Extreme low channel eeg decoding training platform and physical control method based on domestic heterogeneous computing power

Abstract

This invention discloses a method and platform for single- and dual-channel EEG decoding and physical device control based on heterogeneous computing power, relating to the field of brain-computer interfaces. Addressing the shortcomings of existing systems, such as reliance on multi-channel acquisition, limited computing power, and lack of dynamic safety assessment in control, this invention trains a cross-paradigm neural decoding large-scale model based on a domestically produced NPU heterogeneous architecture. Under simplified single- and dual-channel input, it extracts weak EEG features by combining knowledge distillation and spatiotemporal attention mechanisms. The decoding intent confidence is compared with a dynamic threshold generated by an MLP network based on parameters such as risk level. Finally, it converts the decoded intent into machine instructions through a hierarchical instruction interface to control external devices (such as XR, wheelchairs, etc.). This invention overcomes the bottleneck of missing spatial information in low-channel environments, significantly reducing hardware costs while achieving calibration-free high-precision decoding and a low-latency, highly reliable physical collision avoidance control closed loop.

Description

Technical Field

[0001] This invention belongs to the field of brain-computer interface and artificial intelligence computing, specifically relating to a method and platform for single and dual-channel EEG decoding and physical device control based on heterogeneous computing power. Background Technology

[0002] Brain-computer interface (BCI) technology, as a novel human-computer interaction paradigm that directly connects the human brain to the external physical world, shows disruptive application prospects in medical rehabilitation (such as motor assistance for patients with high-level paraplegia), extended reality (XR) spatial computing, and next-generation consumer electronics. Non-invasive BCIs based on scalp electroencephalography (EEG) have become the most commercially viable technology due to their non-invasiveness and high safety. However, as applications move from the laboratory to the real physical world, existing technologies have revealed serious limitations, specifically in the following four aspects: The physical contradiction between minimalist hardware form and decoding accuracy: Traditional high-precision intent decoding systems typically rely on 32-channel or even 64-channel wet electrode EEG caps. These devices are not only expensive in terms of hardware, but also require professionals to inject conductive gel into the subject's scalp to reduce contact impedance. During prolonged wear, the drying of the conductive gel leads to a sharp deterioration in signal quality, completely failing to meet the demands of consumer-grade scenarios (such as everyday XR headsets). However, when the acquisition channels are forcibly reduced to single-channel or dual-channel (e.g., retaining only frontal or occipital lobe nodes) and dry electrodes are used, the spatial topological information of the EEG signal is severely lost due to the "volume conduction effect" of the skull, resulting in an exponential decrease in the signal-to-noise ratio (SNR). Existing convolutional neural networks (CNNs) or recurrent neural networks (RNNs) are prone to overfitting and cannot extract stable and effective weak EEG features.

[0003] Poor cross-subject generalization ability and excessively high calibration costs: Due to significant individual physical differences in the cerebral cortex folding structure, skull thickness, and neural conduction pathways among different subjects, EEG signals exhibit strong non-stationarity and cross-subject variability (session-to-session and subject-to-subject variability). Existing systems typically require paradigm-evoked training (calibration) lasting tens of minutes to several hours when connecting new users to collect sufficient target samples to fine-tune model parameters. This lengthy calibration process, which is prone to causing visual / mental fatigue, completely blocks the commercialization possibility of "plug-and-play" brain-computer interface devices.

[0004] The computing power ecosystem is dependent on foreign entities, and edge deployment is difficult: Currently, mainstream EEG large-scale model training and inference frameworks heavily rely on foreign general-purpose GPU computing power ecosystems (such as the CUDA architecture). In mobile scenarios such as controlling wheelchairs or wearable devices, the high power consumption and large size of traditional GPUs make edge computing deployment impossible; while transmitting the entire signal to the cloud for decoding is subject to network communication fluctuations, resulting in uncontrollable end-to-end latency.

[0005] Physical control lacks a dynamic safety assessment mechanism: When brain-computer interfaces drive physical devices (especially hardware with physical lethality or high-risk attributes such as robotic arms and wheelchairs), traditional control logic often uses static confidence thresholds (such as a fixed setting of 0.8). This rigid judgment method cannot adapt to fluctuations in user state and changes in complex environment, and is prone to triggering erroneous commands under low signal-to-noise ratio, leading to serious physical collisions or safety accidents. Summary of the Invention

[0006] To overcome numerous technical shortcomings in existing technologies, such as low decoding accuracy of low-channel signals, cumbersome device calibration, reliance on foreign computing platforms, and lack of dynamic safety guarantees in physical control, this invention provides a single- and dual-channel EEG decoding and physical device control method and platform based on heterogeneous computing power. This invention aims to achieve high-precision, calibration-free EEG decoding with minimal hardware input through the deep integration of knowledge distillation, the Transformer attention mechanism, and the heterogeneous computing power of a domestically produced NPU, and to construct a closed-loop physical world control system with high real-time performance and a dynamic anti-collision mechanism.

