Low-signal-to-noise-ratio fragment recognition neural network training method and electroencephalogram signal processing method

Abstract

The low-signal-to-noise-ratio fragment recognition neural network training method comprises the steps that electroencephalogram signals are collected; segmenting the electroencephalogram signals; identifying the signal fragments as'inapplicable 'or'applicable'; performing time domain feature calculation on the signal fragments output in the step 20 to obtain time domain feature parameters; performing low signal-to-noise ratio fragment recognition neural network training by using the time domain characteristic parameters and'inapplicable 'or'applicable'; and obtaining a low signal-to-noise ratio fragment identification neural network model or a feature data set. The electroencephalogram signal processing method comprises the following steps: collecting an electroencephalogram signal, and dividing the electroencephalogram signal into signal fragments with set sizes; a-level signal recognition filtering is carried out, a low-signal-to-noise-ratio fragment recognition neural network obtained through a low-signal-to-noise-ratio fragment recognition neural network training method is adopted, whether the signal-to-noise ratio of a signal fragment is larger than a set threshold value or not is analyzed, and if the signal-to-noise ratio of the signal fragment is larger than the set threshold value, the signal fragment is abandoned; and the B-level signal identification and filtering is to filter the electroencephalogram signals after the A-level signal identification and filtering in the time-frequency domain.

Description

Technical Field

[0001] The present application belongs to the technical field of electroencephalogram (EEG) signal processing and analysis, and particularly relates to a neural network training method for identifying low signal-to-noise ratio (SNR) segments and an EEG signal processing method based thereon. Background Art

[0002] Electroencephalogram (EEG) has been widely used in biomedical engineering due to its high temporal resolution and non-invasive data acquisition method for neuronal electrical activity. Related research typically analyzes EEG signals based on a specific number of channels to accomplish various medical tasks. However, during the acquisition process, EEG signals are often interfered with by factors such as eye movement, muscle movement, and environmental noise, resulting in a low signal-to-noise ratio, which seriously affects the accuracy of subsequent signal analysis and interpretation.

[0003] In actual engineering scenarios, due to considerations of equipment portability, cost control, and ease of use, more and more application scenarios tend to use EEG signals with a small number of channels for analysis. Therefore, how to effectively remove artifacts under the condition of a low number of channels has become a key technical problem in EEG signal processing. At present, Independent Component Analysis (ICA) is a more commonly used multi-channel EEG artifact removal method. However, this method relies on the assumption that the number of artifact sources is less than or equal to the number of channels. Therefore, when the number of channels is less than the number of artifact sources, its noise reduction ability is significantly reduced, and it is difficult to meet the noise reduction needs of EEG signals with a small number of channels. Summary of the Invention

[0004] In this application, the technical solution proposed by the inventor first uses a low signal-to-noise ratio segment recognition neural network training method to train the model and obtain a model and data set for filtering out low signal-to-noise ratio segments; and uses the results of the training for EEG signal processing, first filtering out low signal-to-noise ratio segments of the EEG signal, and then performing Class B signal recognition filtering on this basis; the two-level recognition and filtering, especially the first-level low signal-to-noise ratio segment recognition, eliminates those segments that are particularly unsuitable for subsequent signal processing and recognition, and provides very high-quality EEG signals for subsequent truly useful EEG signal analysis. The technical solution of this application is a basic EEG signal processing method that can provide relatively clean EEG data for the subsequent process.

[0005] The technical solution of the present application to solve the above-mentioned technical problems is a low signal-to-noise ratio segment recognition neural network training method, including step C10: collecting EEG signals; step C20: segmenting the EEG signals to obtain signal slices of a set size; step C30: marking the above-mentioned signal slices as "not applicable" or "applicable"; step C40: performing time domain feature calculation on the signal slices output from step 20 to obtain time domain feature parameters; step C50: using the time domain feature parameters and the "not applicable" or "applicable" to perform low signal-to-noise ratio segment recognition neural network training; obtaining a low signal-to-noise ratio segment recognition neural network model or feature data set.

