A method for removing artifacts from electroencephalogram signals and a collection system

Abstract

The application discloses an electroencephalogram artifact removal method and a collection system. The method uses a unified trigger signal to synchronously collect sensor data and electroencephalogram signals of a subject. Time stamp alignment verification is performed on the two types of synchronous data. A motion state quantitative index is determined. If the index does not exceed a preset motion threshold, the collected electroencephalogram signals are output. If the index exceeds the threshold, an abnormal processing procedure is triggered. The procedure includes sequentially performing adaptive filtering and nonlinear compensation on the electroencephalogram signals. The adaptive filtering takes the subject's resting electroencephalogram data as a reference, uses a step-dynamically-adjusted NLMS filter to suppress the linear part of the motion artifact and the baseline noise, and the nonlinear compensation constructs an adaptive filter based on a second-order Volterra series, optimizes kernel coefficients through an RLS algorithm, and processes nonlinear distortion. The output is the electroencephalogram signals after abnormal processing. The method is suitable for nonlinear and non-stationary motion artifact signals, can improve signal synchronicity and quality, and takes into account processing efficiency.

Description

Technical Field

[0001] This invention relates to brain-computer interface (BCI) signal processing technology, specifically to a method and acquisition system for removing artifacts from electroencephalogram (EEG) signals. Background Technology

[0002] Electroencephalography (EEG), as an important neurophysiological detection technique, plays a crucial role in cognitive neuroscience, clinical diagnosis, and brain-computer interfaces. EEG can record the electrical signals generated by the activity of neurons in the brain in real time, providing rich information for the study of brain function and neurological diseases. However, various interference signals, known as EEG artifacts, are inevitably introduced during EEG signal acquisition. These artifacts severely affect the quality of EEG signals and the accuracy of subsequent analysis, potentially leading to misinterpretations of brain activity and hindering the development of related research and applications. Therefore, effectively removing EEG artifacts has become an important topic in EEG research. In-depth research into methods for removing EEG artifacts is of great significance for improving the reliability of EEG data, advancing neuroscience research, and promoting the widespread application of EEG technology in clinical and other fields.

[0003] Electroencephalography (EEG), as an important neurophysiological technique, is widely used in cognitive science, clinical diagnosis, and brain-computer interfaces. However, non-neurogenic electrical interference (i.e., EEG artifacts) is inevitably introduced during signal acquisition, seriously affecting data quality and analytical accuracy. Artifacts originate from three sources: physiological sources (such as eye movement, electrocardiogram, and electromyography), power sources (such as power frequency interference and equipment noise), and environmental sources (such as static electricity and radio frequency signals). Typical artifact types include: eye movement artifacts (high-amplitude, low-frequency spikes, prominent in the frontal region), electromyography artifacts (high-frequency irregular activity, related to muscle movement), and electrocardiogram artifacts (periodic spikes, prominent in the left lead). These artifacts exhibit complex spatiotemporal characteristics: eye movement artifacts present as biphasic waveforms, electromyography artifacts have localized characteristics, and power frequency interference manifests as a sinusoidal wave across all leads. Summary of the Invention

[0004] The technical objective of this application is to address the technical problem that traditional EEG signal artifact removal methods cannot completely decompose motion artifacts or adapt to nonlinear and non-stationary motion artifact signals, thus failing to meet the needs of complex practical applications. This application provides an EEG signal artifact removal method and acquisition system to adapt to nonlinear and non-stationary motion artifact signals and improve the accuracy and reliability of EEG signal processing.

[0005] To achieve the above technical objectives, this application adopts the following technical solution.

[0006] In a first aspect, embodiments of this application provide a method for removing artifacts from electroencephalogram (EEG) signals, including: triggering sensor data of the subject at the same time using a uniform trigger signal, and collecting EEG signals from each lead;

[0007] The acquired sensor data and the synchronized data with the EEG signal are time-stamp aligned and verified, and the following data analysis is performed: the data analysis includes: determining a motion state quantification index based on the sensor data; if the motion state quantification index does not exceed a preset motion threshold, the acquired EEG signal is output; if the motion state quantification index exceeds the preset motion threshold, an anomaly handling process is triggered, and the anomaly-handled EEG signal is output.

[0008] The anomaly handling process includes sequentially performing adaptive filtering and nonlinear compensation on the EEG signal. The adaptive filtering includes using pre-acquired still EEG data of the subject without motion interference as a reference input, and employing an adaptive NLMS filter with dynamically adjustable step size factor to filter the currently acquired EEG signal to suppress the linear part of motion artifacts and base noise. The nonlinear compensation includes using an adaptive nonlinear filter constructed based on a second-order Volterra series model, and using a recursive least squares algorithm to optimize its linear and nonlinear kernel coefficients online to handle the nonlinear distortion introduced by motion interference in the adaptively filtered signal.

[0009] Secondly, embodiments of this application provide an electroencephalogram (EEG) signal acquisition system, comprising:

[0010] An external trigger signal generator is used to generate a unified trigger signal, which is simultaneously transmitted to the EEG signal acquisition module and the sensor data acquisition module, triggering the acquisition of sensor data and EEG signals of each lead of the subject at the same time.

[0011] The EEG signal acquisition module is used to acquire EEG signals from each lead of the subject.

[0012] The sensor data acquisition module is used to collect sensor data from the subject.

[0013] The data processing unit is used to perform timestamp alignment verification on the two types of synchronous data, namely the sensor data and the EEG signal, and to perform the following data analysis: the data analysis includes: determining a motion state quantification index based on the sensor data; if the motion state quantification index does not exceed a preset motion threshold, then outputting the collected EEG signal; if the motion state quantification index exceeds the preset motion threshold, triggering an anomaly handling process, and outputting the EEG signal processed by the anomaly handling process.

[0014] The anomaly handling process sequentially performs adaptive filtering and nonlinear compensation on the EEG signal. The adaptive filtering includes: using pre-acquired still EEG data of the subject without motion interference as a reference input, and employing an adaptive NLMS filter with dynamically adjustable step size factor to filter the currently acquired EEG signal to suppress the linear part of motion artifacts and base noise. The nonlinear compensation includes: employing an adaptive nonlinear filter constructed based on a second-order Volterra series model, and using a recursive least squares algorithm to optimize its linear and nonlinear kernel coefficients online to handle the nonlinear distortion introduced by motion interference in the adaptively filtered signal.

