A method and system for artifact removal from electroencephalogram (EEG) signals

Abstract

This application relates to the field of EEG signal processing technology, and provides a method and system for artifact removal from EEG signals, including: acquiring multi-channel EEG signals to be de-artifacted, and constructing a three-dimensional tensor representation of the multi-channel EEG signals; constructing a channel map to characterize the spatial topological relationship of EEG electrodes based on the spatial distribution relationship between the electrodes corresponding to the multi-channel EEG signals; extracting three-dimensional features from the three-dimensional tensor representation using a three-dimensional convolutional neural network; performing cross-channel feature propagation and fusion based on the channel map during the extraction process; obtaining modulation weights for modulating the original complex spectrum based on the three-dimensional features and performing modulation to obtain the temporal representation corresponding to the modulation result; obtaining fused features based on the multi-channel EEG signals, the channel map, and the temporal representation; processing the fused features to obtain the de-artifacted EEG signal. This application can improve the artifact removal effect of EEG signals.

Description

Technical Field

[0001] This application belongs to the field of electroencephalogram (EEG) signal processing technology, and in particular relates to a method and system for removing artifacts from EEG signals. Background Technology

[0002] Electroencephalography (EEG), as an important physiological signal reflecting brain neural activity, is widely used in brain-computer interfaces, cognitive neuroscience, and clinical diagnostic aids. However, during actual acquisition, EEG signals are highly susceptible to various artifacts, including electrooculography (EOG), electromyography (EMG), power line interference, electrode displacement, and environmental noise. These artifacts typically have large amplitudes, strong non-stationarity, and uncertain durations, and often overlap with real EEG signals in the spectrum, severely reducing the reliability of EEG signal analysis and applications. Therefore, effectively suppressing artifacts is a key technical problem in the field of EEG signal processing.

[0003] In existing technologies, earlier EEG artifact removal methods were mainly based on traditional signal processing and model-driven approaches. For example, methods that use regression models to linearly eliminate eye movement artifacts in EEG using reference EEG signals can suppress eye movement interference to some extent. However, these methods rely on linear assumptions and the quality of the reference channel, and their effectiveness is limited when artifact morphology is complex or there is multi-source interference. Independent component analysis (ICA), as a typical blind source separation technique, is widely used in EEG artifact removal. Its core idea is to decompose multi-channel EEG signals into several statistically independent components, and then remove artifact components through manual or rule-based identification, thereby achieving signal cleanup. Although ICA has a certain ability to separate artifacts such as EEG and EMG artifacts under ideal conditions, it relies heavily on the assumption of statistical independence, the component selection process is subjective, and when artifacts and EEG components overlap, it is prone to the accidental deletion of effective EEG information and structural distortion.

[0004] To enhance the ability to handle non-stationary artifacts, some existing techniques incorporate time-frequency analysis methods into EEG artifact removal tasks. For example, wavelet transform is used to decompose the signal into multiple scales, and thresholding strategies or ICA are combined to suppress artifact components. These methods improve the ability to characterize transient artifacts to some extent, but their effectiveness is highly sensitive to threshold settings and parameter selection, and it is difficult to maintain stable performance across different datasets and application scenarios.

[0005] With the development of deep learning technology, more and more data-driven methods have been introduced into the field of EEG artifact removal. Some studies utilize convolutional neural networks, recurrent neural networks, or Transformer structures to learn the mapping relationship between artifact-containing EEGs and clean EEGs in the time or frequency domain, achieving end-to-end artifact suppression. Recent studies have further attempted to extend deep models to multi-channel EEG scenarios; however, these methods typically still treat multi-channel EEGs as simple stacked inputs along the channel dimension, lacking explicit modeling of electrode spatial distribution and topological relationships, making it difficult to guarantee cross-channel structural consistency during artifact removal.

[0006] In addition, some studies have used methods such as embedding spatial decomposition and generative adversarial networks to separate and reconstruct EEG components and artifact components in EEG signals, in order to improve the model's expressive power and artifact removal performance. However, these methods often do not explicitly introduce electrode spatial topological constraints under multi-channel conditions, which may still disrupt cross-channel spatial continuity during artifact removal.

[0007] In summary, although existing EEG artifact removal techniques have been continuously developing in terms of automation and model representation capabilities, they still have significant shortcomings in explicit modeling of electrode space topology, three-dimensional joint modeling of channels-frequency-time, and maintenance of cross-channel consistency, resulting in poor artifact removal performance of EEG signals. Summary of the Invention

[0008] This application provides a method and system for removing artifacts from electroencephalogram (EEG) signals, which can solve the problem of poor artifact removal effect of EEG signals.

