Classification method, terminal device and medium of steady-state visual evoked potential

Abstract

The embodiment of the present invention provides a classification method, terminal device and medium for steady-state visual evoked potential, which belongs to the field of brain-computer interface technology. The method includes: determining a first SSVEP signal corresponding to an initial object under a first stimulus target and a first stimulus label corresponding to the first stimulus target, and preprocessing the first SSVEP signal to obtain a second SSVEP signal; performing model training according to the second SSVEP signal and the first stimulus label to obtain an initial classification model; obtaining a third SSVEP signal corresponding to the target object under the second stimulus target and a second stimulus label corresponding to the second stimulus target, and preprocessing the third SSVEP signal to obtain a fourth SSVEP signal; fine-tuning the initial classification model according to the fourth SSVEP signal and the second stimulus label to obtain a target classification model corresponding to the target object; obtaining a fifth SSVEP signal corresponding to the target object, and performing type recognition on the fifth SSVEP signal according to the target classification model to obtain a third stimulus label corresponding to the fifth SSVEP signal.

Description

Technical Field

[0001] The present invention relates to the field of brain-computer interface technology, and in particular to a classification method, terminal equipment and medium for steady-state visual evoked potentials. Background Art

[0002] When the human body is subjected to a visual stimulus of a fixed frequency flickering or changing pattern, the potential activity of the cerebral cortex will be modulated, resulting in a continuous response related to the stimulus frequency (at the fundamental frequency or multiple of the stimulus frequency). This response has a periodic rhythm similar to that of the visual stimulus, which is the steady-state visual evoked potential (SSVEP). The SSVEP signal is manifested in the EEG signal as a spectrum peak at the stimulus frequency or harmonics in the power spectrum. By analyzing and detecting the frequency corresponding to the spectrum peak, the stimulus source of the subject's visual gaze can be detected. SSVEP has a high ITR and stable command generation performance, so it has great application value in the field of human-computer interaction.

[0003] Existing SSVEP classification technologies are mainly divided into methods based on traditional machine learning and deep learning. Traditional machine learning methods rely on effective feature extraction, which is usually based on empirical knowledge, which limits the assurance of SSVEP detection performance. In contrast, deep learning methods require a large amount of training data. However, the amount of existing SSVEP data is limited, and there are certain individual differences in SSVEP signals, which seriously restricts the generalization ability and robustness of deep learning algorithms. Therefore, the current SSVEP classification algorithm shows poor performance under the requirements of high precision and real-time performance, and it is difficult to meet the requirements of practical applications. Summary of the invention

[0004] The main purpose of the embodiments of the present invention is to provide a classification method, terminal device and medium for steady-state visual evoked potential, aiming to solve the problem that the SSVEP classification algorithm in the related art exhibits poor performance under the requirements of high precision and real-time performance and is difficult to meet the requirements of practical applications.

[0005] In a first aspect, an embodiment of the present invention provides a method for classifying steady-state visual evoked potentials, comprising:

[0006] Determine a first SSVEP signal corresponding to the initial object under the first stimulation target and a first stimulation label corresponding to the first stimulation target, and preprocess the first SSVEP signal to obtain a second SSVEP signal;

[0007] Performing model training according to the second SSVEP signal and the first stimulus label to obtain an initial classification model;

[0008] Obtaining a third SSVEP signal corresponding to the target object under the second stimulation target and a second stimulation label corresponding to the second stimulation target, and preprocessing the third SSVEP signal to obtain a fourth SSVEP signal;

[0009] Fine-tuning the initial classification model according to the fourth SSVEP signal and the second stimulus label to obtain a target classification model corresponding to the target object;

[0010] A fifth SSVEP signal corresponding to the target object is obtained, and a type of the fifth SSVEP signal is identified according to the target classification model to obtain a third stimulation label corresponding to the fifth SSVEP signal.

[0011] In a second aspect, an embodiment of the present invention further provides a terminal device, comprising a processor, a memory, a computer program stored in the memory and executable by the processor, and a data bus for realizing connection and communication between the processor and the memory, wherein when the computer program is executed by the processor, the steps of any one of the methods for classifying steady-state visual evoked potentials provided in the specification of the present invention are implemented.

[0012] In a third aspect, an embodiment of the present invention further provides a storage medium for computer-readable storage, characterized in that the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of any method for classifying steady-state visual evoked potentials provided in the specification of the present invention.

[0013] The embodiment of the present invention provides a classification method, terminal device and medium for steady-state visual evoked potential, the method comprising: first, obtaining a first SSVEP signal and a corresponding first stimulus label of an initial object under a first stimulus target, and then preprocessing the first SSVEP signal to generate a second SSVEP signal, and then using the second SSVEP signal and the first stimulus label for model training, thereby constructing an initial classification model. The initial classification model provides a basis for subsequent individualized adjustment. Obtain a third SSVEP signal and a corresponding second stimulus label of a target object under a second stimulus target, and preprocess the third SSVEP signal to generate a fourth SSVEP signal, thereby fine-tuning the initial classification model based on the fourth SSVEP signal and the second stimulus label to form a target classification model of the target object. This adjustment enables the target classification model to better adapt to the signal characteristics of the target object, thereby improving the classification accuracy. Further obtain a fifth SSVEP signal of the target object, and classify it using the target classification model, thereby obtaining the third stimulus label corresponding to the fifth SSVEP signal. Through this method, the classification accuracy and stability of the SSVEP signals of the target objects can be significantly improved, and the performance challenges of the SSVEP classification algorithm under high precision and real-time requirements can be effectively addressed, thereby achieving more reliable signal recognition and processing in application scenarios such as brain-computer interfaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0015] Figure 1 A schematic flow chart of a method for classifying steady-state visual evoked potentials provided by an embodiment of the present invention;

[0016] Figure 2 A schematic diagram of the model structure of an initial classification model provided by an embodiment of the present invention;

[0017] Figure 3 A schematic block diagram of the structure of a terminal device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0019] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0020] It should be understood that the terms used in this specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0021] The embodiment of the present invention provides a classification method for steady-state visual evoked potential, a terminal device and a medium. The classification method for steady-state visual evoked potential can be applied to a terminal device, which can be an electronic device such as a tablet computer, a laptop computer, a desktop computer, a personal digital assistant and a wearable device. The terminal device can be a server or a server cluster.

[0022] Some embodiments of the present invention are described in detail below in conjunction with the accompanying drawings. In the absence of conflict, the following embodiments and features in the embodiments can be combined with each other.