[0015] Compared with the prior art, the EEG signal artifact removal method and EEG signal acquisition system provided in this application have the following beneficial technical effects: by using a unified trigger signal to start sensor data and EEG signal acquisition at the same time, combined with timestamp alignment verification, the time synchronization of the two types of signals is ensured, avoiding the time sequence misalignment of motion data and EEG artifacts caused by clock asynchrony in traditional acquisition, providing a reliable time sequence basis for subsequent accurate determination of motion artifacts, and solving the problem of synchronization of cross-type signals (physiological motion data + EEG signal).

[0016] The anomaly handling process includes pre-processing adaptive filtering and post-processing nonlinear filtering. The pre-processing adaptive filtering stage uses static EEG data as a reference sample to dynamically suppress motion artifacts and base noise in the current EEG signal. Post-processing nonlinear compensation involves constructing an adaptive nonlinear filter to compensate for nonlinear distortion introduced by motion interference. This filtering method effectively handles complex motion artifacts, adapts to nonlinear and non-stationary motion artifact signals, and improves the accuracy and reliability of EEG signal processing. Complex filtering algorithms are only activated when necessary, avoiding unnecessary computational resource consumption, balancing computational efficiency and signal quality, and improving the overall performance and stability of the device.

[0017] The abnormal processing flow is dynamically triggered based on the quantitative indicators of motion state, and the original signal is directly output in the static or low-motion state, avoiding the waste of computing power caused by complex filtering throughout the process. Especially in the scenario of portable EEG devices, it can reduce power consumption, balance signal quality and device battery life, and solve the problems of "low efficiency and resource redundancy" in the traditional fixed processing mode.

[0018] The adaptive filtering step size dynamic adjustment and the nonlinear compensation kernel coefficient online optimization are both adapted to the differences in EEG baseline and the non-stationary characteristics of motion artifacts among different subjects. Compared with conventional methods that rely on preset models, they reduce artifact removal bias caused by individual differences and improve signal processing stability in multi-subject experiments.

[0019] It should be understood that the summary section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0020] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of this application in any way. Furthermore, the shapes and scales of the components in the drawings are merely illustrative to aid in understanding this application and do not specifically limit the shapes and scales of the components. Those skilled in the art, guided by the teachings of this application, can select various possible shapes and scales to implement this application according to specific circumstances. In the drawings:

[0021] Figure 1 A schematic diagram of a method for removing artifacts from electroencephalogram (EEG) signals provided in one embodiment;

[0022] Figure 2 A schematic flowchart of a method for removing artifacts from electroencephalogram signals provided in another embodiment;

[0023] Figure 3 This is a schematic diagram of the NLMS filtering process in the embodiment;

[0024] Figure 4 This is a schematic diagram of the adaptive Volterra filter filtering process in the embodiment;

[0025] Figure 5 A schematic diagram of the EEG acquisition system provided in the embodiment;

[0026] Figure 6 A schematic diagram of the EEG signal acquisition module and sensor data acquisition module in the EEG acquisition system provided in the embodiment;

[0027] Figure 7 The confusion matrix results for classifying three subjects using unprocessed EEG data are shown in the example, where (a) is the result for subject 1, (b) is the result for subject 2, and (c) is the result for subject 3.

[0028] Figure 8 The confusion matrix results for classifying three subjects using the traditional artifact removal method are shown in the example, where (a) is the result for subject 1, (b) is the result for subject 2, and (c) is the result for subject 3.

[0029] Figure 9 The classification confusion matrix results of the EEG data of three subjects after processing with ARIMA and the distributed gradient boosting library are shown in the example. (a) is the result of subject 1, (b) is the result of subject 2, and (c) is the result of subject 3.

[0030] Figure 10 The following is a schematic diagram of the artifact removal effect based on the NLMS filter in the embodiment, where (a) is a schematic diagram of the clean signal, (b) is a schematic diagram of the noisy expected response signal, (c) is the error vector after filtering, and (d) is the output signal after filtering.

[0031] Figure 11 The example demonstrates the effect of NLMS filter on artifact removal in the frequency domain.

[0032] Figure 12 The following is an example of the filtering effect of the second-order Volterra filter: (a) is the original EEG signal, (b) is the EEG signal containing motion artifacts, and (c) is the EEG signal after removing artifacts using the second-order Volterra filter (SVF). Detailed Implementation

[0033] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0034] EEG filtering is a core technology that uses frequency response differences to design filters to selectively remove artifacts. It mainly includes the following two categories:

[0035] 1. Classical Frequency Filtering: Based on the frequency difference between artifacts and EEG signals, filters are designed to separate interference in specific frequency bands. This includes: Notch filtering: specifically eliminating power frequency interference (50Hz / 60Hz), often using elliptic filters to achieve a steep transition band.

[0036] Bandpass / bandstop filtering: Low-pass filter removes eye movement artifacts (<1Hz), high-pass filter suppresses electromyography artifacts (>20Hz).

[0037] Classical frequency filtering uses Butterworth filters and Chebyshev filters. Butterworth filters have a flat passband, making them suitable for distortion-sensitive scenarios; Chebyshev filters have a steeper transition band, but their passband exhibits ripples; elliptic filters have a low order, but their design is complex.

[0038] 2. Time-Frequency Analysis Filtering (Wavelet Transform): This method utilizes multi-scale decomposition of wavelet basis functions to locate artifact features in the time-frequency domain. Thresholding is used to suppress specific scale coefficients (such as low-frequency eye-tracking artifacts) while preserving genuine EEG signals. Its advantage lies in its adaptability to non-stationary signals, and the processing effect can be optimized by selecting the appropriate wavelet basis.

[0039] The challenges and limitations of the above two types of EEG filtering methods are as follows:

[0040] Classical filtering relies on prior frequency assumptions, which may inadvertently remove useful signals; wavelet transform requires manually setting thresholds and has limited ability to separate complex artifacts. In practical applications, filtering strategies need to be selected based on artifact characteristics, or integrated with methods such as ICA and machine learning to improve robustness. As research into EEG artifact removal deepens, traditional EEG filtering techniques will face new challenges in dealing with complex motion artifact generation processes. For example, ICA methods, based on the assumption of independent components, cannot completely decompose motion artifacts, while traditional filtering methods struggle to adapt to nonlinear and non-stationary motion artifact signals. A single artifact removal method often cannot meet the complex needs of practical applications. In the future, the integration of multiple methods will become a development trend.

[0041] Example 1 provides a method for removing artifacts from electroencephalogram (EEG) signals, including:

[0042] The sensor data of the subjects and the collection of EEG signals of each lead are triggered at the same time using a unified trigger signal.

