Real-time adaptive electroencephalogram artifact suppression and enhancement embedded system

Abstract

The invention discloses a real-time self-adaptive electroencephalogram artifact suppression and enhancement embedded system. The system aims at overcoming the limitation of an existing portable EEG device in the aspect of artifact processing, various artifacts in EEG signals are recognized and removed in a low-power-consumption, high-precision and real-time mode, and meanwhile the quality of effective EEG signals is optimized and enhanced. One key innovation of the invention lies in breaking through the limitation of single EEG signal processing, and accurate separation and correction of artifacts are carried out through the assistance of EOG and EMG. For example, the EOG signal is used for assisting eye movement artifact removal through synchronous correlation analysis and an adaptive regression filter; the EMG signal is used for assisting myoelectricity artifact correction, and suppression is carried out through spectral analysis and dynamic adjustment of filter parameters. In addition, the system also has the functions of real-time detection and repair of electrode artifacts. According to the system, high-quality EEG signals close to offline processing are provided on portable embedded equipment with limited resources, and the system has universality and high-precision application potential and can be widely applied to a plurality of frontier fields such as medical diagnosis, brain-computer interface (BCI), cognitive ability enhancement, neurological rehabilitation training, sleep disorder monitoring and human-computer interaction.

Description

TECHNICAL FIELD

[0001] The present application relates to advanced processing techniques for electroencephalogram (EEG) signals, particularly focusing on developing an innovative embedded system for real-time, ultra-low power EEG artifact suppression and enhancement. This system provides a revolutionary artifact removal and signal optimization solution for the next generation of portable EEG devices by deeply integrating edge artificial intelligence (AI) techniques, multi-sensor signal fusion strategies, adaptive filtering algorithms, and blind source separation techniques. The core goal is to significantly improve the quality and signal-to-noise ratio (SNR) of EEG signals in complex dynamic environments, thereby exhibiting unprecedented application potential in medical diagnosis, high-end brain-computer interfaces (BCI), cognitive ability enhancement, neurorehabilitation training, sleep disorder monitoring, and human-computer interaction. The technical breakthrough of the present application aims to overcome the limitations of existing portable EEG systems in artifact processing, enabling them to truly move from the laboratory to daily life and achieve universal and high-precision applications. BACKGROUND

[0002] Electroencephalogram (EEG) is a non-invasive neurophysiological monitoring method that records the electrical activity of cortical neurons through scalp electrodes. It is a key tool for studying brain function, diagnosing neurological diseases, and implementing brain-computer interface control. It is widely used in cognitive neuroscience research, epilepsy diagnosis, sleep disorder assessment, coma depth monitoring, and as the core input signal for brain-computer interfaces (BCI). With its millisecond-level ultra-high temporal resolution, EEG signals can accurately capture the rapid changes in brain activity, making it unparalleled in revealing neural dynamics and real-time human-computer interaction.

[0003] However, in actual EEG signal acquisition processes, especially in uncontrolled daily environments or dynamic medical scenarios, the quality of EEG signals is easily severely disturbed by various physiological and non-physiological artifacts. These artifacts usually have amplitudes several or even dozens of times higher than the weak brain electrical signals, greatly reducing the signal-to-noise ratio (SNR) of EEG signals and severely distorting the authenticity of the original brain electrical information, thereby posing great challenges to subsequent data analysis, feature extraction, pattern recognition, and clinical diagnosis.

[0004] The main artifact types include but are not limited to: eye movement artifacts, electromyographic artifacts, electrocardiographic artifacts, power line interference artifacts, and electrode artifacts.

[0005] Currently, there are various methods for EEG artifact removal, mainly divided into signal processing and machine learning. However, these methods face many challenges in actual embedded and real-time applications.

[0006] Many efficient artifact removal algorithms, such as Independent Component Analysis (ICA), Principal Component Analysis (PCA), Canonical Correlation Analysis (CCA), etc., usually require batch processing of long-time data in an offline state. They need to perform complex operations such as iterative optimization or matrix decomposition, which requires high computational resources and takes a long time, making it difficult to meet the millisecond-level delay requirements of real-time applications.

[0007] Traditional complex algorithms and some early machine learning models require powerful Central Processing Units (CPUs), Graphics Processing Units (GPUs), or dedicated Digital Signal Processors (DSPs) when executed. This contradicts the design requirements of portable EEG devices for low power consumption, small size, light weight, and long battery life. Directly deploying these algorithms on resource-constrained embedded platforms often leads to slow processing speed, high energy consumption, device heating, and other problems, which seriously affect user experience and device practicality.

