An electroencephalogram signal acquisition monitoring system

Abstract

This invention relates to electroencephalogram (EEG) signal acquisition, specifically to an EEG signal acquisition and monitoring system. A mobile terminal receives EEG signals, fNIRS hemodynamic signals, and IMU motion data from a signal acquisition device. It employs a physical information neural network method with an adaptive sampling strategy to identify BCG vascular pulsation artifacts in the EEG signal using fNIRS hemodynamic signals. Then, based on the fNIRS hemodynamic signals and IMU motion data, a multiple linear regression model for motion artifacts is established to identify motion artifact segments in the EEG signal. The detected artifact segments are interpolated and repaired to eliminate artifacts in the EEG signal. The EEG is then plotted and displayed in real time based on the artifact-free EEG signal. The technical solution provided by this invention effectively overcomes the shortcomings of effectively eliminating BCG vascular pulsation artifacts and motion artifacts in EEG signals.

Description

Technical Field

[0001] This invention relates to electroencephalogram (EEG) signal acquisition, and more specifically to an EEG signal acquisition and monitoring system. Background Technology

[0002] Electroencephalogram (EEG) signals, as important bioelectrical signals, reflect the neural electrical activity of the brain and have wide applications in neuroscience research, clinical diagnosis, and brain-computer interfaces. However, during the acquisition of EEG signals, they are often affected by various artifacts. These artifacts may originate from the user's physiological activities (such as vascular pulsation) or electromagnetic interference from the external environment, thus seriously affecting the quality of EEG signals and the accuracy of subsequent analysis.

[0003] Ballistocardiogram (BCG) is a common artifact in EEG signals, primarily caused by minute head movements due to heartbeats, introducing periodic interference into the EEG signal. The frequency range of BCG artifacts is similar to certain EEG rhythms (such as alpha and beta waves), making them difficult to remove effectively with simple filtering methods. Furthermore, motion artifacts are also a significant source of interference in EEG signals, especially when the user directly moves their head or engages in body movements that cause changes in the contact resistance between the EEG electrodes and the scalp, resulting in signal distortion.

[0004] To eliminate artifacts in EEG signals, traditional methods include filtering, independent component analysis (ICA), and regression analysis. However, these methods often have limitations. For example, filtering may simultaneously remove valuable EEG components, ICA requires strict assumptions about the independence of artifacts from EEG components, and regression analysis relies on accurate artifact reference signals. In recent years, with the continuous development of machine learning and deep learning technologies, neural network-based methods have shown great potential in EEG artifact removal. However, existing neural network methods still face problems such as low sampling efficiency and insufficient model generalization ability when dealing with complex and variable EEG artifacts. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] In view of the above-mentioned shortcomings of the existing technology, the present invention provides an electroencephalogram (EEG) signal acquisition and monitoring system that can effectively overcome the defects of the existing technology in that it is difficult to effectively eliminate BCG vascular pulsation artifacts and motion artifacts in EEG signals.

[0007] (II) Technical Solution

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] An electroencephalogram (EEG) signal acquisition and monitoring system includes a signal acquisition device and a mobile terminal;

[0010] The signal acquisition device collects EEG electroencephalogram (EEG) signals, fNIRS hemodynamic signals, and IMU motion data, and uploads them to a mobile terminal.

[0011] The mobile terminal receives EEG signals, fNIRS hemodynamic signals, and IMU motion data sent by the signal acquisition device. It uses a physical information neural network method with an adaptive sampling strategy to identify BCG pulsation artifact segments in the EEG signals using fNIRS hemodynamic signals. Then, it establishes a multiple linear regression model for motion artifacts based on fNIRS hemodynamic signals and IMU motion data to identify motion artifact segments in the EEG signals. The detected artifact segments are then interpolated and repaired to eliminate artifacts in the EEG signals. The EEG is then plotted and displayed in real time based on the artifact-free EEG signals.

[0012] Among them, the physical information neural network method with adaptive sampling strategy changes the positions of trial functions and basis functions, so that the network can automatically adjust the sampling position during training, focus on important regions, optimize model performance by dynamically adjusting the sampling strategy, and improve the accuracy and efficiency of signal processing by using an adaptive mechanism.

[0013] Preferably, the signal acquisition device includes the following components:

[0014] EEG brain signal acquisition electrodes use 16-channel electrodes to acquire EEG brain signals from the user's brain.

[0015] fNIRS hemodynamic signal acquisition electrode: 8-channel electrode to acquire fNIRS hemodynamic signals in the user's brain;

[0016] The IMU motion data acquisition module collects IMU motion data from the user's head.

[0017] The preprocessing module preprocesses the EEG electroencephalogram (EEG) signals, fNIRS hemodynamic signals, and IMU motion data, respectively.

[0018] The Bluetooth transmission module uploads pre-processed EEG brain signals, fNIRS hemodynamic signals, and IMU motion data to the mobile terminal via serial Bluetooth, achieving wireless transmission.

[0019] The power supply module uses an external battery and supports continuous recording for 48 hours.

[0020] Preferably, the preprocessing module preprocesses the EEG electroencephalogram (EEG) signals, fNIRS hemodynamic signals, and IMU motion data, respectively, including:

[0021] For EEG signals, a high-precision AD converter is used to convert analog EEG signals into digital EEG signals, and an FIR filter is used for bandpass filtering to remove high-frequency noise and low-frequency drift.