[0043] The acquired sensor data and EEG signals, which are synchronized data, are time-stamp aligned and verified. The following data analysis is performed: the motion state quantification index is determined based on the aligned and verified sensor data. If the motion state quantification index does not exceed the preset motion threshold, the acquired EEG signal is output. If the motion state quantification index exceeds the preset motion threshold, the abnormality handling process is triggered, and the abnormality-handled EEG signal is output.

[0044] The anomaly handling process includes sequentially performing adaptive filtering and nonlinear compensation on the EEG signal. The adaptive filtering includes using still EEG data to suppress the linear part of motion artifacts and base noise, while the nonlinear compensation includes processing the nonlinear distortion introduced by motion interference through an adaptive nonlinear filter.

[0045] In this embodiment, the external trigger signal can be generated by an external trigger signal generator.

[0046] In this embodiment, sensor data may include one or more of physiological state data or motion behavior data.

[0047] In this embodiment, the acquired sensor data (at least one of physiological state data or motion behavior data) and electroencephalogram (EEG) signals are time-stamp aligned and verified. For example, if there is a very small delay, it is calibrated using an algorithm, and data fusion is completed based on the application scenario requirements. Physiological state data may include the subject's heart rate, while motion behavior data may include the subject's step count, acceleration, or angular velocity. In sports health monitoring, bioelectrical signals (such as heart rate) or motion postures (such as step count, acceleration, and angular velocity) are correlated and analyzed, and the processed valid data is finally output for use by upper-layer applications.

[0048] In some embodiments, timestamp alignment verification is performed using interpolation. When the time delay between sensor data and EEG signal is detected to be less than a preset value (e.g., the preset value can be 0.1ms), the corresponding number of data points are inserted into the time-lagging data sequence to make the two types of synchronized data precisely aligned in the time dimension, so as to ensure the accuracy of subsequent data analysis.

[0049] In some embodiments, a triaxial accelerometer can be used to acquire subject acceleration data at a sampling rate of 200 Hz, while simultaneously acquiring EEG signals from each lead (10-20 system) at a sampling rate of 1000 Hz. For example... Figure 1 As shown, the resultant acceleration force F is determined based on triaxial acceleration data, serving as a quantitative indicator of motion state. The formula for calculating the resultant acceleration force F is as follows:

[0050] ;

[0051] Where a x a y a z These are the acceleration values ​​along the x, y, and z axes. If the resultant acceleration force F exceeds a preset empirical threshold T (T can be set empirically), i.e., F>T, then the exception handling process is triggered.

[0052] In this embodiment, motion state is monitored in real time by an accelerometer. When abnormal jitter is detected, static EEG data is first collected as a reference input, and noise in the current EEG signal is suppressed by an adaptive filtering algorithm. Then, a nonlinear compensation model is established using a Volterra filter, and finally, a high-quality compensated EEG signal is output. The entire process adopts a dynamic threshold triggering mechanism to achieve optimized allocation of computing resources.

[0053] In this embodiment, adaptive filtering includes: using pre-acquired still EEG data of the subject without motion interference as a reference input, and employing an adaptive NLMS filter with dynamically adjustable step size factor to filter the currently acquired EEG signal to suppress the linear part of motion artifacts and base noise; nonlinear compensation includes: employing an adaptive nonlinear filter constructed based on a second-order Volterra series model, and using a recursive least squares algorithm to optimize its linear and nonlinear kernel coefficients online to handle the nonlinear distortion introduced by motion interference in the adaptively filtered signal.

[0054] As an example, adaptive filtering utilizes resting EEG data to suppress the linear portion of motion artifacts and base noise, such as Figure 1 As shown, the process includes: acquiring a reference EEG signal (EEG_rest) from the subject within a preset time period (e.g., within 1 second) without motion interference; constructing an adaptive filter using the LMS algorithm; filtering the currently acquired EEG signal (EEG_current) using the reference EEG signal (EEG_rest) as the reference input; the adaptive parameter step size being determined by the difference between the motion state quantification index and the preset motion threshold T; and then constructing an adaptive nonlinear filter using a second-order Volterra series model to output the compensated EEG signal.

[0055] In other embodiments, the adaptive filter is an NLMS filter, which uses still EEG data to suppress the linear portion of motion artifacts and base noise. The core process of the adaptive NLMS filter is to continuously adjust the filter coefficients based on the input signal and the desired response signal to make the output signal as close as possible to the desired response. First, the filter coefficients are initialized. Then, in each iteration, the filter output is calculated, and the difference between it and the desired response is used to obtain the error vector. Based on the error vector and the correlation characteristics of the input signal, the filter coefficients are updated according to the NLMS algorithm rules. This process is iterated until a preset convergence condition is met or the maximum number of iterations is reached, and finally, the filtered signal is output.

[0056] like Figure 3 As shown, the general process for establishing an adaptive NLMS filter is as follows: First, initialize relevant parameters (such as sliding window length N, step size factor mu, etc.), calculate the data length L, and initialize the filter coefficient vector. Given a zero vector of length N, initialize the error vector. Given a zero vector of length LN and an empty list of filtered signals (filtered_signal), the system iterates through a loop to extract the sampled value of the input signal at time n in each iteration. n is the sampling point number;

[0057] Obtain the desired response signal , calculate the current filter output (the EEG signal after non - linear compensation processing), and add it to the list of filtered signals. Then calculate the error vector , and according to the error vector and the sampled value of the input signal at time n update the filter coefficient vector . After completion, update the iteration n until the condition n < L - N for the number of iterations is no longer satisfied, and finally return the error vector and the list of filtered signals filtered_signal

[0058] Prepare an input signal with a preparation time consistent with the sampling frequency and the desired response signal , ensuring the accuracy and integrity of the data, as well as their consistency in terms of time, sampling rate, etc

[0059] In the iterative calculation stage, at the nth iteration, extract a segment of data with a length equal to the sliding window length N from the input signal sequence, and reverse its order to obtain a new input signal for subsequent calculations; according to the current filter coefficient vector and the input signal , calculate the filter output through dot - product operation ; compare the filter output with the desired response signal to obtain the error vector , which reflects the deviation between the current filter output and the desired result. According to the NLMS algorithm rules, combine the error vector , the input signal and dynamically adjust the step - size factor to update the filter coefficient vector , and use the formula to update the filter coefficient vector, where n is the sampling point number. Under the innovative dynamic step - size mechanism, it will change in real - time according to the signal characteristics; add a very small positive number (such as ) to the denominator to prevent the denominator from being zero due to too small input signal energy, ensuring the stability of the algorithm

[0060] In the convergence judgment stage, set the convergence judgment condition. If the convergence condition is met, end the iterative process and output the final filter coefficient vector and the filtered signal (which can be obtained by filtering the input signal throughout with the final filter coefficient vector); if not, return to the iterative calculation stage, continue the next iteration, and continuously optimize the filter coefficient vector until convergence

[0061] For example, when the mean square value of the error vector in several consecutive iterations is less than a certain preset threshold, or when the change in the filter coefficient vector in several consecutive iterations is less than a specific value, the filter is considered to have converged.

