical feature. A learnable entropy power transformation module converts the local statistical features into learnable entropy power transformation features based on a learnable scale power parameter. This allows the EEG scale of different frequency bands and time windows to adaptively adjust according to the task objective, overcoming the limitations of fixed difference EEG models. To address the issue of insufficient expressive power of entropy or fixed log-variance features, this paper utilizes normalization and vectorization modules to expand the feature tensor composed of task-adaptive entropy power transformation features in a set order, and normalizes it into feature vectors based on learnable scaling and translation parameters. The radial basis function network module calculates the similarity of the radial basis function network based on the learnable network center and learnable kernel width parameters. Then, by combining linear output weights and biased output EEG state probabilities, the cognitive state, pathological state, or EEG event category can be determined. This can be implemented in a clear modular process on acquisition hardware and embedded processors, making it suitable for scenarios such as brain-computer interfaces, neurocognitive assessment, and epilepsy detection. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0015] Figure 1This is a schematic diagram of the EEG recognition system based on learnable entropy power transformation and radial basis function network according to an embodiment of the present invention; Figure 2 This is a flowchart of an EEG recognition method based on learnable entropy power transformation and radial basis function network according to an embodiment of the present invention. Detailed Implementation

[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0017] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0019] according to Figure 1 As shown, this embodiment of the invention provides an EEG recognition system based on learnable entropy power transformation and radial basis function network, including: an EEG acquisition module 1 and a processor 2.

[0020] In this embodiment, the EEG acquisition module 1 is used to acquire multi-channel EEG signals of a designated object and form an EEG sample.

[0021] Specifically, the EEG acquisition module 1 includes a multi-channel EEG electrode cap or lead electrodes, a low-noise amplifier, an anti-aliasing filter, an analog-to-digital converter, and a data transmission interface. The multi-channel EEG electrode cap or lead electrodes are applied to the scalp and the acquisition is legally performed with the consent of the subject. Subsequent processing of the multi-channel EEG signals is also performed with the consent of the subject.

[0022] In specific implementation, processor 2 includes the following modules: The time-frequency window selection module 21 is used to divide the EEG samples into multiple time windows and filter them into multiple frequency bands.

[0023] Let the first Each EEG sample or test is represented as a matrix after preprocessing. ,in Indicates the number of brainwave channels. This represents the number of sampling points in an analysis segment. Indicates the sample index. This represents the total number of samples in the training or testing batch. Let the sampling rate be... The time window set is The frequency band set is ,in For the number of time windows, This refers to the number of frequency bands. (For) The The first channel in the The frequency band and the first The signal within each time window is denoted as ,in , , , , Indicates the first The number of sampling points within a time window.

[0024] In the time-frequency window selection module, preferably, for a 4-second event-related EEG segment, three time windows of 2 seconds each with a 50% overlap are used. Second, Seconds and The filter operates within a 0.5 Hz to 40 Hz range, employing a combination of full-band and multi-scale overlapping frequency bands to simultaneously preserve broadband energy information, medium-scale rhythmic information, and local frequency information. Zero-phase digital bandpass filtering can be used to reduce phase distortion introduced by the filter.

[0025] In this embodiment, the local statistical feature extraction module 22 is used to calculate the local standard deviation corresponding to each channel, frequency band, and time window EEG sample as a local statistical feature.

[0026] Specifically, in the local statistical feature extraction module, the local standard deviation for each channel, frequency band, and time window EEG sample is: ; ; ; in, Local standard deviation; It is a non-negative preset constant, and its dimensions are the same as those of the previous one. To prevent numerical instability caused by zero variance, the same approach is preferred. to ; It has locally unbiased variance; It is a local mean; EEG signals; For the first Number of sampling points within a time window; subscript , , and They represent the first The first EEG sample, the first The first channel, the first The frequency band and the first A time window.

[0027] Through the above calculations, the original EEG sample Converted to local standard deviation tensor This tensor significantly reduces the dimensionality at the sample level while preserving the three-dimensional structure of channels, frequency bands, and time windows.

