An electroencephalogram signal automatic feature extraction method based on a convolutional neural network

Abstract

The application discloses a kind of based on convolutional neural network's electroencephalogram automatic feature extraction method, comprising the following steps: obtaining original electroencephalogram data, based on channel spectrum curvature rate of change carries out adaptive artifact suppression and channel amplitude dynamic re-labeling;Frequency band morphing reconstruction is executed, according to spectrum curvature rate of change sequence determines frequency band boundary, constructs frequency morphing mapping function and generates multiple frequency band component tensor;The frequency band component tensor is input into improved ConvNeXt model, extracts multidimensional electroencephalogram feature tensor;Cross-frequency domain attention fusion is executed, and fusion electroencephalogram feature vector is constructed;Task reasoning is carried out, and electroencephalogram task response result is output;According to frequency morphing consistency index and predicted offset degree, feature feedback reconstruction is executed, and morphing convolution kernel parameter is updated jointly.The application improves the precision and stability of electroencephalogram feature extraction, and is suitable for cognitive state recognition, medical auxiliary analysis and other multiple electroencephalogram intelligent processing scenarios.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of electroencephalogram signal processing and artificial intelligence, and in particular to an electroencephalogram signal automatic feature extraction method based on a convolutional neural network. BACKGROUND

[0002] With the wide application of electroencephalogram signals in the fields of neural engineering, brain-computer interface and intelligent medical treatment, the automatic feature extraction method for electroencephalogram signals has attracted widespread attention. The existing methods generally use fixed frequency band filtering, short-time Fourier transform or wavelet decomposition to preprocess and extract features of electroencephalogram signals, and combine convolutional neural networks or recurrent networks for modeling. However, there are still the following deficiencies in practical application:

[0003] Firstly, the traditional artifact processing method mostly uses static threshold or global model, which is difficult to adaptively identify the artifact segment according to the dynamic change of channel and time spectrum characteristics, and is easy to cause feature pollution; secondly, the division of electroencephalogram frequency band generally relies on fixed boundary, which cannot reflect the individualized frequency spectrum structure, and lacks modeling ability for frequency deformation characteristics; in addition, the existing network structure mostly uses fixed sampling method, which is difficult to capture effective features under nonlinear frequency mapping, and has limited response ability to frequency spectrum deformation and rhythm shift.

[0004] At the same time, the existing model lacks modeling of brain region connection structure, and the cooperative relationship between channels cannot be fully integrated into the convolution process; in the multi-band fusion stage, there is also a lack of weighting strategy based on frequency band coupling and channel consistency, which affects the integration effect of high-dimensional features; more importantly, the model is usually a static inference structure, which lacks a feedback mechanism combined with task results, and is difficult to adapt to the state drift and time sequence variation problem of electroencephalogram signals.

[0005] Therefore, how to provide an electroencephalogram signal automatic feature extraction method based on a convolutional neural network is a problem that those skilled in the art need to solve. SUMMARY

[0006] One object of the present application is to provide an electroencephalogram signal automatic feature extraction method based on a convolutional neural network. The present application combines channel spectrum curvature rate analysis, adaptive artifact suppression, frequency band deformation reconstruction and improved ConvNeXt model, constructs a cross-frequency domain deformation convolution structure and a channel routing mechanism driven by a brain region connection graph, fuses neural frequency variation modulation and rhythmic time sequence residual path, and introduces a feature feedback reconstruction mechanism after multi-band feature fusion to enhance the stability of the model, which has the advantages of high feature expression accuracy, strong frequency spectrum structure adaptability and good recognition robustness.

[0007] According to the electroencephalogram signal automatic feature extraction method based on a convolutional neural network of the present application, the following steps are included:

[0008] Raw EEG signal data is acquired, and artifact segments are identified based on an adaptive artifact suppression method using the channel spectrum curvature change rate. Dynamic amplitude recalibration is then performed on each channel to generate an EEG time series tensor.

[0009] Frequency band deformation reconstruction is performed on the EEG time series tensor. Based on the frequency band division, the frequency band boundary points are determined by the spectral curvature rate of change sequence. The original spectral interval is nonlinearly mapped by constructing a frequency deformation mapping function to generate multiple frequency band component tensors.

[0010] Multiple frequency band component tensors are input into the improved ConvNeXt model. The improved ConvNeXt model includes a ConvNeXt convolution backbone structure that introduces cross-frequency domain deformable convolution kernels, and introduces a neural frequency-varying modulation layer, a channel routing mechanism driven by brain region connectivity maps, and a rhythmic temporal residual path mechanism to extract multidimensional EEG feature tensors.

[0011] A cross-frequency domain attention fusion method is applied to the multidimensional EEG feature tensor. The fused EEG feature vector is formed by weighted concatenation of the estimated frequency band coupling strength matrix and the channel structure consistency factor.

[0012] Task inference is performed on the fused EEG feature vectors, and the EEG task response results are output using category distribution mapping according to the preset task.

[0013] Based on the EEG task response results, feature feedback reconstruction is performed according to the frequency deformation consistency index and prediction offset, and the sampling offset parameters of the cross-frequency domain deformation convolution kernel are jointly updated.

[0014] Optionally, the acquisition of raw EEG signal data, identification of artifact segments based on an adaptive artifact suppression method using the channel spectral curvature change rate, and dynamic amplitude recalibration processing of each channel to generate an EEG time series tensor specifically includes:

[0015] The channel spectrum curvature change rate is the channel spectrum curvature change rate formed by differential calculation of the amplitude change trend between adjacent frequency points in the short-time spectrum of a single channel according to continuous time slices. The adaptive artifact suppression method is to statistically analyze the average change level of the channel spectrum curvature change rate in continuous time slices and dynamically update the artifact recognition threshold within the window range.