[0009] In a first aspect, embodiments of this application provide a method for destigmatizing electroencephalogram (EEG) signals, including: Acquire the multi-channel EEG signal that needs to be de-artifacted, and construct a three-dimensional tensor representation of the multi-channel EEG signal; the multi-channel EEG signal includes EEG signals from multiple channels; Based on the spatial distribution relationship between electrodes corresponding to multi-channel EEG signals, a channel diagram is constructed to characterize the spatial topology relationship of EEG electrodes. A three-dimensional convolutional neural network is used to extract three-dimensional features represented by three-dimensional tensors. During the extraction process, cross-channel feature propagation and fusion are performed on the output of each layer of the three-dimensional convolutional neural network based on the channel graph. Modulation weights for modulating the original complex spectrum are obtained based on three-dimensional features; the original complex spectrum is obtained by performing a short-time Fourier transform on the multi-channel EEG signal. The original complex spectrum is modulated based on the modulation weights, and the inverse short-time Fourier transform is performed on the modulated complex spectrum to obtain the time-domain representation. Based on the multi-channel EEG signal, channel plot and time domain representation, the fusion features are obtained, and the fusion features are processed by the reconstruction mapping function to obtain the artifact-free EEG signal.

[0010] Optionally, a three-dimensional tensor representation of the multi-channel EEG signal is constructed, including: Short-time Fourier transform is performed on the EEG signal of each channel in the multi-channel EEG signal to obtain the time-frequency domain representation of the EEG signal of each channel; By stacking the time-frequency amplitude values ​​in the time-frequency domain representations of all channels along the channel dimension, a three-dimensional tensor representation of the multi-channel EEG signal is obtained. ; , Indicates the number of channels. Indicates the number of frequency points. Indicates the number of time frames.

[0011] Optionally, based on the spatial distribution relationship between the electrodes corresponding to the multi-channel EEG signals, a channel map is constructed to characterize the spatial topological relationship of the EEG electrodes, including: Construct a weighted adjacency matrix ; For weighted adjacency matrix Perform normalization processing, and then normalize the weighted adjacency matrix. As a channel diagram used to characterize the spatial topological relationships of EEG electrodes; Among them, the weighted adjacency matrix The Middle Line number Column elements for: ; Indicates the first The electrode corresponding to the first channel and the first The Euclidean distance between the electrodes corresponding to each channel. , Indicates the first The spatial coordinates of the electrodes corresponding to each channel Indicates the first The spatial coordinates of the electrodes corresponding to each channel The scaling parameter represents the Gaussian kernel function. Indicates the relationship with the first The electrode corresponding to each channel is closest to The set of channel indexes corresponding to each electrode.

[0012] Optionally, cross-channel feature propagation and fusion are performed on the outputs of each layer in the 3D convolutional neural network based on the channel map, including: For the output of each layer in the 3D convolutional neural network, perform the following steps: The output of this layer of the network is obtained through the channel diagram. In Propagation is carried out to obtain spatial propagation results. ; express The Middle Each feature channel at time... Cross-channel feature vectors, , Representing feature dimension, Indicates the number of channels. Indicates the length of time; The following residual fusion formula is used to update the first... Each feature channel at time... Cross-channel feature vectors: ; For any feature channel With time index , , Indicates the updated number Each feature channel at time... Cross-channel feature vectors, This represents the learnable fusion coefficient.

[0013] Optionally, modulation weights for modulating the original complex spectrum are obtained based on three-dimensional features, including: The sigmoid function is used to process the three-dimensional features to obtain the modulation weights.

[0014] Optionally, the modulation weights include the modulation weights of each channel at different time-frequency positions, and the original complex spectrum includes the original complex spectrum of each channel at different time-frequency positions; modulating the original complex spectrum based on the modulation weights includes: The original complex spectrum is modulated using the following formula: ; in, Indicates the first Each channel in time-frequency position The complex spectrum obtained by modulation at the point of origin, Indicates frequency index, Indicates the time frame index. Indicates the first Each channel in time-frequency position Modulation weights at the location, Indicates the first Each channel in time-frequency position The original complex spectrum at the location.

[0015] Optionally, based on multi-channel EEG signals, channel plots, and time-domain representations, fused features are obtained, including: Features of multi-channel EEG signals are extracted and used as single-channel time-domain branch features. The channel plot is used as a multi-channel spatial branching feature; Encode the time-domain representation into three-dimensional time-frequency branch features; The single-channel time-domain branch features, multi-channel spatial branch features, and three-dimensional time-frequency branch features are spliced ​​together to obtain the fused features.

[0016] Secondly, embodiments of this application provide a brainwave signal artifact removal system, comprising: The acquisition module is used to acquire the multi-channel EEG signal that needs to be de-artifacted and to construct a three-dimensional tensor representation of the multi-channel EEG signal; the multi-channel EEG signal includes EEG signals from multiple channels; The module is used to construct a channel diagram that characterizes the spatial topology of EEG electrodes based on the spatial distribution relationship between electrodes corresponding to multi-channel EEG signals. The extraction module is used to extract three-dimensional features represented by three-dimensional tensors using a three-dimensional convolutional neural network; during the extraction of three-dimensional features, cross-channel feature propagation and fusion are performed on the output of each layer of the three-dimensional convolutional neural network based on the channel graph; The processing module is used to obtain modulation weights for modulating the original complex spectrum based on three-dimensional features; the original complex spectrum is obtained by performing a short-time Fourier transform on the multi-channel EEG signal. The modulation module is used to modulate the original complex spectrum based on the modulation weights, and to perform an inverse short-time Fourier transform on the modulated complex spectrum to obtain a time-domain representation. The reconstruction module is used to obtain fused features based on multi-channel EEG signals, channel plots, and time-domain representations, and to process the fused features using a reconstruction mapping function to obtain the artifact-free EEG signal.

