Low-noise motor imagery control system with few channels of hairless region electroencephalogram

Abstract

The application discloses a low-noise motor imagery control system of few-channel electroencephalogram in hairless area, and belongs to the field of brain-computer interface. The system comprises a wearable textile electrode electroencephalogram acquisition module, a real-time signal processing and deep learning inference module and a mechanical hand driving execution module. The textile electrode is arranged at four hairless positions in the application, and no conductive paste and skin pretreatment are needed. The acquisition hardware adopts a split type PCB architecture, the power supply and the signal circuit are separated and independently powered, physical isolation noise reduction is realized, a continuous same prediction count filtering method is proposed, and a control instruction is output only when the model continuously outputs the same category for multiple times. The application suppresses noise through specific electrode layout and hardware partition, and suppresses random misclassification jitter in combination with the continuous same prediction count filtering method, realizes stable closed-loop control, and is suitable for home rehabilitation of stroke.

Description

Technical Field

[0001] This invention belongs to the fields of brain-computer interface, wearable bioelectrical data acquisition, motor imagery recognition, and intelligent rehabilitation robot technology, specifically relating to a low-noise motor imagery control system for hairless areas with few channels of brain electrical activity. Background Technology

[0002] Stroke has become one of the leading causes of limb motor dysfunction worldwide. Existing brain-computer interface (BCI) systems based on motor imagery often place electrodes on the hairy areas of the scalp, which are susceptible to hair obstruction, skin impedance fluctuations, and motion artifacts, resulting in insufficient signal stability. The use of wet electrodes, in particular, faces further limitations in home rehabilitation scenarios: the conductive gel dries out during prolonged training, leading to signal attenuation; stroke patients, due to limb dysfunction, often cannot independently complete electrode placement and gel application on the contralateral scalp and must rely on assistance; furthermore, prolonged skin contact with the gel can cause allergies or hygiene problems, resulting in poor wearing comfort and reduced long-term patient compliance. To compensate for signal quality deficiencies, existing systems generally employ multi-channel designs, leading to bulky head-mounted devices, high hardware costs, and poor wearing comfort. In addition, existing wet electrodes require the use of conductive paste and skin pretreatment, while rigid dry electrodes are hard, have poor adhesion, and can cause pressure pain with prolonged wear.

[0003] Traditional EEG acquisition circuits often employ a single-board integrated design, with the power management circuit and signal acquisition circuit sharing the same ground plane. Power ripple and digital switching noise can easily couple to the analog signal path through ground loops, reducing the signal-to-noise ratio of EEG signals. Although existing technologies employ battery power or partial shielding, it is still difficult to effectively isolate electrical crosstalk between the power supply side and the signal side, limiting the stable acquisition of low-noise EEG signals.

[0004] Furthermore, most existing motor imagery brain-computer interface systems rely on offline analysis. In online applications, the non-stationary nature of EEG signals, environmental noise, and fluctuations in user attention make it difficult for offline-trained models to directly adapt to online data streams, resulting in a significant decrease in classification accuracy. Existing systems often employ single-frame inference output, and the random fluctuations in classification results can easily trigger erroneous actuators and cause movement jitter, leading to insufficient control stability and failing to meet the stability requirements of rehabilitation applications. Summary of the Invention

[0005] This application provides a low-noise motor imagery control system for hairless areas with few channels of EEG, including:

[0006] Wearable textile electrode EEG acquisition module, used to collect EEG signals of users performing motor imagery tasks;

[0007] The real-time signal processing and deep learning inference module is used to process and decode the EEG signals in real time and output control commands during the online control phase of the motor imagery task.

[0008] The robotic arm drive execution module is used to execute actions according to the control commands;

[0009] The wearable textile electrode EEG acquisition module includes four active electrodes, one reference electrode and one bias electrode. The electrodes are placed at four hairless locations on the human scalp. The module adopts a split PCB architecture, which separates the power management circuit and the signal acquisition circuit and supplies them with independent power.