[0062] The adaptive NLMS filter simplifies control parameters by using a uniform step size, which simplifies filter parameter settings (filter upper and lower limits, quality factor, etc.) and reduces algorithm complexity. Furthermore, the adaptive NLMS filter employs a dynamic step size adjustment mechanism (referred to as "dynamic adjustment"): abandoning the traditional fixed step size setting and introducing an intelligent dynamic adjustment strategy. Based on the statistical characteristics of the signal, such as the degree of autocorrelation, the magnitude of the current error, and its changing trend, the step size is changed in real time.

[0063] As an example, dynamic adjustment can be based on increasing the step size in stable signal segments to accelerate convergence, while decreasing the step size in areas of signal abrupt change or severe noise interference to improve filtering stability and accuracy, effectively balancing convergence speed and steady-state error. Dynamically adjusting the adaptive filtering step size based on the stationarity of EEG signals can specifically address the problem of traditional fixed step sizes "failing to balance convergence speed and filtering stability." It accelerates convergence in stable segments, improving processing efficiency, enhances stability in abrupt / high-noise segments, reduces signal distortion, adapts to the non-stationary characteristics of EEG signals, optimizes artifact removal accuracy, and reduces computational redundancy.

[0064] In a specific example, when the rate of change of the EEG signal amplitude at M1 consecutive sampling points is less than the set threshold P1, and the fluctuation of the power spectral density in the α band is less than Q1dB, the segment is judged to be a stable signal segment; if the rate of change of the EEG signal amplitude exceeds the preset high threshold P2 within M2 sampling points, or the change of the power spectral density in a certain frequency band exceeds Q2dB, or the kurtosis value of the signal exceeds R% of the standard kurtosis value, it is judged to be a region of signal mutation or severe noise interference.

[0065] The alpha band is one of the core frequency bands of electroencephalogram (EEG), mainly referring to brain electrical activity in the range of 8Hz-13Hz. Brain electrical activity can be captured by attaching multi-channel electrodes to specific locations on the scalp (such as the occipital and parietal lobes) using EEG acquisition equipment, and then the alpha band EEG signals are extracted after signal filtering (usually within the 8-13Hz band).

[0066] For example, if an EEG signal is sampled at 1000Hz, and the amplitude change rate is less than 0.5μV / ms for 100 consecutive sampling points (i.e., within 0.1 seconds), and the power spectral density fluctuation in the α band (8Hz-13Hz) is less than 2dB, it is considered a stable signal range. In this range, increasing the step size factor, such as multiplying the step size factor by a coefficient K1 greater than 1 (K1 > 1, e.g., K1 = 1.5), accelerates the filter convergence speed. For instance, if the amplitude change rate is greater than 5μV / ms within 50 sampling points (0.05 seconds), or the power spectral density change in the β band (13-30Hz) exceeds 10dB, or the kurtosis value exceeds 150% of the standard kurtosis value, this region is identified. In this case, decreasing the step size factor, such as dividing the step size factor by a coefficient K2 greater than 1 (K2 > 1, e.g., K2 = 2), improves filter stability and accuracy.

[0067] In some embodiments, the output expression of the adaptive nonlinear filter constructed based on the second-order Volterra series model is:

[0068] ;

[0069] in The output of the filter is the EEG signal after nonlinear compensation processing, where n is the sampling point number, i.e., the discrete-time index, M is the length of the linear term sliding window, and L is the length of the quadratic term feature window. The weighting coefficients of the linear term (first-order coefficients of the Volterra kernel) are used to characterize the contribution of the linear part of the input signal to the output; i and j are coefficient indices, and n is the sampling point number; Let be the sampled value of the input signal at time ni. H represents the sampled value of the input signal at time nj. i H represents the second-order weighting coefficient of the nonlinear term with coefficient index i. j The second-order weighting coefficients (second-order coefficients of the Volterra kernel) for the nonlinear term with coefficient index j are used to characterize the contribution of the nonlinear interaction component of the input signal to the output. The Volterra kernel coefficients can be optimized using the Recursive Least Squares (RLS) algorithm. H i and H j .

[0070] This expression compensates for nonlinear distortion in motion artifacts by superimposing a linear term (first term) and a second-order nonlinear term (second term). The linear term handles the linear relationship of the signal, while the second-order nonlinear term captures the nonlinear interaction features in the signal.

[0071] The above describes how to remove the nonlinear component in motion artifacts. In this embodiment, an adaptive Volterra filter is designed and implemented to remove both linear and nonlinear terms, and the least mean square (LMS) strategy is used for weight adjustment. In this embodiment, while strictly adhering to the LMS rule, weights are set separately for the linear and nonlinear terms to control the weight update speed of the two different adaptive processes. Normalization is also added to make the gradient descent method NLMS (Normalized Least Mean Square Error), which prevents excessive updates and improves algorithm stability. Finally, a mechanism to prevent gradient explosion updates is introduced by adding a decimal term to the normalization factor to prevent division by zero and excessive gradient updates.

[0072] The adaptive Volterra filter is described in detail below. For example... Figure 4 As shown, the main filtering process of the adaptive Volterra filter can be summarized as follows: Initialize parameters, including: linear term sliding window length M, quadratic term feature window length L, dimension L2 of the second-order feature vector (number of elements in the upper triangular part), step size parameter mu1, step size parameter mu2, and initialization... Matrix, desired response signal Linear weights Nonlinear second-order weighting coefficients Error vector Slide input signal (make) (For an input signal buffer of length M), set to store second-order input characteristics. The matrix for second-order input features Calculate the elements of the upper triangular region A matrix, expressed as: , where vec() is a vectorization operation, This indicates that the second-order input feature of the matrix is... Convert to a column vector, where upper represents the upper triangular portion.