[0006] Alternatively, the time domain characteristic parameters include any one or more of absolute mean, standard deviation, kurtosis, maximum absolute value, and zero-crossing rate.

[0007] Alternatively, the step C30 includes a step C31 of manually marking the signal fragments as “not applicable” or “applicable”.

[0008] Alternatively, the step C30 includes a step C32, wherein an AI artifact recognition model is used to perform AI recognition analysis on the signal slice, and the AI ​​artifact recognition model outputs an analysis result of "not applicable" or "applicable" for the signal slice.

[0009] Alternatively, the training of the AI ​​artifact recognition model includes the following steps: Step D10: Collecting EEG signals; Step D20: Segmenting the EEG signals to obtain signal fragments of a set size; Step D30: Manually labeling the above-mentioned signal fragments as artifact signals; Step D40: Using the labeled signal fragments to perform AI training; and obtaining an AI artifact recognition model or an AI feature data set.

[0010] Alternatively, the identified artifact signal includes any one or more of “electrooculographic artifact,” “myographic artifact,” “electrocardiographic artifact,” “power frequency interference,” and “physiological interference.”

[0011] The technical solution of the present application to solve the above-mentioned technical problems can also be an EEG signal processing method, including step A10: collecting EEG signals and dividing the EEG signals into signal slices of a set size; step A20: Class A signal recognition and filtering; the Class A signal recognition and filtering includes a low signal-to-noise ratio segment recognition neural network, analyzing whether the signal-to-noise ratio of the signal slice is greater than a set threshold, and if it is greater than the set threshold, discarding the signal slice; step A30: Class B signal recognition and filtering; step A40: subsequent processing of the EEG signal after noise reduction; the low signal-to-noise ratio segment recognition neural network is obtained by the low signal-to-noise ratio segment recognition neural network training method according to any one of claims 1 to 5; the Class B signal recognition and filtering is to filter the EEG signal after the Class A signal recognition and filtering in the time-frequency domain.

[0012] Alternatively, in step A30, the Class B signal identification and filtering includes the following steps: step AH10: denoising through short-time Fourier transform, and then restoring the time domain signal through inverse short-time Fourier transform; step AH20: splicing the time domain signals obtained in step AH10; step AH30: performing time domain noise reduction processing again on the time domain signals obtained by splicing in step AH10.

[0013] Alternatively, step AH10 includes the following steps: step HB10: converting it from the time domain to the time-frequency domain through short-time Fourier transform STFT to obtain a time-frequency spectrum matrix in complex form; step HB20: decomposing the time-frequency spectrum matrix into an amplitude spectrum and a phase spectrum; step HB30: performing a convolution operation on the amplitude spectrum input into a two-dimensional convolution layer to obtain an updated amplitude spectrum; the updated amplitude spectrum is recombined with the original phase information to obtain a denoised complex time-frequency spectrum matrix; step B40: restoring the complex time-frequency spectrum matrix obtained in step B30 to a time domain signal through an inverse short-time Fourier transform ISTFT.

[0014] Alternatively, step AH30 includes the following steps: step HC10: performing three one-dimensional convolution operations on the time domain signal obtained by splicing in step AH10; step HC20: inputting the time domain signal obtained by splicing in step AH10 into a long short-term memory recursive neural network operation; step HC30: fusing and splicing the calculation results of the one-dimensional convolution layer and the long short-term memory network, and after passing through the fully connected layer, outputting the time domain signal sequence after noise reduction processing.

[0015] The technical effects of the above technical solution include: the time domain feature parameters and the "not applicable" or "applicable" mark are jointly used as feature elements for the training input of the low signal-to-noise ratio segment recognition neural network, which improves the accuracy of the low signal-to-noise ratio segment recognition neural network model or feature data set, making subsequent use of it for low signal-to-noise ratio segment recognition more accurate.

[0016] The technical effects of the above technical solution include: time domain characteristic parameters can be selected as needed, providing multiple options.