[0010] Furthermore, the wearable textile electrode EEG acquisition module has four hairless locations on the human scalp: AF7, AF8, T9, and T10. The reference electrode is located at the FpZ position, and the bias electrode is located at the earlobe position. AF7 and AF8 are located in the prefrontal cortex and have executive functions, while T9 and T10 are located in the temporal lobe and have language comprehension and memory functions. All four locations can directly and effectively capture EEG features related to motor imagery.

[0011] Optionally, the AF7 and AF8 electrodes adopt a circular structure, the T9 and T10 electrodes adopt an elongated elliptical structure, and the FpZ reference electrode area is 1 cm².

[0012] Optionally, the electrodes of the wearable textile electrode EEG acquisition module are made of silver fiber conductive textile material, which contains silver and nylon material, and has a line resistance of 2.4 to 3.0 Ω / cm. Sponge is embedded inside the electrodes.

[0013] Optionally, the split PCB adopts a left and right split layout behind the head and ears; the left side is the power management PCB board, which integrates the battery and voltage regulation isolation circuit; the right side is the signal acquisition PCB board, which integrates the analog front end and microcontroller peripherals; the analog front end is arranged close to the end of the signal acquisition PCB board.

[0014] Optionally, the real-time signal processing and deep learning inference module needs to define motion imagery tasks, including three types of tasks: clenching a fist, opening a hand, and resting. Users need to perform motion imagery of the corresponding type based on visual cues.

[0015] Optionally, the real-time signal processing and deep learning inference module performs 8-30Hz bandpass filtering, 50Hz power frequency notch filtering, and Z-score normalization on the EEG signal in sequence.

[0016] Optionally, the real-time signal processing and deep learning inference module has a built-in sliding window data augmentation unit that performs overlapping sliding segmentation of EEG signals by setting the time window length and sliding step size.

[0017] Furthermore, the continuous prediction filtering stabilization decision unit discards unstable initial data from the early stages of each trial during the real-time control phase, and selects valid data from subsequent time periods for continuous sliding window inference. A threshold for consecutive identical prediction counts is set; only when multiple consecutive frames of inference output the same motion imagery category result is a valid control command determined and issued.

[0018] Optionally, the execution module is a robotic arm drive execution module, including a main control unit, a PWM drive module, a micro servo motor, and a 3D printed robotic arm body.

[0019] Optionally, the system enables real-time control; the user needs to visualize based on the prompts and tasks (clenching fist, opening hand, and resting) displayed on the PC screen; the wearable textile electrode EEG acquisition module collects EEG signals in real time, the real-time signal processing and deep learning inference module decodes the user's movement intentions online and outputs control commands, the robotic arm drive execution module executes actions in real time according to the commands, and the user obtains visual feedback by observing the robotic arm's movements and adjusts subsequent strategies.

[0020] The beneficial effects of this invention are as follows:

[0021] 1) Textile electrodes are deployed at fixed points in hairless areas with few channels to avoid obstruction by scalp hair and interference from movement friction. No conductive cream or skin pretreatment is required. Deployment is convenient and can be worn independently. It avoids the signal attenuation problem caused by the drying of wet electrode gel, making it suitable for long-term home rehabilitation for stroke patients.

[0022] 2) The left and right split distributed PCB isolated power supply architecture is adopted. The analog circuit and digital circuit are independently regulated and powered, and the strong and weak current grounds are isolated, which greatly reduces the noise and electromagnetic crosstalk of the acquisition circuit and realizes stable acquisition of low-noise EEG signals under the condition of few channels.

[0023] 3) The system integrates a continuous prediction filtering decision mechanism, combined with sliding window data augmentation and online fine-tuning through transfer learning, to effectively adapt to the non-stationary characteristics of EEG signals, suppress random misclassification in real-time reasoning, and improve the real-time performance and operational stability of the motor imagery control system. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the low-noise motor imagery control system for hairless areas with few channels of EEG in this invention.