[0023] Please refer to Figure 1 , Figure 1 A schematic flow chart of a method for classifying steady-state visual evoked potentials provided in an embodiment of the present invention.

[0024] like Figure 1 As shown, the steady-state visual evoked potential classification method includes steps S101 to S105.

[0025] Step S101: determining a first SSVEP signal corresponding to an initial object under a first stimulation target and a first stimulation label corresponding to the first stimulation target, and preprocessing the first SSVEP signal to obtain a second SSVEP signal.

[0026] Exemplarily, the initial object may be a plurality of users who need to collect SSVEP signals, the first stimulation target may be a certain frequency or a specific visual stimulation, and the first stimulation target should be able to make the target object produce an obvious SSVEP signal. The stimulation type of the first stimulation target can be set according to actual needs, and this application does not make specific restrictions.

[0027] Exemplarily, a first SSVEP signal corresponding to the first stimulation target of the initial object is obtained through a signal acquisition system. After obtaining the first SSVEP signal, the first SSVEP signal is marked according to the first stimulation target, so as to obtain a first stimulation label corresponding to the first SSVEP signal. The first stimulation label can be a stimulation type corresponding to the first stimulation target, or a stimulation number corresponding to the first stimulation target, etc. The first stimulation label can be set according to actual needs, and this application does not limit it. Among them, the signal acquisition system can be an electroencephalograph, an EEG system, etc.

[0028] Exemplarily, a bandpass filter is applied to filter the first SSVEP signal, or the amplitude of the first SSVEP signal window is normalized so that the signal intensity remains consistent in different time windows or under different experimental conditions, and the first SSVEP signal is smoothed to reduce high-frequency noise and enhance the characteristics of the signal. The first SSVEP signal can also be resampled so that the data meets the requirements of subsequent analysis and ensures that the main characteristics of the signal are maintained during the resampling process. Thus, a cleaner and more representative second SSVEP signal is obtained, laying the foundation for subsequent analysis and modeling.

[0029] In some embodiments, the preprocessing of the first SSVEP signal to obtain the second SSVEP signal includes: using a notch filter to remove power frequency interference from the first SSVEP signal to obtain a first intermediate signal; performing bandpass filtering on the first intermediate signal to obtain a second intermediate signal; downsampling the second intermediate signal to obtain a third intermediate signal; setting a reference signal, and using a task discrimination pattern analysis algorithm in combination with the reference signal to spatially filter the third intermediate signal to obtain the second SSVEP signal after feature extraction.

[0030] Exemplarily, the first SSVEP signal is preprocessed by first removing power frequency interference from the first SSVEP signal using a notch filter to obtain a first intermediate signal. For example, the first SSVEP signal is preprocessed by removing power frequency interference using a 50 Hz notch filter to obtain the first intermediate signal.

[0031] Exemplarily, the first intermediate signal is subjected to bandpass filtering to obtain the second intermediate signal. For example, a 6-stage Butterworth filter is used to perform bandpass filtering of 6 Hz-80 Hz on the second intermediate signal to obtain the second intermediate signal.

[0032] Exemplarily, the second intermediate signal is downsampled to obtain the third intermediate signal. For example, the second intermediate signal is downsampled to 250 Hz to obtain the third intermediate signal.

[0033] Exemplarily, a reference signal is set, and the third intermediate signal is spatially filtered using a task discrimination pattern analysis algorithm in combination with the reference signal to obtain a second SSVEP signal after feature extraction.

[0034] Exemplarily, the task discriminant pattern analysis algorithm to be used is determined, such as the commonly used independent component analysis (ICA), canonical correlation analysis (CCA) or other spatial filtering algorithms. These algorithms can extract features related to the second intermediate signal according to task requirements. Collect or generate reference signals. These reference signals are usually known, standard signals related to the task, for example, the reference signal of the stimulation frequency is a standard signal, which is used to guide the spatial filtering process. The task discriminant pattern analysis algorithm and the reference signal are used to calculate the spatial filter matrix, and then the calculated spatial filter matrix is ​​applied to the third intermediate signal for spatial filtering processing, which usually includes a matrix multiplication operation to project the signal into a new space, so that the features of the first SSVEP signal are more prominent, and noise and interference are effectively suppressed, thereby obtaining a second SSVEP signal.

[0035] Specifically, preprocessing the first SSVEP signal through the above steps can effectively remove interference, enhance the target signal, reduce the amount of data and improve the accuracy of feature extraction, thereby significantly improving the quality of the signal, and further improving the quality of the signal and the accuracy of feature extraction, laying a solid foundation for further data analysis and model training.

[0036] In some embodiments, setting the reference signal includes: determining a stimulation frequency corresponding to a target task and a harmonic range corresponding to the stimulation frequency; setting a corresponding sine wave signal in the harmonic range of the stimulation frequency according to the target task and setting a corresponding cosine wave signal in the harmonic range of the stimulation frequency according to the target task; and determining the reference signal based on the sine wave signal and the cosine wave signal.

[0037] Exemplarily, the characteristics of the target task are determined to determine the corresponding stimulation frequency. For example, some visual stimulation may involve specific frequencies, such as 10 Hz, 15 Hz, etc. According to the stimulation frequency, its harmonic range is determined. Harmonics are integer multiples of the stimulation frequency, so the harmonic range includes the stimulation frequency and its multiples. For example, if the stimulation frequency is 10 Hz, then 10 Hz, 20 Hz, 30 Hz, etc. may be of interest.

[0038] Exemplarily, a sine wave signal is set according to the determined stimulation frequency and its harmonic range. For example, if the stimulation frequency is 10 Hz, the frequency of the sine wave signal can be set to 10 Hz, 20 Hz, 30 Hz, etc. A sine wave signal is generated based on these frequencies. Similar to the sine wave, a cosine wave signal is set according to the stimulation frequency and its harmonic range. For example, if the stimulation frequency is 10 Hz, the frequency of the cosine wave signal can be set to 10 Hz, 20 Hz, 30 Hz, etc. A cosine wave signal is generated based on these frequencies.

[0039] Exemplarily, the generated sine wave signal and cosine wave signal are synthesized into a reference signal. This reference signal will be used for subsequent signal processing and feature extraction. The reference signal can include multiple frequency components, which are set according to the requirements of the target task.