[0016] The adaptively updated artifact detection threshold is applied to the channel spectrum curvature change rate sequence in the time dimension. When the channel spectrum curvature change rate of consecutive time slices is higher than the artifact detection threshold, the current consecutive time slice is marked as an artifact segment. When the channel spectrum curvature change rate of consecutive time slices is near the artifact detection threshold and the change amplitude is lower than the preset stable change range, the current consecutive time slice is marked as a stable interval.

[0017] The original EEG signal data within the stable interval is subjected to dynamic amplitude recalibration processing. The dynamic amplitude recalibration processing involves obtaining the maximum and minimum amplitude of each channel signal within a set time window, using the maximum and minimum amplitudes to form the original amplitude interval, and converting the original amplitude interval into a new amplitude interval using a linear interval mapping method.

[0018] After artifact identification, stationary interval division and amplitude dynamic recalibration, the signals of each channel are arranged in the order of acquisition channel and sampling time. A multi-dimensional data structure is constructed with time slice as the time dimension and channel number as the channel dimension to form an EEG time series tensor.

[0019] Optionally, the step of performing frequency band deformation reconstruction on the EEG time series tensor, determining the frequency band boundary points based on the frequency band division using the spectral curvature change rate sequence, and generating multiple frequency band component tensors by constructing a frequency deformation mapping function to perform nonlinear mapping on the original spectral interval, specifically includes:

[0020] Based on the EEG time series tensor, a windowed fast Fourier transform is performed on the signals of each channel within a frequency domain analysis window of a set length to form the power spectral density curve corresponding to each window.

[0021] Calculate the frequency band division using the spectral curvature change rate sequence on the power spectral density curve;

[0022] The frequency band division is performed by performing extreme value detection on the spectrum curvature change rate sequence, and the position that meets the change rate peak condition is taken as the frequency band boundary point, and multiple original frequency band intervals are formed by adjacent frequency band boundary points;

[0023] A frequency deformation mapping function is constructed within the original frequency band interval. The frequency deformation mapping function is a nonlinear continuous function defined on the original frequency interval, including: an exponential nonlinear compression mapping function and a Gaussian reweighted mapping function based on the mean and variance of the power spectrum within the interval.

[0024] The frequency deformation mapping function is applied to each original frequency band interval, and a nonlinear transformation is performed on the frequency index and frequency domain interpolation resampling is performed to obtain the deformation tensor corresponding to multiple spectra.

[0025] Channel dimension alignment and amplitude normalization are performed on the deformation tensors of all channels in the same frequency band to form frequency band component tensors. Each frequency band component tensor constitutes multiple frequency band component tensors in the order of frequency bands.

[0026] Optionally, the improved ConvNeXt model includes a ConvNeXt convolutional backbone structure with cross-frequency domain deformable convolutional kernels, a neural frequency-varying modulation layer, a channel routing mechanism driven by brain region connectivity maps, and a rhythmic temporal residual path mechanism.

[0027] Multiple frequency band component tensors are respectively input into the ConvNeXt convolutional backbone structure in the improved ConvNeXt model. The ConvNeXt convolutional backbone structure is composed of multiple depthwise separable convolutional modules stacked together. Each depthwise separable convolutional module consists of a channel-wise convolutional layer, a layer normalization layer and a GELU activation function to generate the initially encoded frequency band feature tensor.

[0028] In the ConvNeXt convolutional backbone structure, a cross-frequency domain deformation convolution kernel is set. The sampling position of the cross-frequency domain deformation convolution kernel is obtained by two factors: one is the energy distribution curve in the frequency dimension of the frequency band component tensor, and the other is the frequency deformation mapping function. The sampling position is obtained by multiplying the current frequency index by the scaling factor output by the frequency deformation mapping function and then weighted and shifting. Linear interpolation is used to perform frequency domain convolution sampling at non-integer positions, and the frequency deformation-aware intermediate feature tensor is output.

[0029] The intermediate feature tensor is input into the neural frequency-modulation layer, which includes a set of trainable sigmoid compression functions. Each sigmoid compression function acts on the frequency amplitude vector of a channel, outputs a modulation factor vector, and performs a frequency-wise dot product operation between the modulation factor vector and the intermediate feature tensor to output the frequency-modulated modulation feature tensor.

[0030] The modulation feature tensor is input into a channel routing mechanism driven by a brain region connectivity graph. The brain region connectivity graph is a weighted graph structure constructed based on the spatial positional relationship of channels. Channels are graph nodes, and the edge weights between nodes are generated by the physical connection distance and the signal covariance. The channel routing mechanism uses the brain region connectivity graph to generate a routing weight matrix between channels, performs a weighted reconstruction operation on the channel dimension on the modulation feature tensor, and outputs a structural feature tensor that integrates the brain region connectivity relationship.

[0031] The structural feature tensor is input into the rhythmic temporal residual path mechanism, which is implemented based on a bidirectional gated unit (GRU) structure. It performs sequential modeling in the time dimension and introduces the rhythmic control vector estimated from the principal periodic component of the power spectrum of each segment in the frequency band component tensor. The GRU state transition matrix is ​​applied through an exponential weighting function to dynamically adjust the update amplitude and frequency periodic response of the temporal residual path, and outputs a multidimensional EEG feature tensor containing rhythmic dynamics.

[0032] Optionally, the method of performing cross-frequency domain attention fusion on the multidimensional EEG feature tensor, which involves weighted concatenation of the estimated frequency band coupling strength matrix and channel structure consistency factor to form a fused EEG feature vector, specifically includes:

[0033] The multidimensional EEG feature tensor is split according to the frequency band dimension, and the channel-frequency two-dimensional feature matrix corresponding to each frequency band is extracted.

[0034] Based on the channel-frequency two-dimensional feature matrix of each frequency band, the mean vector and standard deviation vector of the full channel feature corresponding to each frequency band are calculated respectively. The mean vectors of all frequency bands are combined pairwise, and the frequency band coupling strength matrix is ​​constructed according to the product of the corresponding covariance degree and standard deviation.