[0017] The above-mentioned solution in this application has the following beneficial effects: In this embodiment, a three-dimensional tensor representation of a multi-channel EEG signal is constructed. Based on the spatial distribution relationship between electrodes corresponding to the multi-channel EEG signal, a channel map is constructed to characterize the spatial topological relationship of EEG electrodes. Then, a three-dimensional convolutional neural network is used to extract the three-dimensional features of the three-dimensional tensor representation. During the extraction of the three-dimensional features, cross-channel feature propagation and fusion are performed based on the channel map to enhance cross-channel consistency. Next, modulation weights are obtained based on the three-dimensional features to indicate the suppression intensity of different channels, frequencies, and time positions. Based on the modulation weights, time-frequency modulation and artifact suppression are performed on the original complex spectrum corresponding to the multi-channel EEG signal. Finally, the features of the multi-channel EEG signal, the channel map, and the time-domain representation corresponding to the modulated complex spectrum are combined to obtain the fused features. The fused features are then processed using a reconstruction mapping function to obtain the artifact-free EEG signal. Since the three-dimensional tensor representation can jointly characterize the EEG signal features in the three dimensions of channel, frequency, and time, and explicitly introduce the spatial distribution relationship of electrodes, the artifact removal process can fully perceive the cross-channel structural correlation. Meanwhile, by adaptively modulating different time-frequency and channel positions, the system achieves targeted suppression of complex and irregular artifact components, prioritizing the expressive power of the artifact removal process for artifact-prominent regions, thereby improving the overall artifact removal effect.

[0018] Other beneficial effects of this application will be described in detail in the following detailed description section. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 A flowchart of an EEG signal artifact removal method based on electrode spatial topology guided by an embodiment of this application; Figure 2 The image shows the power spectral density of the artifact-removed signal and the clean signal in an experiment of this application. Figure 3 This is a residual plot of the clean signal and the artifact removal signal in an experiment of this application; Figure 4a This is a time-frequency diagram of the cleaning signal in an experiment of this application; Figure 4b This is a time-frequency diagram of the pollution signal in an experiment of this application; Figure 4c This is a time-frequency diagram of the artifact-removed signal in an experiment of this application; Figure 5 This is a time-domain waveform comparison diagram of the contaminated signal, clean signal, and artifact removal signal in an experiment of this application. Detailed Implementation

[0021] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0022] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0023] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0024] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0025] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0026] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0027] To address the poor artifact removal performance of current EEG signals, this application provides a method for EEG signal artifact removal. This method constructs a three-dimensional tensor representation of a multi-channel EEG signal and, based on the spatial distribution relationship between electrodes corresponding to the multi-channel EEG signal, constructs a channel map to characterize the spatial topological relationship of the EEG electrodes. Then, a three-dimensional convolutional neural network is used to extract three-dimensional features from the three-dimensional tensor representation. During the extraction process, cross-channel feature propagation and fusion are performed based on the channel map to enhance cross-channel consistency. Next, modulation weights indicating the suppression intensity at different channels, frequencies, and time positions are obtained based on these three-dimensional features. Time-frequency modulation and artifact suppression are then applied to the original complex spectrum corresponding to the multi-channel EEG signal based on these modulation weights. Finally, the features of the multi-channel EEG signal, the channel map, and the time-domain representation corresponding to the modulated complex spectrum are combined to obtain fused features. A reconstruction mapping function is used to process the fused features to obtain the artifact-removed EEG signal. The three-dimensional tensor representation can jointly characterize EEG signal features in three dimensions: channel, frequency, and time, and explicitly introduces the spatial distribution relationship of electrodes, thus enabling the artifact removal process to fully perceive cross-channel structural relationships. Meanwhile, by adaptively modulating different time-frequency and channel positions, the system achieves targeted suppression of complex and irregular artifact components, prioritizing the expressive power of the artifact removal process for artifact-prominent regions, thereby improving the overall artifact removal effect.

[0028] The method for removing artifacts from electroencephalogram (EEG) signals provided in this application will be illustrated below with reference to specific embodiments.

[0029] like Figure 1 As shown, the method for removing artifacts from electroencephalogram (EEG) signals provided in this application includes the following steps: Step 11: Obtain the multi-channel EEG signal that needs to be de-artifacted, and construct a three-dimensional tensor representation of the multi-channel EEG signal; the multi-channel EEG signal includes EEG signals from multiple channels.