[0017] The technical effects of the above technical solution include: manual marking as "not applicable" or "applicable" can introduce human expert experience and improve the accuracy of recognition.

[0018] The technical effects of the above technical solution include: marking as "not applicable" or "applicable" based on the AI ​​artifact recognition model, further improving the efficiency of recognition.

[0019] The technical effects of the above technical solution include: training of the AI ​​artifact recognition model provides a more accurate means of AI artifact recognition, which can improve the efficiency and accuracy of artifact recognition.

[0020] The technical effects of the above technical solution include: there are multiple types of artifact signals, which can eliminate the influence of multiple artifacts and improve the accuracy of the model and data set.

[0021] The technical effects of the above technical solution include: Class A signal recognition filtering can filter out low signal-to-noise ratio segments, improve the pertinence of subsequent algorithms, and eliminate the interference of low signal-to-noise ratio segments on subsequent signal processing.

[0022] The technical effects of the above technical solution include: the combination of Class A signal recognition filtering and Class B signal recognition filtering provides a higher-quality EEG signal processing method, which can integrate the advantages of the two types of filtering, making Class B signal recognition filtering more effective, and reducing the computing power consumption and signal interference caused by low signal-to-noise ratio fragments in various links of EEG signal processing.

[0023] The technical effects of the above technical solution include: Class B signal identification and filtering, and effective interference signal filtering in the time and frequency domains.

[0024] The technical effects of the above technical solution include: time domain signal splicing, providing continuous time domain signals to facilitate subsequent continuous processing.

[0025] The technical effects of the above technical solution include: performing one-dimensional convolution layers multiple times in step AH30 can perform specific filtering processing to make the output signal have higher quality.

[0026] The technical effects of the above technical solution include: fusing and splicing the calculation results of the one-dimensional convolutional layer and the long short-term memory network, and outputting a time domain signal sequence after noise reduction processing after passing through the fully connected layer. This can provide a subsequent EEG signal with better noise filtering, reduce the computing power consumption of subsequent algorithms in low signal-to-noise ratio segments and artifact processing, and allow subsequent algorithms to focus more on specific application analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of the EEG signal processing method flow Figure 1 ;

[0028] Figure 2 This is a schematic diagram of the EEG signal processing method flow Figure 2 ;

[0029] Figure 3 This is a schematic diagram of the EEG signal processing method flow Figure 3 ;

[0030] Figure 4 This is a schematic diagram of the training process of the neural network for low signal-to-noise ratio segment recognition. Figure 1 ;

[0031] Figure 5 This is a schematic diagram of the training process of the neural network for low signal-to-noise ratio segment recognition. Figure 2 ;

[0032] Figure 6 This is a schematic diagram of the training process of the neural network for low signal-to-noise ratio segment recognition. Figure 3 ;

[0033] Figure 7 This is a diagram of the training process of the AI ​​artifact recognition model Figure 1 ;

[0034] Figure 8 This is a diagram of the training process of the AI ​​artifact recognition model Figure 2 ;

[0035] Figure 9 This is a schematic diagram of a neural network for identifying low signal-to-noise ratio segments;

[0036] Figure 10 It is a flow chart of Class B signal identification and filtering;

[0037] Figure 11 This is a flow chart of step AH10 in the Class B signal identification and filtering. DETAILED DESCRIPTION

[0038] The contents of this application are further described in detail below in conjunction with the accompanying drawings. It should be noted that the following is a description of the preferred embodiments of the present invention and does not constitute any limitation to the present invention. The description of the preferred embodiments of the present invention is only an illustration of the general principles of the present invention. The numbers such as "first", "second" and "A" and "B" involved in the present invention are only for the convenience of explanation and do not represent the order relationship in time or space. The combination of letters and numbers "TA", "TB" and "H" involved in the present invention are only for the convenience of explanation, and the specific meaning is determined by what they refer to.