[0040] For example, a sine wave signal and a cosine wave signal containing five harmonic ranges corresponding to the stimulation frequency of each target task are used as reference signals, as shown in Formula 1. is the reference signal corresponding to a specific frequency visual stimulus, is the stimulation frequency, t is the sampling time, is the number of sampling points, F s is the sampling frequency after downsampling. Then, the task discriminant pattern analysis algorithm is used to perform spatial filtering on the third intermediate signal to obtain the second SSVEP signal after feature extraction, wherein the number of delay points of the task discriminant pattern analysis algorithm (TDCA algorithm) is 5 and the number of principal components is 8.

[0041] (Formula 1)

[0042] (Formula 2)

[0043] Specifically, clarifying the stimulation frequency of the target task and the corresponding harmonic range can help focus on the target frequency and its related frequency bands in signal processing. This ensures that the analysis of the steady-state visual evoked potential signal can be optimized for the frequency related to the target task, reducing the interference of other frequency components. Setting sine and cosine wave signals corresponding to the stimulation frequency and its harmonics can provide a benchmark for signal processing, thereby helping to separate the components related to the target stimulation frequency from the first VEP signal and enhance the detection ability of these components.

[0044] Step S102: Perform model training according to the second SSVEP signal and the first stimulus label to obtain an initial classification model.

[0045] Exemplarily, the second SSVEP signal and the first stimulus label are combined into a training data set. The training data set is divided into a training set and a validation set. The training set is used to train the classification model, and the validation set is used to evaluate the performance of the classification model.

[0046] Exemplarily, a model suitable for the classification task is selected. For example, a traditional classification algorithm such as a support vector machine, logistic regression, or a more complex model such as a neural network can be selected. For example, the classification model is a neural network classification model, and the number of layers of the neural network and the number of neurons in each layer are determined, or the classification model can select the kernel function and regularization parameter of a support vector machine.

[0047] Exemplarily, the training set is input into the selected classification model for training to obtain an initial classification model, and the model parameters are adjusted by optimization algorithms such as gradient descent to improve the classification performance. The loss function and performance indicators such as accuracy during the training process are monitored to ensure good training of the model. In addition, the performance of the initial classification model can also be evaluated on the validation set. Check the model's accuracy, recall rate, precision and other indicators to determine whether the initial classification model reaches the expected performance level. When the performance of the initial classification model evaluated on the validation set meets the actual needs, the obtained initial classification model is saved and then used for subsequent practical applications or further analysis.

[0048] In some embodiments, the initial classification model includes a channel expansion module, a first feature extraction module, a second feature extraction module, and a type output module, and the initial classification model is obtained by performing model training according to the second SSVEP signal and the first stimulus label, including: using the channel expansion module to perform channel expansion on the second SSVEP signal in a layer-by-layer convolution manner to obtain a first feature signal; using the first feature extraction module to perform spatial feature extraction on the first feature signal to obtain a second feature signal; using the second feature extraction module to perform temporal feature extraction on the second feature signal to obtain a third feature signal; using the type output module to determine the label probability corresponding to the third feature signal and the target stimulus, and determining the predicted stimulus label corresponding to the second SSVEP signal according to the label probability; using the cross entropy loss function to perform loss calculation on the predicted stimulus label and the first stimulus label to obtain a loss value; when the loss value does not meet the preset condition, continuing to perform model training according to the second SSVEP signal and the first stimulus label until the loss value meets the preset condition; when the loss value meets the preset condition, obtaining the target model parameters, and determining the initial classification model according to the target model parameters.

[0049] Exemplarily, the initial classification model includes a channel expansion module, a first feature extraction module, a second feature extraction module, and a type output module.

[0050] Exemplarily, the channel expansion module is used to perform channel expansion on the second SSVEP signal in a layer-by-layer convolution manner to obtain the first characteristic signal, for example, the channel expansion module determines the type of convolution layer to be used. Layer-by-layer convolution usually involves multiple convolution layers, and each layer uses a different convolution kernel to process the signal. The parameters of each layer of convolution are selected, including the size, stride, padding, and number of output channels of the convolution kernel, and then the convolution kernel under each layer of convolution parameters is used to perform convolution processing and channel expansion on the second SSVEP signal to obtain the corresponding first characteristic signal.

[0051] Exemplarily, the first feature extraction module includes a convolution layer and a feature merging layer, etc. The convolution layer of the first feature extraction module is used to slide the first feature signal in the spatial dimension through the convolution kernel to extract local features and obtain multiple feature maps, and different feature maps represent different spatial features. Thereby, the feature maps processed by the convolution layer are stacked together through the feature merging layer, or the feature maps are flattened into a feature vector, and then the integrated feature representation is represented as the second feature signal. The second feature signal contains the spatial features extracted from the first feature signal for further analysis, classification or other processing tasks.

[0052] Exemplarily, the purpose of temporal feature extraction is to extract features with time correlation from the second feature signal. This generally involves identifying and capturing the pattern, trend and dynamic change of the signal in time. The second feature extraction module can be a temporal convolutional network, a recurrent neural network or its variants such as a long short-term memory network LSTM, a gated recurrent unit GRU, etc. Thus, the second feature extraction module is used to extract the temporal features of the second feature signal to obtain the third feature signal.

[0053] Exemplarily, the target stimulus is the sum of all first stimulus labels. For example, if there are 12 first stimulus targets, there will be 12 corresponding first stimulus labels, and the target stimulus is the 12 first stimulus labels, that is, the target stimulus is all first stimulus labels in the training set.

[0054] Exemplarily, the Softmax activation function is used in the type output module to generate the label probability corresponding to the target stimulus for the third feature signal, and then the stimulus label corresponding to the maximum probability in the label probability is determined as the predicted stimulus label corresponding to the second SSVEP signal.

[0055] Exemplarily, the cross entropy loss function is used to calculate the error between the predicted stimulus label and the first stimulus label, so that the loss value is obtained by using the label probability corresponding to the predicted stimulus label and the first stimulus label for loss calculation through the cross entropy loss function. A higher loss value means that there is a large difference between the model's prediction and the true label, indicating that the model's performance may be poor. A lower loss value indicates that the model's prediction is closer to the true label and the model performance is better.

[0056] Exemplarily, the preset condition is a threshold set according to expert experience or historical experience, and when the loss value is greater than or equal to the threshold, the model training continues according to the second SSVEP signal and the first stimulus label until the loss value is less than the threshold; when the loss value is less than the threshold, the target model parameters corresponding to the loss value are obtained, and the target model parameters are saved to obtain the initial classification model.

