Personalized model training and common feature extraction system based on multi-channel EEG

Abstract

The invention relates to the technical field of brain-computer interfaces, and discloses a multi-channel EEG-based personalized model training and common feature extraction system, which comprises a three-channel EEG sensor electrode, a triaxial accelerometer, an electro-oculogram auxiliary electrode, a signal conditioning and acquisition circuit module, a local processing unit, a wireless communication module and a power management module. The invention also provides a tag generation method of the multi-channel EEG-based personalized model training and common feature extraction system, and the method comprises the following steps: S1, multi-modal signal synchronous collection: through cooperation of all parts, bio-electricity signals and behavior signals of a user in different states can be monitored and captured in real time, and the multi-modal signal synchronous collection is realized; in addition, the local model can actively align cross-user feature distribution while optimizing task performance, inter-domain differences are effectively reduced, and the generalization ability of the global model is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of brain-computer interface, in particular to a personalized model training and common feature extraction system based on multi-channel EEG. BACKGROUND

[0002] Brain-computer interface technology aims to establish a direct communication link between the human brain and external devices, and is widely used in medical rehabilitation, neural regulation, intelligent interaction and other fields. Non-invasive EEG is the mainstream research direction due to its high safety and convenient operation, but the core challenge is how to achieve high-precision signal analysis and user intention recognition under limited channel conditions.

[0003] At present, commercial brain-computer interface systems generally adopt a high-density EEG acquisition scheme (>=64 channels), combined with a professional shielding environment and a precision amplifier, to capture the subtle potential changes of the cerebral cortex. Typical applications of such systems include: (1) clinical diagnosis of neurological diseases; (2) consciousness task decoding in laboratory environment; (3) brain function atlas construction in high-end scientific research scenarios. The technical principle is based on the rich topological information brought by high-density spatial resolution, and through complex feature engineering such as frequency domain power spectrum analysis and spatial filtering, signal interpretation is achieved.

[0004] However, such a scheme has the following defects:

[0005] ①Hardware is bulky: heavy electrode cap needs to be worn, wiring is complex, and it is difficult to meet the portable needs of daily scenarios;

[0006] ②Environment-dependent: susceptible to electromagnetic interference, requires professional operators to assist in debugging;

[0007] ③Poor real-time performance: there is a second-level delay from signal acquisition to decision output, which cannot support dynamic task interaction.

[0008] The simplified version of BCI devices launched for the consumer market has been reduced to 8-16 channels, but still follows the traditional paradigm of pattern matching through pre-recorded standard action templates, lacking the ability to adapt to individual physiological differences, resulting in a significant decrease in recognition accuracy over time. Therefore, a personalized model training and common feature extraction system based on multi-channel EEG is proposed. SUMMARY

[0009] (I) Technical problems solved

[0010] In view of the deficiencies of the prior art, the present application provides a personalized model training and common feature extraction system based on multi-channel EEG, which solves the problems of hardware bulkiness, strong environment dependence, poor real-time performance, pattern matching through pre-recorded standard action templates, lack of adaptability to individual physiological differences, and significant decrease in recognition accuracy over time.

[0011] (II) Technical Solution

[0012] To achieve the above object, the present application provides the following technical solutions:

[0013] The personalized model training and common feature extraction system based on multi-channel EEG comprises the following components:

[0014] The three-channel EEG sensor electrode, three-axis accelerometer, auxiliary electrode for electrooculogram, signal conditioning and acquisition circuit module, local processing unit, wireless communication module and power management module; the signal conditioning and acquisition circuit module comprises a preamplifier, band-pass filter, notch filter and analog-to-digital converter;

[0015] The bioelectric signals collected by the three-channel EEG sensor electrode and the auxiliary electrode for electrooculogram are first amplified and filtered by the preamplification and filtering circuit, and the analog signals processed are converted into digital signals by the analog-to-digital converter and transmitted to the microcontroller; at the same time, the three-axis accelerometer directly communicates with the microcontroller through the I 2 C or SPI interface to realize the collection of motion data; the microcontroller integrates the two types of data and uploads them to the cloud server through the wireless communication module, during which all modules are connected in common and shielded to reduce electromagnetic interference.