[0035] Based on the frequency band coupling strength matrix, a frequency band attention allocation matrix is ​​generated through an attention mechanism. The integer row vector of the current frequency band in the frequency band coupling strength matrix is ​​extracted as the spectral coupling expression, and the coupling discreteness distance is calculated with the average coupling strength vector of all frequency bands. The coupling discreteness is mapped to the attention score by the inverse normalization strategy. The attention scores of all frequency bands are normalized so that the sum is 1, thus forming the frequency band attention allocation matrix.

[0036] For the channel-frequency two-dimensional feature matrix in each frequency band, the channel structure consistency factor is calculated based on the brain region connectivity map. The channel-frequency two-dimensional feature matrix of each frequency band and the corresponding structure consistency factor are then subjected to element-wise weighted fusion processing to obtain the channel structure enhancement feature matrix.

[0037] Based on the attention weights corresponding to each frequency band in the frequency band attention allocation matrix, the channel structure enhancement feature matrices after processing all frequency bands are weighted and superimposed, and average pooling is performed on the frequency dimension to form a fused EEG feature vector by concatenating them according to the channel dimension.

[0038] Optionally, the step of performing task inference on the fused EEG feature vectors and outputting the EEG task response result according to the preset task using a category distribution mapping specifically includes:

[0039] Perform a linear mapping operation on the fused EEG feature vector to convert it into an intermediate feature representation;

[0040] The intermediate feature representation is linearly expanded. The linear expansion method is to construct a target vector of a set length based on the total number of categories of the preset task, and then transform it again by constructing a new set of linear weights and bias parameters to generate a category score vector consistent with the number of task categories.

[0041] Each value in the category score vector is input into the exponential transformation function. The results of all exponential transformation functions are normalized. After the sum is calculated, each item is divided by the sum to obtain the task category response probability distribution vector.

[0042] Based on the category number corresponding to the maximum probability value in the task category response probability distribution vector, the EEG task response result indicated by the category code is output as the final classification result;

[0043] If the preset task objective is a multi-class prediction task, then the complete task category response probability distribution vector is output as the EEG task response result.

[0044] If the task is a binary classification or single-label prediction task, only the category with the highest probability value and its corresponding probability value will be output as the EEG task response result.

[0045] Optionally, the step of performing feature feedback reconstruction based on the EEG task response results, according to the frequency deformation consistency index and prediction offset, and jointly updating the sampling offset parameters of the cross-frequency domain deformable convolution kernel, specifically includes:

[0046] The EEG task response results are compared with historical response labels, the degree of consistency is statistically analyzed, and the response consistency index is calculated.

[0047] The prediction bias is calculated by comparing the category response probability in the EEG task response results with the historical average response probability of the corresponding category.

[0048] A stability score for deformation is constructed by jointly using the response consistency index and the prediction offset. If the stability score is lower than a set threshold, a feedback reconstruction process is initiated.

[0049] Based on the current distribution of the predicted offset in the frequency band, the mapping interval of the frequency deformation mapping function in the corresponding frequency band is fine-tuned using a local interval sampling density adjustment strategy.

[0050] Based on the channel distribution with low response consistency, the sampling offset parameters of the corresponding channels and frequency bands in the cross-frequency domain deformation convolution kernel are adjusted. The adjustment method is to perform weighted accumulation based on the response offset weight in the most recent batches to generate the offset correction amount and update the dynamic sampling strategy of the convolution kernel.

[0051] The updated frequency deformation mapping function and sampling offset parameters participate in the next round of EEG signal frequency band reconstruction and convolution processing, forming a feature extraction path with enhanced stability.

[0052] The beneficial effects of this invention are:

[0053] This invention addresses the problems of insufficient artifact suppression, rigid frequency band modeling, lack of spatial connectivity information, and unstable response to signal temporal drift in existing EEG signal feature extraction methods. It proposes an automatic feature extraction method with structural innovation, which has significant theoretical value and engineering application significance. This invention introduces a channel spectral curvature change rate index in the preprocessing stage, enabling dynamic identification and amplitude recalibration of artifact segments and enhancing the stability of signal input quality. Through frequency band deformation reconstruction and frequency deformation mapping functions, it effectively breaks the limitations of fixed frequency band division, improving the accuracy of spectral structure modeling. An improved ConvNeXt model is introduced, integrating cross-frequency domain deformation convolution kernels, a channel routing mechanism driven by brain region connectivity maps, and a rhythmic temporal residual path mechanism, achieving for the first time the collaborative extraction of spatial connectivity relationships, spectral deformation features, and rhythmic dynamic information of EEG within a unified structure. An attention fusion strategy combining frequency band coupling strength and channel structure consistency factors achieves precise weighted integration of multi-frequency band information. Simultaneously, by introducing a feature feedback reconstruction mechanism driven by frequency deformation consistency and prediction offset, the model possesses adaptive adjustment capabilities to signal fluctuations and state drift. In summary, this invention not only significantly improves the accuracy and stability of EEG feature extraction but also enhances the model's generalization ability and robustness in complex application scenarios, possessing broad engineering application value. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Fig. 1 This is a flowchart of an automatic feature extraction method for electroencephalogram (EEG) signals based on a convolutional neural network proposed in this invention.

[0056] Fig. 2 This is a schematic diagram of an automatic feature extraction method for electroencephalogram (EEG) signals based on a convolutional neural network proposed in this invention.

[0057] Fig. 3 This is a framework diagram of the improved ConvNeXt model in the automatic feature extraction method for EEG signals based on convolutional neural networks proposed in this invention. Detailed Implementation

[0058] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0059] refer to Figs. 1-3 An automatic feature extraction method for electroencephalogram (EEG) signals based on convolutional neural networks includes the following steps:

[0060] Raw EEG signal data is acquired, and artifact segments are identified based on an adaptive artifact suppression method using the channel spectrum curvature change rate. Dynamic amplitude recalibration is then performed on each channel to generate an EEG time series tensor.