[0057] Specifically, by performing channel expansion on the second SSVEP signal through layer-by-layer convolution, richer features can be extracted from the second SSVEP signal. This expansion can increase the dimension of the feature map, so that the network can learn more feature representations, thereby improving the effectiveness of subsequent processing. Spatial feature extraction of the expanded features helps to capture the spatial pattern and structural information in the signal. By extracting spatial features, important spatial features in the signal can be effectively identified, and the model's understanding of spatial information can be enhanced. Temporal feature extraction of spatial features can capture the time series characteristics of the signal. By analyzing the temporal dynamics, the temporal change pattern of the signal can be identified, and the modeling ability of time series information can be improved. This provides support for obtaining a high-quality initial classification model in the future.

[0058] In some embodiments, the channel extension module includes a first convolution layer, a first normalization layer, and a first ELU layer. The channel extension module is used to perform channel expansion on the second SSVEP signal in a layer-by-layer convolution manner to obtain a first characteristic signal, including: using the L2 norm and spectral normalization to constrain the parameters of the first convolution layer, and using the first convolution kernel of the first convolution layer to convolve the second SSVEP signal to obtain a first convolution result; using the first normalization layer to batch normalize the first convolution result to obtain a first normalized result; using the first ELU layer to activate the first normalized result to obtain the first characteristic signal.

[0059] Exemplarily, the channel expansion module includes a first convolutional layer, a first normalization layer, and a first ELU layer.

[0060] Exemplarily, the L2 norm, also known as the Euclidean norm, is used to measure the size of the weight matrix. By applying the L2 norm constraint to the weights of the convolutional layer, the maximum value of its weights can be limited, overfitting can be prevented, and the stability of the model can be maintained. Spectral normalization is a method for normalizing the weight matrix. By constraining the spectral norm of the weight matrix, i.e., the maximum singular value, the weight matrix of the convolutional layer has good numerical stability and training efficiency. Thus, the parameters of the first convolutional layer are constrained using the L2 norm and spectral normalization.

[0061] Exemplarily, a first convolution kernel to which parameter constraints have been applied is obtained from the first convolution layer. This convolution kernel is a filter for specific feature extraction, and its parameters have been constrained by L2 norm and spectral normalization. The first convolution kernel is applied to the second SSVEP signal. The convolution operation involves sliding the convolution kernel on the input signal and calculating the dot product of the convolution kernel and the signal area. Through the convolution operation, a first convolution result can be obtained.

[0062] Exemplarily, the first convolution result is batch normalized using the first normalization layer to obtain a first normalized result, and the first ELU layer is used to perform an activation operation on the first normalized result to obtain a first feature signal.

[0063] Specifically, parameter constraints through L2 norm and spectral normalization ensure the stability and efficiency of the training process; convolution processing and feature extraction steps improve the initial classification model's ability to learn the features of the second SSVEP signal; batch normalization and activation operations further improve the training speed and expressiveness of the initial classification model, thereby providing support for subsequent improvements in the performance, stability and generalization ability of the initial classification model.

[0064] In some embodiments, the first feature extraction module includes a second convolution layer, a second normalization layer, a second ELU layer, and a first random inactivation layer, and the use of the first feature extraction module to perform spatial feature extraction on the first feature signal to obtain the second feature signal includes: using L2 norm and spectral normalization to perform parameter constraints on the second convolution layer, and using the second convolution kernel of the second convolution layer to convolve the first feature signal to obtain a second convolution result; using the second normalization layer to perform batch normalization on the second convolution result to obtain a second normalized result; using the second ELU layer to perform an activation operation on the second normalized result to obtain a first activation result; using the first random inactivation layer to perform a random discarding operation on the first activation result to obtain the second feature signal.

[0065] Exemplarily, the first feature extraction module includes a second convolutional layer, a second normalization layer, a second ELU layer, and a first random inactivation layer.

[0066] Exemplarily, the L2 norm and spectral normalization are used to perform parameter constraints on the second convolution layer, and then a second convolution kernel with parameter constraints applied is obtained from the second convolution layer to perform convolution processing on the first feature signal, and a second convolution result is generated.

[0067] Exemplarily, the second convolution result is batch normalized using the second normalization layer to obtain the second normalized result, and the second convolution result is batch normalized using the second normalization layer. For example, the mean and variance of each feature map are calculated, and the value of each feature map is subtracted from the mean and divided by the standard deviation to perform standardization so that the input of each layer has zero mean and unit variance, thereby obtaining the second normalized result.

[0068] Exemplarily, the second normalized result after batch normalization is applied to the second ELU layer, so as to obtain the first activation result by performing an activation operation through the ELU function in the second ELU layer, wherein the ELU function provides linear activation in the positive region and is processed by smooth exponential activation in the negative region, which can effectively alleviate the problem of gradient disappearance.

[0069] Exemplarily, a random discard operation is performed on the first activation result, for example, a random discard ratio is set, such as discarding 50% of the activation units. This means that in each training iteration, a portion of the activation values ​​will be randomly discarded. Some activation values ​​in the first activation result are randomly discarded according to the set ratio. The discard operation helps prevent overfitting and enhance the generalization ability of the model, thereby obtaining the second feature signal.

[0070] Specifically, by effectively constraining the parameters of the second convolutional layer, extracting features, standardizing and activating features, and finally applying random dropout to reduce overfitting, it will help build a more robust and efficient convolutional neural network, thereby improving the training effect and generalization ability of the initial classification model.

[0071] In some embodiments, the second feature extraction module includes a third convolution layer, a third normalization layer, a third ELU layer, a second random inactivation layer, a fourth convolution layer, a fourth normalization layer, a RELU layer, and a third random inactivation layer. The use of the second feature extraction module to extract time features from the second feature signal to obtain a third feature signal includes: using L2 norm and spectral normalization to constrain parameters of the third convolution layer, and using the third convolution kernel of the third convolution layer to convolve the second feature signal to obtain a third convolution result; using the third normalization layer to batch normalize the third convolution result to obtain a third normalized result. ; Using the third ELU layer to activate the third normalization result to obtain a second activation result; Using the second random inactivation layer to randomly discard the second activation result to obtain an intermediate feature signal; Using the fourth convolution kernel of the fourth convolution layer to convolve the intermediate feature signal to obtain a fourth convolution result; Using the fourth normalization layer to batch normalize the fourth convolution result to obtain a fourth normalization result; Using the RELU layer to activate the fourth normalization result to obtain a third activation result; Using the third random inactivation layer to randomly discard the third activation result to obtain the third feature signal.