[0016] The local processing module amplifies and filters the bioelectric signals before analog-to-digital conversion, and calibrates, filters and data compression pre-processes the motion data before communication and integration with the microcontroller

[0017] Further, the electrode positions of the three-channel EEG sensor electrode are arranged according to the international 10-20 system: FP1 is the left frontal pole region, FP2 is the right frontal pole region, and Cz is the top of the central region; dry electrodes or semi-dry electrodes are used, the material is silver, silver chloride or conductive rubber, and the impedance is controlled within the range of 10kΩ-50kΩ; the electrodes are connected to the signal conditioning circuit module through flexible wires or PCB traces.

[0018] On the basis of the foregoing scheme, the three-axis accelerometer is installed at the central position of the headband, close to the position of the Cz electrode, for detecting the acceleration changes in the forward, left, right and upward directions of the head; the sampling frequency is set to 50Hz-100Hz, and the dynamic range is ±2g; the output signal is used to identify head nodding, shaking and other fatigue-related actions.

[0019] As a further scheme of the present application, the auxiliary electrode for electrooculogram is arranged in a single channel below the lateral canthus of the left eye, and the reference electrode shares FP1 or FP2; it is used to capture horizontal or vertical eye movement and blinking signals; the sampling frequency is synchronized with the EEG and is 250Hz.

[0020] Further, the signal conditioning and acquisition circuit module comprises a preamplifier with a gain of 100-1000 times; a band-pass filter with a frequency of 0.3-100 Hz; a notch filter with a 50 or 60 Hz power frequency suppression; and an analog-to-digital converter with a resolution of ≥16 bits and a sampling rate of ≥250 Hz.

[0021] The EEG, accelerometer, and EOG signals are respectively conditioned through independent channels and then collected by a microcontroller in time synchronization.

[0022] On the basis of the foregoing scheme, the local processing unit adopts a low-power embedded processor with floating-point operation capability, and a built-in real-time operating system responsible for performing data preprocessing, label generation, model inference, and training tasks.

[0023] As a further scheme of the present application, the wireless communication module supports Bluetooth Low Energy and Wi-Fi protocols for uploading the locally trained model parameters to a cloud server and receiving global model updates; and the power management module is powered by a lithium polymer battery equipped with a charging management IC, with a continuous use time of ≥8 hours.

[0024] Further, a label generation method for a personalized model training and common feature extraction system based on multi-channel EEG is also proposed, comprising the following steps:

[0025] S1: Multi-modal signal synchronous acquisition, simultaneously collecting FP1, FP2, and Cz three-channel EEG signals, three-axis acceleration signals, and EOG signals; all signals are time-stamped with a unified time reference to ensure cross-modal synchronization.

[0026] S2: Event detection and label triggering, blink detection, fatigue state detection, and focused state detection.

[0027] S3: Label alignment and encoding, aligning the detected label events with the 2-second EEG data segments before and after them; each 2-second data segment is labeled with a three-digit binary encoding label.

[0028] On the basis of the foregoing scheme, the specific operations in S2 are as follows:

[0029] a. Blink detection: differential processing of the EOG signal to detect voltage mutations; if the amplitude exceeds the threshold and the duration is <0.5 seconds, it is determined as a "blink" event; output binary label blink=1, and the duration corresponds to the EEG data segment;

[0030] b. Fatigue state detection: primary judgment, when the accelerometer detects periodic nodding motion of the head in the vertical direction, i.e., frequency 0.5-1.5 Hz, amplitude >0.3g, and duration >3s, it is marked as "fatigue candidate";

[0031] Secondary verification: Simultaneously analyze whether the theta wave power of FP1 and FP2 channels increases significantly, with an increase of ≥30% relative to the baseline, and whether the alpha wave decreases; if both conditions are met, confirm that the "fatigue" label fatigue=1, otherwise remove the label;

[0032] c. Focus state detection: When the EOG signal is stable, there is no blinking, the accelerometer shows no significant movement (RMS < 0.1g), and the Cz channel α wave power is enhanced or the β wave frequency is 13–30Hz, it is marked as a "focus" state (focus = 1).

[0033] (III) Beneficial Effects

[0034] Compared with existing technologies, this invention provides a personalized model training and common feature extraction system based on multi-channel EEG, which has the following beneficial effects:

[0035] 1. In this invention, through the collaborative cooperation of various components, the bioelectrical signals and behavioral signals of users in different states are monitored and captured in real time. Based on these data, time-synchronized tags are generated, and the data is analyzed and preprocessed by the local processing unit. Finally, the data is uploaded to the cloud through the wireless communication module for further data training and model updates, thereby achieving personalized feature extraction and accurate modeling.