[0061] Frequency band deformation reconstruction is performed on the EEG time series tensor. Based on the frequency band division, the frequency band boundary points are determined by the spectral curvature rate of change sequence. The original spectral interval is nonlinearly mapped by constructing a frequency deformation mapping function to generate multiple frequency band component tensors.

[0062] Multiple frequency band component tensors are input into the improved ConvNeXt model. The improved ConvNeXt model includes a ConvNeXt convolution backbone structure that introduces cross-frequency domain deformable convolution kernels, and introduces a neural frequency-varying modulation layer, a channel routing mechanism driven by brain region connectivity maps, and a rhythmic temporal residual path mechanism to extract multidimensional EEG feature tensors.

[0063] A cross-frequency domain attention fusion method is applied to the multidimensional EEG feature tensor. The fused EEG feature vector is formed by weighted concatenation of the estimated frequency band coupling strength matrix and the channel structure consistency factor.

[0064] Task inference is performed on the fused EEG feature vectors, and the EEG task response results are output using category distribution mapping according to the preset task.

[0065] Based on the EEG task response results, feature feedback reconstruction is performed according to the frequency deformation consistency index and prediction offset, and the sampling offset parameters of the cross-frequency domain deformation convolution kernel are jointly updated.

[0066] In this embodiment, the acquisition of raw EEG signal data, the identification of artifact segments based on an adaptive artifact suppression method using the channel spectral curvature change rate, and the dynamic amplitude recalibration processing of each channel to generate an EEG time series tensor specifically include:

[0067] The channel spectrum curvature change rate is the channel spectrum curvature change rate formed by differential calculation of the amplitude change trend between adjacent frequency points in the short-time spectrum of a single channel according to continuous time slices. It is used to characterize the local bending of the spectrum. The adaptive artifact suppression method is to statistically analyze the average change level of the channel spectrum curvature change rate in continuous time slices and dynamically update the artifact recognition threshold within the window range, so that the judgment criteria for artifact segments are updated with the change of the input signal.

[0068] The adaptively updated artifact detection threshold is applied to the channel spectrum curvature change rate sequence in the time dimension. When the channel spectrum curvature change rate of consecutive time slices is higher than the artifact detection threshold, the current consecutive time slice is marked as an artifact segment. When the channel spectrum curvature change rate of consecutive time slices is near the artifact detection threshold and the change amplitude is lower than the preset stable change range, the current consecutive time slice is marked as a stable interval.

[0069] The original EEG signal data within the stable interval is subjected to dynamic amplitude recalibration processing. The dynamic amplitude recalibration processing involves acquiring the maximum and minimum amplitude values ​​of each channel signal within a set time window, using the maximum and minimum amplitude values ​​to form the original amplitude interval, and converting the original amplitude interval into a new amplitude interval using a linear interval mapping method. The linear interval mapping method involves linearly converting each amplitude value of the original signal according to the ratio of the interval start point and the interval span between the original amplitude interval and the target amplitude interval, thereby generating a continuous amplitude sequence after dynamic amplitude recalibration.

[0070] After artifact identification, stationary interval division and amplitude dynamic recalibration, the signals of each channel are arranged in the order of acquisition channel and sampling time. A multi-dimensional data structure is constructed with time slice as the time dimension and channel number as the channel dimension to form an EEG time series tensor.

[0071] In this embodiment, the step of performing frequency band deformation reconstruction on the EEG time series tensor, determining the frequency band boundary points based on the frequency band division using the spectral curvature change rate sequence, and generating multiple frequency band component tensors by constructing a frequency deformation mapping function to perform nonlinear mapping on the original spectral interval, specifically includes:

[0072] Based on the EEG time series tensor, a windowed fast Fourier transform is performed on the signals of each channel within a frequency domain analysis window of a set length to form the power spectral density curve corresponding to each window.

[0073] The frequency band division spectrum curvature change rate sequence is calculated on the power spectral density curve. The frequency band division spectrum curvature change rate is a change rate index obtained by weighting and superimposing the first-order difference change and the second-order difference change of adjacent frequency points of the power spectral density on the frequency axis. It is used to characterize the trend change of the overall energy distribution structure in the frequency domain.

[0074] The frequency band division is performed by performing extreme value detection on the spectrum curvature change rate sequence, and the position that meets the change rate peak condition is taken as the frequency band boundary point, and multiple original frequency band intervals are formed by adjacent frequency band boundary points;

[0075] A frequency deformation mapping function is constructed within the original frequency band interval. The frequency deformation mapping function is a nonlinear continuous function defined on the original frequency interval, including: an exponential nonlinear compression mapping function for performing exponential decay remapping based on the frequency index position within the interval, and a Gaussian reweighted mapping function based on the mean and variance of the power spectrum within the interval, for reordering the frequency index according to the normalized weight.

[0076] The frequency deformation mapping function is applied to each original frequency band interval, and a nonlinear transformation is performed on the frequency index and frequency domain interpolation resampling is performed to obtain the deformation tensor corresponding to multiple spectra.

[0077] Channel dimension alignment and amplitude normalization are performed on the deformation tensors of all channels in the same frequency band to form frequency band component tensors. Each frequency band component tensor constitutes multiple frequency band component tensors in the order of frequency bands. The number of frequency bands is determined by the frequency band division using the rate of change of spectral curvature. The frequency division structure is nonlinearly reconstructed by the frequency deformation mapping function.

[0078] In this embodiment, the improved ConvNeXt model includes a ConvNeXt convolution backbone structure with cross-frequency domain deformable convolution kernels, a neural frequency-varying modulation layer, a brain region connectivity map-driven channel routing mechanism, and a rhythmic temporal residual path mechanism.

[0079] Multiple frequency band component tensors are respectively input into the ConvNeXt convolutional backbone structure in the improved ConvNeXt model. The ConvNeXt convolutional backbone structure is composed of multiple depthwise separable convolutional modules stacked together. Each depthwise separable convolutional module consists of a channel-wise convolutional layer, a layer normalization layer, and a GELU activation function, keeping the spatial dimension and frequency dimension structure of the frequency band component tensors unchanged, and generating the initially encoded frequency band feature tensors.