[0022] For the fNIRS hemodynamic signal, optical density conversion is performed and bandpass filtering is performed using an FIR filter to extract hemodynamic fluctuations;

[0023] For IMU motion data, Kalman filtering is applied to smooth the user's head movement trajectory.

[0024] Preferably, the mobile terminal includes the following components:

[0025] The Bluetooth control module, based on the Android API, enables connection and interaction with the signal acquisition device.

[0026] The data splicing and storage module uses a Java native interface combined with C++ to achieve efficient data splicing and storage to an SD card;

[0027] The BCG vascular pulsation artifact recognition module uses a physical information neural network method with an adaptive sampling strategy to identify BCG vascular pulsation artifact segments in EEG signals using fNIRS hemodynamic signals.

[0028] The motion artifact recognition module establishes a multiple linear regression model for motion artifacts based on fNIRS hemodynamic signals and IMU motion data to identify motion artifact segments in EEG signals.

[0029] The interpolation and repair module performs interpolation and repair on the detected artifact segments to eliminate artifacts in the EEG brain signals.

[0030] The graphics drawing and display module draws and displays the EEG in real time based on the EEG signals after artifact removal.

[0031] Preferably, the BCG vascular pulsation artifact recognition module employs a physical information neural network method with an adaptive sampling strategy to identify BCG vascular pulsation artifact segments in EEG signals using fNIRS hemodynamic signals, including:

[0032] S11. In the time-space domain of the fNIRS hemodynamic signal, initial sampling points are randomly or uniformly selected. These sampling points correspond to the time and space locations in the signal and are used for subsequent network training.

[0033] S12. Construct a physical information neural network. The network input is the location information of the sampling point, including time and spatial coordinates. The network output is the predicted signal strength at the sampling point location.

[0034] The S13, BCG vascular pulsation artifact is caused by systemic hemodynamic changes due to cardiac pulsation. It is modeled as a physical equation, and the residual of the physical equation is defined.

[0035] S14. In order to make the physical information neural network pay more attention to important regions with large residuals containing BCG pulsation artifacts, an adaptive weight function is introduced to allocate weights according to the residual size at the sampling point. The adaptive mechanism is used to improve the accuracy and efficiency of signal processing.

[0036] S15. Define a loss function that includes residual terms, data matching terms, and boundary condition terms, and consider an adaptive weight function to optimize the parameters of the physical information neural network by minimizing the loss function.

[0037] S16. During network training, the sampling point positions are dynamically adjusted according to the residual size. Regions with an absolute residual value greater than 0.05 are considered regions with large residuals, and the number of sampling points is increased for these regions. Regions with an absolute residual value not greater than 0.05 are considered regions with small residuals, and the number of sampling points is reduced for these regions. By changing the positions of the trial function and basis function, the network automatically adjusts the sampling position during training, focusing on important regions with large residuals containing BCG vascular pulsation artifacts. By dynamically adjusting the sampling strategy, the model performance is optimized, and the adaptive mechanism is further utilized to improve the accuracy and efficiency of signal processing.

[0038] S17. Input the new fNIRS hemodynamic signal into the trained physical information neural network, identify whether there is a BCG vascular pulsation artifact segment in the fNIRS hemodynamic signal based on the network output, and determine the BCG vascular pulsation artifact segment in the EEG signal.

[0039] Preferably, the residuals of the physical equations defined in S13 include:

[0040] The residual of the physical equation is defined as the difference between the predicted and expected signal strength values ​​output by the physical information neural network at the sampling point location:

[0041] ;

[0042] Where R(x) is the residual of the physical equation at sampling point x. This represents the predicted signal strength output by the physical information neural network at sampling point x. F(x) represents the parameters of the physical information neural network, and F(x) represents the expected value or value of the physical equation at the sampling point x under known conditions. In practical applications, this value is defined or estimated as needed.

[0043] In S14, to make the physical information neural network pay more attention to important regions with large residuals containing BCG vascular pulsation artifacts, an adaptive weight function is introduced to allocate weights based on the magnitude of the residuals at the sampling points, including:

[0044] The adaptive weight function is expressed as follows:

[0045] ;

[0046] in, The weight at sampling point x, The squared 2-norm of the residual R(x) of the physical equation at sampling point x. Parameters used to control the rate of weight decay;

[0047] S15 defines a loss function that includes residual terms, data matching terms, and boundary condition terms, and considers an adaptive weight function. The parameters of the physical information neural network are optimized by minimizing this loss function, including:

[0048] The loss function is expressed as follows:

[0049] ;

[0050] Where L is the loss function, For sampling point x i The weight of the position, The physical equation at sampling point x i The residual R(x) at the location i The square of the 2-norm of ), where N is the number of sampling points;

[0051] For physical information neural networks at data matching point x j The predicted signal strength at ' ', y j For data matching point x j 'The corresponding true value, where M is the number of data matching points,' This indicates the calculation of the 2-norm square. The weights of the data matching items;

[0052] For physical information neural networks at boundary condition point x k The predicted signal strength at '', b k For boundary condition point x k The corresponding boundary value, where K is the number of boundary condition points. The weights of the boundary condition terms;

[0053] In S17, the new fNIRS hemodynamic signal is input into the trained physical information neural network. Based on the network output, the presence of BCG vascular pulsation artifacts in the fNIRS hemodynamic signal is identified, and the BCG vascular pulsation artifacts in the EEG signal are determined, including:

[0054] The new fNIRS hemodynamic signal is input into the trained physical information neural network. If there is a large deviation between the network output and the expected signal, and this deviation is consistent with the characteristics of BCG vascular pulsation artifact, then the fNIRS hemodynamic signal in the corresponding time period is identified as a BCG vascular pulsation artifact segment, and the EEG signal in the same time period is used as a BCG vascular pulsation artifact segment.