[0007] The technical solution adopted by this invention to solve its technical problem is: providing a method for single and dual-channel EEG decoding and physical device control based on heterogeneous computing power, characterized by including the following steps: (1) Data aggregation and preprocessing steps: Aggregate EEG datasets and heterogeneous data, establish a unified EEG spatiotemporal feature cleaning and alignment pool, and perform noise reduction, artifact removal, resampling and standardization on the original EEG signals. (2) Heterogeneous computing power model training steps: Based on the domestic neural network processing unit (NPU) and heterogeneous computing architecture, the preprocessed data is used to train a large neural decoding model, which is configured to support feature learning across subjects and paradigms; (3) Single and dual channel feature extraction steps: Under single or dual channel EEG signal input conditions, the trained neural decoding big model is used to extract weak EEG features under the steady-state visual evoked potential (SSVEP) and / or motor imagery (MI) paradigm, and the signal-to-noise ratio is enhanced through the spatiotemporal attention mechanism; (4) Intent decoding and confidence assessment steps: Input the extracted features into the decoding network, output the user intent category and its confidence, and trigger a control command when the confidence is higher than a preset dynamic threshold; (5) Physical device feedback control steps: The triggered control command is converted into a standardized machine command, and the real-time control of the external controlled hardware device is realized through the physical control interface. The real-time control meets the response time of the preset real-time requirements.

[0008] Preferably, in step (2), the neural decoding big model adopts the Transformer architecture, including: a multi-head self-attention layer for capturing the temporal dependencies of EEG signals; a cross-paradigm feature fusion module for learning shared representations between the SSVEP and MI paradigms; and a personalized adaptation layer configured to achieve zero-calibration generalization for new subjects.

[0009] Preferably, in step (3), the single / dual channel feature extraction step further includes: performing time-frequency domain transformation on the single-channel or dual-channel EEG signal to extract energy features of multiple frequency bands; using the knowledge distillation technique of the pre-trained model, using the model trained with multi-channel EEG data as the teacher model and the model with the single / dual channel input as the student model, and transferring the spatial feature distribution knowledge of the multi-channel model to the student model through output alignment of the feature layer or logic layer; and using data augmentation strategies, including signal superposition, noise injection and time-domain perturbation, to expand the training samples of the single / dual channels.

[0010] Preferably, in step (4), the preset dynamic threshold is dynamically generated by calling a preset multidimensional parameter mapping table or a multilayer perceptron (MLP) network; the input features of the multidimensional parameter mapping table or MLP network include at least: the number of training samples of the current user, the historical decoding accuracy, and the risk level of the physical control task.

[0011] Preferably, in step (5), the physical control interface adopts a hierarchical instruction mapping mechanism, specifically including: a first layer: mapping intent categories to standardized intermediate instructions; a second layer: converting the intermediate instructions into device-specific control protocols according to the target device type; and a third layer: adding safety constraints and collision detection logic.

[0012] Preferably, the method further includes an online incremental learning step: collecting EEG signals and feedback tags during user usage; incrementally updating the model on the NPU edge computing device; and synchronizing the updated model parameters to the cloud model library.

[0013] Preferably, the external controlled hardware device includes one or more of the following: an extended reality (XR) spatial interaction module, a wheelchair, a smart home controller, and a robotic arm; the response time that meets the preset real-time requirements is specifically less than 2 seconds. Beneficial effects

[0014] Compared with the prior art, the present invention has achieved the following significant technical progress and beneficial effects: Breaking through the bottleneck of low-channel spatial information and significantly reducing the hardware deployment threshold: This invention innovatively introduces a Transformer-based knowledge distillation architecture, utilizing spatial prior knowledge learned from massive amounts of data by a high-channel (e.g., 32 / 64-channel) teacher model to guide the training of single- and dual-channel student models. Combined with a multi-head self-attention mechanism to enhance the focus on time-frequency domain features, it successfully overcomes the technical challenge of extremely low signal-to-noise ratio under low-channel dry electrode input. While ensuring decoding accuracy approaching that of high-channel devices, it reduces hardware cost and power consumption by more than 80%, enabling the brain-computer interface module to be easily integrated into everyday XR headsets or portable headbands.