[0080] A cross-frequency domain deformation convolution kernel is set in the ConvNeXt convolution backbone structure to realize non-uniform sampling and frequency structure awareness in the frequency axis direction. The sampling position of the cross-frequency domain deformation convolution kernel is obtained by two factors: one is the energy distribution curve in the frequency dimension of the frequency band component tensor, and the other is the frequency deformation mapping function. The sampling position is obtained by multiplying the current frequency index by the scaling factor output by the frequency deformation mapping function and then weighted and shifting. Linear interpolation is used to perform frequency domain convolution sampling at non-integer positions to output the intermediate feature tensor of frequency deformation awareness.

[0081] The intermediate feature tensor is input into the neural frequency-modulation layer, which includes a set of trainable sigmoid compression functions. Each sigmoid compression function acts on the frequency amplitude vector of a channel, outputs a modulation factor vector, and performs a frequency-wise dot product operation between the modulation factor vector and the intermediate feature tensor to achieve dynamic enhancement of the expressive power of different frequency ranges, and outputs the frequency-modulated modulation feature tensor.

[0082] The modulation feature tensor is input into a channel routing mechanism driven by a brain region connectivity graph. The brain region connectivity graph is a weighted graph structure constructed based on the spatial positional relationship of channels. Channels are graph nodes, and the edge weights between nodes are generated by the physical connection distance and the signal covariance. The channel routing mechanism uses the brain region connectivity graph to generate a routing weight matrix between channels, performs a weighted reconstruction operation on the channel dimension on the modulation feature tensor, and outputs a structural feature tensor that integrates the brain region connectivity relationship.

[0083] The structural feature tensor is input into the rhythmic temporal residual path mechanism, which is implemented based on a bidirectional gated unit (GRU) structure. It performs sequential modeling in the time dimension and introduces the rhythmic control vector estimated from the principal periodic component of the power spectrum of each segment in the frequency band component tensor. The GRU state transition matrix is ​​applied through an exponential weighting function to dynamically adjust the update amplitude and frequency periodic response of the temporal residual path, and outputs a multidimensional EEG feature tensor containing rhythmic dynamics.

[0084] In this embodiment, the method of performing cross-frequency domain attention fusion on the multidimensional EEG feature tensor, which involves weighted concatenation of the estimated frequency band coupling strength matrix and channel structure consistency factor to form a fused EEG feature vector, specifically includes:

[0085] The multidimensional EEG feature tensor is split according to the frequency band dimension, and the channel-frequency two-dimensional feature matrix corresponding to each frequency band is extracted. The channel-frequency two-dimensional feature matrix is ​​a two-dimensional feature expression structure composed of different channels and frequency points under each EEG frequency band.

[0086] Based on the channel-frequency two-dimensional feature matrix of each frequency band, the mean vector and standard deviation vector of the full channel feature corresponding to each frequency band are calculated respectively. The mean vectors of all frequency bands are combined pairwise, and the frequency band coupling strength matrix is ​​constructed according to the product of the corresponding covariance degree and standard deviation. The frequency band coupling strength matrix is ​​a two-dimensional symmetric matrix that measures the cooperative relationship between different frequency bands. Each element represents the joint similarity between the two corresponding frequency bands in terms of energy mean and frequency response mode, which is used to characterize the spectral coupling strength between frequency bands.

[0087] Based on the frequency band coupling strength matrix, a frequency band attention allocation matrix is ​​generated through an attention mechanism. The integer row vector of the current frequency band in the frequency band coupling strength matrix is ​​extracted as the spectral coupling expression, and the coupling discreteness distance is calculated with the average coupling strength vector of all frequency bands. The coupling discreteness distance is the average of the absolute values ​​of the differences between the components of the vector. The inverse normalization strategy is used to map the coupling discreteness to the attention score. The attention scores of all frequency bands are normalized so that the sum is 1, thus forming the frequency band attention allocation matrix.

[0088] For the channel-frequency two-dimensional feature matrix in each frequency band, the channel structure consistency factor is calculated based on the brain region connectivity map. The structure consistency factor represents the similarity of feature vectors between the current channel and the topologically adjacent channels. It is obtained by statistically analyzing the mean cosine similarity of features between channel groups and is used to measure the stability and feature consistency of the current channel in spatial structure. The channel-frequency two-dimensional feature matrix of each frequency band and the corresponding structure consistency factor are subjected to element-wise weighted fusion processing to obtain the channel structure enhancement feature matrix.

[0089] Based on the attention weights corresponding to each frequency band in the frequency band attention allocation matrix, the channel structure enhancement feature matrices after processing all frequency bands are weighted and superimposed, and average pooling is performed on the frequency dimension to reduce the feature dimension. The feature vectors are then concatenated according to the channel dimension to form a fused EEG feature vector.

[0090] In this embodiment, the step of performing task reasoning on the fused EEG feature vectors and outputting the EEG task response result using category distribution mapping according to the preset task specifically includes:

[0091] A linear mapping operation is performed on the fused EEG feature vector. The linear mapping operation is achieved by constructing a matrix multiplication between the weight vector and the fused EEG feature vector and superimposing a bias term, thereby converting it into an intermediate feature representation.

[0092] The intermediate feature representation is linearly expanded. The linear expansion method is to construct a target vector of a set length based on the total number of categories of the preset task, and then transform it again by constructing a new set of linear weights and bias parameters to generate a category score vector consistent with the number of task categories.

[0093] Each value in the category score vector is input into the exponential transformation function, which is an exponential function with the natural constant as the base. The results of all exponential transformation functions are normalized, and after the sum is calculated, each item is divided by the sum to obtain the task category response probability distribution vector.