[0039] like Figure 4 A method for training a neural network for identifying low signal-to-noise ratio segments includes step C10: collecting electroencephalogram (EEG) signals; step C20: segmenting the EEG signals to obtain signal segments of a set size; step C30: marking the signal segments as "not applicable" or "applicable"; step C40: performing time domain feature calculations on the signal segments output from step 20 to obtain time domain feature parameters; step C50: training a neural network for identifying low signal-to-noise ratio segments using the time domain feature parameters and the "not applicable" or "applicable" value; and obtaining a neural network model or feature data set for identifying low signal-to-noise ratio segments.

[0040] like Figure 4 , a low signal-to-noise ratio segment recognition neural network training method, the time domain feature parameters include any one or more of absolute mean, standard deviation, kurtosis, maximum absolute value, and zero-crossing rate.

[0041] like Figure 6, a low signal-to-noise ratio segment recognition neural network training method, the step C30 includes step C31, manually marking the above signal segment as "not applicable" or "applicable".

[0042] like Figure 5 , a low signal-to-noise ratio segment identification neural network training method, the step C30 includes step C32, using an AI artifact recognition model to perform AI recognition analysis on the signal fragment, and the AI ​​artifact recognition model outputs an analysis result of "not applicable" or "applicable" for the signal fragment.

[0043] like Figure 7 and Figure 8 A low signal-to-noise ratio segment recognition neural network training method, the training of the AI ​​artifact recognition model includes the following steps: step D10: collecting EEG signals; step D20: segmenting the EEG signals to obtain signal slices of a set size; step D30: manually marking the above signal slices as artifact signals; step D40: using the marked signal slices to perform AI training; obtaining an AI artifact recognition model or an AI feature data set.

[0044] like Figure 8 A low signal-to-noise ratio segment recognition neural network training method is disclosed. The identified artifact signals include any one or more of "electrooculographic artifacts", "electromyographic artifacts", "electrocardiographic artifacts", "power frequency interference" and "physiological interference".

[0045] like Figures 1 to 3 A method for processing EEG signals includes step A10: collecting EEG signals and dividing them into signal segments of a set size; step A20: class A signal recognition and filtering; the class A signal recognition and filtering includes a low signal-to-noise ratio segment recognition neural network, analyzing whether the signal-to-noise ratio of the signal segment is greater than a set threshold, and if so, discarding the signal segment; step A30: class B signal recognition and filtering; step A40: subsequent processing of the EEG signal after noise reduction; the low signal-to-noise ratio segment recognition neural network is obtained by adopting a low signal-to-noise ratio segment recognition neural network training method; the class B signal recognition and filtering is to filter the EEG signal after the class A signal recognition and filtering in the time-frequency domain.

[0046] like Figure 3 and Figure 10A method for processing EEG signals, wherein in step A30, the Class B signal recognition and filtering is performed using an EEG noise reduction neural network; the EEG noise reduction neural network is an end-to-end EEG noise reduction neural network. The Class B signal recognition and filtering includes the following steps: Step AH10: De-noising using a short-time Fourier transform, followed by restoring the time domain signal using an inverse short-time Fourier transform; Step AH20: Concatenating the time domain signals obtained in step AH10; and Step AH30: Further performing time domain noise reduction on the time domain signals obtained by concatenating the time domain signals in step AH10.

[0047] like Figure 11 , a method for processing EEG signals, step AH10 includes the following steps: step HB10: short-time Fourier transform STFT converts it from the time domain to the time-frequency domain to obtain a time-frequency spectrum matrix in complex form; step HB20: decomposes the time-frequency spectrum matrix into an amplitude spectrum and a phase spectrum; step HB30: performs a convolution operation on the amplitude spectrum input into a two-dimensional convolution layer to obtain an updated amplitude spectrum; the updated amplitude spectrum is recombined with the original phase information to obtain a denoised complex time-frequency spectrum matrix; step HB40: restores the complex time-frequency spectrum matrix obtained in step B30 to a time domain signal through an inverse short-time Fourier transform ISTFT.