[0072] Exemplarily, the second feature extraction module includes a third convolutional layer, a third normalization layer, a third ELU layer, a second random inactivation layer, a fourth convolutional layer, a fourth normalization layer, a RELU layer, and a third random inactivation layer.

[0073] Exemplarily, the L2 norm and spectral normalization are used to perform parameter constraints on the third convolution layer, and then a third convolution kernel with parameter constraints applied is obtained from the third convolution layer to perform convolution processing on the second feature signal, and a third convolution result is generated.

[0074] Exemplarily, the third convolution result is batch normalized using the third normalization layer to obtain the third normalized result. The third convolution result is batch normalized using the third normalization layer, for example, the mean and variance of each feature map are calculated, and the value of each feature map is normalized by subtracting the mean and dividing it by the standard deviation so that the input of each layer has zero mean and unit variance, thereby obtaining the third normalized result.

[0075] Exemplarily, the third normalization result after batch normalization is applied to the third ELU layer, so as to obtain the second activation result by performing an activation operation through the ELU function in the third ELU layer, wherein the ELU function provides linear activation in the positive region and is processed by smooth exponential activation in the negative region, which can effectively alleviate the problem of gradient disappearance.

[0076] Exemplarily, a second random dropout layer is used to perform a random dropout operation on the second activation result, for example, a random dropout ratio is set, for example, 30% of the activation units are dropped. This means that in each training iteration, a portion of the activation values ​​will be randomly dropped. Part of the activation values ​​in the second activation result are randomly dropped according to the set ratio. The dropout operation helps prevent overfitting and enhance the generalization ability of the model, thereby obtaining intermediate feature signals.

[0077] Exemplarily, a fourth convolution kernel is extracted from the fourth convolution layer. This convolution kernel is a filter for processing the intermediate feature signal. The fourth convolution kernel is applied to the intermediate feature signal. This operation involves sliding the convolution kernel on the intermediate feature signal and calculating the dot product of the convolution kernel and the signal area to generate a fourth convolution result. The fourth convolution result is input to the fourth normalization layer for batch normalization. This includes calculating the mean and variance of the batch, and normalizing the fourth convolution result so that it has zero mean and unit variance. The standardized result can also be adjusted by a learnable scaling factor and offset to obtain a fourth normalized result. The RELU function in the RELU layer is then applied to the fourth normalized result. The RELU function sets all negative values ​​to zero and leaves positive values ​​unchanged. This helps introduce nonlinearity and increase the expressiveness of the model. Finally, the proportion of random discarding is determined, for example, discarding a portion of the activation values. This means that in each training iteration, a portion of the activation values ​​will be randomly discarded. Thus, a portion of the activation values ​​in the third activation result are randomly discarded to obtain the third feature signal.

[0078] Specifically, through the above steps, a second feature extraction module with strong generalization ability can be effectively constructed. The parameter constraints of L2 norm and spectral normalization help stabilize the training process and reduce overfitting; convolution and batch normalization operations improve feature extraction and training speed; different activation functions (ELU and RELU) enhance the nonlinear expression ability and training efficiency of the model; random dropout operation further improves the generalization ability of the model, which will help improve the performance and stability of the second feature extraction module according to the above steps.

[0079] In some embodiments, the type output module includes a flattening layer, a linear layer, and a Softmax layer. The use of the type output module to determine the label probability corresponding to the third feature signal and the target stimulus includes: using the flattening layer to flatten the third feature signal to obtain a flattened result; using the linear layer to weighted combine the flattened results to obtain a characterization result corresponding to the first stimulus label; and using the Softmax layer combined with the characterization result to determine the label probability corresponding to the second SSVEP signal and the target stimulus.

[0080] Exemplarily, the type output module includes a flattening layer, a linear layer, and a Softmax layer.

[0081] Exemplarily, the third feature signal is converted into a one-dimensional vector using a flattening layer. The main function of the flattening layer is to flatten the multidimensional feature map into a one-dimensional array for subsequent fully connected layer processing. This is usually a conversion between a convolutional layer and a fully connected layer.

[0082] Exemplarily, the flattened result is input to a linear layer, which contains a set of trainable weights and biases. The role of the linear layer is to perform a weighted combination of the flattened result, that is, matrix multiplication of the flattened vector with the weight matrix, and then add the bias term to generate a new vector, which is used to represent different stimulus labels.

[0083] Exemplarily, the output of the linear layer, i.e., the representation result, is input to the Softmax layer. The Softmax layer converts the output of the linear layer into a probability distribution, where each value represents the probability of the corresponding first stimulus label. The Softmax function calculates the label probability of each target stimulus by taking the exponential of each value output by the linear layer and then dividing these values ​​by their sum.

[0084] Specifically, the flattening layer converts the multi-dimensional features into one-dimensional data suitable for processing by the fully connected layer, simplifies the data structure and prepares for weighted combination, so that the Softmax layer is used to combine the representation results to determine the corresponding label probability between the second SSVEP signal and the target stimulus, which can significantly improve the classification ability and prediction performance of the model.

[0085] The model structure of the initial classification model is as follows Figure 2 As shown, the initial classification model includes a channel expansion module, a first feature extraction module, a second feature extraction module, and a type output module, wherein the channel expansion module includes a first convolutional layer, a first normalization layer, and a first ELU layer, the first feature extraction module includes a second convolutional layer, a second normalization layer, a second ELU layer, and a first random inactivation layer, the second feature extraction module includes a third convolutional layer, a third normalization layer, a third ELU layer, a second random inactivation layer, a fourth convolutional layer, a fourth normalization layer, a RELU layer, and a third random inactivation layer, and the type output module includes a flattening layer, a linear layer, and a Softmax layer. Thus, the above model structure can effectively improve the accuracy and reliability of model classification.

[0086] Step S103: obtaining a third SSVEP signal corresponding to the target object under the second stimulation target and a second stimulation label corresponding to the second stimulation target, and preprocessing the third SSVEP signal to obtain a fourth SSVEP signal.

[0087] Exemplarily, the target object is a user who needs to obtain an accurate classification model, and then the third SSVEP signal corresponding to the target object under the second stimulation target and the second stimulation label corresponding to the second stimulation target are obtained.