[0073] Sliding second-order input features Calculate second-order nonlinear input features As a weighted sum of second-order features, ;calculate As a synthesis of the input signal and second-order features, Calculate the filter output , ; Calculate the error vector , Update linear weights , Update the nonlinear second-order weighting coefficients , n is the sampling point index; store the error vector. The process iterates until the end of the data pool, then exits the computation.

[0074] In some embodiments, constructing an adaptive Volterra filter includes the following steps:

[0075] 1. Initialization phase parameters:

[0076] M is the length of the sliding window for the linear term;

[0077] L is the length of the quadratic term feature window;

[0078] : The dimension of the second-order eigenvector (the number of elements in the upper triangular part).

[0079] The input signal is a linear input vector (current and past). (Number of input values).

[0080] Second-order input characteristics (constructed from the input signal).

[0081] Linear weights.

[0082] Error vector.

[0083] 2. Iterative updates are performed using the following formula:

[0084] ; here It's a sliding window that updates with the latest version in each iteration. , The input signal is sampled at time n, with the oldest value removed.

[0085] 3. Construct second-order features, including constructing the outer product matrix: ;in, Extract the upper triangular portion as a second-order feature vector: And store it in the sliding window.

[0086] 4. Calculate the second-order nonlinear input characteristics. ;here It is composed of second-order input feature vectors Through nonlinear second-order weighting coefficients The calculated nonlinear contribution.

[0087] 5. Calculate the synthesis of the input signal and the second-order nonlinear characteristic: ;in It is a weighted input, which includes: a linear input signal Nonlinear second-order input feature vector .

[0088] 6. Calculate intermediate forecast values: ;in It is the predicted output of the filter, calculated through a weighted sum.

[0089] 7. Calculate the error vector: Error vector It is the expected response signal and predicted output The error between them.

[0090] 8. Linear weight update: Here, the LMS (Least Mean Square) rule is used for adaptive weight update, employing the error vector. and input signal Make corrections. It is a small number ( () is used to prevent division by zero.

[0091] 9. Update of nonlinear second-order weighting coefficients: .

[0092] Here It is a weighted sum of second-order features. Normalization factor. Used for stable updates, || represents the norm of the vector.

[0093] 10. Save the error vector Determine whether all iterations have been completed.

[0094] 11. If all iterations are completed, the code will eventually return an error vector. If not all iterations are completed, return to step 2 to continue iterating.

[0095] In some embodiments, such as Figure 2As shown, the anomaly handling process includes ARIMA processing of the processed EEG signal, specifically including: preliminary data cleaning through bandpass filtering (preserving the EEG signal frequency band, such as 1-40Hz), baseline drift removal, etc.; performing stationarity testing to determine the ARIMA model parameters (p, d, q), where p is the order of the autoregressive (AR) term, representing the number of lagged observations used in the ARIMA model, i.e., the correlation between the current value and the observations of the past p time steps; d is the difference order, representing the number of differences required to transform a non-stationary time series into a stationary series (usually 0, 1, or 2); and q is the order of the moving average (MA) term, representing the number of lagged error terms used in the model, i.e., the correlation between the current value and the error terms of the past q time steps. The ARIMA model is trained to learn the temporal characteristics of normal EEG signals. The trained ARIMA model is used to predict the expected output value of the EEG signal. The deviation between the expected output value of the EEG signal and the actual EEG signal value after adaptive filtering and nonlinear compensation is detected. If the deviation exceeds the preset range, it is determined that there are artifact segments in the signal. The artifact segments are corrected based on the expected output value of the EEG signal (such as replacing with predicted values ​​or interpolating for repair).

[0096] This application utilizes a unified trigger signal to simultaneously initiate sensor data and EEG signal acquisition, ensuring time synchronization between the two types of signals. Compared to traditional acquisition methods, this hardware-level clock bus synchronization mechanism avoids data errors caused by asynchronous acquisition, laying the foundation for subsequent accurate data analysis, improving signal synchronization quality, and enabling the analysis results to more accurately reflect the subject's true physiological or motor state. By performing timestamp alignment verification and data analysis on the acquired sensor data and EEG signals, the accuracy and consistency of the data are ensured.

[0097] This application employs a two-stage filtering architecture, including pre-stage adaptive filtering and post-stage nonlinear filtering. The pre-stage adaptive filtering uses resting EEG data (EEG_rest) as a reference sample to dynamically suppress motion artifacts and base noise in the current EEG signal. The post-stage nonlinear compensation involves constructing an adaptive nonlinear filter (such as an adaptive nonlinear filter based on Volterra series) to compensate for nonlinear distortion introduced by motion interference. This two-stage filtering approach effectively addresses complex motion artifacts, adapts to nonlinear and non-stationary motion artifact signals, and improves the accuracy and reliability of EEG signal processing.

[0098] By monitoring motion status in real time using sensor data, a quantitative indicator of motion status is determined based on the real-time sensor data and compared with a preset motion threshold. A complex filtering algorithm is only activated when abnormal jitter is detected. This mechanism can respond promptly to changes in motion status, ensuring stable signal quality under different motion conditions. Simultaneously, activating the complex filtering algorithm only when necessary avoids unnecessary consumption of computational resources, balancing computational efficiency and signal quality, and improving the overall performance and stability of the device.

[0099] Example 2: Based on the same inventive concept as the EEG signal artifact removal method provided in the above examples, this example provides an EEG signal acquisition system, including an external trigger signal generator, an EEG signal acquisition module, a sensor data acquisition module, and a data processing unit.

[0100] An external trigger signal generator is used to generate a unified trigger signal, which is simultaneously transmitted to the EEG signal acquisition module and the sensor data acquisition module, triggering the acquisition of sensor data and EEG signals from each lead of the subject at the same time.

[0101] The EEG signal acquisition module is used to acquire EEG signals from each lead of the subject.

[0102] The sensor data acquisition module is used to collect sensor data from the subjects.

[0103] The data processing unit is used to perform timestamp alignment verification on the two types of synchronous data, namely the acquired sensor data and the EEG signal, and to perform the following data analysis: The data analysis includes: determining the motion state quantification index based on the sensor data; if the motion state quantification index does not exceed the preset motion threshold, the acquired EEG signal is output; if the motion state quantification index exceeds the preset motion threshold, the abnormality handling process is triggered, and the abnormality-handled EEG signal is output.