[0055] Among them, the 95th percentile of the residual of the physical information neural network is regarded as a large deviation. Fourier transform or wavelet analysis is performed on the residual signal to calculate its power spectral density. If the energy in the 0.5~3Hz frequency band accounts for more than 50% of the total energy and the energy in other frequency bands is evenly distributed, it is consistent with the characteristics of BCG vascular pulsation artifact.

[0056] Preferably, the physical information neural network includes an input layer, a hidden layer, and an output layer. An activation function is used to introduce nonlinearity in the hidden layer. The parameters of the physical information neural network... This represents the weights and biases of the input layer, hidden layer, and output layer.

[0057] Preferably, the motion artifact recognition module establishes a multiple linear regression model for motion artifacts based on fNIRS hemodynamic signals and IMU motion data to identify motion artifact segments in EEG signals, including:

[0058] S21. Extract the features of fNIRS hemodynamic signals and IMU motion data, and linearly combine the features of fNIRS hemodynamic signals and IMU motion data to model a multiple linear regression model of motion artifacts.

[0059] S22. Estimate the regression coefficients by minimizing the sum of squared residuals;

[0060] S23. Use regression coefficients to calculate motion artifacts in order to identify motion artifact segments in EEG signals.

[0061] Preferably, in S21, features of fNIRS hemodynamic signals and IMU motion data are extracted, and the features of fNIRS hemodynamic signals and IMU motion data are linearly combined to model a multiple linear regression model of motion artifacts, including:

[0062] For fNIRS hemodynamic signals, the mean, slope, and power spectral density of HbO2 oxyhemoglobin and HbR deoxyhemoglobin signals were extracted.

[0063] For IMU motion data, extract the absolute value, variance, and frequency domain features of acceleration and angular velocity;

[0064] The features of fNIRS hemodynamic signals and IMU motion data are linearly combined to model a multiple linear regression model of motion artifacts:

[0065] ;

[0066] Where Artifact(t) is the multiple linear regression model of motion artifacts, X m (t) represents the m-th feature. For the m-th feature X m The regression coefficients of (t), where P is the number of features. For the intercept term, Here, t represents the residual term, and t represents time.

[0067] S22 estimates the regression coefficients by minimizing the sum of squared residuals, including:

[0068] The regression coefficients are estimated using the following formula:

[0069] ;

[0070] in, For the regression coefficient vector, EEG(t) represents the EEG brain signal, and Q represents the number of time points.

[0071] S23 uses regression coefficients to calculate motion artifacts to identify motion artifact segments in EEG signals, including:

[0072] ;

[0073] in, For motion artifacts, For the m-th feature X m The estimated values ​​of the regression coefficients of (t), This is an estimate of the intercept term.

[0074] Preferably, the interpolation repair module performs interpolation repair on the detected artifact segments to eliminate artifacts in the EEG signal, including:

[0075] S31. Mark the start and end times of the BCG vascular pulsation artifact and motion artifact segments in the EEG signal;

[0076] S32. Select n clean sampling points at each end of the BCG vascular pulsation artifact segment and the motion artifact segment, and construct a cubic spline function:

[0077] ;

[0078] Where f(t) is a cubic spline function, , , The first The start and end times of each sub-interval, a l b l c l d l For the first The spline coefficients of each subinterval are determined by solving a system of linear equations to ensure that the first and second derivatives of the cubic spline function f(t) are continuous at the sampling points;

[0079] S33. Determine Gaussian white noise with a noise level equivalent to that of EEG brainwave signals:

[0080] ;

[0081] in, Represents Gaussian white noise Follows a mean of 0 and a variance of Gaussian distribution, Gaussian white noise The standard deviation is estimated by calculating the noise level of non-artifact segments in the EEG signal;

[0082] S34. Replace the BCG vascular pulsation artifact segment and motion artifact segment in the EEG signal with cubic spline interpolation and Gaussian white noise signal:

[0083] ;

[0084] Among them, EEG new (t) represents the EEG signal after artifact removal, and t1 and t2 are the start and end times of the BCG pulsation artifact segment or motion artifact segment in the EEG signal, respectively.

[0085] (III) Beneficial Effects

[0086] Compared with the prior art, the EEG signal acquisition and monitoring system provided by the present invention has the following beneficial effects:

[0087] 1) Improve the quality of EEG brainwave signals

[0088] a. Accurately identify BCG vascular pulsation artifacts and motion artifacts:

[0089] BCG vascular pulsation artifact identification: By utilizing the correlation between fNIRS hemodynamic signals and EEG signals, and through a physical information neural network method with an adaptive sampling strategy, BCG vascular pulsation artifact segments in EEG signals can be accurately identified. This method overcomes the limitation of traditional methods that may remove valuable EEG components while removing BCG vascular pulsation artifacts.