[0088] Exemplarily, the third SSVEP signal is preprocessed, and first, the third SSVEP signal is processed by notch filtering to remove power frequency interference to obtain a fourth intermediate signal. For example, the third SSVEP signal is notched filtered at 50 Hz to remove power frequency interference to obtain a fourth intermediate signal. The fourth intermediate signal is then bandpass filtered to obtain a fifth intermediate signal, for example, a 6-stage Butterworth filter is used to bandpass filter the fourth intermediate signal at 6-80 Hz to obtain a fifth intermediate signal. The fifth intermediate signal is then downsampled to obtain a sixth intermediate signal, for example, the fifth intermediate signal is downsampled to 250 Hz to obtain a sixth intermediate signal.

[0089] Exemplarily, a reference signal is set, and the sixth intermediate signal is spatially filtered using a task discrimination pattern analysis algorithm in combination with the reference signal to obtain a fourth SSVEP signal after feature extraction.

[0090] Step S104: fine-tune the initial classification model according to the fourth SSVEP signal and the second stimulus label to obtain a target classification model corresponding to the target object.

[0091] Exemplarily, the initial classification model is incrementally trained by taking the fourth SSVEP signal and the corresponding second stimulus label as input data. This process includes data preparation, selecting and configuring the initial classification model, performing incremental training, verifying and adjusting the model performance, and finally obtaining an optimized target classification model. This target classification model can accurately identify and classify SSVEP signals related to the target object, thereby providing reliable classification results for the SSVEP signals related to the target object in practical applications.

[0092] Step S105: obtaining a fifth SSVEP signal corresponding to the target object, and performing type recognition on the fifth SSVEP signal according to the target classification model to obtain a third stimulation label corresponding to the fifth SSVEP signal.

[0093] Exemplarily, a fifth SSVEP signal corresponding to the target object is obtained, and then the fifth SSVEP signal is preprocessed to obtain a sixth SSVEP signal corresponding to the target object, and then the sixth SSVEP signal is type-identified according to the target classification model to obtain a third stimulation label corresponding to the fifth SSVEP signal.

[0094] Exemplarily, the initial classification model is fine-tuned through the fourth SSVEP signal and the second stimulus label to obtain the target classification model corresponding to the target object, and then when the fifth SSVEP signal corresponding to the target object is obtained, the target classification model corresponding to the target object can be used for accurate identification, thereby providing a reliable classification result for the fifth SSVEP signal of the target object in practical applications, thereby providing good support for subsequent applications.

[0095] The embodiment of the present invention provides a classification method, terminal device and medium for steady-state visual evoked potential, the method comprising: first, obtaining a first SSVEP signal and a corresponding first stimulus label of an initial object under a first stimulus target, and then preprocessing the first SSVEP signal to generate a second SSVEP signal, and then using the second SSVEP signal and the first stimulus label for model training, thereby constructing an initial classification model. The initial classification model provides a basis for subsequent individualized adjustment. Obtain a third SSVEP signal and a corresponding second stimulus label of a target object under a second stimulus target, and preprocess the third SSVEP signal to generate a fourth SSVEP signal, thereby fine-tuning the initial classification model based on the fourth SSVEP signal and the second stimulus label to form a target classification model of the target object. This adjustment enables the target classification model to better adapt to the signal characteristics of the target object, thereby improving the classification accuracy. Further obtain a fifth SSVEP signal of the target object, and classify it using the target classification model, thereby obtaining the third stimulus label corresponding to the fifth SSVEP signal. Through this method, the classification accuracy and stability of the SSVEP signals of the target objects can be significantly improved, and the performance challenges of the SSVEP classification algorithm under high precision and real-time requirements can be effectively addressed, thereby achieving more reliable signal recognition and processing in application scenarios such as brain-computer interfaces.

[0096] See also Figure 3 , Figure 3 A schematic block diagram of the structure of a terminal device provided in an embodiment of the present invention.

[0097] like Figure 3 As shown, the terminal device 300 includes a processor 301 and a memory 302 , and the processor 301 and the memory 302 are connected via a bus 303 , such as an I2C (Inter-integrated Circuit) bus.

[0098] Specifically, the processor 301 is used to provide computing and control capabilities to support the operation of the entire terminal device. The processor 301 can be a central processing unit (CPU), and the processor 301 can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0099] Specifically, the memory 302 may be a Flash chip, a read-only memory (ROM) disk, an optical disk, a USB flash drive, or a mobile hard disk.

[0100] Those skilled in the art will understand that Figure 3 The structure shown in the figure is only a block diagram of a partial structure related to the embodiment of the present invention, and does not constitute a limitation on the terminal device to which the embodiment of the present invention is applied. The specific server may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0101] The processor is used to run a computer program stored in the memory, and implement any one of the steady-state visual evoked potential classification methods provided by the embodiments of the present invention when executing the computer program.

[0102] In one embodiment, the processor is used to run a computer program stored in the memory, and implements the following steps when executing the computer program:

[0103] Determine a first SSVEP signal corresponding to the initial object under the first stimulation target and a first stimulation label corresponding to the first stimulation target, and preprocess the first SSVEP signal to obtain a second SSVEP signal;

[0104] Performing model training according to the second SSVEP signal and the first stimulus label to obtain an initial classification model;

[0105] Obtaining a third SSVEP signal corresponding to the target object under the second stimulation target and a second stimulation label corresponding to the second stimulation target, and preprocessing the third SSVEP signal to obtain a fourth SSVEP signal;

[0106] Fine-tuning the initial classification model according to the fourth SSVEP signal and the second stimulus label to obtain a target classification model corresponding to the target object;

[0107] A fifth SSVEP signal corresponding to the target object is obtained, and a type of the fifth SSVEP signal is identified according to the target classification model to obtain a third stimulation label corresponding to the fifth SSVEP signal.

[0108] In some implementations, the processor 301, during the process of preprocessing the first SSVEP signal to obtain the second SSVEP signal, performs:

[0109] Using notch filtering to remove power frequency interference from the first SSVEP signal to obtain a first intermediate signal;

[0110] Performing bandpass filtering on the first intermediate signal to obtain a second intermediate signal;

[0111] downsampling the second intermediate signal to obtain a third intermediate signal;

[0112] A reference signal is set, and the third intermediate signal is spatially filtered using a task discrimination pattern analysis algorithm in combination with the reference signal to obtain the second SSVEP signal after feature extraction.