[0008] The type, intensity, and source of artifacts are dynamically changing in practical applications, such as varying blink frequency and varying muscle tension due to tasks or emotions. Many existing methods use fixed parameters or models, making it difficult to perform real-time and robust adaptive processing of constantly changing artifacts. This means that the same set of artifact removal parameters may not work well in different individuals, different environments, or different tasks, requiring frequent manual adjustments, which reduces the level of automation and intelligence of the system.

[0009] Most traditional artifact removal methods only rely on the EEG signal itself for processing. However, due to the overlap between the signal characteristics of different physiological artifacts (such as electrooculography and electromyography) and real brain electrical signals, it is sometimes difficult to accurately distinguish and effectively separate artifacts from brain electrical activity based solely on internal information from the EEG signal. For example, high-frequency brain electrical activity caused by certain cognitive tasks may be confused with electromyographic artifacts, leading to excessive removal or incomplete removal.

[0010] Many artifact removal methods, especially those based on signal decomposition, require operators to manually identify and select artifact components for removal based on the topological graph, time series, and spectral characteristics of the components after separating independent components. This not only consumes time and effort, but also highly depends on the experience and expertise of the operator, introducing subjectivity and limiting automation and popular application.

[0011] In view of the above challenges, the current technology urgently needs an innovative solution that can overcome these limitations. This solution should be able to achieve real-time, efficient, and adaptive suppression of various EEG artifacts on low-power embedded platforms, while simultaneously enhancing the quality of useful EEG signals to meet the growing demand for portable, ubiquitous, and high-precision EEG monitoring applications. SUMMARY

[0012] The present invention aims to address the many challenges faced by real-time, low-power, and high-precision EEG artifact suppression and enhancement in the prior art. The present invention provides an innovative real-time adaptive EEG artifact suppression and enhancement embedded system that builds a unified framework for collaborative work and intelligent decision-making through deep integration of edge artificial intelligence (AI) technology, multi-modal biosignal sensor fusion strategy, adaptive signal filtering, and advanced source separation algorithms. The core advantage lies in its ability to accurately and real-time identify and remove various artifacts in EEG signals, while simultaneously optimizing and enhancing the quality of effective EEG signals, ensuring ultra-low power and ultra-low latency real-time processing in resource-constrained portable EEG devices, greatly improving the usability and reliability of EEG data.

[0013] The working principle of the present invention is as follows:

[0014] 1. Edge AI-driven lightweight deep learning artifact identification and restoration model

[0015] The present invention deploys a highly optimized and lightweight deep learning model within the embedded device. This model can analyze input raw EEG data streams in real-time and efficiently, accurately identifying and classifying various main types of artifacts.

[0016] 2. Multi-sensor signal fusion and collaborative correction mechanism

[0017] A key innovation of the present invention is to break through the limitations of single EEG signal processing and introduce a miniature multi-modal biosensor fusion strategy. In addition to the core EEG sensor array, the system also integrates miniature electrooculography (EOG) sensors and miniature electromyography (EMG) sensors. This multi-sensor fusion method provides rich and complementary information for accurate artifact identification, source localization, and collaborative correction, significantly improving the accuracy and reliability of artifact removal.

[0018] 3. Collaborative optimization of adaptive filtering and advanced source separation algorithms

[0019] The application combines traditional and proven signal processing methods (such as Adaptive ICA, Regression Filtering, Adaptive Notch Filtering, etc.) with deep learning technology to form a dynamic adjustment and collaborative optimization artifact removal strategy.

[0020] Adaptive parameter adjustment under deep learning guidance: The application innovatively uses the real-time identification results of artifact type, intensity, and dynamic characteristics by edge AI model to dynamically adjust the parameters of these traditional algorithms. For example, the AI model can intelligently select the order, step size of the adaptive filter, or the convergence threshold and learning rate of ICA according to the type and amplitude of the current artifact. This AI-driven parameter adaptability enables the system to adapt to changing artifact environments in real time, ensuring optimal artifact suppression in different situations and avoiding the tediousness and inefficiency of manual intervention.

[0021] The system can adopt a hierarchical artifact suppression strategy. First, coarse-grained artifact detection and classification are performed through multi-sensor fusion and AI model; second, adaptive regression filtering or adaptive ICA is used for separation and removal of core artifacts (such as electrooculogram and electromyogram); finally, specific artifacts (such as power frequency interference and baseline drift) are removed by precise adaptive filters for fine processing. This hierarchical collaborative mechanism ensures the comprehensiveness and efficiency of artifact removal. The application not only focuses on artifact suppression, but also emphasizes the enhancement of useful brain electrical signals. By accurately removing artifacts, the signal-to-noise ratio is improved, which itself is a kind of signal enhancement.