[0036] 2. This invention enables the local model to actively align with cross-user feature distributions while optimizing task performance, effectively reducing inter-domain differences and improving the generalization ability of the global model.

[0037] 3. By limiting the range of personalized fine-tuning, it retains user specificity, prevents overfitting, and accelerates convergence, making it suitable for deployment on edge devices.

[0038] 3. This invention synchronously collects and processes multimodal information such as EEG signals, accelerometer data, and EOG signals, enabling preliminary data processing and model training to be completed locally, significantly reducing reliance on a central server. At the same time, this system achieves low-latency interaction with the cloud server through Bluetooth Low Energy and Wi-Fi protocols, improving the system's flexibility and application breadth, and meeting the needs for EEG data processing and analysis in different scenarios. Attached Figure Description

[0039] Fig. 1 This is a schematic diagram of the structure of the personalized model training and common feature extraction system based on multi-channel EEG proposed in this invention;

[0040] Fig. 2 This is a schematic diagram of the label generation method for a personalized model training and common feature extraction system based on multi-channel EEG proposed in this invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] Example 1

[0043] Reference Figs. 1-2 The personalized model training and common feature extraction system based on multi-channel EEG includes the following components: a three-channel EEG sensor electrode, with the electrode positions arranged according to the international 10-20 system: FP1 (left of the forehead region), FP2 (right of the forehead region), and Cz (top of the central region); dry or semi-dry electrode design is adopted, with materials such as silver, silver chloride, or conductive rubber, and the impedance is controlled within the range of 10kΩ-50kΩ; the electrodes are connected to the signal conditioning circuit module through flexible wires or PCB traces.

[0044] The triaxial accelerometer is installed in the center of the headband, near the Cz electrode, to detect changes in acceleration in the forward, backward, left, right, and up / down directions of the head; the sampling frequency is set to 50Hz-100Hz, with a dynamic range of ±2g; the output signal is used to identify fatigue-related movements such as head nodding and shaking.

[0045] An electrooculogram (EOG) auxiliary electrode is arranged in a single channel below the outer canthus of the left eye, with the reference electrode sharing FP1 or FP2; it is used to capture horizontal or vertical eye movements and blink signals; the sampling frequency is synchronized with the EEG and is 250Hz.

[0046] The signal conditioning and acquisition circuit module includes a preamplifier with a gain of 100-1000 times; a bandpass filter with a frequency of 0.3–100Hz; a notch filter with 50 or 60Hz power frequency suppression; an analog-to-digital converter with a resolution ≥16 bits and a sampling rate ≥250Hz; EEG, accelerometer, and EOG signals are conditioned through independent channels and then acquired synchronously by a microcontroller (MCU) with a timestamp alignment accuracy better than ±10ms.

[0047] The local processing unit uses a low-power embedded processor with floating-point operation capabilities; it has a built-in real-time operating system (RTOS) responsible for performing data preprocessing, label generation, model inference and training tasks.

[0048] The wireless communication module supports Bluetooth Low Energy and Wi-Fi protocols. It is used to upload the locally trained model parameters to the cloud server and receive global model updates. The low-latency interaction with the cloud server is achieved through Bluetooth Low Energy and Wi-Fi protocols, which improves the system's flexibility and application breadth and meets the needs of EEG data processing and analysis in different scenarios.

[0049] The power management module supplies power to the lithium polymer battery (3.7V) and is equipped with a charging management IC, providing a battery life of ≥8 hours.

[0050] The bioelectrical signals acquired by the three-channel EEG sensor electrodes and electrooculography (EOG) auxiliary electrodes are first amplified and noise-filtered by a preamplifier and filter circuit. The processed analog signals are then converted into digital signals by an analog-to-digital converter and transmitted to the microcontroller. Simultaneously, the triaxial accelerometer transmits signals via I... 2 The C or SPI interface communicates directly with the microcontroller to acquire motion data. After integrating the two types of data, the microcontroller uploads the data to the cloud server via a wireless communication module. During this process, all modules are connected to a common ground and shielded to reduce electromagnetic interference. The local processing module amplifies and filters the bioelectrical signals before analog-to-digital conversion, and performs calibration, filtering, and data compression preprocessing on the motion data before communication and integration with the microcontroller. It synchronously acquires and processes multimodal information such as EEG signals, accelerometer data, and EOG signals, enabling preliminary data processing and model training to be completed locally, significantly reducing dependence on a central server.