[0030] The aforementioned multi-channel EEG signals can be EEG signals acquired by EEG acquisition equipment, specifically the EEG signals of the target subject (such as a patient). During the acquisition process, the EEG acquisition equipment has several (e.g., 16, 32) electrodes and reference leads. Typically, the potential of one electrode relative to the reference lead is considered as one channel's EEG signal. Understandably, different channels correspond to different electrodes.

[0031] For example, multi-channel EEG signals can be used. Represented as: ; In the above formula, This indicates the number of EEG electrode channels (i.e., the number of channels). Indicates the number of time sampling points. Indicates the first Each channel is in The potential value at a given time.

[0032] Because EEG signals are easily affected by factors such as baseline drift, amplifier DC bias, and individual differences during acquisition, directly inputting the raw signal into the subsequent modeling network can easily lead to unstable feature distribution, affecting model convergence and artifact removal performance. Therefore, after acquiring the multi-channel EEG signal (also known as the raw multi-channel EEG signal), preprocessing is required. Preprocessing includes removing the DC component of each channel and normalizing the amplitude of each channel. After this preprocessing, a standardized multi-channel EEG time-series signal matrix is ​​obtained. : That is, the three-dimensional tensor representation and the multi-channel EEG signals used later are all preprocessed multi-channel EEG signals (i.e., multi-channel EEG time-series signal matrices). ).

[0033] EEG signals exhibit significant non-stationary characteristics, with their spectral structure dynamically changing over time. Analysis in only the time or frequency domain is insufficient to fully characterize these signal features. Therefore, this application performs a time-frequency transformation on the preprocessed multi-channel EEG signal to characterize its local variations in time and frequency. In this embodiment, the time-frequency transformation is performed on the multi-channel EEG signal by constructing a three-dimensional tensor representation.

[0034] In some embodiments of this application, the specific implementation of constructing a three-dimensional tensor representation of a multi-channel EEG signal is as follows: First, a short-time Fourier transform is performed on the EEG signal of each channel in the multi-channel EEG signal to obtain the time-frequency domain representation of the EEG signal of each channel; then, the time-frequency amplitude values ​​in the time-frequency domain representations of all channels are stacked along the channel dimension to obtain the three-dimensional tensor representation of the multi-channel EEG signal. . , Indicates the number of channels. Indicates the number of frequency points. This represents the number of time frames. The short-time Fourier transform is performed to map the one-dimensional time-series signal to a time-frequency domain representation, thereby obtaining the spectral amplitude information of each channel at different time frames and frequency positions.

[0035] The aforementioned three-dimensional tensor representation is used to jointly characterize the features of multi-channel EEG signals across three dimensions: channel, frequency, and time. It is worth noting that the three-dimensional tensor representation can place the three coupled dimensions of a multi-channel EEG signal (i.e., channel, frequency, and time) in the same coordinate system, allowing the model to simultaneously learn the joint spatial (channel)-spectrum-temporal patterns.

[0036] Step 12: Based on the spatial distribution relationship between the electrodes corresponding to the multi-channel EEG signals, construct a channel diagram to characterize the spatial topology relationship of the EEG electrodes.

[0037] It should be noted that the electrodes corresponding to each EEG channel have a clear spatial distribution relationship on the scalp surface, and the EEG information of adjacent electrodes generally has a strong correlation. Ignoring this spatial structure relationship can easily disrupt cross-channel consistency during artifact removal. Therefore, in this application, each EEG channel is considered as a node in a graph structure, and the graph structure of the channels is constructed based on the spatial distribution relationship between the electrodes. Specifically, the channel graph characterizing the spatial topological relationship of EEG electrodes can be constructed through the following steps: Step 12.1, construct the weighted adjacency matrix .

[0038] Among them, the weighted adjacency matrix The Middle Line number Column elements for: ; In the above formula, Indicates the first The electrode corresponding to the first channel and the first The Euclidean distance between the electrodes corresponding to each channel. , Indicates the first The spatial coordinates of the electrodes corresponding to each channel Indicates the first The spatial coordinates of the electrodes corresponding to each channel can be considered as the electrode's coordinates in the world coordinate system. For Gaussian kernel function, The scaling parameter (i.e. bandwidth parameter) of the Gaussian kernel function. Indicates the relationship with the first The electrode corresponding to each channel is closest to The set of channel indices corresponding to each electrode, else represents other.

[0039] Step 12.2, for the weighted adjacency matrix Perform normalization processing, and then normalize the weighted adjacency matrix. As a channel diagram used to characterize the spatial topological relationships of EEG electrodes, this channel diagram is used to constrain the subsequent cross-channel feature propagation process.

[0040] Step 13: Extract the three-dimensional features represented by the three-dimensional tensor using a three-dimensional convolutional neural network; during the extraction of the three-dimensional features, perform cross-channel feature propagation and fusion of the output of each layer of the three-dimensional convolutional neural network based on the channel graph.