[0015] Achieving cross-subject calibration-free generalization and a "plug-and-play" user experience: Through a unified cross-paradigm feature fusion module and incremental learning mechanism, the model can extract domain-invariant features from the SSVEP and MI paradigms. New subjects can use the system directly without undergoing a lengthy and tiring baseline data acquisition and calibration process (zero-calibration), greatly improving the system's versatility and commercial feasibility.

[0016] The computing power ecosystem is completely independent and controllable, meeting the extreme real-time requirements of edge computing: The underlying model operators and training inference framework of this invention have been deeply optimized for heterogeneous computing on domestic neural network processing units (NPUs, such as Huawei Ascend architecture). By utilizing the INT8 quantization and automatic mixed precision characteristics of the NPU, the massive cloud model is effectively lightweighted and pushed down to edge computing nodes, completely eliminating dependence on foreign GPU ecosystems, and strictly compressing the end-to-end control latency from intent generation to physical device action to less than 2 seconds, meeting the timeliness requirements of high-frequency physical interaction.

[0017] A unique dynamic confidence level and layered protection network ensure absolute safety in physical interactions: Abandoning the traditional, dangerous static threshold mechanism, this invention introduces a dynamic confidence level assessment model driven by a multilayer perceptron (MLP). This model can adaptively calculate decision thresholds in real time based on the user's historical performance, incremental data volume, and the physical risk level of the current task (e.g., the risk of manipulating a robotic arm is far greater than scrolling on a virtual interface). Simultaneously, combined with a three-layer mapping of the physical control interface and torque / boundary anti-collision logic, a dual safety barrier is constructed from the algorithm layer to the execution layer, completely eliminating the risk of accidental operation in low signal-to-noise ratio environments. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort. Figure 1 The flowchart illustrates the single / dual-channel EEG decoding and physical device control method based on heterogeneous computing power provided in this embodiment of the invention.

[0019] Figure 2 This is a mapping diagram of the network structure for single and dual-channel EEG signal feature extraction and knowledge distillation provided in an embodiment of the present invention.

[0020] Figure 3 The diagram shows the logic flow of dynamic confidence threshold generation and MLP decision-making provided in the embodiments of the present invention.

[0021] Figure 4 This is a schematic diagram of the hierarchical instruction mapping and anti-collision mechanism of the physical control interface provided in an embodiment of the present invention. The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. The specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] Example 1: Single and Dual Channel EEG Decoding and XR Spatial Interaction Control Based on Ascend Heterogeneous Computing Power. This example provides an optimal implementation scheme based on the domestic Ascend NPU computing power architecture (such as the Ascend 910 training cluster and Ascend 310 edge inference chip), which is mainly applied to object rotation (SSVEP paradigm) and virtual hand grasping (MI paradigm) in extended reality (XR) space.

[0023] Step S1 (Data Aggregation and Preprocessing): The system acquires single-channel or dual-channel EEG signals from the subjects (SSVEP uses Oz or O1+O2 channels, MI uses C3+C4 channels, using minimally sized dry electrode caps). The signals are bandpass filtered from 0.5Hz to 50Hz and power frequency notch filtered to remove baseline drift and artifacts, and then uniformly downsampled to 250Hz before Z-score standardization.

[0024] Steps S2 and S3 (Network Mapping for Feature Extraction and Knowledge Distillation): To overcome the severe lack of spatial topological information in EEG signals under single and dual channels, this embodiment deploys a Transformer model on the Ascend CANN architecture and employs a knowledge distillation technique based on hint-based learning. First, a large Transformer model pre-trained on 32-channel full open-source data is used as the teacher model. Subsequently, a linear projection layer is introduced into the hidden layer of the student model (single / dual channel input) to align its output feature dimension with that of the teacher model. The spatial features of the intermediate layer are transferred by calculating the mean squared error (MSE) loss; simultaneously, the soft target is aligned using the Kullback-Leibler Divergence (KL) loss in the output layer. Combined with the Transformer's multi-head self-attention mechanism to automatically focus on high-weight time-frequency bands, the simplified channel model successfully achieves feature representation capabilities approaching those of high-channel devices.