[0048] like Figure 10 A method for processing electroencephalogram (EEG) signals, wherein step AH30 includes the following steps: step HC10: performing three one-dimensional convolution operations on the time domain signal obtained by splicing in step AH10; step HC20: inputting the time domain signal obtained by splicing in step AH10 into a long short-term memory recursive neural network operation; step HC30: fusing and splicing the calculation results of the one-dimensional convolution layer and the long short-term memory network, and outputting a time domain signal sequence after noise reduction processing after passing through a fully connected layer.

[0049] In some embodiments of the present application, the first step is step 1, which is time series signal preprocessing and label generation for EEG noise reduction.

[0050] A batch of raw EEG signals is obtained from the dataset and preprocessed for subsequent artifact labeling, low signal-to-noise ratio segment identification, and noise reduction model training. The collected signals should cover common EEG activity states and common artifact interference to ensure the breadth and representativeness of the data.

[0051] Manual labeling is used to screen signals for interference features such as abnormally increased amplitude and baseline drift. Specifically, the start and end times of each segment with a low signal-to-noise ratio are recorded and a label is assigned to indicate whether the segment is suitable for subsequent processing.

[0052] The target channel data is then extracted from the original multi-channel signal and segmented using a fixed-length sliding window approach to generate a set of EEG time segments of equal length. Windows containing "not applicable" segments are given an overall label of "not applicable," while the rest are marked as "applicable."

[0053] This step ultimately outputs a data set containing EEG signals and their corresponding labels (i.e., whether they are suitable for subsequent processing) for training the "low signal-to-noise ratio segment recognition" model. The time periods marked as "not applicable" are eliminated, and only the time periods that meet the requirements are retained. The retained time periods are spliced ​​in the time domain to generate a continuous and stable sequence of processed signals. The independent component analysis (ICA) algorithm can be used to perform artifact processing on the continuous and stable EEG signals generated in the above steps to construct the "label signal" required for supervised learning. The target channel data is extracted from the pure label signal and the original signal, and divided according to a fixed window length to generate multiple groups of EEG time segments of equal length. This step ultimately outputs the EEG signal and its corresponding "artifact removal reference label", which constitutes the data set for training the "end-to-end EEG artifact removal" model.

[0054] In some embodiments of the present application, step 2: constructing a neural network for identifying low signal-to-noise ratio segments; this step aims to construct a neural network model that can automatically identify low signal-to-noise ratio EEG segments, providing a pre-screening mechanism for subsequent artifact removal.

[0055] The specific process is as follows Figure 4 As shown, the explanation is as follows: First, the time domain features of the multiple EEG segments obtained in step 1 are extracted to quantify the signal quality. The extracted features include: Mean Absolute Value (MAV), Standard Deviation (STD), Kurtosis (Kurtosis), Maximum Absolute Value (MAX_ABS) and Zero Crossing Rate (ZCR). These features reflect the amplitude, volatility, morphological characteristics and frequency changes of the signal and other information, and have strong distinguishing ability. Assume that the input EEG data is a multi-channel time series with a channel number of C and a length of N. The signal value of the i-th sampling point of the c-th channel is recorded as x c,i The calculation methods of the above five eigenvalues ​​are shown in formulas (1) to (5).

[0056]

[0057] in, is the mean value of the signal, l[·] is the indicator function, which takes the value 1 when the condition in the brackets is met, otherwise it takes the value 0. Figure 9 After the time domain feature calculation is completed, each segment is represented as a feature vector with a dimension of C×5. Next, the constructed feature vector is input into a shallow fully connected neural network for classification. The network consists of two fully connected layers and outputs a probability value of "not applicable" (that is, it does not meet the requirements of subsequent processing). The network uses training data with manually annotated labels and optimizes parameters through supervised learning. The loss function is selected as binary cross entropy to measure the deviation between the model's predicted probability and the true label.