[0090] Motion artifact recognition: Based on fNIRS hemodynamic signals and IMU motion data, a multiple linear regression model for motion artifacts is established, which can accurately identify motion artifact segments in EEG signals caused by head movements. This multimodal data fusion method effectively improves the accuracy and robustness of motion artifact recognition.

[0091] b. Artifact Repair:

[0092] Interpolation repair of detected artifacts effectively eliminated artifacts in EEG signals, significantly improved signal quality, and provided a reliable data foundation for subsequent EEG analysis.

[0093] 2) Improve the accuracy and efficiency of signal processing.

[0094] The physical information neural network method with adaptive sampling strategy automatically adjusts the sampling position during training by changing the positions of trial functions and basis functions, focusing on important regions. This adaptive mechanism can dynamically adjust the sampling strategy, optimize model performance, and improve the accuracy of signal processing. At the same time, the adaptive sampling strategy reduces unnecessary computation and improves the efficiency of signal processing, enabling the system to process large amounts of data in real time and meet the needs of real-time monitoring.

[0095] 3) Enhance system usability and user experience

[0096] a. Real-time rendering and display of electroencephalograms:

[0097] Based on the EEG signals after artifact removal, the system can generate and display EEG in real time, providing doctors and researchers with an intuitive display of brain activity, facilitating timely diagnosis and analysis.

[0098] b. Multimodal data fusion:

[0099] By combining data from three modalities—EEG, fNIRS, and IMU—the system can more comprehensively reflect the physiological state and motor activity of the brain, improving its accuracy and practicality.

[0100] 4) Promote the development of EEG signal processing technology

[0101] a. Innovative methods:

[0102] The proposed physical information neural network method with adaptive sampling strategy provides a new technical approach for the field of EEG signal processing and helps to promote technological progress in related fields.

[0103] b. Broad application prospects:

[0104] The technical solution of this application is not only applicable to clinical diagnosis and neuroscience research, but can also be applied to multiple fields such as brain-computer interfaces and rehabilitation therapy, and has broad application prospects. Attached Figure Description

[0105] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0106] Figure 1 This is a schematic diagram of the system of the present invention;

[0107] Figure 2 This is a schematic diagram of the process by which the BCG vascular pulsation artifact recognition module in this invention identifies BCG vascular pulsation artifact segments in EEG signals.

[0108] Figure 3 This is a schematic diagram of the process by which the motion artifact recognition module in this invention identifies motion artifact segments in EEG signals.

[0109] Figure 4 This is a schematic diagram of the process by which the interpolation repair module in this invention eliminates artifacts in EEG signals;

[0110] Figure 5 This is a schematic diagram of the signal acquisition device in this invention. Detailed Implementation

[0111] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0112] A brainwave signal acquisition and monitoring system, such as Figure 1 As shown, it includes a signal acquisition device and a mobile terminal;

[0113] The signal acquisition device collects EEG electroencephalogram (EEG) signals, fNIRS hemodynamic signals, and IMU motion data, and uploads them to a mobile terminal.

[0114] like Figure 1 and Figure 5 As shown, the signal acquisition device includes the following components:

[0115] EEG brain signal acquisition electrodes use 16-channel electrodes to acquire EEG brain signals from the user's brain.

[0116] fNIRS hemodynamic signal acquisition electrode: 8-channel electrode to acquire fNIRS hemodynamic signals in the user's brain;

[0117] The IMU motion data acquisition module collects IMU motion data from the user's head.

[0118] The preprocessing module preprocesses the EEG electroencephalogram (EEG) signals, fNIRS hemodynamic signals, and IMU motion data, respectively.

[0119] The Bluetooth transmission module uploads pre-processed EEG brain signals, fNIRS hemodynamic signals, and IMU motion data to the mobile terminal via serial Bluetooth, achieving wireless transmission.

[0120] The power supply module uses an external battery and supports continuous recording for 48 hours.

[0121] Specifically, the preprocessing module preprocesses the EEG electroencephalogram (EEG) signals, fNIRS hemodynamic signals, and IMU motion data, including:

[0122] For EEG signals, a high-precision AD converter (using a 24-bit ADS1298 chip) is used to convert analog EEG signals into digital EEG signals, and an FIR filter is used for bandpass filtering to remove high-frequency noise and low-frequency drift.

[0123] For the fNIRS hemodynamic signal, optical density conversion is performed and bandpass filtering is performed using an FIR filter to extract hemodynamic fluctuations;

[0124] For IMU motion data, Kalman filtering is applied to smooth the user's head movement trajectory.

[0125] The mobile terminal receives EEG signals, fNIRS hemodynamic signals, and IMU motion data from the signal acquisition device. It employs a physical information neural network method with an adaptive sampling strategy to identify BCG vascular pulsation artifacts in the EEG signals using fNIRS hemodynamic signals. Then, based on the fNIRS hemodynamic signals and IMU motion data, it establishes a multiple linear regression model for motion artifacts to identify these artifacts. The detected artifacts are then interpolated and repaired to eliminate artifacts in the EEG signals. Finally, the EEG is plotted and displayed in real time based on the artifact-free EEG signals.