[0025] Step S4 (Dynamic Confidence Assessment Mechanism): In this embodiment, the preset dynamic threshold is generated in real time by a multilayer perceptron (MLP) network deployed in on-chip memory. The input layer of the MLP receives a three-dimensional feature vector V = [N, A, R]. Where N is the number of incremental samples for the current user; A is the historical time window decoding accuracy; and R is the risk level scalar of the current XR interaction task (set to a low-risk value of 0.3). After the vector is transformed by the nonlinear transformation of the hidden layer of the MLP and activated by the Sigmoid function, a continuous dynamic threshold (e.g., 0.65) is output. The decoding intent can only be triggered when the maximum Softmax probability value output by the Transformer is greater than 0.65. Step S5 (Physical World Layered Control): The decoding intent enters the WABO Interface physical interface. The first layer maps the intent to a standardized intermediate instruction (e.g., ROTATE_RIGHT); the second layer converts it into the WebSocket JSON control protocol of the XR device; and the third layer performs a boundary security check of the virtual space and then sends it to the XR rendering engine (e.g., Unity / Unreal). In the tests of the above embodiments, the present invention achieved significant technical effects: (1) High generalization: Under zero-calibration conditions, the single-channel accuracy of SSVEP reached 75% and the dual-channel accuracy reached 80%. After providing only 30 seconds of fine-tuning data, the single-channel accuracy jumped to 82% and the dual-channel accuracy reached 87%. (2) Low latency: Thanks to the acceleration of NPU's INT8 model quantization, the end-to-end average latency from EEG feature acquisition to XR object rotation is stable at around 1100ms (standard deviation 150ms), which fully meets the preset real-time requirement of less than 2 seconds. (3) Hardware cost and wearable experience: Compared with the traditional 32-channel system, the hardware cost of the simplified channel system of the present invention is reduced by about 96% (from 50,000 yuan to about 2,000 yuan), and the user's wear preparation time is greatly reduced from the traditional 5 minutes to less than 10 seconds.

[0026] Example 2: High-Risk Motion Control for Industrial Robotic Arms / Smart Homes Based on Edge Heterogeneous Computing This example focuses on the hierarchical mapping of physical commands and anti-collision logic supporting multiple types of high-risk hardware. The controlled device is an industrial robotic arm with physical lethality. Dynamic Anti-Collision Network Linkage: For the robotic arm grasping task, the MLP network in step S4 sets the task risk scalar R to an extremely high value (e.g., 0.9), thereby outputting an extremely stringent dynamic confidence threshold (e.g., 0.95). Commands are only issued when the model is extremely confident in the user's intent, eliminating false triggers caused by low signal-to-noise ratios from the source. Physical Interface Security Interception: In the WABO Interface, the first layer maps the intent to 6-DoF space intermediate commands; the second layer converts them into Modbus TCP control messages for the robotic arm servo motors. The key lies in the third layer (safety constraint layer): before the command actually drives the motor, the system calls the inverse kinematics module to check whether the target coordinates exceed the maximum working radius of the robotic arm (spatial boundary constraint); at the same time, combined with the real-time feedback of the end torque sensor of the robotic arm, if the system predicts that the command will cause the collision torque to exceed the safety tolerance value, the physical interface will forcibly discard the EEG decoding command and trigger the robotic arm to hover in place, thus building a double absolute safety moat from the algorithm side to the physical side.

[0027] Example 3: Wheelchair Control Based on Online Incremental Learning This example addresses the motion assistance scenario for patients with high-level paraplegia. Since the wheelchair is a mobile controlled terminal, the system employs a low-power Ascend 310 edge computing module. The EEG acquisition end retains only a single-channel Fpz node in the forehead to maximize the patient's wearing dignity and comfort. The system continuously runs an online incremental learning process in the background, collecting feedback labels of the patient's EEG signal characteristics and actual movement trajectory during wheelchair operation. Backpropagation is performed at the NPU edge, only fine-tuning and updating the parameters of the Transformer's personalized adaptation layer, and then synchronizing to the cloud when the network is idle. In the control layer, the third layer of the WABO Interface is configured for mandatory hard binding with the millimeter-wave radar / ultrasound sensors mounted on the wheelchair chassis. When the sensor detects a dynamic or static obstacle within 1 meter in front, even if the EEG decodes a high-confidence "forward" intention, the system will automatically implement braking to ensure the safety of the disabled user.

[0028] 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 method for single- and dual-channel EEG decoding and physical device control based on heterogeneous computing power, characterized in that, The method includes the following steps: (1) Data aggregation and preprocessing step: aggregate EEG datasets and heterogeneous data, establish a unified EEG spatiotemporal feature cleaning and alignment pool, and perform denoising, artifact removal, resampling and standardization processing on the original EEG signals; (2) Heterogeneous computing power model training step: based on the domestic neural network processing unit (NPU) and heterogeneous computing architecture, train a large neural decoding model using the preprocessed data, and configure the model to support feature learning across subjects and paradigms; (3) Single and dual channel feature extraction step: under single or dual channel EEG signal input conditions, use the trained large neural decoding model to extract weak EEG features under the steady-state visual evoked potential (SSVEP) and / or motor imagery (MI) paradigms, and enhance the signal-to-noise ratio through spatiotemporal attention mechanism; (4) Intent decoding and confidence assessment step: input the extracted features into the decoding network, output the user intent category and its confidence, and trigger a control command when the confidence is higher than the preset dynamic threshold; (5) Physical device feedback control steps: The triggered control command is converted into a standardized machine command, and the real-time control of the external controlled hardware device is realized through the physical control interface. The real-time control meets the response time of the preset real-time requirements.