[0058] Through this low signal-to-noise ratio segment recognition network, the system can eliminate segments of poor quality in the initial stage of the input signal, avoiding resource waste and misprocessing in the subsequent artifact removal process, and effectively improving the accuracy and stability of the overall processing flow.

[0059] In some embodiments of the present application, step 3: end-to-end EEG denoising network construction; this step aims to build an end-to-end EEG denoising neural network model to automatically perform time-frequency joint modeling and denoising processing on the screened EEG segments to reconstruct high-quality EEG signals.

[0060] The specific process is as follows Figure 10 and Figure 11 As shown, the explanation is as follows: First, for the multi-channel EEG segment obtained in step 1 (3), the short-time Fourier transform (STFT) is used to convert it from the time domain to the time-frequency domain to obtain a complex time-frequency spectrum matrix. The matrix can be further decomposed into an amplitude spectrum and a phase spectrum. Among them, a two-dimensional convolutional layer with residual connection is used to perform preliminary noise reduction learning on the amplitude spectrum while keeping the phase information unchanged. Then, the updated amplitude spectrum is recombined with the original phase information to obtain a complex time-frequency spectrum matrix after preliminary denoising. Subsequently, it is restored to a time domain signal through the inverse short-time Fourier transform (ISTFT).

[0061] like Figure 10 and Figure 11As shown in the figure, after completing the initial denoising, the recovered time-domain signal is used for further refined denoising. Specifically, the signal recovered by the ISFT is concatenated with the original EEG signal in the channel dimension and used as the input for the subsequent network. Next, a one-dimensional convolutional layer with residual connections and a long short-term memory (LSTM) network are used to jointly model the time-domain features. The convolutional layer is used to extract local features, while the LSTM is used to capture global temporal dependencies.

[0062] Ultimately, the model outputs a denoised EEG time series with the same dimensions as the original input. During training, the mean square error (MSE) is used as the loss function to measure the difference between the denoised signal and the ideal target signal and guide the optimization of network parameters.

[0063] The neural network constructed in this step can achieve end-to-end EEG signal artifact removal, taking into account the modeling capabilities of both frequency and time domains, and provide a high-quality input basis for subsequent analysis tasks.

[0064] In step 4 of some embodiments of this application: Through the above steps, an EEG noise reduction system with the ability to identify and automatically clean low signal-to-noise ratio EEG noise is completed. In practical applications, this system can achieve online real-time processing of EEG signals, significantly improving data quality and reducing manual intervention.

[0065] System operation process is as follows Figure 3 As shown in the figure, the system uses a sliding window to receive new EEG segments of length T in real time and feeds them into a low signal-to-noise ratio segment identification neural network. Based on a pretrained classification model, the network outputs a probability value for the current segment being a "low signal-to-noise ratio" signal. If this probability value falls below a set threshold, the segment is considered of acceptable quality and suitable for subsequent artifact removal. Otherwise, the segment is considered unusable and discarded, thus preventing interference with subsequent analysis results.

[0066] All EEG segments identified as suitable are fed into an end-to-end EEG denoising neural network for further processing. This network, based on the STFT and ISTFT transformation mechanisms, combines convolutional layers with an LSTM network structure to complete a two-stage denoising process in both the time-frequency and time domains, ultimately outputting a high-quality denoised EEG signal.

[0067] The system can effectively identify and remove low-quality EEG segments, improve the accuracy and robustness of subsequent analysis, and has good application prospects and scalability.

[0068] The technical solution of this application can effectively improve the efficiency of EEG data processing. By introducing a low signal-to-noise ratio segment identification module in the preprocessing stage, the quality of the input EEG segments can be effectively screened, severely interfering signals can be eliminated at the source, and resource waste and misprocessing in the subsequent artifact removal process can be avoided, thereby improving the reliability and efficiency of the overall signal processing process.

[0069] High-precision artifact removal: The constructed EEG denoising neural network model adopts an end-to-end architecture, integrating a two-stage modeling strategy in the time and time-frequency domains. The introduction of STFT and ISTFT enables the model to fully utilize the time-frequency characteristics of the signal. The combination of convolutional networks and LSTM further enhances the model's ability to model both local and global information, achieving more detailed and precise artifact suppression.