[0104] The anomaly handling process sequentially performs adaptive filtering and nonlinear compensation on the EEG signal. The adaptive filtering includes: using pre-acquired still EEG data of the subject without motion interference as a reference input, and employing an adaptive NLMS filter with dynamically adjustable step size factor to filter the currently acquired EEG signal to suppress the linear part of motion artifacts and base noise. The nonlinear compensation includes: using an adaptive nonlinear filter constructed based on a second-order Volterra series model, and using a recursive least squares algorithm to optimize its linear and nonlinear kernel coefficients online to handle the nonlinear distortion introduced by motion interference in the adaptively filtered signal.

[0105] The embodiment employs a synchronous bus triggering system. The external trigger signal generator is the core driving source for the entire hardware synchronous acquisition, responsible for generating a unified trigger signal and transmitting it to the trigger interfaces of the EEG signal acquisition module and the sensor data acquisition module, ensuring that the acquisition actions of both are initiated at the same time.

[0106] In some embodiments, see Figure 5 The EEG signal acquisition module includes a trigger interface, a trigger selection unit, an acquisition engine, and an ADC data FIFO for the EEG signal acquisition chip. The trigger selection unit of the EEG signal acquisition chip is used to receive trigger signals sent by an external trigger signal generator, parse the trigger event type, drive the acquisition engine to activate the internal ADC channel to start synchronous acquisition of EEG signals, and store the acquired ADC data into the ADC data FIFO.

[0107] The sensor data acquisition module includes a sensor trigger interface, a sensor trigger selection unit, an internal sensor group (in some embodiments, the sensor group is an IMU), and a sensor data FIFO. The sensor trigger selection unit is used to receive trigger signals sent by an external trigger signal generator, identify trigger events, drive the internal sensor group to start collecting sensor data from the subject, and store the collected data in the sensor data FIFO. The data processing unit, as a terminal processing module, is used to read the ADC data FIFO and sensor data FIFO through hardware or software interfaces, and perform data analysis to provide a processing foundation for subsequent applications (such as sports health monitoring, bioelectrical signal and posture correlation analysis).

[0108] In some embodiments, such as Figure 5 As shown, the sensor group inside the sensor data acquisition module includes an IMU sensor, which includes at least one of an accelerometer and a gyroscope. The accelerometer is used to acquire acceleration data, and the gyroscope is used to acquire angular velocity data.

[0109] Please see Figure 5 and Figure 6 An external trigger signal generator generates a trigger signal of a specific frequency or pulse, which is simultaneously transmitted to the external trigger interface of the EEG signal acquisition chip and the trigger interface of the IMU through hardware circuitry, triggering the acquisition actions of both and ensuring time synchronization.

[0110] In this embodiment, the triggering and acquisition process of the EEG signal acquisition chip is as follows:

[0111] Trigger reception: The EEG signal acquisition chip receives external trigger signals through the trigger interface.

[0112] Trigger Selection: The signal enters the "Trigger Selection" unit, where the trigger event type is parsed.

[0113] Start Acquisition: Trigger the event-driven "Acquisition Engine" to activate the internal ADC channels (Channel A, Channel B, Channel C, and Channel D) to begin synchronous acquisition of bioelectrical signals.

[0114] Data caching: The acquired ADC data is stored in the "ADC data FIFO" and awaits transmission to the data processing unit.

[0115] The sensor data acquisition module includes an IMU. The IMU triggering and acquisition process is as follows: Trigger reception: The IMU receives external trigger signals through the trigger interface.

[0116] Trigger Selection: The signal is processed by the "Trigger Selection" unit to identify the trigger event.

[0117] Start data acquisition: Triggering an event drives the accelerometer and gyroscope inside the IMU to start acquiring motion data.

[0118] Data caching: The collected IMU data is stored in the "IMU data FIFO" to prepare for subsequent processing.

[0119] The data processing unit acts as the terminal hub, reading the ADC data FIFO and IMU data FIFO of the EEG signal acquisition chip through hardware or software interfaces. It performs timestamp alignment verification on the two types of synchronized data and completes data fusion based on application scenario requirements.

[0120] To verify the classification results of the EEG signals obtained by the EEG signal artifact removal method of this application, EEG feature extraction and classification models were performed on the unprocessed raw EEG data and the EEG signals after adaptive filtering and nonlinear compensation, respectively, to obtain the classification results.

[0121] EEG feature extraction includes extracting time-domain features (calculating mean, variance, peak value, etc.), frequency-domain features (extracting the energy proportion of each frequency band (δ, θ, α, β, γ) through FFT transformation), and time-frequency domain features (using wavelet transform to obtain the time-frequency feature matrix).

[0122] Building a classification model involves: selecting a classification algorithm (such as SVM, random forest, deep learning models such as CNN-LSTM, and EEGNET), and dividing the model into training and test sets. The model is then trained using input feature data, and hyperparameters (such as the kernel function and learning rate of SVM) are optimized to improve classification accuracy.

[0123] Classification prediction: Extract features from the artifact-corrected EEG signal, input them into the trained classification model, and output the classification results (such as motor imagery type, emotional state, etc.).

[0124] Results validation: The model performance was evaluated using metrics such as confusion matrix, accuracy, and recall. Cross-validation was used to ensure the reliability of the results, and finally, a stable classification conclusion was output.

[0125] Figure 7 The results of confusion matrices for classifying three subjects using unprocessed raw EEG data are presented. Each matrix visually shows the correspondence between the actual and predicted categories. It is clearly visible in the figures that the diagonal (correctly classified) values ​​of the confusion matrices for each subject are low, while the off-diagonal (misclassified) values ​​are high. Figure 7 In the table, (a) shows the results for subject 1, (b) shows the results for subject 2, and (c) shows the results for subject 3. The average accuracy rate for subject 1 was 56.25%, for subject 2 it was 58.33%, and for subject 3 it was 60.42%. The average accuracy rate of the three subjects was less than 60%, which fully demonstrates that the classifier was severely affected by noise and motion artifacts when processing the raw data.

[0126] Figure 8 The results of the confusion matrix obtained by classifying the EEG data of three subjects after processing with traditional artifact removal methods (ICA + time-frequency analysis) are presented. Figure 8 In the diagram, (a) represents the results of subject 1, (b) represents the results of subject 2, and (c) represents the results of subject 3; the mean accuracy rate for subject 1 was 66.66%, for subject 2 it was 70.83%, and for subject 3 it was 72.92%. Figure 7 In comparison, the diagonal values ​​of the confusion matrices for each participant showed improvement, indicating that the traditional method reduced noise interference and improved the correct classification rate to some extent. However, significant confusion between some categories can still be observed. For example, in the first participant's confusion matrix, although the prediction accuracy for label 0 reached 83.33%, the prediction accuracy for label 1 was only 50% binary probability.