2. The method according to claim 1, characterized in that, In step (2), the neural decoding big model adopts the Transformer architecture, including: a multi-head self-attention layer for capturing the temporal dependencies of EEG signals; a cross-paradigm feature fusion module for learning shared representations between the SSVEP and MI paradigms; and a personalized adaptation layer configured to achieve calibration-free generalization for new subjects.

3. The method according to claim 1, characterized in that, In step (3), the single-channel and dual-channel feature extraction step further includes: performing time-frequency domain transformation on the single-channel or dual-channel EEG signal to extract energy features of multiple frequency bands; using the knowledge distillation technique of the pre-trained model, using the model trained with multi-channel EEG data as the teacher model and the model with the single-channel and dual-channel input as the student model, and transferring the spatial feature distribution knowledge of the multi-channel model to the student model through output alignment of the feature layer or logic layer; and using data augmentation strategies, including signal superposition, noise injection and temporal perturbation, to expand the training samples of the single-channel and dual-channel models.

4. The method according to claim 1, characterized in that, In step (4), the preset dynamic threshold is dynamically generated by calling a preset multidimensional parameter mapping table or a multilayer perceptron (MLP) network; The input features of the multidimensional parameter mapping table or MLP network include at least: the number of training samples for the current user, historical decoding accuracy, and the risk level of the physical control task.

5. The method according to claim 1, characterized in that, In step (5), the physical control interface adopts a hierarchical instruction mapping mechanism, specifically including: First layer: mapping intent categories to standardized intermediate instructions; Second layer: converting the intermediate instructions into device-specific control protocols according to the target device type; Third layer: adding safety constraints and collision detection logic.

6. The method according to claim 1, characterized in that, The method also includes an online incremental learning step: collecting EEG signals and feedback tags during user usage; incrementally updating the model on the NPU edge computing device; and synchronizing the updated model parameters to the cloud model library.

7. The method according to claim 1, characterized in that, The external controlled hardware device includes one or more of the following: extended reality (XR) spatial interaction module, wheelchair, smart home controller, and robotic arm; the response time that meets the preset real-time requirements is specifically less than 2 seconds.

8. A single / dual-channel EEG decoding training platform based on heterogeneous computing power, characterized in that, The platform includes: (1) a data aggregation module, used to aggregate EEG datasets and heterogeneous data to establish a unified EEG spatiotemporal feature cleaning and alignment pool; (2) a heterogeneous computing training module, based on a domestic neural network processing unit (NPU) and a heterogeneous computing architecture, used to train a large neural decoding model with cross-subject and cross-paradigm capabilities; (3) a single-channel and dual-channel feature extraction module, used to extract weak EEG features under the SSVEP and / or MI paradigms under single-channel or dual-channel EEG signal input conditions; (4) an intent decoding module, used to decode the extracted features into user intents and output confidence assessment results; (5) a physical control interface module, used to convert high-confidence intent decoding results into standardized machine instructions to realize real-time control of external controlled hardware devices, with response time meeting preset requirements.

9. The platform according to claim 8, characterized in that, The heterogeneous computing training module includes: an NPU training server cluster for large-scale model training; a distributed training framework that supports data parallelism and model parallelism; and an automatic mixed-precision training component configured to improve training efficiency and reduce GPU memory usage.

10. The platform according to claim 8, characterized in that, The platform also includes: an edge inference module, deployed on the NPU edge computing device, for real-time EEG decoding; a cloud-edge collaboration module, for collaborative optimization of cloud-based model training and edge-based model deployment; and a user management module, for managing personalized model parameters for multiple users.

11. The platform according to claim 8, characterized in that, The single / dual-channel feature extraction module includes: a signal preprocessing unit for filtering, denoising, and normalizing single / dual-channel EEG signals; a time-frequency feature extraction unit for extracting energy features and temporal features across multiple frequency bands; and a feature enhancement unit that utilizes attention mechanisms and knowledge distillation techniques to improve feature quality.

Images