[0113] In some implementations, the processor 301, during the process of setting the reference signal, executes:

[0114] Determine a stimulation frequency corresponding to a target task and a harmonic range corresponding to the stimulation frequency;

[0115] Setting a corresponding sine wave signal in the harmonic range of the stimulation frequency according to the target task and setting a corresponding cosine wave signal in the harmonic range of the stimulation frequency according to the target task;

[0116] The reference signal is determined according to the sine wave signal and the cosine wave signal.

[0117] In some implementations, the initial classification model includes a channel expansion module, a first feature extraction module, a second feature extraction module, and a type output module. In the process of obtaining the initial classification model by model training according to the second SSVEP signal and the first stimulus label, the processor 301 executes:

[0118] Using the channel expansion module to perform channel expansion on the second SSVEP signal in a layer-by-layer convolution manner to obtain a first characteristic signal;

[0119] Using the first feature extraction module to perform spatial feature extraction on the first feature signal to obtain a second feature signal;

[0120] Using the second feature extraction module to perform time feature extraction on the second feature signal to obtain a third feature signal;

[0121] Determine the label probability corresponding to the third characteristic signal and the target stimulus by using the type output module, and determine the predicted stimulus label corresponding to the second SSVEP signal according to the label probability;

[0122] Using a cross entropy loss function to perform loss calculation on the predicted stimulus label and the first stimulus label to obtain a loss value;

[0123] When the loss value does not meet the preset condition, continue to perform model training according to the second SSVEP signal and the first stimulus label until the loss value meets the preset condition;

[0124] When the loss value satisfies the preset condition, the target model parameters are obtained, and the initial classification model is determined according to the target model parameters.

[0125] In some implementations, the channel extension module includes a first convolution layer, a first normalization layer, and a first ELU layer. In the process of obtaining the first characteristic signal by performing channel extension on the second SSVEP signal in a layer-by-layer convolution manner using the channel extension module, the processor 301 executes:

[0126] Performing parameter constraints on the first convolution layer using L2 norm and spectral normalization, and performing convolution processing on the second SSVEP signal using the first convolution kernel of the first convolution layer to obtain a first convolution result;

[0127] Using the first normalization layer to perform batch normalization processing on the first convolution result to obtain a first normalized result;

[0128] The first normalization result is activated by using the first ELU layer to obtain the first characteristic signal.

[0129] In some implementations, the first feature extraction module includes a second convolution layer, a second normalization layer, a second ELU layer, and a first random inactivation layer. In the process of using the first feature extraction module to extract spatial features from the first feature signal to obtain the second feature signal, the processor 301 executes:

[0130] Performing parameter constraints on the second convolution layer by using an L2 norm and spectral normalization, and performing convolution processing on the first feature signal by using a second convolution kernel of the second convolution layer to obtain a second convolution result;

[0131] Using the second normalization layer to perform batch normalization processing on the second convolution result to obtain a second normalized result;

[0132] Using the second ELU layer to perform an activation operation on the second normalized result to obtain a first activation result;

[0133] The first random inactivation layer is used to perform a random discard operation on the first activation result to obtain the second characteristic signal.

[0134] In some implementations, the second feature extraction module includes a third convolution layer, a third normalization layer, a third ELU layer, a second random inactivation layer, a fourth convolution layer, a fourth normalization layer, a RELU layer, and a third random inactivation layer. In the process of using the second feature extraction module to extract time features from the second feature signal to obtain a third feature signal, the processor 301 executes:

[0135] Performing parameter constraints on the third convolution layer using L2 norm and spectral normalization, and performing convolution processing on the second feature signal using a third convolution kernel of the third convolution layer to obtain a third convolution result;

[0136] Using the third normalization layer to perform batch normalization processing on the third convolution result to obtain a third normalized result;

[0137] Using the third ELU layer to perform an activation operation on the third normalization result to obtain a second activation result;

[0138] Using the second random deactivation layer to perform a random discard operation on the second activation result to obtain an intermediate feature signal;

[0139] Using the fourth convolution kernel of the fourth convolution layer to perform convolution processing on the intermediate feature signal to obtain a fourth convolution result;

[0140] Using the fourth normalization layer to perform batch normalization processing on the fourth convolution result to obtain a fourth normalized result;

[0141] Using the RELU layer to perform an activation operation on the fourth normalization result to obtain a third activation result;

[0142] The third activation result is randomly discarded using the third random inactivation layer to obtain the third characteristic signal.

[0143] In some implementations, the type output module includes a flattening layer, a linear layer, and a Softmax layer, and the processor 301, in the process of using the type output module to determine the label probability corresponding to the third feature signal and the target stimulus, executes:

[0144] Using the flattening layer to perform a flattening operation on the third characteristic signal to obtain a flattening result;

[0145] Using the linear layer to perform weighted combination on the flattened results to obtain a representation result corresponding to the first stimulus label;

[0146] The Softmax layer is used in combination with the characterization result to determine the label probability corresponding to the second SSVEP signal and the target stimulus.

[0147] It should be noted that those skilled in the art can clearly understand that, for the sake of convenience and brevity of description, the specific working process of the terminal device described above can refer to the corresponding process in the aforementioned steady-state visual evoked potential classification method embodiment, and will not be repeated here.

[0148] An embodiment of the present invention further provides a storage medium for computer-readable storage, wherein the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of any method for classifying steady-state visual evoked potentials provided in the description of the embodiment of the present invention.

[0149] The storage medium may be an internal storage unit of the terminal device in the aforementioned embodiment, such as a hard disk or memory of the terminal device. The storage medium may also be an external storage device of the terminal device, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the terminal device.

[0150] It will be appreciated by those skilled in the art that all or some of the steps, systems, and functional modules / units in the methods disclosed above may be implemented as software, firmware, hardware, and appropriate combinations thereof. In a hardware embodiment, the division between the functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed by several physical components in cooperation. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or implemented as hardware, or implemented as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on a computer-readable medium, which may include a computer storage medium (or non-transitory medium) and a communication medium (or transient medium). As is known to those skilled in the art, the term computer storage medium includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media include, but are not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tapes, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. In addition, it is well known to those skilled in the art that communication media typically contain computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0151] It should be understood that the term "and / or" used in the present specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, including these combinations. It should be noted that, in this article, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "including a..." does not exclude the presence of other identical elements in the process, method, article or system including the element.