[0094] Based on the category number corresponding to the maximum probability value in the task category response probability distribution vector, the EEG task response result indicated by the category code is output as the final classification result;

[0095] If the preset task objective is a multi-class prediction task, then the complete task category response probability distribution vector is output as the EEG task response result.

[0096] If the task is a binary classification or single-label prediction task, only the category with the highest probability value and its corresponding probability value will be output as the EEG task response result.

[0097] In this embodiment, the step of performing feature feedback reconstruction based on the EEG task response results, according to the frequency deformation consistency index and prediction offset, and jointly updating the sampling offset parameters of the cross-frequency domain deformable convolution kernel, specifically includes:

[0098] The EEG task response results are compared with historical response labels, the degree of consistency is statistically analyzed, and the response consistency index is calculated. The response consistency index is the consistency rate between a number of consecutive classification outputs and the true labels in the current time period.

[0099] The prediction offset is calculated by comparing the category response probability in the EEG task response results with the historical average response probability of the corresponding category. The prediction offset reflects the degree of fluctuation of the current sample prediction probability in the time series.

[0100] A stability score for deformation is constructed by jointly using the response consistency index and the prediction offset. If the stability score is lower than a set threshold, a feedback reconstruction process is initiated.

[0101] Based on the current distribution of predicted offset in the frequency band, the sampling density fine-tuning strategy within the local interval is adopted for the mapping interval of the frequency deformation mapping function in the corresponding frequency band to enhance the resampling accuracy of the high offset frequency band.

[0102] Based on the channel distribution with low response consistency, the sampling offset parameters of the corresponding channels and frequency bands in the cross-frequency domain deformation convolution kernel are adjusted. The adjustment method is to perform weighted accumulation based on the response offset weight in the most recent batches to generate the offset correction amount and update the dynamic sampling strategy of the convolution kernel.

[0103] The updated frequency deformation mapping function and sampling offset parameters participate in the frequency band reconstruction and convolution processing of the next round of EEG signals, forming a feature extraction path with enhanced stability, thereby improving the overall response consistency and temporal adaptability of EEG analysis.

[0104] Example 1:

[0105] To verify the feasibility of this invention in practice, it was applied to a multi-channel EEG signal classification experiment based on a visual stimulus task, collected in a university's neuroinformation processing laboratory. This experiment aimed to analyze differences in brain region activity under different stimulus conditions and improve EEG classification accuracy. The data source consisted of 64-channel EEG signals collected from 16 subjects viewing graphic stimuli with their eyes closed, at a sampling rate of 1000 Hz for a total duration of 5 minutes. The raw data contained varying degrees of electromyographic artifacts and electrode offset noise, and there were strong individual differences in EEG rhythms. Traditional feature extraction models had limited classification performance when processing such non-stationary high-dimensional signals.

[0106] First, the preprocessing procedure of this invention is performed on the EEG data of all subjects. For each channel, the rate of change of channel spectral curvature within a 100ms sliding window is calculated. This rate of change is combined with the average rate of change within the window to construct a dynamic threshold for artifact detection, identifying and eliminating artifact segments. For data marked as stationary intervals, dynamic amplitude recalibration is performed, linearly mapping their original amplitude range to 0-1 to ensure the comparability of amplitudes across channels.

[0107] Frequency band deformation reconstruction is performed on the preprocessed signal. The rate of change of spectral curvature for frequency band division is calculated, and the frequency band boundary points are extracted through its local maxima to divide the spectrum into multiple frequency bands. Within each frequency band, a frequency deformation mapping function is generated using an exponential decay compression function or a Gaussian reweighting function. After frequency index nonlinear transformation and resampling, a spectral deformation tensor is constructed and synthesized into a multi-band component tensor with a unified structure.

[0108] Next, the aforementioned multi-band component tensors are input into the improved ConvNeXt model. In this model, the ConvNeXt backbone structure extracts preliminary features based on depthwise separable convolutional layers, while simultaneously setting cross-frequency domain deformable convolutional kernels. The sampling offset is driven by both the frequency deformation function and energy distribution, and convolutional sampling is completed through linear interpolation. The intermediate feature tensors are then fed into a neural frequency modulation layer for frequency enhancement, and spatial structural relationships are fused using a brain region connectivity map channel routing mechanism. Finally, rhythmic dynamics are modeled through a rhythmic temporal residual path mechanism, outputting a multidimensional EEG feature tensor.

[0109] Subsequently, cross-frequency domain attention fusion processing was performed on the multidimensional EEG feature tensor. First, a frequency band coupling strength matrix and a structural consistency factor were constructed. Then, an additive attention mechanism was used to generate a frequency band attention allocation matrix. All frequency bands were summed and fused according to weighted weights to obtain the fused EEG feature vector.

[0110] Finally, the EEG feature vectors are integrated into a linear mapping process for EEG task response inference, outputting a response probability distribution based on preset categories, and extracting prediction results. If the historical prediction consistency index is low or the current prediction bias is significant, a feature feedback reconstruction mechanism is triggered to fine-tune and optimize the frequency deformation mapping function and convolution kernel sampling bias to enhance the model's adaptability to subsequent signals. Table 1 shows some experimental results:

[0111] Table 1. Comparison of EEG signal preprocessing effects based on average per channel.

[0112]

[0113] Table 2 Comparison of Classification Accuracy and Robustness

[0114]

[0115] Table 3. Changes in the average performance of the 5 iterations before and after the intervention of the feedback reconstruction module.

[0116]

[0117] Table 1 shows that the adaptive artifact recognition algorithm guided by the channel spectrum curvature change rate can significantly improve the recognition rate of stationary signals and enhance amplitude comparability. Table 2 shows that the improved ConvNeXt, after introducing structure-aware and frequency modulation mechanisms, significantly improves classification accuracy and has higher robustness to artifact interference. Table 3 indicates that the feedback reconstruction mechanism can effectively improve prediction stability and consistency and optimize feature representation.