[0041] In some embodiments of this application, the specific implementation of cross-channel feature propagation and fusion of the output of each layer in a three-dimensional convolutional neural network based on the channel graph is as follows: For the output of each layer in the 3D convolutional neural network, perform steps one and two respectively: Step 1: Analyze the output of this layer of the network using the channel diagram. In Propagation is carried out to obtain spatial propagation results. . express The Middle Each feature channel at time... Cross-channel feature vectors, , Representing feature dimension, Indicates the number of channels. Indicates the length of time. , It should be noted that the set of cross-channel feature vectors for each feature channel at all time points is the output of that layer of the network. .

[0042] Step two, update the first residual using the following residual fusion formula. Each feature channel at time... Cross-channel feature vectors: .

[0043] For any feature channel With time index , , Indicates the updated number Each feature channel at time... The cross-channel feature vector, denoted as the first... Each feature channel at time... The updated cross-channel feature vectors are then processed. It should be noted that the set of updated cross-channel feature vectors for each feature channel at all time points is the final output of this layer of the network. . Represents the learnable fusion coefficient. Used to control the proportion of neighborhood information in the output. The larger the value, the more neighborhood information it integrates; The smaller the value, the closer the output is to the original channel features. Through this mechanism, the model can achieve an adaptive balance between preserving the differences in channel features and enhancing cross-channel consistency.

[0044] Understandably, the final output of the last layer of the network is the three-dimensional feature in step 13.

[0045] It is worth mentioning that 3D convolutional neural networks can perform convolution operations simultaneously in three dimensions: channel, frequency, and time. This allows for joint modeling of the spectral structure characteristics, temporal evolution, and cross-channel correlations of multi-channel EEG signals within a unified space. Through multi-layer 3D convolution and nonlinear mapping, 3D feature representations at different scales are extracted step by step, providing high-level semantic features for subsequent artifact suppression and signal reconstruction. Simultaneously, during the 3D feature extraction process, explicit electrode spatial topological constraints are introduced, and features are propagated and fused across channels using the channel graph structure. This ensures that feature updates follow the spatial topological relationships of EEG electrodes, thereby enhancing the connections between channels.

[0046] Step 14: Obtain the modulation weights for modulating the original complex spectrum based on the three-dimensional features; the original complex spectrum is obtained by performing a short-time Fourier transform on the multi-channel EEG signal.

[0047] The aforementioned raw complex spectrum includes the raw complex spectrum of each channel at different time-frequency positions.

[0048] In some embodiments of this application, the sigmoid function can be used to process the three-dimensional features to obtain modulation weights. Specifically, the three-dimensional features can be input into the sigmoid function for calculation to obtain a modulation weight that indicates the suppression strength at different channels, frequencies, and time positions. That is, the modulation weight output by the sigmoid function includes the modulation weight of each channel at different time-frequency positions, and this weight is used for adaptive modulation of the signal in the time-frequency-channel space.

[0049] Step 15: Modulate the original complex spectrum based on the modulation weights, and perform an inverse short-time Fourier transform on the modulated complex spectrum to obtain the time-domain representation.

[0050] In some embodiments of this application, the original complex spectrum can be modulated using the following formula: ; in, Indicates the first Each channel in time-frequency position The complex spectrum obtained by modulation at the point of origin, Indicates frequency index, Indicates the time frame index. Indicates the first Each channel in time-frequency position Modulation weights at the location, Indicates the first Each channel in time-frequency position The original complex spectrum at the location.

[0051] It is worth mentioning that element-wise weighted representation of the original complex spectrum is effective for the th... Each channel at frequency Time frame The components at the location are adaptively suppressed / preserved. By modulating in this way, it is possible to achieve adaptive suppression of complex and irregular artifact components without relying on manual thresholds or explicit artifact annotations.

[0052] Step 16: Based on the multi-channel EEG signal, channel diagram and time domain representation, obtain the fusion features, and use the reconstruction mapping function to process the fusion features to obtain the artifact-free EEG signal.

[0053] In some embodiments of this application, the specific process for obtaining the above-mentioned fusion features includes the following steps 16.1 to 16.4: Step 16.1: Extract the features of the multi-channel EEG signal and use the features of the multi-channel EEG signal as the single-channel time-domain branch features.

[0054] In some alternative embodiments, features of multi-channel EEG signals can be extracted using a convolutional neural network.

[0055] Step 16.2: Use the channel map as a multi-channel spatial branch feature.

[0056] Step 16.3: Encode the time-domain representation into three-dimensional time-frequency branch features.

[0057] In some alternative embodiments, a convolutional encoder can be used to encode the time-domain representation to obtain three-dimensional time-frequency branch features.

[0058] Step 16.4: The single-channel time-domain branch features, multi-channel spatial branch features, and three-dimensional time-frequency branch features are spliced ​​together to obtain the fused features.

[0059] In some embodiments of this application, single-channel temporal branch features, multi-channel spatial branch features, and three-dimensional time-frequency branch features can be aligned into the same tensor form, and then channel dimension concatenation (Concat) can be performed to obtain fused features. Finally, a reconstruction mapping function is used to regress the fused features into a temporal EEG signal, which is the artifact-free EEG signal, that is, the artifact-free EEG signal corresponding to the multi-channel EEG signal in step 11.