[0022] 4. Special hardware acceleration unit and low-power real-time processing architecture

[0023] To meet the stringent requirements of portable EEG devices for ultra-low power consumption and real-time performance, the application introduces a special hardware acceleration unit in the system architecture. This is the key to achieving efficient operation of the system's core functions on an embedded platform: the hardware acceleration unit can be implemented using a Field-Programmable Gate Array (FPGA) or an Application-Specific Integrated Circuit (ASIC).

[0024] FPGA: With the advantages of reconfigurability and strong parallel processing capability, it is suitable for rapid prototyping, algorithm iteration, and small batch production. Core signal processing algorithms (such as matrix multiplication, convolution operation, FFT, etc.) and the core logic of adaptive filters are directly mapped to hardware logic gates, achieving high parallelism.

[0025] ASIC: Once the algorithm and architecture are determined, ASIC can provide higher performance, lower power consumption and smaller size, which is the ideal choice for final mass production. It directly "hard codes" the algorithm into the chip through customized circuit design, achieving extreme efficiency.

[0026] Through the organic combination of the above innovations, the present application provides an unprecedented real-time adaptive EEG artifact suppression and enhancement solution, which can provide high-quality EEG signals close to offline processing on portable and low-power embedded devices, laying a solid foundation for the wide application of EEG technology in clinical diagnosis, BCI, cognitive science and other fields. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is the overall structure flowchart of the present application, which shows the relationship between the sensor module, the analog front-end and signal conditioning unit, the edge AI processing unit and the hardware acceleration unit, and the output module, and the signal transmission direction.

[0028] Figure 2 is a module display of the present application, which shows the composition of the edge computing module and the adaptive and source separation module and the relationship between the small modules. DETAILED DESCRIPTION

[0029] The embedded system of the present application is composed of the following core modules, which work together to form a complete signal acquisition, processing and output closed loop:

[0030] Sensor module: This is the input end of the system, responsible for high-precision acquisition of physiological electrical signals.

[0031] EEG sensor array: contains multiple high-precision, low-noise dry electrodes or wet electrodes, distributed in key areas of the scalp (such as the international 10-20 or 10-10 system), used to acquire the potential activity of the cerebral cortex. The sensor should have high common mode rejection ratio (CMRR) and low input noise to ensure the purity of the original signal. Active electrode design can be used, with built-in preamplifier and buffer circuit to reduce input impedance and transmission noise.

[0032] EOG sensor: contains at least 2 or 4 microelectrodes, usually placed above, below or on both sides of the eyes (such as the outer corner of the eye), used to capture the potential changes produced by eye movement (horizontal and vertical directions) and blinking. These electrodes are co-located with EEG electrodes to ensure timing synchronization and signal correlation.

[0033] EMG sensor: contains at least 2 or 4 microelectrodes, usually placed in areas prone to muscle artifact, such as the temple, forehead, neck muscles or masseter, used to monitor the electrical signals produced by muscle activity.

[0034] All sensors are synchronously collected by high-precision, multi-channel analog-to-digital converters (ADCs), ensuring the accurate alignment of the time stamps of EEG, EOG, and EMG signals, laying the foundation for subsequent multi-sensor fusion processing. The ADCs should have high sampling rates (e.g., at least 250 Hz for EEG, up to 1000 Hz or higher) and high bit resolution (e.g., 16 or 24 bits) to capture subtle changes in signals and provide a large enough dynamic range.

[0035] Analog front-end and signal conditioning unit:

[0036] Pre-amplifier: The first stage of amplification for weak bioelectric signals, usually with a gain of hundreds to thousands of times.

[0037] Band-pass filter: Filters out DC drift and high-frequency noise (such as RF interference).

[0038] Notch filter: Selectively filters out power frequency interference (50 / 60 Hz), which can use hardware fixed filters or software configurable adaptive notch.

[0039] Edge AI processing unit:

[0040] This is the core intelligent processing module of the system, responsible for deploying and running lightweight deep learning models.

[0041] It is equipped with embedded AI accelerators (such as NPU, Neural Processing Unit), such as processors based on ARM Cortex-M series or RISC-V architecture, combined with specially optimized AI inference engines. These processors provide sufficient computing power while maintaining extremely low power consumption. Pre-trained, pruned, and quantized deep learning models are deployed to this unit. The model receives synchronized digital signals (EEG, EOG, EMG) from the sensor module for real-time inference.