[0118] The following beneficial effects can be obtained from the embodiments: The EEG feature extraction method based on frequency deformation modeling and multidimensional structure perception proposed in this invention can effectively improve the expression accuracy and classification performance of EEG signals in non-stationary and interference environments. In particular, it shows significant advantages in accuracy and response stability in task-driven EEG decoding, and has strong practical application value and promotion potential.

[0119] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An automatic feature extraction method for electroencephalogram (EEG) signals based on convolutional neural networks, characterized in that, Includes the following steps: Raw EEG signal data is acquired, and artifact segments are identified based on an adaptive artifact suppression method using the channel spectrum curvature change rate. Dynamic amplitude recalibration is then performed on each channel to generate an EEG time series tensor. Frequency band deformation reconstruction is performed on the EEG time series tensor. Based on the frequency band division, the frequency band boundary points are determined by the spectral curvature rate of change sequence. The original spectral interval is nonlinearly mapped by constructing a frequency deformation mapping function to generate multiple frequency band component tensors. Multiple frequency band component tensors are input into the improved ConvNeXt model. The improved ConvNeXt model includes a ConvNeXt convolution backbone structure that introduces cross-frequency domain deformable convolution kernels, and introduces a neural frequency-varying modulation layer, a channel routing mechanism driven by brain region connectivity maps, and a rhythmic temporal residual path mechanism to extract multidimensional EEG feature tensors. A cross-frequency domain attention fusion method is applied to the multidimensional EEG feature tensor. The fused EEG feature vector is formed by weighted concatenation of the estimated frequency band coupling strength matrix and the channel structure consistency factor. Task inference is performed on the fused EEG feature vectors, and the EEG task response results are output using category distribution mapping according to the preset task. Based on the EEG task response results, feature feedback reconstruction is performed according to the response consistency index and prediction offset, and the sampling offset parameters of the cross-frequency domain deformable convolution kernel are jointly updated. The improved ConvNeXt model includes a ConvNeXt convolutional backbone structure with cross-frequency domain deformable convolutional kernels, a neural frequency-varying modulation layer, a brain region connectivity map-driven channel routing mechanism, and a rhythmic temporal residual path mechanism. Multiple frequency band component tensors are respectively input into the ConvNeXt convolutional backbone structure in the improved ConvNeXt model. The ConvNeXt convolutional backbone structure is composed of multiple depthwise separable convolutional modules stacked together. Each depthwise separable convolutional module consists of a channel-wise convolutional layer, a layer normalization layer and a GELU activation function to generate the initially encoded frequency band feature tensor. In the ConvNeXt convolutional backbone structure, a cross-frequency domain deformation convolution kernel is set. The sampling position of the cross-frequency domain deformation convolution kernel is obtained by two factors: one is the energy distribution curve in the frequency dimension of the frequency band component tensor, and the other is the frequency deformation mapping function. The sampling position is obtained by multiplying the current frequency index by the scaling factor output by the frequency deformation mapping function and then weighted and shifting. Linear interpolation is used to perform frequency domain convolution sampling at non-integer positions, and the frequency deformation-aware intermediate feature tensor is output. The intermediate feature tensor is input into the neural frequency-modulation layer, which includes a set of trainable sigmoid compression functions. Each sigmoid compression function acts on the frequency amplitude vector of a channel, outputs a modulation factor vector, and performs a frequency-wise dot product operation between the modulation factor vector and the intermediate feature tensor to output the frequency-modulated modulation feature tensor. The modulation feature tensor is input into a channel routing mechanism driven by a brain region connectivity graph. The brain region connectivity graph is a weighted graph structure constructed based on the spatial positional relationship of channels. Channels are graph nodes, and the edge weights between nodes are generated by the physical connection distance and the signal covariance. The channel routing mechanism uses the brain region connectivity graph to generate a routing weight matrix between channels, performs a weighted reconstruction operation on the channel dimension on the modulation feature tensor, and outputs a structural feature tensor that integrates the brain region connectivity relationship. The structural feature tensor is input into the rhythmic temporal residual path mechanism, which is implemented based on the bidirectional gated unit (GRU) structure. It performs sequential modeling in the time dimension and introduces the rhythmic control vector estimated by the principal periodic component of the power spectrum of each segment in the frequency band component tensor. The GRU state transition matrix is ​​applied through an exponential weighting function to dynamically adjust the update amplitude and frequency periodic response of the temporal residual path, and outputs a multidimensional EEG feature tensor containing rhythmic dynamics. The method for performing cross-frequency domain attention fusion on the multidimensional EEG feature tensor involves weighted concatenation of the estimated frequency band coupling strength matrix and channel structure consistency factor to form a fused EEG feature vector, specifically including: The multidimensional EEG feature tensor is split according to the frequency band dimension, and the channel-frequency two-dimensional feature matrix corresponding to each frequency band is extracted. Based on the channel-frequency two-dimensional feature matrix of each frequency band, the mean vector and standard deviation vector of the full channel feature corresponding to each frequency band are calculated respectively. The mean vectors of all frequency bands are combined pairwise, and the frequency band coupling strength matrix is ​​constructed according to the product of the corresponding covariance degree and standard deviation. Based on the frequency band coupling strength matrix, a frequency band attention allocation matrix is ​​generated through an attention mechanism. The integer row vector of the current frequency band in the frequency band coupling strength matrix is ​​extracted as the spectral coupling expression, and the coupling discreteness distance is calculated with the average coupling strength vector of all frequency bands. The coupling discreteness is mapped to the attention score by the inverse normalization strategy. The attention scores of all frequency bands are normalized so that the sum is 1, thus forming the frequency band attention allocation matrix. For the channel-frequency two-dimensional feature matrix in each frequency band, the channel structure consistency factor is calculated based on the brain region connectivity map. The channel-frequency two-dimensional feature matrix of each frequency band and the corresponding structure consistency factor are then subjected to element-wise weighted fusion processing to obtain the channel structure enhancement feature matrix. Based on the attention weights corresponding to each frequency band in the frequency band attention allocation matrix, the channel structure enhancement feature matrices after processing all frequency bands are weighted and superimposed, and average pooling is performed on the frequency dimension to form a fused EEG feature vector by concatenating them according to the channel dimension. The step involves performing a feature feedback reconstruction operation based on the EEG task response results, according to the response consistency index and prediction offset, and jointly updating the sampling offset parameters of the cross-frequency domain deformable convolution kernel. Specifically, this includes: The EEG task response results are compared with historical response labels, the degree of consistency is statistically analyzed, and the response consistency index is calculated. The prediction bias is calculated by comparing the category response probability in the EEG task response results with the historical average response probability of the corresponding category. A stability score for deformation is constructed by jointly using the response consistency index and the prediction offset. If the stability score is lower than a set threshold, a feedback reconstruction process is initiated. Based on the current distribution of the predicted offset in the frequency band, the mapping interval of the frequency deformation mapping function in the corresponding frequency band is fine-tuned using a local interval sampling density adjustment strategy. Based on the channel distribution with low response consistency, the sampling offset parameters of the corresponding channels and frequency bands in the cross-frequency domain deformation convolution kernel are adjusted. The adjustment method is to perform weighted accumulation based on the response offset weight in the most recent batches to generate the offset correction amount and update the dynamic sampling strategy of the convolution kernel. The updated frequency deformation mapping function and sampling offset parameters participate in the next round of EEG signal frequency band reconstruction and convolution processing, forming a feature extraction path with enhanced stability.