[0127] Figure 9 The results of the classification confusion matrices of EEG data from three subjects after processing with ARIMA and a distributed gradient boosting library are presented. After processing with this method, the diagonal regions of the confusion matrices of each subject are significantly highlighted, and the off-diagonal misclassified regions are significantly shrunk, indicating a qualitative improvement in the classifier's ability to distinguish signals of different categories. Specifically, the confusion matrices of the first two subjects exhibit typical diagonal distribution characteristics, with a significant improvement in classification accuracy; the confusion matrix of the third subject is almost entirely diagonally filled, with the vast majority of samples being accurately classified. Overall, the method of this invention significantly improves the separability of EEG signals by effectively suppressing motion artifacts and optimizing feature extraction, greatly reducing the confusion between categories, and fully verifying the significant advantages of this method in improving EEG classification performance.

[0128] The following section describes in detail the effect of the NLMS (Normalized Least Mean Square) filter, combining code execution data and generated figures. Figure 10and Figure 11 .

[0129] The code generates a pure signal from Figure 10 As shown in (a), the clean signal is a standard sine wave with an amplitude that periodically fluctuates between -1 and 1 within a time interval of 0 to 2 seconds. The frequency is 50 Hz in the first second, 100 Hz in the middle 0.5 seconds, and 123 Hz in the last 0.5 seconds. The signal waveform is smooth and regular, without any interference.

[0130] Figure 10 Figure (b) shows the noisy desired response signal. Based on a set signal-to-noise ratio (1dB here), the noise intensity is calculated using a formula, and Gaussian white noise is superimposed on the clean signal to obtain the noisy desired response signal. It is clear from the graph that compared to the clean signal, the waveform becomes chaotic, the noise interference is significant, the amplitude fluctuation range is increased, and the original sine wave regularity is disrupted, making it difficult to directly distinguish the frequency and amplitude information of the original signal.

[0131] Figure 10 Figure (c) shows the error vector after filtering. The error vector is obtained after filtering the noisy signal using the NLMS function. As can be seen from the graph, the error vector still exhibits significant fluctuations. This indicates that although the NLMS filter continuously adjusts its coefficients to reduce the gap between the output and the desired response signal, it cannot completely eliminate noise due to the complexity of the noise and the limitations of the filter itself, leaving some residuals. However, these fluctuations also reflect the filter's gradual correction of noise during the iteration process, and the error fluctuations are within a certain range.

[0132] Figure 10 Figure (d) shows the filtered signal. After processing by the NLMS filter, the waveform of the signal is significantly closer to the clean signal compared to the noisy desired response signal. Although there are some transitional fluctuations in the initial stage, possibly due to the filter not yet fully converging, the subsequent waveform largely follows the frequency and amplitude changes of the clean signal over time. This indicates that the NLMS filter effectively removes most of the noise, resulting in significant signal recovery and improvement. (See frequency domain diagram). Figure 11 As can be seen, high-frequency signals are well suppressed, while effective low-frequency signals are well preserved.

[0133] The filtering results of the adaptive Volterra filter are visible. Figure 12. Figure 12 Figure (a) shows the filtering effect on the original EEG signal. A 10Hz sine wave was generated as the original EEG signal, with an amplitude range of -0.5 to 0.5, without motion artifacts. The curve in the figure is smooth and regular, presenting a standard sine wave shape, reflecting the pure characteristics of the EEG signal under ideal conditions, and serving as the benchmark for subsequent comparisons.

[0134] Figure 12 Figure (b) shows the filtering effect on an EEG signal with motion artifacts. A simulated IMU motion artifact (a 1Hz sine wave superimposed with noise) was generated and superimposed onto the first channel of the EEG signal. Compared to the original signal, the waveform changed significantly, exhibiting regular superposition and altered extreme values, clearly demonstrating the contamination of the EEG signal by motion artifacts, thus necessitating artifact removal.

[0135] Figure 12 Figure (c) shows the filtering effect on the EEG signal after artifact removal using a second-order Volterra filter (SVF). The code uses the artifact removal function to process the EEG containing artifacts using SVF, suppressing motion artifacts. Motion artifacts are significantly reduced. Although the waveform still differs from the original signal (possibly due to the nonlinear processing of the signal by the filter), the regularity is clearly restored compared to the signal containing artifacts, verifying the effectiveness of the SVF algorithm in suppressing motion artifacts.

[0136] The code reproduces the process of motion interference and recovery of EEG signals through the workflow of "generating the original signal → adding artifacts → removing artifacts". The charts visually represent this process, clearly demonstrating the role of the SVF algorithm in motion artifact removal and its effectiveness in improving EEG signal quality.

[0137] It should be fully understood that the user information involved in this application (including but not limited to user physiological information, user personal information, etc.) is information and data authorized by the user or fully authorized by all parties. The use of user information shall comply with privacy policies and practices that are generally considered to meet or exceed industry or government requirements for maintaining user privacy. The collection, use and processing of related data shall comply with relevant laws, regulations and standards, and provide corresponding operation access points for users to choose to authorize or refuse.

[0138] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or physical entities, or by products with certain functions. A typical implementation device is a computer. Specifically, the computer can be, for example, a personal computer, a laptop computer, or a tablet computer, or any combination of these devices.

[0139] The above provides a detailed description of the EEG signal artifact removal method and acquisition system provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the concept of this application and should not be construed as limiting the scope of protection of this application.

Claims

1. A method for removing artifacts from electroencephalogram (EEG) signals, characterized in that, include: The sensor data of the subjects and the collection of EEG signals of each lead are triggered at the same time using a unified trigger signal. The acquired sensor data and the electroencephalogram (EEG) signal are time-stamp aligned and verified, and the following data analysis is performed: the data analysis includes: determining a motion state quantification index based on the aligned and verified sensor data; if the motion state quantification index does not exceed a preset motion threshold, the acquired EEG signal is output; if the motion state quantification index exceeds the preset motion threshold, an anomaly handling process is triggered, and the anomaly-handled EEG signal is output. The abnormality handling process includes sequentially performing adaptive filtering and nonlinear compensation on the electroencephalogram (EEG) signal; The adaptive filtering includes: using pre-acquired still EEG data of the subject without motion interference as reference input, and employing an adaptive NLMS filter with dynamically adjustable step size factor to filter the currently acquired EEG signal to suppress the linear part of motion artifacts and base noise; the nonlinear compensation includes: employing an adaptive nonlinear filter constructed based on a second-order Volterra series model, and using a recursive least squares algorithm to optimize its linear and nonlinear kernel coefficients online to handle the nonlinear distortion introduced by motion interference in the adaptively filtered signal.

2. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 1, characterized in that, The timestamp alignment verification is calibrated using an interpolation method. When the time delay between the sensor data and the EEG signal is detected to be less than a preset value, the corresponding number of data points are inserted into the time-lagging data sequence to make the two types of synchronized data accurately aligned in the time dimension.

3. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 1, characterized in that, The output expression of the adaptive nonlinear filter constructed based on the second-order Volterra series model is as follows: ； in The output of the filter is the EEG signal after nonlinear compensation processing, where M is the length of the linear term sliding window and L is the length of the quadratic term feature window. Here are the weight coefficients for the linear terms, i and j are the coefficient indices, and n is the sampling point number; Let be the sampled value of the input signal at time ni. H represents the sampled value of the input signal at time nj. i H represents the second-order weighting coefficient of the nonlinear term with coefficient index i. j For the second-order weighting coefficient of the nonlinear term with coefficient index j; The Volterra kernel coefficients were optimized using the recursive least squares algorithm. H i and H j .

4. The method for removing EEG signal artifacts according to claim 1, characterized in that, The abnormality handling process further includes ARIMA processing of the processed EEG signals, including: The processed EEG signal is subjected to bandpass filtering and baseline drift removal; then a stationarity test is performed to determine the ARIMA model parameters, and the ARIMA model is trained to learn the temporal characteristics of normal EEG signals. The trained ARIMA model is used to predict the expected output value of the EEG signal, and the deviation between the expected output value of the EEG signal and the actual EEG signal value after adaptive filtering and nonlinear compensation is detected. If the deviation exceeds a preset range, the signal is judged to have artifact segments, and the artifact segments are corrected based on the expected output value of the EEG signal.

5. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 1, characterized in that, An adaptive NLMS filter with a dynamically adjustable step size factor is used to filter the currently acquired EEG signal, including: initialization phase, iterative calculation phase, and convergence judgment phase; The initialization phase includes: setting the sliding window length and step size factor, and initializing the filter coefficient vector to a zero vector; preparing the input signal and the desired response signal, wherein the input signal and the desired response signal have the same time and sampling frequency; The iterative calculation stage includes: in the nth iteration, extracting a data segment with the same length as the sliding window from the input signal and reversing its order to obtain a new input signal; calculating the filter output by dot product operation based on the current filter coefficient vector and the new input signal; comparing the filter output with the expected response signal to obtain the error vector; and updating the filter coefficient vector according to the NLMS algorithm rules, combining the error vector, the new input signal, and the dynamically adjusted step size factor. The convergence judgment stage sets convergence judgment conditions. If the convergence conditions are met, the iteration process ends and the final filter coefficient vector and the filtered signal are output. If the conditions are not met, the iteration calculation stage is returned and the next iteration continues, continuously optimizing the filter coefficient vector until convergence.

6. The method for removing artifacts from electroencephalogram (EEG) signals according to claim 1, characterized in that, The basis for the dynamic adjustment is: in the signal stable segment, increase the step size to speed up the convergence speed of the filter; in the region of signal abrupt change or severe noise interference, decrease the step size to improve the stability and accuracy of the filter. Among them, when the rate of change of the amplitude of the EEG signal at M1 consecutive sampling points is less than the set threshold P1, and the fluctuation of the power spectral density in the α band is less than Q1dB, the segment is judged to be a signal stationary segment. If, within M2 sampling points, the rate of change of EEG signal amplitude exceeds the preset high threshold P2, or the change of power spectral density in a certain frequency band exceeds Q2dB, or the kurtosis value of the signal exceeds R% of the standard kurtosis value, then it is judged as a region of signal mutation or severe noise interference.

7. A brainwave signal acquisition system, characterized in that, include: An external trigger signal generator is used to generate a unified trigger signal, which is simultaneously transmitted to the EEG signal acquisition module and the sensor data acquisition module, triggering the acquisition of sensor data and EEG signals of each lead of the subject at the same time. The EEG signal acquisition module is used to acquire EEG signals from each lead of the subject. The sensor data acquisition module is used to collect sensor data from the subject. The data processing unit is used to perform timestamp alignment verification on the two types of synchronous data, namely the sensor data and the electroencephalogram (EEG) signal, and to perform the following data analysis. The data analysis includes: determining a motion state quantification index based on the aligned and verified sensor data; if the motion state quantification index does not exceed a preset motion threshold, then outputting the collected EEG signal; if the motion state quantification index exceeds the preset motion threshold, triggering an anomaly handling process and outputting the EEG signal processed by the anomaly handling process; wherein, the anomaly handling process sequentially performs adaptive filtering and nonlinear compensation on the EEG signal. The adaptive filtering includes: using pre-acquired still EEG data of the subject without motion interference as reference input, and employing an adaptive NLMS filter with dynamically adjustable step size factor to filter the currently acquired EEG signal to suppress the linear part of motion artifacts and base noise; the nonlinear compensation includes: employing an adaptive nonlinear filter constructed based on a second-order Volterra series model, and using a recursive least squares algorithm to optimize its linear and nonlinear kernel coefficients online to handle the nonlinear distortion introduced by motion interference in the adaptively filtered signal.

8. The EEG signal acquisition system according to claim 7, characterized in that, The EEG signal acquisition module includes a trigger interface, a trigger selection unit, an acquisition engine, and an ADC data FIFO for the EEG signal acquisition chip. The trigger selection unit of the EEG signal acquisition chip is used to receive the trigger signal sent by the external trigger signal generator, parse the trigger event type, drive the acquisition engine to activate the internal ADC channel to start synchronously acquiring EEG signals, and store the acquired ADC data into the ADC data FIFO. The sensor data acquisition module includes a sensor trigger interface, a sensor trigger selection unit, an internal sensor group, and a sensor data FIFO. The sensor trigger selection unit is used to receive the trigger signal sent by the external trigger signal generator, identify the trigger event, drive the internal sensor group to start collecting sensor data from the subject, and store the collected data in the sensor data FIFO. The data processing unit is used to read the ADC data FIFO and sensor data FIFO through a hardware or software interface, and to perform the data analysis.

9. The EEG signal acquisition system according to claim 8, characterized in that, The sensor group includes an IMU sensor, which includes at least one of an accelerometer and a gyroscope. The accelerometer is used to collect acceleration data, and the gyroscope is used to collect angular velocity data.

Images