[0025] Figure 2 This is a schematic diagram of the split-type distributed PCB hardware structure in this invention;

[0026] Figure 3 This is a schematic diagram showing the electrode placement in the hairless area of ​​the scalp in this invention;

[0027] Figure 4 This is a flowchart of the continuous prediction filtering process in this invention;

[0028] Figure 5 This is a schematic diagram of motion visualization and closed-loop control in this invention. Detailed Implementation

[0029] To more clearly describe the inventive objectives, advantages, and technical solutions of this invention, the invention will be described in detail below with reference to the accompanying drawings and implementation methods. However, it should be understood that the implementation methods described below are only for explaining the invention and are not intended to limit the invention. The invention will be described in detail below, and through these details, those skilled in the art can fully understand the invention.

[0030] In this embodiment, as Figure 1 As shown, the wearable textile electrode EEG acquisition module is based on the ADS1299 analog front-end design, which integrates a high-resolution ADC, built-in programmable gain, and a wide sampling rate. The module uses an ESP32 as the microcontroller and Wi-Fi module. The PCB design includes an onboard UART debugging system and a battery charging system with power regulation. The module is powered by a 3.7V lithium-ion battery, with three independent regulators supplying power to the analog and digital components respectively. The front-end uses a ±2.5V bipolar power supply, while the digital power supply uses a separate 3.3V supply, isolated from the ESP32's 3.3V power supply to reduce noise. The UART system is powered via USB during use to minimize power consumption.

[0031] Furthermore, such as Figure 2 As shown, the entire system is integrated onto two separate PCBs to achieve weight balance between the left and right sides of the headband and to prevent noise interference from the analog front-end from the boost regulator and power supply system. These two PCBs are designed to be positioned on either side of the head, forming a circle behind the ears. The left PCB contains all the power supply components, including the battery, battery charging circuitry, and voltage regulator. The right PCB contains the rest of the system and positions the analog front-end at one end of the PCB to reduce electromagnetic interference.

[0032] The electrodes are hand-embroidered using conductive silver fiber thread. This thread is composed of 18% silver and 80% nylon, with a resistance of 2.4-3.0 Ω / cm. The embroidery method involves starting from the center and working outwards in a circular pattern to the outer perimeter of the electrode. The end of the EEG cable is stripped open, and the exposed wire is finely unfolded on the inner side and embroidered simultaneously with the thread, ensuring good contact and integration with the outer surface of the textile electrode. Sponge is embedded inside the electrode to generate pressure for better electrode-skin contact and comfort.

[0033] like Figure 3As shown, this embodiment uses the electrode placement method of the international 10-20 system in the hairless area, employing a total of six electrodes: four active electrodes, one reference electrode, and one bias electrode. The active electrodes are placed at positions AF7, AF8, T9, and T10. The reference electrode is placed at position FpZ. The textile electrodes at positions AF7 and AF8 are circular with an area of ​​1.8 cm²; the textile electrodes at positions T9 and T10 are elongated elliptical with an area of ​​2.8 cm², to accommodate different head shapes. The FpZ reference electrode has a smaller area of ​​1 cm² to minimize signal attenuation to adjacent active electrodes. The bias electrode is located on the right earlobe, using an ear clip electrode for better fixation.

[0034] like Figure 4 As shown, the real-time signal processing and deep learning inference module includes:

[0035] S410: Set up motor imagery tasks and complete signal preprocessing; set up three types of motor imagery tasks: open hand, clench fist, and rest, including four stages: fixation, cueing, imagery, and rest. EEG data are first filtered by 8-30Hz bandpass filtering and 50Hz notch filtering, and then Z-score normalization is applied independently to the four channels. The normalization parameters are calculated and saved only from the training data and are used to transform new samples in the fine-tuning stage and real-time control.

[0036] S420: Employs a sliding window to segment data for offline data augmentation; sets the time window length to correspond to the sampling points and the sliding step length to correspond to 125 sampling points for overlapping sliding segmentation; extracts several seconds of pure effective EEG data from a single motor imagery test, and splices them into long-time signals according to the categories of clenched fist, open hand, and resting hand. Through overlapping sliding segmentation, it simultaneously achieves real-time data stream frame-by-frame slicing and sample augmentation, providing a unified standard input sample for model training and online fine-tuning.

[0037] S430: Construct and train a one-dimensional convolutional neural network; the input shape is 500×4, containing five sequential convolutional layers, a max pooling layer, a global max pooling layer, a dropout layer with a dropout rate of 0.5, and a Softmax activation function; the model uses the Adam optimizer, combined with early stopping and learning rate decay strategies. Due to the greater variability and noise in real-time EEG signals.