[0028] In this embodiment, the learnable entropy power transformation module 23 is used to convert local statistical features into task-adaptive entropy power transformation features based on the learnable scale power parameter.

[0029] For continuous random variables Entropy power is defined as : ; in express The differential entropy.

[0030] Each locally filtered EEG segment is approximated as a Gaussian variable. Then its differential entropy is: ; Will Substituting the expression ,get This invention does not strictly limit local EEG to a Gaussian distribution, but rather uses this approximation to establish a learnable relationship between entropy power and local amplitude scale.

[0031] Further, a learnable scale power parameter is introduced. The generalized power transform representation of entropy power, i.e., the entropy power transformation characteristic of task adaptation, is as follows: ; ; in, Entropy power transformation characteristics adapted for the task; Local standard deviation; These are unconstrained parameters; For the Sigmoid function; and These represent the preset lower and upper limits, respectively; subscript , , and They represent the first The first EEG sample, the first The first channel, the first The frequency band and the first A time window.

[0032] Preferably, It shares resources at the channel level but learns independently at the frequency band and time window levels; that is, all channels within the same frequency band and time window use the same... This setup significantly reduces the number of trainable parameters and makes the scale power parameter primarily reflect the task-related scale modulation of the time-frequency components, rather than random fluctuations on a single electrode.

[0033] In other embodiments, the following methods may also be used: In any form, as long as it can be learned and used... All scale power parameters can fall within the scope of this invention.

[0034] In this embodiment, the normalization and vectorization module 24 is used to transform the feature tensor composed of the task-adaptive entropy power transformation features. Expanded into one-dimensional vectors according to a set order, and then normalized into feature vectors based on learnable scaling and translation parameters. ,in .

[0035] right Execution layer normalization is obtained : ; in, and They are vectors The mean and variance, and For layer-normalized learnable scaling and translation parameters, This represents element-wise multiplication. The normalized stability constant is preferably... .

[0036] In this embodiment, the radial basis function network module 25 is used to calculate the radial basis function network similarity based on the learnable network center and learnable kernel width parameters, and then combine the linear output weights and the biased output EEG state probability.

[0037] Radial basis function network module includes A learning network center Learnable kernel width parameter Linear output weights and bias ,in .

[0038] In the radial basis function network module, the process of calculating the probability of EEG states is as follows: according to Calculate the first The squared Euclidean distance between the centers of a learnable network; according to Calculate the Gaussian radial basis response ; all Composition of network similarity vectors ; according to and Calculate the probability of brainwave states ; in, These are normalized feature vectors; For the first A learning network center For learnable kernel width parameters; The output vector of the output layer of the radial basis functions; For bias.

[0039] In this embodiment, the result output module 26 is used to output the cognitive state, pathological state, or brainwave event category according to the probability of the brainwave state.

[0040] Probability of EEG state Let be the binary classification probability. When When, the result output module outputs the first EEG state; when At that time, the second EEG state is output. Threshold The value can be set to 0.5, or it can be set according to the sensitivity of medical alarms, false alarm rate, or specific brain-computer interface task requirements. For EEG recognition of familiar / unfamiliar faces, the first EEG state can be EEG evoked by familiar faces, and the second EEG state can be EEG evoked by unfamiliar faces; for EEG detection of epilepsy, the first EEG state can be EEG during an attack or suspected attack, and the second EEG state can be EEG during a non-attack attack.

[0041] In one or more embodiments, the processor further includes a training calibration module for updating learnable scale power parameters, layer-normalized learnable scaling and translation parameters, learnable network center, learnable kernel width parameter, linear output weights, and biases using labeled EEG training samples.

[0042] In the training and calibration module, the total loss uses a binary cross-entropy loss with logit, and a center constraint is applied to the learnable scale power parameter. The expression for the total loss is: ; in, Total loss; Represents the binary cross-entropy loss. For learnable scale power-law regularization strength, The initial center value of the learnable scale power parameter, To output the weight regularization coefficients; Linear output weights; For the number of time windows, Number of frequency bands; For the corresponding number The frequency band and the first The learnable scale power parameter for each time window.