[0152] The serial numbers of the embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. The above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for classifying steady-state visual evoked potentials, characterized in that: The method comprises: Determine a first SSVEP signal corresponding to the initial object under the first stimulation target and a first stimulation label corresponding to the first stimulation target, and preprocess the first SSVEP signal to obtain a second SSVEP signal; Performing model training according to the second SSVEP signal and the first stimulus label to obtain an initial classification model; Obtaining a third SSVEP signal corresponding to the target object under the second stimulation target and a second stimulation label corresponding to the second stimulation target, and preprocessing the third SSVEP signal to obtain a fourth SSVEP signal; Fine-tuning the initial classification model according to the fourth SSVEP signal and the second stimulus label to obtain a target classification model corresponding to the target object; Obtaining a fifth SSVEP signal corresponding to the target object, and performing type recognition on the fifth SSVEP signal according to the target classification model to obtain a third stimulus label corresponding to the fifth SSVEP signal; The initial classification model includes a channel expansion module, a first feature extraction module, a second feature extraction module, and a type output module. The initial classification model is obtained by performing model training according to the second SSVEP signal and the first stimulus label, including: Using the channel expansion module to perform channel expansion on the second SSVEP signal in a layer-by-layer convolution manner to obtain a first characteristic signal; Using the first feature extraction module to perform spatial feature extraction on the first feature signal to obtain a second feature signal; Using the second feature extraction module to perform time feature extraction on the second feature signal to obtain a third feature signal; Determine the label probability corresponding to the third characteristic signal and the target stimulus by using the type output module, and determine the predicted stimulus label corresponding to the second SSVEP signal according to the label probability; Using a cross entropy loss function to perform loss calculation on the predicted stimulus label and the first stimulus label to obtain a loss value; When the loss value does not meet the preset condition, continue to perform model training according to the second SSVEP signal and the first stimulus label until the loss value meets the preset condition; When the loss value satisfies the preset condition, the target model parameters are obtained, and the initial classification model is determined according to the target model parameters.

2. The method according to claim 1, characterized in that The preprocessing of the first SSVEP signal to obtain a second SSVEP signal includes: Using notch filtering to remove power frequency interference from the first SSVEP signal to obtain a first intermediate signal; Performing bandpass filtering on the first intermediate signal to obtain a second intermediate signal; downsampling the second intermediate signal to obtain a third intermediate signal; A reference signal is set, and the third intermediate signal is spatially filtered using a task discrimination pattern analysis algorithm in combination with the reference signal to obtain the second SSVEP signal after feature extraction.

3. The method according to claim 2, characterized in that The setting of the reference signal comprises: Determine a stimulation frequency corresponding to a target task and a harmonic range corresponding to the stimulation frequency; Setting a corresponding sine wave signal in the harmonic range of the stimulation frequency according to the target task and setting a corresponding cosine wave signal in the harmonic range of the stimulation frequency according to the target task; The reference signal is determined according to the sine wave signal and the cosine wave signal.

4. The method according to claim 1, characterized in that: The channel extension module includes a first convolution layer, a first normalization layer, and a first ELU layer. The channel extension module is used to perform channel extension on the second SSVEP signal in a layer-by-layer convolution manner to obtain a first characteristic signal, including: Performing parameter constraints on the first convolution layer using L2 norm and spectral normalization, and performing convolution processing on the second SSVEP signal using the first convolution kernel of the first convolution layer to obtain a first convolution result; Using the first normalization layer to perform batch normalization processing on the first convolution result to obtain a first normalized result; The first normalization result is activated by using the first ELU layer to obtain the first characteristic signal.

5. The method according to claim 1, characterized in that: The first feature extraction module includes a second convolution layer, a second normalization layer, a second ELU layer, and a first random inactivation layer. The using the first feature extraction module to perform spatial feature extraction on the first feature signal to obtain a second feature signal includes: Performing parameter constraints on the second convolution layer by using an L2 norm and spectral normalization, and performing convolution processing on the first feature signal by using a second convolution kernel of the second convolution layer to obtain a second convolution result; Using the second normalization layer to perform batch normalization processing on the second convolution result to obtain a second normalized result; Using the second ELU layer to perform an activation operation on the second normalized result to obtain a first activation result; The first random inactivation layer is used to perform a random discard operation on the first activation result to obtain the second characteristic signal.

6. The method according to claim 1, characterized in that The second feature extraction module includes a third convolution layer, a third normalization layer, a third ELU layer, a second random inactivation layer, a fourth convolution layer, a fourth normalization layer, a RELU layer, and a third random inactivation layer. The method of using the second feature extraction module to perform time feature extraction on the second feature signal to obtain a third feature signal includes: Performing parameter constraints on the third convolution layer using L2 norm and spectral normalization, and performing convolution processing on the second feature signal using a third convolution kernel of the third convolution layer to obtain a third convolution result; Using the third normalization layer to perform batch normalization processing on the third convolution result to obtain a third normalized result; Using the third ELU layer to perform an activation operation on the third normalization result to obtain a second activation result; Using the second random deactivation layer to perform a random discard operation on the second activation result to obtain an intermediate feature signal; Using the fourth convolution kernel of the fourth convolution layer to perform convolution processing on the intermediate feature signal to obtain a fourth convolution result; Using the fourth normalization layer to perform batch normalization processing on the fourth convolution result to obtain a fourth normalized result; Using the RELU layer to perform an activation operation on the fourth normalization result to obtain a third activation result; The third activation result is randomly discarded using the third random inactivation layer to obtain the third characteristic signal.

7. The method according to claim 1, characterized in that The type output module includes a flattening layer, a linear layer, and a Softmax layer, and the method of using the type output module to determine the label probability corresponding to the third feature signal and the target stimulus includes: Using the flattening layer to perform a flattening operation on the third characteristic signal to obtain a flattening result; Using the linear layer to perform weighted combination on the flattened results to obtain a representation result corresponding to the first stimulus label; The Softmax layer is used in combination with the characterization result to determine the label probability corresponding to the second SSVEP signal and the target stimulus.

8. A terminal device, characterized in that: The terminal device includes a processor and a memory; The memory is used to store computer programs; The processor is configured to execute the computer program and implement the steady-state visual evoked potential classification method according to any one of claims 1 to 7 when executing the computer program.

9. A computer storage medium for computer storage, characterized in that: The computer storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of the steady-state visual evoked potential classification method according to any one of claims 1 to 7.

Images