[0042] The model analyzes the input data, identifies the time-domain, frequency-domain, and spatial features of artifacts, and classifies them into eye movement, electromyography, power line, or other types. Not only does it identify the presence or absence of artifacts, but it also assesses their strength and impact range, providing a basis for subsequent adaptive processing.

[0043] Hardware acceleration unit:

[0044] Designed specifically for accelerating computationally intensive algorithms, achieving ultra-low latency and high energy efficiency.

[0045] FPGA Implementation: Suitable for flexible development and low-to-medium batch production. The core iterative calculations of Fast Fourier Transform (FFT), Independent Component Analysis (ICA), and iterative updates of adaptive filters (such as LMS, RLS algorithms) can be directly integrated into the logic units of FPGA. The parallel programmable logic array of FPGA can realize large-scale parallel computing, significantly improving processing speed.

[0046] ASIC Implementation: Suitable for mass production, providing the highest performance and lowest power consumption. Through customized design, the entire artifact suppression and enhancement algorithm calculation process is fixed on the silicon chip, achieving extreme efficiency and integration.

[0047] Adaptive Filtering and Source Separation:

[0048] According to the artifact type and intensity information provided by the AI unit, dynamically adjust the parameters of the adaptive filter (such as step size, filter order) to track and suppress dynamic changes in artifacts in real time.

[0049] Using adaptive ICA or Non-negative Matrix Factorization (NMF) algorithms, combined with prior information from AI models and multi-sensor data, EEG signals are decomposed into independent or sparse source signals, and the components corresponding to artifact sources are accurately identified and removed. After artifact removal, the pure EEG signal is further optimized, for example, through denoising algorithms to improve the signal-to-noise ratio, or through feature extraction modules to prepare data for subsequent applications.

[0050] Output Module:

[0051] Output the processed high-quality EEG signal to external devices or for local storage. Such as Bluetooth, Wi-Fi or UWB (Ultra Wide Band), used to transmit processed EEG data in real time to smartphones, tablets, PCs or other cloud platforms for display, storage or further analysis. Or SD card or eMMC, used to cache or record data when there is no external connection.

[0052] 2. Artifact accurate separation and correction based on multi-sensor fusion

[0053] Multi-sensor fusion is the key to the invention's ability to achieve high-precision artifact removal.

[0054] EOG-assisted eye movement artifact removal:

[0055] Synchronization correlation analysis: EOG signals have high synchronization and correlation with eye movement artifacts in frontal and prefrontal EEG channels. The system calculates the correlation coefficient between EOG and EEG channels in real time, and uses AI models to identify the occurrence period of eye movement artifacts. The EOG signal is used as a reference input, and the linear or nonlinear contribution of EOG to EEG is learned through adaptive regression filters (such as LMS or RLS algorithm), and is subtracted from the EEG signal. AI models can help determine the optimal order and learning rate of the regression model. For transient and high amplitude blinking artifacts, the system can use EOG signals to accurately mark blinking events, and then use AI models for interpolation or reconstruction-based signal recovery to avoid direct truncation or removal of data.

[0056] EMG-assisted electromyographic artifact correction:

[0057] EMG artifacts are usually manifested as high-frequency broadband noise (above 20 Hz). The system accurately locks the frequency range and intensity of electromyographic artifacts through spectral analysis of EEG and EMG signals combined with AI identification. The energy or features of EMG signals are used as control parameters for adaptive filters to dynamically adjust the cutoff frequency or gain of the filter to suppress electromyographic artifacts in the corresponding frequency band.

[0058] Detection and repair of electrode artifacts:

[0059] The system monitors electrode impedance in real time. When the impedance is abnormal, the AI model can make a judgment combined with signal features such as baseline drift, high amplitude spikes or signal loss. After identifying the abnormal electrode channel, the system can selectively interpolate the channel (such as spherical spline interpolation) or temporarily ignore the channel according to the AI model's suggestion to avoid contaminating other channels.

[0060] The above-mentioned specific embodiments of the present application realize real-time, intelligent and high-precision suppression and enhancement of electroencephalogram artifacts, and are completely suitable for resource-constrained portable embedded devices. It not only significantly improves the quality of EEG signals, but also provides a solid technical foundation for future brain-computer interface, neurorehabilitation, cognitive enhancement and other fields of innovative applications.