[0043] In the training and calibration module, the learnable network center is determined using a farthest point initialization strategy, and the process is as follows: From the training candidate feature set Select one feature as the first learnable network center. and in the selected After finding the learnable network centers, the candidate feature with the largest minimum squared Euclidean distance to the set of selected centers is chosen as the first... A learnable network center is initialized and continues to be updated during training. It is a positive integer greater than 1.

[0044] When selected When the center is, the first The center is determined by the following formula: ; The above formula ensures that the initial network center covers the training feature space as much as possible, reducing the risk of the network concentrating in local areas. After initialization, During training, it continues to be updated as a learnable parameter, thereby enabling the network similarity space to be jointly adapted to the final EEG recognition target.

[0045] The following is a preferred implementation for the EEG recognition task of familiar / unfamiliar faces: A 64-channel EEG electrode cap was used, with 59 scalp EEG recording channels. Electrode placement followed the Extended International 10-20 system. The raw sampling rate of the EEG acquisition hardware was 1000Hz, and the analog front-end included a low-noise amplifier, anti-aliasing filter, and safety isolation circuitry. The acquired raw EEG data was converted by an ADC and transmitted to an embedded processor, which then resampled the data. Hz, and a 4-second EEG segment during the face image presentation phase is extracted as a sample, therefore , .

[0046] In this embodiment, each trial includes a 2-second preparation phase, a 4-second face presentation phase and a 3-second rest phase; a trigger marker is written when the presentation of face images starts, and only the 4-second data after the trigger marker is intercepted. Settings of the time-frequency window selection module , that is, seconds, seconds, seconds, each time window sampling points. The frequency band set includes a full frequency band [0.5,40] Hz, 19 overlapping frequency bands with 4 Hz scale Hz, ; 9 overlapping frequency bands with 8 Hz scale Hz, ; and 4 overlapping frequency bands with 16 Hz scale Hz, . Accordingly , each sample forms a local standard deviation tensor with a size of .

[0047] The learnable entropy power transformation module adopts channel-shared, time-frequency independent scale power parameters , the initial central value is set to . The radial basis function network module is provided with network centers, and the expanded feature dimension is . During training, five-fold one-data-block-out cross-validation is adopted, with one experimental block as test data and the remaining data blocks as training data in each fold; the training batch size is 160, the number of training rounds is 800, the learning rate for scale power parameters is , and the learning rates for layer normalization, radial basis function network layer and output layer are .

[0048] In this embodiment, when the output probability , the system outputs "familiar face-induced electroencephalogram"; when , the system outputs "unfamiliar face-induced electroencephalogram". The output terminal displays the classification result and probability value simultaneously, and writes the time stamp, trial number and recognition result into the memory. This embodiment reflects the system-level application of the present invention in small-sample, multi-channel, event-related electroencephalogram recognition.

[0049] Taking a 4-second familiar / unfamiliar face trial as an example, it is explained how a section of specific multi-channel electroencephalogram signal passes through each module of the system of the present invention and obtains the final output. It is assumed that after resampling, the trial obtains , with the channel unit being microvolt. The processor performs zero-phase filtering and three-segment time window division on it. For the -th channel, the -th frequency band and the Taking a time window as an example, this frequency band can correspond to 12.5Hz to 20.5Hz. seconds, number of sampling points .

[0050] Assume the mean of this local segment after filtering is The local variance was calculated. ,Pick And then calculate If the scale power parameter after training is Then the output of this local component is calculated: ; The above calculations are for , , After executing all combinations, we get For example, if there are other trials in the same trial... , ,but ;like , ,but These figures demonstrate that the learnable entropy power transformation module does not simply preserve the original variance or standard deviation, but rather performs differentiated scaling on different time windows and frequency bands through learnable scale powers.