[0070] In some embodiments, training and application are not differentiated, and training models and data sets can be iteratively updated while actually being applied. The technical solution of this application can support real-time processing and has good engineering feasibility: combining the sliding window mechanism and modular neural network design, the entire system architecture is adapted to real-time streaming data processing scenarios, with fast reasoning and low-latency response capabilities, and is suitable for the deployment requirements of edge computing platforms such as wearable devices and portable EEG collectors.

[0071] After completing the initial noise reduction, this application uses the recovered time domain signal to further perform refined denoising. The specific approach is to splice the signal restored by ISTFT with the original EEG signal in the channel dimension as the input of the subsequent network. Then, a one-dimensional convolutional layer with residual connection and a long short-term memory network (Long Short-Term Memory, LSTM) are used to jointly model the time domain features, where the convolutional layer is used to extract local features and the LSTM is used to capture global temporal dependencies.

[0072] Ultimately, the model outputs a denoised EEG time series with the same dimensions as the original input. During training, the mean square error (MSE) is used as the loss function to measure the difference between the denoised signal and the ideal target signal and guide the optimization of network parameters.

[0073] The neural network constructed in this application can achieve end-to-end EEG signal artifact removal, taking into account the modeling capabilities of both frequency domain and time domain, providing a high-quality input basis for subsequent analysis tasks.

[0074] This application overcomes the existing problems of high dependency and low denoising accuracy in EEG signal denoising. It provides a method and system for EEG denoising using a time-frequency domain dual-modal deep neural network. This method can adaptively identify and remove artifacts from EEG signals with a limited number of channels, effectively improving the signal-to-noise ratio and signal quality, thereby enhancing the accuracy and reliability of subsequent tasks such as feature extraction and disease identification.

[0075] Introducing a low-SNR segment identification module: In traditional EEG denoising processes, all segments are typically fed into the denoising network for processing, which can lead to resource waste and incorrect cleaning. In this paper, by extracting time-domain statistical features from EEG segments and determining their SNR levels using a shallow fully connected neural network, unusable segments can be effectively screened out, improving data processing quality from the source.

[0076] A dual-stage denoising strategy integrating frequency and time domains: A structured model combining STFT and ISTFT is proposed: the amplitude spectrum is denoised in the time-frequency domain, the signal is reconstructed while keeping the phase information unchanged, and then further refined through time domain modeling, significantly improving the noise reduction accuracy and signal restoration.

[0077] End-to-end neural network architecture combines convolution and recurrent units for EEG noise reduction: Combining a two-dimensional convolution (for time-frequency domain modeling), a one-dimensional convolution (for time-domain local modeling), and a LSTM (for time-domain global dependency modeling) neural network architecture, it effectively removes EEG artifacts while maintaining input-output consistency.

[0078] The technical solution in the present application can also be a computing and processing device or a detection device, including all or part of the devices for running the above-mentioned method; the memory of the computing and processing device includes the above-mentioned feature data set or model.

[0079] The technical solution in the present application can also be a data storage device that stores all or part of the program code for executing the above method and stores the above data set or model.

[0080] Although the present invention is illustrated and described based on the preferred embodiment and several alternatives, the invention is not limited by the specific description in this specification. Other additional replacement or equivalent components can also be used to practice the present invention.

Claims

1. A neural network training method for identifying low signal-to-noise ratio segments, characterized in that: include Step C10: collecting EEG signals; Step C20: segmenting the EEG signal to obtain signal segments of a set size; Step C30: marking the signal fragments as "not applicable" or "applicable"; Step C40: performing time domain feature calculation on the signal slices outputted in step 20 to obtain time domain feature parameters; Step C50: Using the time domain feature parameters and the “not applicable” or “applicable”, a neural network training is performed to identify low signal-to-noise ratio segments; and a neural network model or feature data set for identifying low signal-to-noise ratio segments is obtained.