[0051] This invention also proposes a label generation method for a personalized model training and common feature extraction system based on multi-channel EEG, comprising the following steps:

[0052] S1: Multimodal signal synchronous acquisition, simultaneously acquiring three-channel EEG signals (FP1, FP2, Cz), three-axis acceleration signals, and EOG signals; all signals are timestamped with a unified time base (UTC or local clock) to ensure cross-modal synchronization.

[0053] S2: Event detection and tag triggering, the specific operations are as follows:

[0054] a. Blink Detection: Differential processing is performed on the EOG signal to detect voltage changes; if the amplitude exceeds the threshold and the duration is <0.5 seconds, it is determined as a "blink" event; output binary tag blink=1, and the duration corresponds to the EEG data segment;

[0055] b. Fatigue state detection: Initial judgment: when the accelerometer detects periodic nodding motion of the head in the vertical direction, i.e., frequency 0.5-1.5Hz, amplitude >0.3g, duration >3s, it is marked as "fatigue candidate";

[0056] Secondary verification: Simultaneously analyze whether the theta wave power of FP1 and FP2 channels increases significantly, with an increase of ≥30% relative to the baseline, and whether the alpha wave decreases; if both conditions are met, confirm that the "fatigue" label fatigue=1, otherwise remove the label;

[0057] c. Focus state detection: When the EOG signal is stable, there is no blinking, the accelerometer shows no significant movement (RMS < 0.1g), and the Cz channel α wave power is enhanced or the β wave frequency is 13–30Hz, it is marked as a "focus" state (focus = 1).

[0058] S3: Tag alignment and encoding, aligning the detected tag event with the EEG data segments 2 seconds before and after it; each 2-second data segment is tagged with a three-bit binary code tag. In this invention, through the collaborative cooperation of various components, the bioelectrical signals and behavioral signals of users in different states are monitored and captured in real time. Based on these data, time-synchronized tags are generated, and the tags are analyzed and preprocessed by the local processing unit. Finally, the tags are uploaded to the cloud through the wireless communication module for further data training and model updates, achieving personalized feature extraction and accurate modeling.

[0059] Example 2

[0060] Reference Figs. 1-2 The head-mounted EEG device of the present invention is in the form of a wearable headband or headphones, and its core components include:

[0061] 1. Three-channel EEG sensor electrodes:

[0062] The electrode positions are arranged according to the international 10-20 system: FP1 (left of the frontal region), FP2 (right of the frontal region), Cz (top of the central region);

[0063] It adopts a dry electrode or semi-dry electrode design, and the material is silver / silver chloride (Ag / AgCl) or conductive rubber, with the impedance controlled in the range of 10kΩ-50kΩ;

[0064] The electrodes are connected to the signal conditioning circuit module via flexible wires or PCB traces.

[0065] 2. Triaxial accelerometer:

[0066] Installed in the center of the headband (near the Cz electrode), it is used to detect changes in acceleration of the head in the front-back, left-right, and up-down directions;

[0067] The sampling frequency is set to 50Hz-100Hz, and the dynamic range is ±2g.

[0068] The output signal is used to identify fatigue-related actions such as nodding and shaking.

[0069] 3. Electrooculography (EOG) Assist Electrode:

[0070] A single channel is positioned below the outer canthus of the left eye, with either FP1 or FP2 serving as the reference electrode;

[0071] Used to capture horizontal or vertical eye movements and blinking signals.

[0072] The sampling frequency is synchronized with the EEG and is 250Hz.

[0073] 4. Signal conditioning and acquisition circuit module:

[0074] Includes preamplifiers (gain 100-1000 times), bandpass filters (0.3–100Hz), notch filters (50 / 60Hz power frequency suppression), and analog-to-digital converters (ADCs, resolution ≥16 bits, sampling rate ≥250Hz);

[0075] The EEG, accelerometer, and EOG signals are conditioned through independent channels and then time-synchronized by a microcontroller (MCU), with timestamp alignment accuracy better than ±10ms.

[0076] 5. Local processing unit:

[0077] It employs a low-power embedded processor with floating-point operation capabilities; and has a built-in real-time operating system (RTOS) responsible for performing data preprocessing, label generation, model inference, and training tasks.