[0051] The normalization and vectorization module will Expand as Assume the mean of the expanded vector in this trial is... The standard deviation is Furthermore, for ease of explanation, the layer normalization parameter is taken. The corresponding component is 1. If the corresponding component is 0, then the normalized example values ​​of the above three components are approximately 0.95, -0.71, and 2.29, respectively. In a real system, and It learns automatically from training.

[0052] Radial basis function network module settings Let the kernel width parameter obtained after training be... The squared distances between the normalized feature of this trial and the 17th, 41st, and 88th network centers are respectively , , The corresponding RBF (radial basis function) response is: , , The remaining network responses are calculated in the same way, forming... .

[0053] Assuming the linear output layer weights all 140 RBF responses, we get... Then we get When the threshold At that time, the system outputs the first EEG state, namely "familiarity with a face evoked EEG", and displays "familiar, probability 0.774" on the display terminal. This embodiment fully demonstrates the entire process of a multi-channel EEG sample from acquisition matrix, local time-frequency statistical features, LEPST (Learnable Entropy Power Transform), RBF similarity calculation to final binary classification output.

[0054] This embodiment illustrates the system implementation of the present invention in epilepsy EEG detection. This embodiment does not require a facial stimulus presentation module; the system uses continuously acquired clinical or wearable EEG data as input. The EEG acquisition module employs a 19-channel international 10-20 system lead, with an initial sampling rate of 512Hz. The processor resamples the continuous data to... Hz, and samples are constructed using sliding analysis segments of 4 seconds in length, so each segment , .

[0055] The preprocessing module performs power frequency notch filtering, 0.5 Hz to 40 Hz bandpass filtering, and amplitude limiting protection; the time-frequency window selection module can still use... A 2-second time window and Multiple scale frequency bands are available, and the number of frequency bands can also be set according to the computing power of clinical equipment. In a preferred implementation, the system employs the same 33 frequency bands as described above to maintain the ability to distinguish features near low-frequency slow waves, mid-frequency rhythms, and higher-frequency electromyographic interference.

[0056] In this embodiment, the training labels are set as follows: first EEG state "epilepsy or suspected seizure segment" and second EEG state "non-seizure segment". Since the number of channels becomes 19, the expanded feature dimension is... The preferred number of RBF networks is set to... Scale power initial center scale power learning rate Learning rate of classification module After training is complete, the model parameters are saved in the memory. , , , and .

[0057] In an online detection, a 4-second segment was input into the system, and the output vector of the radial basis function network was obtained. To obtain the probability of an attack If the medical alarm threshold is set to... If the probability of a segment is 0.18, the output module will output "Suspected epileptic seizure, probability 0.910" and send an alarm signal to the monitoring terminal via audio-visual or network interface. If the output probability of another segment is 0.18, the system will output "Non-seizure segment". This embodiment illustrates that the EEG recognition system structure based on learnable entropy power transformation and radial basis function network of the present invention is not limited to face recognition tasks, but can be used as a multi-channel EEG state recognition system with hardware acquisition and real-time inference capabilities for pathological EEG event detection.

[0058] The following describes one implementation of the present invention in individualized calibration and embedded deployment: When an EEG recognition system based on learnable entropy power transformation and radial basis function networks is used for the first time on a subject or patient, the training calibration module constructs a training set from the labeled EEG segments. ,in This indicates the second or first EEG state. Each sample in the training set first undergoes time-frequency window selection, local standard deviation extraction, and initial learnable entropy power transformation to obtain a candidate feature set. .

[0059] Perform farthest point initialization: for example, when At that time, the processor first randomly selects a candidate feature as... Then, it successively selects the candidate feature with the largest minimum distance to the currently selected center set as the new network center, until it obtains... This process can be completed offline on a workstation or edge computing box, or it can be executed when the embedded processor is idle. After initialization, the system iteratively updates all learnable parameters using the Adam optimizer, with the learning rate of the scale-power parameter set to [value missing] during training. Other parameters, such as the learning rate, are set to... The batch size is 160.