Claims

1. An embedded system for real-time adaptive EEG artifact suppression and enhancement, characterized in that, include: The sensor module is used to acquire physiological electrical signals with high precision. The sensor module includes: an electroencephalogram (EEG) sensor array for acquiring the potential activity of the cerebral cortex; a miniature electrooculogram (EOG) sensor for capturing potential changes generated by eye movements and blinking; a miniature electromyogram (EMG) sensor for monitoring electrical signals generated by muscle activity; and a multi-channel analog-to-digital converter (ADC) for simultaneously acquiring EEG, EOG, and EMG signals and performing precise timestamp alignment. The analog front-end and signal conditioning unit are connected to the sensor module and are used to pre-amplify, band-pass filter and notch filter the physiological electrical signal; The edge AI processing unit, connected to the analog front-end and signal conditioning unit, is used to deploy and run lightweight deep learning models, analyze the input raw EEG, EOG, and EMG data streams in real time, identify and classify various major types of artifacts, and assess their strength and impact range. A hardware acceleration unit, connected to the edge AI processing unit, is used to accelerate computationally intensive algorithms, achieving ultra-low latency and high energy efficiency. The hardware acceleration unit can be implemented using a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC). The adaptive filtering and source separation module is connected to the edge AI processing unit and the hardware acceleration unit. Based on the artifact type and intensity information provided by the edge AI processing unit, it dynamically adjusts the parameters of the adaptive filter and uses algorithms such as adaptive independent component analysis (ICA) or non-negative matrix factorization (NMF) in combination with the prior information of the AI ​​model and multi-sensor data to decompose the EEG signal into independent or sparse source signals, and identify and remove the components corresponding to the artifact sources. The output module, connected to the adaptive filtering and source separation module, is used to output the processed high-quality EEG signal to an external device or for local storage.

2. The real-time adaptive EEG artifact suppression and enhancement embedded system according to claim 1, characterized in that: The EEG sensor array uses high-precision, low-noise dry or wet electrodes and features high common-mode rejection ratio (CMRR) and low input noise.

3. The real-time adaptive EEG artifact suppression and enhancement embedded system according to claim 1, characterized in that: The EOG sensor contains at least two or four microelectrodes placed above, below, or to the sides of the eyes; the EMG sensor contains at least two or four microelectrodes placed near the temporalis, frontalis, neck muscles, or masseter muscles.

4. The real-time adaptive EEG artifact suppression and enhancement embedded system according to claim 1, characterized in that: The ADC has a sampling rate of not less than 1000 Hz and a high bit resolution of not less than 12 bits.

5. The real-time adaptive EEG artifact suppression and enhancement embedded system according to claim 1, characterized in that: The edge AI processing unit is equipped with an embedded AI accelerator, such as a processor based on the ARM Cortex-M series or RISC-V architecture, and is paired with a specially optimized AI inference engine.

6. The real-time adaptive EEG artifact suppression and enhancement embedded system according to claim 1, characterized in that: The adaptive filtering and source separation module employs a hierarchical artifact suppression strategy. First, it performs coarse-grained artifact detection and classification through multi-sensor fusion and AI models. Second, it uses adaptive regression filtering or adaptive ICA to separate and remove core artifacts. Finally, it uses precise adaptive filters to refine the processing of remaining specific artifacts.

7. The real-time adaptive EEG artifact suppression and enhancement embedded system according to claim 1, characterized in that: The system uses EOG signals to assist in the removal of eye-movement artifacts. This includes real-time calculation of the correlation coefficient between the EOG channel and the EEG channel, using an AI model to identify the time periods in which eye-movement artifacts occur, using the EOG signal as a reference input, learning the contribution of EOG to EEG through an adaptive regression filter, and subtracting it from the input.

8. The real-time adaptive EEG artifact suppression and enhancement embedded system according to claim 8, characterized in that: The system uses an AI model to determine the optimal order and learning rate of the regression model.

9. The real-time adaptive EEG artifact suppression and enhancement embedded system according to claim 1, characterized in that: The system uses EMG signals to assist in the correction of electromyographic artifacts. This includes performing spectral analysis on EEG and EMG signals, combining AI recognition to accurately pinpoint the frequency range and intensity of electromyographic artifacts, and using the energy or characteristics of EMG signals as control parameters for an adaptive filter to dynamically adjust the filter's cutoff frequency or gain in order to suppress electromyographic artifacts in the corresponding frequency band.

10. The real-time adaptive EEG artifact suppression and enhancement embedded system according to claim 1, characterized in that: The system monitors electrode impedance in real time. When the impedance is abnormal, the AI ​​model makes a judgment based on the signal characteristics and selectively performs channel interpolation or temporarily ignores the abnormal channel.

Images