[0078] 6. Wireless communication module:

[0079] It supports Bluetooth Low Energy (BLE 5.0) or Wi-Fi protocols for uploading locally trained model parameters to a cloud server and receiving global model updates.

[0080] 7. Power Management Module:

[0081] Powered by a lithium polymer battery (3.7V), equipped with a charging management IC (such as TP4056), with a battery life of ≥8 hours.

[0082] The connection relationships between the components are as follows:

[0083] EEG / EOG electrode → preamplifier and filter circuit → ADC → MCU;

[0084] Accelerometer → I 2 C / SPI interface → MCU;

[0085] MCU → BLE / Wi-Fi module → Cloud server;

[0086] All modules share a common ground and are shielded to reduce electromagnetic interference.

[0087] II. Automatic Tag Generation Mechanism

[0088] This invention also proposes a real-time synchronous tag generation method that requires no manual intervention, the specific implementation steps of which are as follows:

[0089] Step 1: Synchronous acquisition of multimodal signals

[0090] Simultaneously acquire three-channel EEG signals (FP1, FP2, and Cz), three-axis acceleration signals, and EOG signals;

[0091] All signals are timestamped with a uniform time base (UTC or local clock) to ensure cross-modal synchronization;

[0092] Step 2: Event Detection and Tag Triggering

[0093] 1. Blink detection:

[0094] Differential processing is performed on the EOG signal to detect voltage abrupt changes.

[0095] If the amplitude exceeds the threshold (e.g., ±50μV) and the duration is less than 0.5 seconds, it is determined as a "blink" event.

[0096] Output binary tag blink=1, duration corresponds to EEG data segment.

[0097] 2. Fatigue status detection:

[0098] Preliminary assessment: When the accelerometer detects periodic head nodding motions in the vertical direction (frequency 0.5–1.5Hz, amplitude >0.3g, duration >3s), it is marked as a "fatigue candidate";

[0099] Secondary verification: Simultaneously analyze whether the theta wave (4–8Hz) power of the FP1 / FP2 channels increases significantly (≥30% increase relative to the baseline) and whether the alpha wave (8–13Hz) decreases;

[0100] If both conditions are met, then the "fatigue" label is confirmed as 1; otherwise, the label is removed.

[0101] 3. Attention Detection:

[0102] When the EOG signal is stable (no blinking), the accelerometer shows no significant movement (RMS < 0.1g), and the Cz channel α wave power is enhanced (especially when the eyes are closed) or β wave (13–30Hz) activity is enhanced (during open-eye tasks), it is marked as a "focused" state with focus = 1.

[0103] Step 3: Label Alignment and Encoding

[0104] Align the detected tag events with the EEG data segments 2 seconds before and after them;

[0105] Each 2-second data segment is labeled with a three-bit binary code (focus, fatigue, blink), for example (1,0,0) indicates focus, no fatigue, and no blinking.

[0106] The tag generation delay is controlled to <50ms, which is far lower than the 1-2 seconds of traditional manual labeling.

[0107] III. Personalized Model Training Methods

[0108] (I) Data Preprocessing Flow

[0109] 1. Bandpass filtering: The original EEG signal is processed by a digital filter to retain the 0.5–40Hz frequency band (covering δ, θ, α, β, and low γ waves). The filter type is a 4th-order Butterworth, and the cutoff frequency is adjustable from (0.3–0.8Hz) to (35–45Hz).

[0110] 2. Segmented windowing: Divide the continuous signal into 2-second non-overlapping or 50% overlapping segments, each containing 500 sampling points (at a sampling rate of 250Hz);

[0111] 3. Label alignment: The automatic labeling results are bound to the corresponding EEG segments to form (X_i, y_i) training sample pairs.

[0112] (II) Model Architecture Design

[0113] A hybrid structure of shared backbone + personalized adaptation layer is adopted:

[0114] 1. Backbone network (shared portion):

[0115] 1D-CNN layer: 3 convolutional blocks, each containing:

[0116] Kernel size: 16, 32, 64

[0117] Number of convolution kernels: 16, 32, 64

[0118] Step size: 1

[0119] Activation function: ReLU

[0120] Batch Normalization (BatchNorm) and Dropout (0.3)

[0121] The output feature map is then fed in after global average pooling:

[0122] LSTM layer: Bidirectional LSTM, 128 hidden units, capturing dynamic time series data.