[0060] After calibration, the inference phase only requires performing fixed steps: filtering, standard deviation calculation, power transform, normalization, RBF distance calculation, and linear output. , , , For example, the input for each trial is... Each sampled value is compressed into 5841-dimensional statistical features; the radial basis function layer output of the radial basis function network module is only a 140-dimensional similarity vector, so the computational and storage requirements are suitable for deployment on portable brain-computer interface terminals composed of ARM processors, DSPs, GPUs or NPUs.

[0061] On familiar / unfamiliar face EEG recognition datasets, the preferred embodiments of this invention employ a leave-one-out cross-validation protocol for verification. Table 1 lists a comparison of the average accuracy of representative EEG feature extraction and classification methods. This table is only used to illustrate the technical effects of the embodiments of this invention and does not constitute a limitation on the application scenarios, number of channels, number of frequency bands, or number of networks described in the claims.

[0062] Table 1. Performance Comparison of Representative EEG Recognition Methods with Embodiments of the Present Invention

[0063] according to Figure 2 This invention also provides an EEG recognition method based on learnable entropy power transformation and radial basis function networks, which is based on... Figure 1 The EEG recognition system shown, based on learnable entropy power transformation and radial basis function networks, executes the following steps within the processor: Step S1: Acquire multi-channel EEG signals of the target object and form an EEG sample; Step S2: Divide the EEG samples into multiple time windows and filter them into multiple frequency bands; Step S3: Calculate the local standard deviation for each channel, frequency band, and time window EEG sample as a local statistical feature; Step S4: Based on the learnable scale power parameter, convert the local statistical features into task-adaptive entropy power transformation features; Step S5: Expand the feature tensor composed of the task-adaptive entropy power transformation features into a one-dimensional vector in a set order, and normalize it according to the learnable scaling parameter and the learnable translation parameter to obtain the feature vector. Step S6: Calculate the radial basis function network similarity based on the learnable network center and learnable kernel width parameter, and then combine it with the linear output weights and the biased output EEG state probability; Step S7: Output the cognitive state, pathological state, or brainwave event category based on the probability of the EEG state.

[0064] The steps in the EEG recognition method based on learnable entropy power transformation and radial basis function network in this embodiment of the invention are the same as the specific implementation process of each module in the processor of the EEG recognition system based on learnable entropy power transformation and radial basis function network described above, and will not be described in detail here.

[0065] In one or more embodiments, a computer-readable storage medium is also provided, having a computer program stored thereon that, when executed by a processor, implements as follows: Figure 2 The steps in the EEG recognition method based on learnable entropy power transformation and radial basis function network are shown.

[0066] In one or more embodiments, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements... Figure 2 The steps in the EEG recognition method based on learnable entropy power transformation and radial basis function network are shown.

[0067] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including functions for executing... Figure 1 The program code for the method shown. In such an embodiment, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by the central processing unit, it performs the various functions defined in the apparatus of this application.

[0068] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0069] The above are merely preferred embodiments of the present invention and are not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A brainwave recognition system based on learnable entropy power transformation and radial basis function networks, characterized in that, include: EEG acquisition module and processor; The EEG acquisition module is used to acquire multi-channel EEG signals of a specified object and form an EEG sample; The processor includes: The time-frequency window selection module is used to divide EEG samples into multiple time windows and filter them into multiple frequency bands; The local statistical feature extraction module is used to calculate the local standard deviation corresponding to each channel, frequency band, and time window EEG sample, as a local statistical feature. The learnable entropy power transformation module is used to convert local statistical features into task-adaptive entropy power transformation features based on learnable scale power parameters. The normalization and vectorization module is used to expand the feature tensor composed of the task-adaptive entropy power transformation features into a one-dimensional vector in a set order, and perform normalization processing according to the learnable scaling parameter and the learnable translation parameter to obtain the feature vector. The radial basis function network module is used to calculate the radial basis function network similarity based on the learnable network center and learnable kernel width parameters, and then combine the linear output weights and the biased output EEG state probability. The result output module is used to output the cognitive state, pathological state, or brain event category based on the probability of the EEG state.