[0038] S440: Based on transfer learning, it uses real-time acquired data to fine-tune the pre-trained model; it adopts a global parameter unfreezing method for offline pre-trained one-dimensional convolutional neural networks. During the online fine-tuning stage, it reduces the learning rate and adjusts the training batch; it maintains the window size, sliding step size, and signal preprocessing rules completely consistent with the offline training stage throughout the process, adapting to the non-stationary distribution shift of EEG signals in actual application scenarios and improving the accuracy of online motor imagery decoding.

[0039] S450: Real-time data stream acquisition, completing continuous prediction filtering; after entering the real-time control stage, this embodiment discards the initial unstable data in the early stage of each trial to avoid data stream transmission jitter and initial fluctuations in EEG response; selects the effective steady-state EEG data for the next few seconds, and performs continuous frame-by-frame inference according to the fixed sliding window rule, generating multiple sets of continuous prediction results per trial; sets a threshold for consecutive identical prediction counts, and only when multiple consecutive frames of inference output the same motor imagery category result is it determined as a valid control command and sent down, suppressing random misclassification in a single frame and avoiding unnecessary shaking and accidental triggering of actions by the robotic arm. S460: Execute control commands; the effective control commands are sent to the robotic arm drive execution module, which maps the command type to a preset servo angle sequence, controls the micro servo to rotate through the PWM drive module, and drives the robotic arm to perform the corresponding opening or clenching fist action; after a single command is executed, the robotic arm maintains the current posture until the next effective control command is received.

[0040] The robotic arm drive module utilizes a 3D-printed robotic arm. Each finger is driven by a pair of SG90 micro servos to simulate basic flexion and extension movements. An Arduino Due is used as the robotic arm's microcontroller, connecting to a computer via USB to receive brainwave motor imagery predictions and converting them into preset angle commands for each servo channel. These commands are transmitted via I2C to a PCA9685 16-channel PWM driver. The PCA9685 PWM driver expands the microcontroller's output capabilities and takes over time-intensive tasks related to servo control. The PCA9685 is powered by a separate 5V, 5000mAh rechargeable battery to ensure circuit stability and prevent high-current surges from damaging the microcontroller's logic circuitry.

[0041] The system enables real-time control, allowing users to control the robotic arm based on text prompts displayed on the screen while observing its feedback actions. EEG signals are transmitted in real-time via a Lab Streaming Layer (LSL) and prepared for input into the model. The online streaming data undergoes the same filtering, signal processing, and sliding window algorithms as in the offline training phase before being input into the model. Initially, the prompt text is displayed; subsequently, the subject begins to imagine and continues to control the robotic arm until the end of the trial, such as... Figure 5As shown, the first few seconds of data in each trial are discarded to allow event correlation to synchronize to its lowest peak and stabilize LSL jitter. Data for the remaining time period is extracted and continuously input into the model. Once the first window of data is filled, the model's output prediction automatically begins. This invention employs a continuous prediction counting filtering algorithm, which sends a command to the robot only when the number of consecutive predictions belonging to the same category reaches a threshold (e.g., several consecutive and identical predictions). Real-time data, prediction results, and corresponding real labels are then saved separately. Compared to existing technologies, this invention improves control accuracy, with input reference noise as low as 0.117µV RMS, and an average accuracy of 89% for three types of motion visualization tasks.

[0042] It should be noted that the continuous prediction filtering decision mechanism of this invention is independent of the specific classifier type used. The described one-dimensional convolutional neural network structure and transfer learning fine-tuning strategy are merely a preferred embodiment of this invention and are not essential for its implementation. Those skilled in the art, based on the guidance of this specification, can implement the continuous prediction filtering decision mechanism of this invention using other classifiers or without transfer learning, and these alternative embodiments all fall within the protection scope of this invention.