[0126] Among them, the physical information neural network method with adaptive sampling strategy changes the positions of trial functions and basis functions, so that the network can automatically adjust the sampling position during training, focus on important regions, optimize model performance by dynamically adjusting the sampling strategy, and improve the accuracy and efficiency of signal processing by using an adaptive mechanism.

[0127] like Figure 1 As shown, the mobile terminal includes the following components:

[0128] The Bluetooth control module, based on Android APIs (such as BluetoothAdapter and BluetoothSocket), enables connection and interaction with the signal acquisition device.

[0129] The data splicing and storage module uses Java native interfaces (such as JNI) combined with C++ to achieve efficient data splicing and storage to the SD card;

[0130] The BCG vascular pulsation artifact recognition module uses a physical information neural network method with an adaptive sampling strategy to identify BCG vascular pulsation artifact segments in EEG signals using fNIRS hemodynamic signals.

[0131] The motion artifact recognition module establishes a multiple linear regression model for motion artifacts based on fNIRS hemodynamic signals and IMU motion data to identify motion artifact segments in EEG signals.

[0132] The interpolation and repair module performs interpolation and repair on the detected artifact segments to eliminate artifacts in the EEG brain signals.

[0133] The graphics drawing and display module draws and displays the EEG in real time based on the EEG signals after artifact removal.

[0134] ① The BCG vascular pulsation artifact recognition module employs an adaptive sampling strategy and a physical information neural network method to identify BCG vascular pulsation artifact segments in EEG signals using fNIRS hemodynamic signals, such as... Figure 2 As shown, it includes:

[0135] S11. In the time-space domain of the fNIRS hemodynamic signal, initial sampling points are randomly or uniformly selected. These sampling points correspond to the time and space locations in the signal and are used for subsequent network training.

[0136] S12. Construct a physical information neural network. The network input is the location information of the sampling point, including time and spatial coordinates. The network output is the predicted signal strength at the sampling point location.

[0137] The S13, BCG vascular pulsation artifact is caused by systemic hemodynamic changes due to cardiac pulsation. It is modeled as a physical equation, and the residual of the physical equation is defined.

[0138] S14. In order to make the physical information neural network pay more attention to important regions with large residuals containing BCG pulsation artifacts, an adaptive weight function is introduced to allocate weights according to the residual size at the sampling point. The adaptive mechanism is used to improve the accuracy and efficiency of signal processing.

[0139] S15. Define a loss function that includes residual terms, data matching terms, and boundary condition terms, and consider an adaptive weight function to optimize the parameters of the physical information neural network by minimizing the loss function.

[0140] S16. During network training, the sampling point positions are dynamically adjusted according to the residual size. Regions with an absolute residual value greater than 0.05 are considered regions with large residuals, and the number of sampling points is increased for these regions. Regions with an absolute residual value not greater than 0.05 are considered regions with small residuals, and the number of sampling points is reduced for these regions. By changing the positions of the trial function and basis function, the network automatically adjusts the sampling position during training, focusing on important regions with large residuals containing BCG vascular pulsation artifacts. By dynamically adjusting the sampling strategy, the model performance is optimized, and the adaptive mechanism is further utilized to improve the accuracy and efficiency of signal processing.

[0141] S17. Input the new fNIRS hemodynamic signal into the trained physical information neural network, identify whether there is a BCG vascular pulsation artifact segment in the fNIRS hemodynamic signal based on the network output, and determine the BCG vascular pulsation artifact segment in the EEG signal.

[0142] Specifically, S13 defines the residuals of the physical equations, including:

[0143] The residual of the physical equation is defined as the difference between the predicted and expected signal strength values ​​output by the physical information neural network at the sampling point location:

[0144] ;

[0145] Where R(x) is the residual of the physical equation at sampling point x. This represents the predicted signal strength output by the physical information neural network at sampling point x. F(x) represents the parameters of the physical information neural network, and F(x) represents the expected value or value of the physical equation at the sampling point x under known conditions. In practical applications, this value is defined or estimated as needed.

[0146] In S14, to make the physical information neural network pay more attention to important regions with large residuals containing BCG vascular pulsation artifacts, an adaptive weight function is introduced to allocate weights based on the magnitude of the residuals at the sampling points, including:

[0147] The adaptive weight function is expressed as follows:

[0148] ;

[0149] in, The weight at sampling point x, The squared 2-norm of the residual R(x) of the physical equation at sampling point x. Parameters used to control the rate of weight decay;

[0150] S15 defines a loss function that includes residual terms, data matching terms, and boundary condition terms, and considers an adaptive weight function. The parameters of the physical information neural network are optimized by minimizing this loss function, including:

[0151] The loss function is expressed as follows:

[0152] ;

[0153] Where L is the loss function, For sampling point x i The weight of the position, The physical equation at sampling point x i The residual R(x) at the location i The square of the 2-norm of ), where N is the number of sampling points;

[0154] For physical information neural networks at data matching point x j The predicted signal strength at ' ', y j For data matching point x j 'The corresponding true value, where M is the number of data matching points,' This indicates the calculation of the 2-norm square. The weights of the data matching items;

[0155] For physical information neural networks at boundary condition point x k The predicted signal strength at '', b k For boundary condition point x k The corresponding boundary value, where K is the number of boundary condition points. The weights of the boundary condition terms;

[0156] In S17, the new fNIRS hemodynamic signal is input into the trained physical information neural network. Based on the network output, the presence of BCG vascular pulsation artifacts in the fNIRS hemodynamic signal is identified, and the BCG vascular pulsation artifacts in the EEG signal are determined, including:

[0157] The new fNIRS hemodynamic signal is input into the trained physical information neural network. If there is a large deviation between the network output and the expected signal, and this deviation is consistent with the characteristics of BCG vascular pulsation artifact, then the fNIRS hemodynamic signal in the corresponding time period is identified as a BCG vascular pulsation artifact segment, and the EEG signal in the same time period is used as a BCG vascular pulsation artifact segment.