2. The EEG recognition system based on learnable entropy power transformation and radial basis function network as described in claim 1, characterized in that, The processor also includes: The training calibration module is used to update the learnable scale power parameter, the layer-normalized learnable scaling and translation parameters, the learnable network center, the learnable kernel width parameter, the linear output weights, and the bias using labeled EEG training samples.

3. The EEG recognition system based on learnable entropy power transformation and radial basis function network as described in claim 2, characterized in that, In the training and calibration module, the total loss uses a binary cross-entropy loss with logit, and a center constraint is applied to the learnable scale power parameter. The expression for the total loss is: ； in, Total loss; Represents the binary cross-entropy loss. For learnable scale power-law regularization strength, The initial center value of the learnable scale power parameter, To output the weight regularization coefficients; Linear output weights; For the number of time windows, Number of frequency bands; For the corresponding number The frequency band and the first The learnable scale power parameter for each time window.

4. The EEG recognition system based on learnable entropy power transformation and radial basis function network as described in claim 2, characterized in that, In the training and calibration module, the learnable network center is determined using a farthest point initialization strategy, and the process is as follows: Select a feature from the set of training candidate features as the first learnable network center, and in the selected... After finding the learnable network centers, the candidate feature with the largest minimum squared Euclidean distance to the set of selected centers is chosen as the first... A learnable network center is initialized and continues to be updated during training. It is a positive integer greater than 1.

5. The EEG recognition system based on learnable entropy power transformation and radial basis function network as described in claim 1, characterized in that, In the learnable entropy power transformation module, the task-adaptive entropy power transformation characteristics are as follows: ； ； in, Entropy power transformation characteristics adapted for the task; Local standard deviation; These are unconstrained parameters; For the Sigmoid function; and These represent the preset lower and upper limits, respectively; subscript , , and They represent the first The first EEG sample, the first The first channel, the first The frequency band and the first A time window.

6. The EEG recognition system based on learnable entropy power transformation and radial basis function network as described in claim 1, characterized in that, In the radial basis function network module, the process of calculating the probability of EEG states is as follows: according to Calculate the first The squared Euclidean distance between the centers of a learnable network; according to Calculate the Gaussian radial basis response ; all Composition of network similarity vectors ; according to and Calculate the probability of brainwave states ; in, These are normalized eigenvectors; For the first A Learnable Network Center For learnable kernel width parameters; The output vector of the output layer of the radial basis functions; For bias.

7. The EEG recognition system based on learnable entropy power transformation and radial basis function network as described in claim 1, characterized in that, In the local statistical feature extraction module, the local standard deviation for each channel, frequency band, and time window EEG sample is: ； ； ； in, Local standard deviation; It is a non-negative preset constant; It has locally unbiased variance; It is a local mean; EEG signals; For the first Number of sampling points within a time window; subscript , , and They represent the first The first EEG sample, the first The first channel, the first The frequency band and the first A time window.

8. A brainwave recognition method based on learnable entropy power transformation and radial basis function network, characterized in that, Based on the EEG recognition system based on learnable entropy power transformation and radial basis function network as described in any one of claims 1-7, the following steps are performed within the processor: Acquire multi-channel EEG signals from a specified object and form an EEG sample; The EEG samples were divided into multiple time windows and filtered into multiple frequency bands; Calculate the local standard deviation for each channel, frequency band, and time window EEG sample as a local statistical feature; Based on the learnable scale power parameter, local statistical features are transformed into task-adaptive entropy power transformation features. The feature tensor composed of the task-adaptive entropy power transformation features is expanded into a one-dimensional vector in a set order, and then normalized according to the learnable scaling parameter and the learnable translation parameter to obtain the feature vector. The radial basis function network similarity is calculated based on the learnable network center and learnable kernel width parameter, and then combined with the linear output weights and the biased output EEG state probability. Based on the probability of the EEG state, the cognitive state, pathological state, or EEG event category is output.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the steps in the EEG recognition method based on learnable entropy power transformation and radial basis function network as described in claim 8.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the EEG recognition method based on learnable entropy power transformation and radial basis function network as described in claim 8.

Images