[0060] The essence of the above reconstruction mapping function is a series of sequentially connected convolutional layers. Specifically, the fused features can be input into this reconstruction mapping function for processing to obtain the EEG signal after artifact removal.

[0061] It is worth mentioning that after completing time-frequency modulation and artifact suppression based on three-dimensional features, the multi-channel spatial branch features, single-channel temporal branch features, and three-dimensional time-frequency branch features are fused, stitched together in the feature channel dimension, and input into the reconstruction module. By comprehensively utilizing cross-channel spatial consistency, temporal morphological information, and structural information after time-frequency suppression, the accuracy of the reconstructed signal can be effectively improved.

[0062] It should be noted that the data processing in steps 12 to 16 can be regarded as an EEG artifact removal model. In order to improve the artifact removal effect, the model can be trained before the actual artifact removal is performed so that the parameters of each function and machine learning module in steps 12 to 16 reach the optimal state.

[0063] Specifically, during the training phase, commonly used supervised learning methods can be employed for end-to-end training of the multi-channel EEG artifact removal model. Specifically, the model takes the artifact-laden EEG signal as input and the corresponding clean EEG signal as the target output, optimizing the model parameters by minimizing the reconstruction error.

[0064] During training, let the artifact-free EEG temporal reconstruction result output by the model be... The corresponding clean EEG signal is Then the loss function Defined as: ; In the above formula, Indicates the number of channels. Indicates the number of time sampling points. Indicates the channel index. Indicates the time index.

[0065] By jointly optimizing the parameters of the 3D convolutional neural network, the channel graph structure propagation module, and the reconstruction module through backpropagation, the model can learn the artifact suppression rules in the time, frequency, and spatial dimensions in a holistic sense. During the model inference phase, only the multi-channel EEG signal containing artifacts needs to be input to output the corresponding artifact-free EEG signal representation.

[0066] The following is an exemplary description of the EEG signal artifact removal method provided in the embodiments of this application, with reference to specific experiments.

[0067] In the experiment, the aforementioned EEG signal artifact removal method was used to remove artifacts from EEG signals. The results showed that this method can effectively suppress complex artifact components while maintaining the consistency of the time and frequency domain structures of the EEG signals during artifact removal. Specific experimental results are as follows: Figure 2 The power spectral density (PSD) of the signal before and after artifact removal is shown. As can be seen from the figure, the artifact removal method successfully reduces noise in the signal, especially in the high-frequency range. The PSD of the artifact-removed signal (i.e., the denoised signal) is similar to that of the clean signal (i.e., the clean signal), demonstrating that the method can effectively remove noise components in the frequency domain.

[0068] In time-domain analysis, Figure 3 The residuals between the clean signal and the artifact-removed signal are shown. As can be seen from the figure, the difference between the artifact-removed signal and the clean signal is very small, with the residuals approaching zero. This indicates that the artifact removal method successfully removed artifacts while preserving as many characteristics of the original signal as possible, verifying the effectiveness of the method in the time domain.

[0069] Figure 4 shows the time-frequency plots of the clean signal, the contaminated signal, and the artifact-removed signal. As can be seen from these three figures, the time-frequency distribution of the contaminated signal is relatively chaotic, with irregular frequency component distribution; while the time-frequency distribution of the artifact-removed signal is more stable and smooth, especially in the high-frequency range, where the artifact-removed signal significantly reduces noise fluctuations and exhibits a smoother spectrum. This indicates that the artifact removal method not only optimizes the frequency characteristics of the signal but also achieves significant results in removing spectral noise, thus improving the overall signal quality.

[0070] Figure 5 The waveforms of the contaminated, clean, and artifact-removed signals in the time domain are compared. The artifact-removed signal is very close to the clean signal in the time domain, while the contaminated signal exhibits larger fluctuations, demonstrating that the artifact removal method can effectively restore the true waveform of the signal.

[0071] Furthermore, in comparative experiments, the method of this application was compared with the existing ASTI-Net artifact removal model, and the performance results of the model were quantitatively evaluated using relative root mean square error (RRMSE), correlation coefficient (CC), and signal-to-noise ratio (SNR). The specific comparison of the metrics is shown in Table 1. Table 1

[0072] As can be seen, the method in this application is significantly better than the comparative method (ASTI-Net) in terms of artifact removal. This application also shows superior performance in terms of correlation coefficient, relative root mean square error, and signal-to-noise ratio.

[0073] The above experimental results demonstrate the feasibility of the method proposed in this application and show its significant technical advantages over existing technologies.