2. The low signal-to-noise ratio segment recognition neural network training method according to claim 1, characterized in that: The time domain characteristic parameters include any one or more of absolute mean, standard deviation, kurtosis, maximum absolute value, and zero-crossing rate.

3. The low signal-to-noise ratio segment recognition neural network training method according to claim 1, characterized in that: The step C30 includes a step C31 of manually marking the signal segments as "not applicable" or "applicable".

4. The low signal-to-noise ratio segment recognition neural network training method according to claim 1, characterized in that: The step C30 includes a step C32, wherein an AI artifact recognition model is used to perform AI recognition analysis on the signal slice, and the AI ​​artifact recognition model outputs an analysis result of "not applicable" or "applicable" for the signal slice.

5. The low signal-to-noise ratio segment recognition neural network training method according to claim 4, characterized in that: The training of the AI ​​artifact recognition model includes the following steps: Step D10: collecting EEG signals; Step D20: segmenting the EEG signal to obtain signal segments of a set size; Step D30: manually marking the signal fragments as artifact signals; Step D40: Perform AI training using the identified signal fragments to obtain an AI artifact recognition model or an AI feature data set.

6. The low signal-to-noise ratio segment recognition neural network training method according to claim 5, characterized in that: The identified artifact signals include any one or more of "electrooculographic artifacts", "myographic artifacts", "electrocardiographic artifacts", "power frequency interference", and "physiological interference".

7. A method for processing electroencephalogram signals, characterized in that: include Step A10: collecting EEG signals and dividing the EEG signals into signal slices of a set size; Step A20: Class A signal identification and filtering; the Class A signal identification and filtering includes a low signal-to-noise ratio segment identification neural network, analyzing whether the signal-to-noise ratio of the signal segment is greater than a set threshold, and if so, discarding the signal segment; Step A30: Class B signal identification and filtering; Step A40: subsequent processing of the EEG signal after noise reduction; The low signal-to-noise ratio segment recognition neural network is obtained by using the low signal-to-noise ratio segment recognition neural network training method according to any one of claims 1 to 5; The B-level signal recognition and filtering is to filter the EEG signal after the A-level signal recognition and filtering in the time-frequency domain.

8. The method for processing electroencephalogram signals according to claim 7, wherein: In step A30, the Class B signal identification and filtering includes the following steps: Step AH10: De-noise the signal by performing short-time Fourier transform, and then restore the signal to the time domain by performing inverse short-time Fourier transform. Step AH20: splicing the time domain signals obtained in step AH10; Step AH30: The time domain signals obtained by splicing in step AH10 are subjected to time domain noise reduction processing again.

9. The method for processing EEG signals according to claim 8, wherein: Step AH10 includes the following steps: Step HB10: Short-time Fourier transform (STFT) is used to convert the data from the time domain to the time-frequency domain to obtain a complex time-frequency matrix. Step HB20: decomposing the time-frequency spectrum matrix into amplitude spectrum and phase spectrum; Step HB30: performing a convolution operation on the amplitude spectrum input into the two-dimensional convolution layer to obtain an updated amplitude spectrum; recombining the updated amplitude spectrum with the original phase information to obtain a denoised complex time-frequency spectrum matrix; Step B40: Restore the complex time-frequency spectrum matrix obtained in step B30 to a time domain signal through inverse short-time Fourier transform (ISTFT).

10. The method for processing EEG signals according to claim 8, wherein: Step AH30 includes the following steps: Step HC10: performing a convolution operation on the time domain signal obtained by splicing in step AH10 three times through one-dimensional convolution layers; Step HC20: The time domain signal obtained by the concatenation in step AH10 is input into the long short-term memory recurrent neural network for operation; Step HC30: The calculation results of the one-dimensional convolutional layer and the long short-term memory network are fused and spliced, and after passing through the fully connected layer, the denoised time domain signal sequence is output.

Images