[0123] 2. Personalized Adaptation Layer (User-Specific):

[0124] A fully connected layer (FC) is added after the LSTM output, with the dimension changing from 64 to 32;

[0125] The weights of this layer are initialized to the global model weights, but during local training, only this layer and subsequent classification layers (accounting for less than 5% of the total parameters) are updated, while the rest of the backbone network is frozen or the learning rate is reduced by 10 times.

[0126] Classification output layer: Softmax, outputs the probabilities of three states (focus, fatigue, blinking).

[0127] (III) Training Methods

[0128] Use the cross-entropy loss function L_task;

[0129] The optimizer is Adam (learning rate 1e-4);

[0130] Each user trains locally for 50-100 rounds, with a batch size of 32.

[0131] By limiting the range of personalized fine-tuning, user specificity is preserved, overfitting is prevented, and convergence is accelerated, making it suitable for edge device deployments.

[0132] IV. Common Feature Extraction and Federated Learning Mechanism

[0133] Federated Learning Framework Design

[0134] (1) Client: Each user's head-mounted device acts as an independent client, performing local personalized model training;

[0135] (2) Server: Deploy a federated learning aggregator in the cloud, responsible for aggregating and distributing model parameters.

[0136] (3) Parameter aggregation strategy

[0137] Let the weight of the local model of the i-th user be Wlocal(i), its local data volume be Ni, and the total data volume be N = ∑Ni.

[0138] The global model update formula is:

[0139] Wglobal=∑i=1K(NiN·Wlocal(i))Wglobal=i=1∑K(NNi·Wloc al(i))

[0140] The aggregation cycle is triggered once every 5 rounds of local training.

[0141] (4) Feature decoupling and domain adaptation module

[0142] To overcome the domain shift problem caused by physiological differences among different users, this invention introduces Maximum Mean Discrepancy (MMD) as a regularization term during the local training phase:

[0143] In the feature space output by the LSTM layer, calculate the MMD distance between the current user feature distribution and the historical global average feature distribution:

[0144] MMD2=∥1n∑i=1nφ(xilocal)-1m∑j=1mφ(xjglobal)∥H2MMD2=n1i=1∑nφ(xilocal)-m1j=1∑mφ(xjglobal)H2

[0145] Where φ(·) is the RKHS kernel mapping (RBF kernel is recommended), xlocal is the local feature, and xglobal is the cross-user average feature of the server cache;

[0146] The total loss function is defined as:

[0147] L=α·Ltask+β·Ldomain=α·LCE+β·MMD2L=α·Ltask+β·Ldom ain=α·LCE+β·MMD2

[0148] Where α = 1.0, β ∈ [0.1, 0.5], the preferred value is 0.3.

[0149] Workflow summary:

[0150] 1. Once the user starts using the device, the system automatically collects EEG, acceleration, and EOG signals;

[0151] 2. Generate status labels in real time and use them for local model training;

[0152] 3. Every so often (e.g., every 30 minutes), the device uploads local model parameters to the cloud;

[0153] 4. The server performs weighted aggregation and MMD regularization to generate an updated global model;

[0154] 5. The global model is distributed to all clients to initialize the next round of local training;

[0155] Through iterative cycles, we achieve synergistic evolution of "personalization + commonality".

[0156] In the description herein, it should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

Claims

1. A personalized model training and common feature extraction system based on multi-channel EEG, characterized in that, It includes the following components: a three-channel EEG sensor electrode, a three-axis accelerometer, an electrooculogram auxiliary electrode, a signal conditioning and acquisition circuit module, a local processing unit, a wireless communication module, and a power management module; the signal conditioning and acquisition circuit module includes a preamplifier, a bandpass filter, a notch filter, and an analog-to-digital converter; The bioelectrical signals acquired by the three-channel EEG sensor electrodes and electrooculography auxiliary electrodes are first amplified and noise-filtered by a preamplifier and filter circuit. The processed analog signals are then converted into digital signals by an analog-to-digital converter and transmitted to the microcontroller. Simultaneously, the triaxial accelerometer transmits the signals via I... 2 The C or SPI interface communicates directly with the microcontroller to acquire motion data. After the microcontroller integrates the two types of data, it uploads them to the cloud server through the wireless communication module. During this process, all modules are connected to a common ground and are shielded to reduce electromagnetic interference. The local processing module amplifies and filters the bioelectrical signals before analog-to-digital conversion, and performs calibration, filtering, and data compression preprocessing on the motion data before communication and integration with the microcontroller.