[0158] Among them, the 95th percentile of the residual of the physical information neural network is regarded as a large deviation. Fourier transform or wavelet analysis is performed on the residual signal to calculate its power spectral density. If the energy in the 0.5~3Hz frequency band accounts for more than 50% of the total energy and the energy in other frequency bands is evenly distributed, it is consistent with the characteristics of BCG vascular pulsation artifact.

[0159] In this application's technical solution, the physical information neural network includes an input layer, a hidden layer, and an output layer. The hidden layer uses an activation function to introduce nonlinearity. The parameters of the physical information neural network... This represents the weights and biases of the input layer, hidden layer, and output layer.

[0160] The motion artifact recognition module establishes a multiple linear regression model for motion artifacts based on fNIRS hemodynamic signals and IMU motion data to identify motion artifact segments in EEG signals, such as... Figure 3 As shown, it includes:

[0161] S21. Extract the features of fNIRS hemodynamic signals and IMU motion data, and linearly combine the features of fNIRS hemodynamic signals and IMU motion data to model a multiple linear regression model of motion artifacts.

[0162] S22. Estimate the regression coefficients by minimizing the sum of squared residuals;

[0163] S23. Use regression coefficients to calculate motion artifacts in order to identify motion artifact segments in EEG signals.

[0164] Specifically, in S21, features of fNIRS hemodynamic signals and IMU motion data are extracted, and these features are linearly combined to model a multiple linear regression model of motion artifacts, including:

[0165] For fNIRS hemodynamic signals, the mean, slope, and power spectral density of HbO2 oxyhemoglobin and HbR deoxyhemoglobin signals were extracted.

[0166] For IMU motion data, extract the absolute value, variance, and frequency domain features of acceleration and angular velocity;

[0167] The features of fNIRS hemodynamic signals and IMU motion data are linearly combined to model a multiple linear regression model of motion artifacts:

[0168] ;

[0169] Where Artifact(t) is the multiple linear regression model of motion artifacts, X m (t) represents the m-th feature. For the m-th feature X m The regression coefficients of (t), where P is the number of features. For the intercept term, Here, t represents the residual term, and t represents time.

[0170] S22 estimates the regression coefficients by minimizing the sum of squared residuals, including:

[0171] The regression coefficients are estimated using the following formula:

[0172] ;

[0173] in, For the regression coefficient vector, EEG(t) represents the EEG brain signal, and Q represents the number of time points.

[0174] S23 uses regression coefficients to calculate motion artifacts to identify motion artifact segments in EEG signals, including:

[0175] ;

[0176] in, For motion artifacts, For the m-th feature X m The estimated values ​​of the regression coefficients of (t), This is an estimate of the intercept term.

[0177] The interpolation and repair module performs interpolation and repair on detected artifact segments to eliminate artifacts in EEG signals, such as... Figure 4 As shown, it includes:

[0178] S31. Mark the start and end times of the BCG vascular pulsation artifact and motion artifact segments in the EEG signal;

[0179] S32. Select n clean sampling points at each end of the BCG vascular pulsation artifact segment and the motion artifact segment, and construct a cubic spline function:

[0180] ;

[0181] Where f(t) is a cubic spline function, , , The first The start and end times of each sub-interval, a l b l c l d l For the first The spline coefficients of each subinterval are determined by solving a system of linear equations to ensure that the first and second derivatives of the cubic spline function f(t) are continuous at the sampling points;

[0182] S33. Determine Gaussian white noise with a noise level equivalent to that of EEG brainwave signals:

[0183] ;

[0184] in, Represents Gaussian white noise Follows a mean of 0 and a variance of Gaussian distribution, Gaussian white noise The standard deviation is estimated by calculating the noise level of non-artifact segments in the EEG signal;

[0185] S34. Replace the BCG vascular pulsation artifact segment and motion artifact segment in the EEG signal with cubic spline interpolation and Gaussian white noise signal:

[0186] ;

[0187] Among them, EEG new (t) represents the EEG signal after artifact removal, and t1 and t2 are the start and end times of the BCG pulsation artifact segment or motion artifact segment in the EEG signal, respectively.

[0188] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A brainwave signal acquisition and monitoring system, characterized in that: Including signal acquisition devices and mobile terminals; The signal acquisition device collects EEG electroencephalogram (EEG) signals, fNIRS hemodynamic signals, and IMU motion data, and uploads them to a mobile terminal. The mobile terminal receives EEG signals, fNIRS hemodynamic signals, and IMU motion data from the signal acquisition device. It employs a physical information neural network method with an adaptive sampling strategy to identify BCG vascular pulsation artifacts in the EEG signals using fNIRS hemodynamic signals. Then, based on the fNIRS hemodynamic signals and IMU motion data, it establishes a multiple linear regression model for motion artifacts to identify these artifacts. The detected artifacts are then interpolated and repaired to eliminate artifacts in the EEG signals. Finally, the EEG is plotted and displayed in real time based on the artifact-free EEG signals.