[0043] The above-described low-noise motor imagery control system for hairless areas with few channels of EEG, as described in this invention, achieves the technical effect of low-noise and stable control through the synergistic effect of three core features: electrode layout, split PCB, and continuous predictive filtering.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A low-noise motor imagery control system for hairless areas with few channels of EEG, the control system comprising: Wearable textile electrode EEG acquisition module, used to collect EEG signals of users performing motor imagery tasks; The real-time signal processing and deep learning inference module is used to process and decode the EEG signals in real time and output control commands during the online control phase of the motor imagery task. The robotic arm drive execution module is used to execute actions according to the control commands; The wearable textile electrode EEG acquisition module includes four active electrodes, one reference electrode and one bias electrode. The electrodes are placed at four hairless locations on the human scalp. The module adopts a split PCB architecture, which separates the power management circuit and the signal acquisition circuit and supplies them with independent power. The real-time signal processing and deep learning inference module has a built-in continuous prediction filtering stable decision unit. This unit performs continuous counting and voting on the inference results of multiple frames of the same category. Only when multiple consecutive frames output the same motion image category will a valid control command be output.

2. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 1, characterized in that, The wearable textile electrode EEG acquisition module uses only four hairless locations for electrode placement: AF7, AF8, T9, and T10 on the human scalp. The reference electrode is placed at the FpZ position, and the bias electrode is placed at the earlobe position. Among them, AF7 and AF8 are located in the prefrontal lobe and have executive functions, while T9 and T10 are located in the temporal lobe and have language comprehension and memory functions. All four locations can directly and effectively capture EEG features related to motor imagery.

3. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 2, characterized in that, The AF7 and AF8 electrodes adopt a circular structure, while the T9 and T10 electrodes adopt an elongated elliptical structure. The area of ​​the FpZ reference electrode is 1 cm².

4. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 1, characterized in that, The electrodes of the wearable textile electrode EEG acquisition module are made of silver fiber conductive textile material, which contains silver and nylon material, and has a line resistance of 2.4 to 3.0 Ω / cm. Sponge is embedded inside the electrodes.

5. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 1, characterized in that, The split PCB adopts a left-right split layout behind the head and ears, including a first PCB board and a second PCB board; the left side is the power management PCB board, which integrates the battery and voltage regulation isolation circuit; the right side is the signal acquisition PCB board, which integrates the analog front end and microcontroller peripherals; the analog front end is arranged close to the end of the signal acquisition PCB board.

6. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 1, characterized in that, The real-time signal processing and deep learning inference module needs to define motion imagery tasks, including three types of tasks: clenching a fist, opening a hand, and resting. Users need to perform the corresponding type of motion imagery based on visual cues.

7. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 1, characterized in that, The real-time signal processing and deep learning inference module sequentially performs 8-30Hz bandpass filtering, 50Hz power frequency notch filtering, and Z-score normalization on the EEG signal.

8. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 1, characterized in that, The real-time signal processing and deep learning inference module has a built-in sliding window data augmentation unit that sets the time window length and sliding step size to perform overlapping sliding segmentation of EEG signals.

9. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 7, characterized in that, The continuous prediction filtering stable decision unit discards unstable initial data in the early stage of each trial during the real-time control phase, selects effective data in subsequent time periods for continuous frame-by-frame inference, and generates multiple sets of continuous prediction results per trial. It sets a continuous identical prediction count threshold, and only when the inference outputs the same motion imagination category result for multiple consecutive frames is it determined as a valid control command and sent down, thereby suppressing random misclassification in a single frame and avoiding the robot arm shaking and triggering actions for no reason.

10. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 1, characterized in that, The execution module is a robotic arm drive execution module, which includes a main control unit, a PWM drive module, a micro servo motor, and a 3D printed robotic arm body.

11. The low-noise motor imagery control system for hairless areas with few channels of EEG according to claim 1, characterized in that, The system enables real-time control; users need to visualize based on the prompts and tasks (clenching fist, opening hand, and resting) displayed on the PC screen; the wearable textile electrode EEG acquisition module collects EEG signals in real time, the real-time signal processing and deep learning inference module decodes the user's movement intentions online and outputs control commands, the robotic arm drive execution module executes actions in real time according to the commands, and the user obtains visual feedback by observing the robotic arm's movements and adjusts subsequent strategies.

Images