2. The personalized model training and common feature extraction system based on multi-channel EEG according to claim 1, characterized in that, The electrode positions of the three-channel EEG sensor electrodes are arranged according to the international 10-20 system: FP1 is the left of the frontal electrode region, FP2 is the right of the frontal electrode region, and Cz is the top of the central region. The electrodes are designed as dry or semi-dry electrodes, and the materials are silver, silver chloride, or conductive rubber. The impedance is controlled within the range of 10kΩ-50kΩ. The electrodes are connected to the signal conditioning circuit module through flexible wires or PCB traces.

3. The personalized model training and common feature extraction system based on multi-channel EEG according to claim 1, characterized in that, The triaxial accelerometer is installed in the center of the headband, near the Cz electrode, and is used to detect changes in acceleration in the front-back, left-right, and up-down directions of the head. The sampling frequency is set to 50Hz-100Hz, and the dynamic range is ±2g. The output signal is used to identify fatigue-related actions such as nodding and shaking.

4. The personalized model training and common feature extraction system based on multi-channel EEG according to claim 1, characterized in that, The single-channel electrooculography (EOG) auxiliary electrode is positioned below the outer canthus of the left eye, with reference electrodes sharing FP1 or FP2; it is used to capture horizontal or vertical eye movements and blink signals; the sampling frequency is synchronized with the EEG and is 250Hz.

5. The personalized model training and common feature extraction system based on multi-channel EEG according to claim 1, characterized in that, The signal conditioning and acquisition circuit module includes a preamplifier with a gain of 100-1000 times; a bandpass filter with a frequency of 0.3-100Hz; a notch filter with 50 or 60Hz power frequency suppression; and an analog-to-digital converter with a resolution ≥16 bits and a sampling rate ≥250Hz. The EEG, accelerometer, and EOG signals are conditioned through independent channels and then time-synchronized by the microcontroller.

6. The personalized model training and common feature extraction system based on multi-channel EEG according to claim 5, characterized in that, The local processing unit uses a low-power embedded processor and has floating-point operation capabilities; It has a built-in real-time operating system responsible for performing data preprocessing, label generation, model inference, and training tasks.

7. The personalized model training and common feature extraction system based on multi-channel EEG according to claim 1, characterized in that, The wireless communication module supports Bluetooth Low Energy and Wi-Fi protocols, and is used to upload locally trained model parameters to the cloud server and receive global model updates. The power management module is powered by a lithium polymer battery and equipped with a charging management IC, providing a battery life of ≥8 hours.

8. The label generation method for the personalized model training and common feature extraction system based on multi-channel EEG according to claim 1, characterized in that, Includes the following steps: S1: Multimodal signal synchronous acquisition, simultaneously acquiring three-channel EEG signals (FP1, FP2, Cz), three-axis acceleration signals, and EOG signals; all signals are timestamped with a unified time base to ensure cross-modal synchronization. S2: Event detection and tag triggering, including blink detection, fatigue detection, and focus detection. S3: Tag alignment and encoding, aligning the detected tag event with the EEG data segments 2 seconds before and after it; Each 2-second data segment is labeled with a three-bit binary code.

9. The label generation method for the personalized model training and common feature extraction system based on multi-channel EEG according to claim 8, characterized in that, The specific operations in S2 are as follows: a. Blink detection: Differential processing is performed on the EOG signal to detect voltage sudden changes; if the amplitude exceeds the threshold and the duration is <0.5 seconds, it is determined as a "blink" event; Output binary tag blink=1, duration corresponds to EEG data segment; b. Fatigue state detection: Initial judgment: when the accelerometer detects periodic nodding motion of the head in the vertical direction, i.e., frequency 0.5-1.5Hz, amplitude >0.3g, duration >3s, it is marked as "fatigue candidate"; Secondary verification: Simultaneously analyze whether the theta wave power of FP1 and FP2 channels increases significantly, with an increase of ≥30% relative to the baseline, and whether the alpha wave decreases; if both conditions are met, confirm that the "fatigue" label fatigue=1, otherwise remove the label; c. Focus state detection: When the EOG signal is stable, there is no blinking, the accelerometer shows no significant movement (RMS < 0.1g), and the Cz channel α wave power is enhanced or the β wave frequency is 13–30Hz, it is marked as a "focus" state (focus = 1).

Images