2. The EEG signal acquisition and monitoring system according to claim 1, characterized in that: The signal acquisition device includes the following components: EEG brain signal acquisition electrodes use 16-channel electrodes to acquire EEG brain signals from the user's brain. fNIRS hemodynamic signal acquisition electrode: 8-channel electrode to acquire fNIRS hemodynamic signals in the user's brain; The IMU motion data acquisition module collects IMU motion data from the user's head. The preprocessing module preprocesses the EEG electroencephalogram (EEG) signals, fNIRS hemodynamic signals, and IMU motion data, respectively. The Bluetooth transmission module uploads pre-processed EEG brain signals, fNIRS hemodynamic signals, and IMU motion data to the mobile terminal via serial Bluetooth, achieving wireless transmission. The power supply module uses an external battery and supports continuous recording for 48 hours.

3. The EEG signal acquisition and monitoring system according to claim 2, characterized in that: The preprocessing module preprocesses the EEG electroencephalogram (EEG) signals, fNIRS hemodynamic signals, and IMU motion data, including: For EEG signals, a high-precision AD converter is used to convert analog EEG signals into digital EEG signals, and an FIR filter is used for bandpass filtering to remove high-frequency noise and low-frequency drift. For the fNIRS hemodynamic signal, optical density conversion is performed and bandpass filtering is performed using an FIR filter to extract hemodynamic fluctuations; For IMU motion data, Kalman filtering is applied to smooth the user's head movement trajectory.

4. The EEG signal acquisition and monitoring system according to claim 2, characterized in that: The mobile terminal includes the following components: The Bluetooth control module, based on the Android API, enables connection and interaction with the signal acquisition device. The data splicing and storage module uses a Java native interface combined with C++ to achieve efficient data splicing and storage to an SD card; The BCG vascular pulsation artifact recognition module uses a physical information neural network method with an adaptive sampling strategy to identify BCG vascular pulsation artifact segments in EEG signals using fNIRS hemodynamic signals. The motion artifact recognition module establishes a multiple linear regression model for motion artifacts based on fNIRS hemodynamic signals and IMU motion data to identify motion artifact segments in EEG signals. The interpolation and repair module performs interpolation and repair on the detected artifact segments to eliminate artifacts in the EEG brain signals. The graphics drawing and display module draws and displays the EEG in real time based on the EEG signals after artifact removal.

5. The EEG signal acquisition and monitoring system according to claim 4, characterized in that: The BCG vascular pulsation artifact recognition module employs an adaptive sampling strategy and a physical information neural network method to identify BCG vascular pulsation artifact segments in EEG signals using fNIRS hemodynamic signals, including: S11. In the time-space domain of the fNIRS hemodynamic signal, initial sampling points are randomly or uniformly selected. These sampling points correspond to the time and space locations in the signal and are used for subsequent network training. S12. Construct a physical information neural network. The network input is the location information of the sampling point, including time and spatial coordinates. The network output is the predicted signal strength at the sampling point location. The S13, BCG vascular pulsation artifact is caused by systemic hemodynamic changes due to cardiac pulsation. It is modeled as a physical equation, and the residual of the physical equation is defined. S14. In order to make the physical information neural network pay more attention to important regions with large residuals containing BCG pulsation artifacts, an adaptive weight function is introduced to allocate weights according to the residual size at the sampling point. The adaptive mechanism is used to improve the accuracy and efficiency of signal processing. S15. Define a loss function that includes residual terms, data matching terms, and boundary condition terms, and consider an adaptive weight function to optimize the parameters of the physical information neural network by minimizing the loss function. S16. During network training, the position of sampling points is dynamically adjusted according to the residual size. Regions with an absolute residual value greater than 0.05 are considered regions with large residuals, and the number of sampling points is increased for these regions. Regions with an absolute residual value not greater than 0.05 are considered regions with small residuals, and the number of sampling points is reduced for these regions. S17. Input the new fNIRS hemodynamic signal into the trained physical information neural network, identify whether there is a BCG vascular pulsation artifact segment in the fNIRS hemodynamic signal based on the network output, and determine the BCG vascular pulsation artifact segment in the EEG signal.