2. The automatic feature extraction method for electroencephalogram (EEG) signals based on convolutional neural networks according to claim 1, characterized in that, The process of acquiring raw EEG signal data, identifying artifact segments based on an adaptive artifact suppression method using the channel spectral curvature change rate, and performing dynamic amplitude recalibration processing on each channel to generate an EEG time series tensor specifically includes: The channel spectrum curvature change rate is the channel spectrum curvature change rate formed by differential calculation of the amplitude change trend between adjacent frequency points in the short-time spectrum of a single channel according to continuous time slices. The adaptive artifact suppression method is to statistically analyze the average change level of the channel spectrum curvature change rate in continuous time slices and dynamically update the artifact recognition threshold within the window range. The adaptively updated artifact detection threshold is applied to the channel spectrum curvature change rate sequence in the time dimension. When the channel spectrum curvature change rate of consecutive time slices is higher than the artifact detection threshold, the current consecutive time slice is marked as an artifact segment. When the channel spectrum curvature change rate of consecutive time slices is near the artifact detection threshold and the change amplitude is lower than the preset stable change range, the current consecutive time slice is marked as a stable interval. The original EEG signal data within the stable interval is subjected to dynamic amplitude recalibration processing. The dynamic amplitude recalibration processing involves obtaining the maximum and minimum amplitude of each channel signal within a set time window, using the maximum and minimum amplitudes to form the original amplitude interval, and converting the original amplitude interval into a new amplitude interval using a linear interval mapping method. After artifact identification, stationary interval division and amplitude dynamic recalibration, the signals of each channel are arranged in the order of acquisition channel and sampling time. A multi-dimensional data structure is constructed with time slice as the time dimension and channel number as the channel dimension to form an EEG time series tensor.

3. The automatic feature extraction method for electroencephalogram (EEG) signals based on convolutional neural networks according to claim 1, characterized in that, The process of performing frequency band deformation reconstruction on the EEG time series tensor involves determining the frequency band boundary points based on the frequency band division using a spectral curvature change rate sequence, and generating multiple frequency band component tensors by constructing a frequency deformation mapping function to perform nonlinear mapping on the original spectral interval. Based on the EEG time series tensor, a windowed fast Fourier transform is performed on the signals of each channel within a frequency domain analysis window of a set length to form the power spectral density curve corresponding to each window. Calculate the frequency band division using the spectral curvature change rate sequence on the power spectral density curve; The frequency band division is performed by performing extreme value detection on the spectrum curvature change rate sequence, and the position that meets the change rate peak condition is taken as the frequency band boundary point, and multiple original frequency band intervals are formed by adjacent frequency band boundary points; A frequency deformation mapping function is constructed within the original frequency band interval. The frequency deformation mapping function is a nonlinear continuous function defined on the original frequency interval, including: an exponential nonlinear compression mapping function and a Gaussian reweighted mapping function based on the mean and variance of the power spectrum within the interval. The frequency deformation mapping function is applied to each original frequency band interval, and a nonlinear transformation is performed on the frequency index and frequency domain interpolation resampling is performed to obtain the deformation tensor corresponding to multiple spectra. Channel dimension alignment and amplitude normalization are performed on the deformation tensors of all channels in the same frequency band to form frequency band component tensors. Each frequency band component tensor constitutes multiple frequency band component tensors in the order of frequency bands.

4. The automatic feature extraction method for electroencephalogram (EEG) signals based on convolutional neural networks according to claim 1, characterized in that, The step of performing task inference on the fused EEG feature vectors, and outputting the EEG task response results according to the preset task using category distribution mapping, specifically includes: Perform a linear mapping operation on the fused EEG feature vector to convert it into an intermediate feature representation; The intermediate feature representation is linearly expanded. The linear expansion method is to construct a target vector of a set length based on the total number of categories of the preset task, and then transform it again by constructing a new set of linear weights and bias parameters to generate a category score vector consistent with the number of task categories. Each value in the category score vector is input into the exponential transformation function. The results of all exponential transformation functions are normalized. After the sum is calculated, each item is divided by the sum to obtain the task category response probability distribution vector. Based on the category number corresponding to the maximum probability value in the task category response probability distribution vector, the EEG task response result indicated by the category code is output as the final classification result; If the preset task objective is a multi-class prediction task, then the complete task category response probability distribution vector is output as the EEG task response result. If the task is a binary classification or single-label prediction task, only the category with the highest probability value and its corresponding probability value will be output as the EEG task response result.

Images