[0074] Overall, the EEG signal artifact removal method of this application provides a multi-channel EEG artifact removal paradigm of "spatial topological constraints + three-dimensional time-frequency joint modeling," which can solve key defects in existing methods such as insufficient utilization of electrode spatial structure, inadequate characterization of time-frequency dynamic features, and easy disruption of cross-channel consistency, achieving more accurate suppression of complex and irregular artifacts. Specifically, the EEG signal artifact removal method of this application addresses the problems in the prior art from the following three aspects: Explicit Modeling of Electrode Spatial Structure: Overcoming the shortcomings of existing methods that generally treat multi-channel EEG as "channel-independent" or simply splice the input, this paper proposes a method that explicitly models EEG channels as graph structure nodes and constructs the topological relationship between channels based on the distribution of electrode space. Furthermore, it introduces cross-channel graph structure constraints and information propagation mechanisms into the network, so that feature updates follow the electrode space adjacency relationship, thereby enhancing cross-channel consistency.

[0075] Three-dimensional time-frequency joint characterization and modeling: Overcoming the problem that it is difficult to characterize the non-stationary characteristics of EEG by modeling only in the time domain or frequency domain, multi-channel EEG signals are mapped to a three-dimensional representation space of channel-frequency-time frame, and a three-dimensional joint feature modeling method is adopted to simultaneously capture cross-channel coupling relationship, spectral structure characteristics and time evolution law in a unified space, providing a more expressive basis for the discrimination and suppression of complex and irregular artifacts.

[0076] Adaptive artifact suppression and high-fidelity reconstruction: Based on three-dimensional joint features, an adaptive modulation mechanism oriented towards channel-frequency-time position is constructed to achieve continuous value suppression of artifact components rather than coarse-grained, indiscriminate processing; and through multi-branch feature fusion and reconstruction mapping, while comprehensively utilizing spatial consistency, temporal morphology and time-frequency structure information, a more stable artifact-free temporal reconstruction result is obtained.

[0077] In summary, the EEG signal artifact removal method of this application can jointly characterize EEG signal features in three dimensions: channel, frequency, and time, and explicitly introduce the spatial distribution relationship of electrodes, enabling the model to fully perceive cross-channel structural correlations during artifact removal. Based on this, through adaptive modulation of different time-frequency and channel locations, targeted suppression of complex and irregular artifact components is achieved, thereby prioritizing the model's expressive power for artifact-prominent regions and improving the overall artifact removal effect.

[0078] This application also provides an EEG signal artifact removal system, which includes an acquisition module, a construction module, an extraction module, a processing module, a modulation module, and a reconstruction module.

[0079] The acquisition module is used to acquire the multi-channel EEG signal that needs to be de-artifacted and to construct a three-dimensional tensor representation of the multi-channel EEG signal; the multi-channel EEG signal includes EEG signals from multiple channels.

[0080] The module is used to construct a channel diagram that characterizes the spatial topology of EEG electrodes based on the spatial distribution relationship between the electrodes corresponding to multi-channel EEG signals.

[0081] The extraction module is used to extract three-dimensional features represented by three-dimensional tensors using a three-dimensional convolutional neural network. During the extraction of three-dimensional features, cross-channel feature propagation and fusion are performed on the output of each layer of the three-dimensional convolutional neural network based on the channel graph.

[0082] The processing module is used to obtain modulation weights for modulating the original complex spectrum based on three-dimensional features; the original complex spectrum is obtained by performing a short-time Fourier transform on the multi-channel EEG signal.

[0083] The modulation module is used to modulate the original complex spectrum based on the modulation weights, and to perform an inverse short-time Fourier transform on the modulated complex spectrum to obtain a time-domain representation.

[0084] The reconstruction module is used to obtain fused features based on multi-channel EEG signals, channel plots, and time-domain representations, and to process the fused features using a reconstruction mapping function to obtain the artifact-free EEG signal.

[0085] In the embodiments of this application, the EEG signal artifact removal system is a system corresponding to the above-mentioned EEG signal artifact removal method, which can solve the problem of poor artifact removal effect of EEG signals.

[0086] It should be noted that the information interaction and execution process between the above modules / units are based on the same concept as the method embodiments of this application. For details on their specific functions and technical effects, please refer to the method embodiments section, and they will not be repeated here.

[0087] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0088] The above description is the preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for removing artifacts from electroencephalogram (EEG) signals, characterized in that, include: Acquire the multi-channel EEG signal that needs to be de-artifacted, and construct a three-dimensional tensor representation of the multi-channel EEG signal; the multi-channel EEG signal includes EEG signals from multiple channels; Based on the spatial distribution relationship between the electrodes corresponding to the multi-channel EEG signals, a channel diagram is constructed to characterize the spatial topology relationship of the EEG electrodes; The three-dimensional features represented by the three-dimensional tensor are extracted using a three-dimensional convolutional neural network; In the process of extracting three-dimensional features, cross-channel feature propagation and fusion are performed on the output of each layer of the three-dimensional convolutional neural network based on the channel map; Modulation weights for modulating the original complex spectrum are obtained based on the three-dimensional features; the original complex spectrum is obtained by performing a short-time Fourier transform on the multi-channel EEG signal. The original complex spectrum is modulated based on the modulation weights, and the modulated complex spectrum is subjected to inverse short-time Fourier transform to obtain a time-domain representation. Based on the multi-channel EEG signal, the channel diagram, and the time-domain representation, a fusion feature is obtained, and the fusion feature is processed using a reconstruction mapping function to obtain an artifact-free EEG signal.

2. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 1, characterized in that, The construction of the three-dimensional tensor representation of the multi-channel EEG signal includes: A short-time Fourier transform is performed on the EEG signal of each channel in the multi-channel EEG signal to obtain the time-frequency domain representation of the EEG signal of each channel. By stacking the time-frequency amplitude values ​​in the time-frequency domain representations of all channels along the channel dimension, a three-dimensional tensor representation of the multi-channel EEG signal is obtained. ; , Indicates the number of channels. Indicates the number of frequency points. Indicates the number of time frames.

3. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 1, characterized in that, The step of constructing a channel map to characterize the spatial topology of EEG electrodes based on the spatial distribution relationship between the electrodes corresponding to the multi-channel EEG signals includes: Construct a weighted adjacency matrix ; For weighted adjacency matrix Perform normalization processing, and then normalize the weighted adjacency matrix. As a channel diagram used to characterize the spatial topological relationships of EEG electrodes; Among them, the weighted adjacency matrix The Middle Line number Column elements for: ； Indicates the first The electrode corresponding to the first channel and the first The Euclidean distance between the electrodes corresponding to each channel. , Indicates the first The spatial coordinates of the electrodes corresponding to each channel Indicates the first The spatial coordinates of the electrodes corresponding to each channel The scaling parameter represents the Gaussian kernel function. Indicates the relationship with the first The electrode corresponding to each channel is closest to The set of channel indexes corresponding to each electrode.

4. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 3, characterized in that, Based on the channel map, cross-channel feature propagation and fusion are performed on the outputs of each layer in the three-dimensional convolutional neural network, including: For the output of each layer in the three-dimensional convolutional neural network, perform the following steps: The output of this layer network is obtained through the channel diagram. In Propagation is carried out to obtain spatial propagation results. ; express The Middle Each feature channel at time... Cross-channel feature vectors, , Representing feature dimension, Indicates the number of channels. Indicates the length of time; The following residual fusion formula is used to update the first... Each feature channel at time... Cross-channel feature vectors: ; For any feature channel With time index , , Indicates the updated number Each feature channel at time... Cross-channel feature vectors, This represents the learnable fusion coefficient.

5. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 1, characterized in that, The step of obtaining modulation weights for modulating the original complex spectrum based on the three-dimensional features includes: The 3D features are processed using the sigmoid function to obtain modulation weights.

6. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 1, characterized in that, The modulation weights include the modulation weights of each channel at different time-frequency positions, and the original complex spectrum includes the original complex spectrum of each channel at different time-frequency positions; the modulation of the original complex spectrum based on the modulation weights includes: The original complex spectrum is modulated using the following formula: ； in, Indicates the first Each channel in time-frequency position The complex spectrum obtained by modulation at the point of origin, Indicates frequency index, Indicates the time frame index. Indicates the first Each channel in time-frequency position Modulation weights at the location, Indicates the first Each channel in time-frequency position The original complex spectrum at the location.

7. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 1, characterized in that, The fusion features obtained based on the multi-channel EEG signal, the channel diagram, and the time-domain representation include: Extract the features of the multi-channel EEG signal and use the features of the multi-channel EEG signal as single-channel time-domain branch features; The channel diagram is used as a multi-channel spatial branch feature; The time-domain representation is encoded into a three-dimensional time-frequency branch feature; The single-channel time-domain branch feature, the multi-channel spatial branch feature, and the three-dimensional time-frequency branch feature are spliced ​​together to obtain the fused feature.

8. A brainwave signal artifact removal system, characterized in that, include: The acquisition module is used to acquire the multi-channel EEG signal that needs to be de-artifacted and to construct a three-dimensional tensor representation of the multi-channel EEG signal; the multi-channel EEG signal includes EEG signals from multiple channels. The construction module is used to construct a channel diagram to characterize the spatial topology of EEG electrodes based on the spatial distribution relationship between the electrodes corresponding to the multi-channel EEG signals. An extraction module is used to extract the three-dimensional features represented by the three-dimensional tensor using a three-dimensional convolutional neural network; In the process of extracting three-dimensional features, cross-channel feature propagation and fusion are performed on the output of each layer of the three-dimensional convolutional neural network based on the channel map; The processing module is used to obtain modulation weights for modulating the original complex spectrum based on the three-dimensional features; the original complex spectrum is obtained by performing a short-time Fourier transform on the multi-channel EEG signal. The modulation module is used to modulate the original complex spectrum based on the modulation weights, and to perform an inverse short-time Fourier transform on the modulated complex spectrum to obtain a time-domain representation. The reconstruction module is used to obtain fused features based on the multi-channel EEG signal, the channel diagram and the time-domain representation, and to process the fused features using a reconstruction mapping function to obtain the artifact-free EEG signal.

Images