6. The EEG signal acquisition and monitoring system according to claim 5, characterized in that: The residuals of the physical equations defined in S13 include: The residual of the physical equation is defined as the difference between the predicted and expected signal strength values ​​output by the physical information neural network at the sampling point location: ； Where R(x) is the residual of the physical equation at sampling point x. This represents the predicted signal strength output by the physical information neural network at sampling point x. F(x) represents the parameters of the physical information neural network, and F(x) represents the expected value or value of the physical equation at the sampling point x under known conditions. In practical applications, this value is defined or estimated as needed. In S14, to make the physical information neural network pay more attention to important regions with large residuals containing BCG vascular pulsation artifacts, an adaptive weight function is introduced to allocate weights based on the magnitude of the residuals at the sampling points, including: The adaptive weight function is expressed as follows: ； in, The weight at sampling point x, The squared 2-norm of the residual R(x) of the physical equation at sampling point x. Parameters used to control the rate of weight decay; S15 defines a loss function that includes residual terms, data matching terms, and boundary condition terms, and considers an adaptive weight function. The parameters of the physical information neural network are optimized by minimizing this loss function, including: The loss function is expressed as follows: ； Where L is the loss function, For sampling point x i The weight of the position, The physical equation at sampling point x i The residual R(x) at the location i The square of the 2-norm of ), where N is the number of sampling points; For physical information neural networks at data matching point x j The predicted signal strength at ' ', y j For data matching point x j 'The corresponding true value, where M is the number of data matching points,' This indicates the calculation of the 2-norm square. The weights of the data matching items; For physical information neural networks at boundary condition point x k The predicted signal strength at position '', b k For boundary condition point x k The corresponding boundary value, where K is the number of boundary condition points. The weights of the boundary condition terms; In S17, the new fNIRS hemodynamic signal is input into the trained physical information neural network. Based on the network output, the presence of BCG vascular pulsation artifacts in the fNIRS hemodynamic signal is identified, and the BCG vascular pulsation artifacts in the EEG signal are determined, including: The new fNIRS hemodynamic signal is input into the trained physical information neural network. If there is a large deviation between the network output and the expected signal, and this deviation is consistent with the characteristics of BCG vascular pulsation artifact, then the fNIRS hemodynamic signal in the corresponding time period is identified as a BCG vascular pulsation artifact segment, and the EEG signal in the same time period is used as a BCG vascular pulsation artifact segment. Among them, the 95th percentile of the residual of the physical information neural network is regarded as a large deviation. Fourier transform or wavelet analysis is performed on the residual signal to calculate its power spectral density. If the energy in the 0.5~3Hz frequency band accounts for more than 50% of the total energy and the energy in other frequency bands is evenly distributed, it is consistent with the characteristics of BCG vascular pulsation artifact.

7. The EEG signal acquisition and monitoring system according to claim 5 or 6, characterized in that: The physical information neural network includes an input layer, a hidden layer, and an output layer. An activation function is used in the hidden layer to introduce nonlinearity. The parameters of the physical information neural network... This represents the weights and biases of the input layer, hidden layer, and output layer.

8. The EEG signal acquisition and monitoring system according to claim 5, characterized in that: The motion artifact recognition module establishes a multiple linear regression model for motion artifacts based on fNIRS hemodynamic signals and IMU motion data to identify motion artifact segments in EEG signals, including: S21. Extract the features of fNIRS hemodynamic signals and IMU motion data, and linearly combine the features of fNIRS hemodynamic signals and IMU motion data to model a multiple linear regression model of motion artifacts. S22. Estimate the regression coefficients by minimizing the sum of squared residuals; S23. Use regression coefficients to calculate motion artifacts in order to identify motion artifact segments in EEG signals.

9. The EEG signal acquisition and monitoring system according to claim 8, characterized in that: S21 extracts features from fNIRS hemodynamic signals and IMU motion data, and linearly combines these features to model a multiple linear regression model for motion artifacts, including: For fNIRS hemodynamic signals, the mean, slope, and power spectral density of HbO2 oxyhemoglobin and HbR deoxyhemoglobin signals were extracted. For IMU motion data, extract the absolute value, variance, and frequency domain features of acceleration and angular velocity; The features of fNIRS hemodynamic signals and IMU motion data are linearly combined to model a multiple linear regression model of motion artifacts: ； Where Artifact(t) is the multiple linear regression model of motion artifacts, X m (t) represents the m-th feature. For the m-th feature X m The regression coefficients of (t), where P is the number of features. For the intercept term, The term represents the residual, and t represents time. S22 estimates the regression coefficients by minimizing the sum of squared residuals, including: The regression coefficients are estimated using the following formula: ； in, For the regression coefficient vector, EEG(t) represents the EEG brain signal, and Q represents the number of time points. S23 uses regression coefficients to calculate motion artifacts to identify motion artifact segments in EEG signals, including: ； in, For motion artifacts, For the m-th feature X m The estimated values ​​of the regression coefficients of (t), This is an estimate of the intercept term.

10. The EEG signal acquisition and monitoring system according to claim 8, characterized in that: The interpolation repair module performs interpolation repair on the detected artifact segments to eliminate artifacts in the EEG signal, including: S31. Mark the start and end times of the BCG vascular pulsation artifact and motion artifact segments in the EEG signal; S32. Select n clean sampling points at each end of the BCG vascular pulsation artifact segment and the motion artifact segment, and construct a cubic spline function: ； Where f(t) is a cubic spline function, , , The first The start and end times of each sub-interval , , , For the first The spline coefficients of each subinterval are determined by solving a system of linear equations to ensure that the first and second derivatives of the cubic spline function f(t) are continuous at the sampling points; S33. Determine Gaussian white noise with a noise level equivalent to that of EEG brainwave signals: ； in, Represents Gaussian white noise Follows a mean of 0 and a variance of Gaussian distribution, Gaussian white noise The standard deviation is estimated by calculating the noise level of non-artifact segments in the EEG signal; S34. Replace the BCG vascular pulsation artifact segment and motion artifact segment in the EEG signal with cubic spline interpolation and Gaussian white noise signal: ； Among them, EEG new (t) represents the EEG signal after artifact removal, and t1 and t2 are the start and end times of the BCG pulsation artifact segment or motion artifact segment in the EEG signal, respectively.

Images