CN-122018684-A - Self-adaptive brain-computer interface feedback calibration method and system

CN122018684ACN 122018684 ACN122018684 ACN 122018684ACN-122018684-A

Abstract

The application belongs to the field of brain-computer interfaces, and particularly relates to a self-adaptive brain-computer interface feedback calibration method and system. The method comprises the steps of obtaining multichannel electroencephalogram signals in real time, calculating comprehensive quality factors of all channels based on multidimensional indexes such as time domain features, frequency band energy, artifact identification and impedance, dynamically distributing channel weights based on the factors and coupling the channel weights to covariance calculation of a spatial filtering algorithm to achieve self-adaptive spatial filtering according to signal quality so as to extract features, monitoring output confidence and signal quality of a classifier in real time, judging that physiological states of users drift and triggering model increment update when the signal quality is continuously good and the confidence is reduced, and suspending update when the signal quality is too low so as to protect a model. The application can dynamically inhibit noise interference, accurately identify physiological drift and realize self-adaptive calibration under reliable conditions, and remarkably improve the identification precision and user experience of the brain-computer interface system in a long-time and non-steady environment.

Inventors

TANG WEI
ZHOU JINGJIE
ZHANG MING

Assignees

中国矿业大学

Dates

Publication Date: 20260512
Application Date: 20260126

Claims (10)

1. An adaptive brain-computer interface feedback calibration method, comprising: s1, acquiring multichannel electroencephalogram signals in real time, and calculating to obtain comprehensive quality factors of all channels based on multi-dimensional evaluation indexes at least comprising time domain features, frequency band energy distribution features, artifact identification results based on a lightweight deep convolutional neural network and electrode contact impedance; S2, dynamically distributing initial weights of all channels based on the comprehensive quality factors to form weight vectors, coupling the weight vectors into a covariance matrix calculation process of a spatial filtering algorithm to generate a spatial projection matrix capable of adaptively shifting according to real-time signal quality, and performing spatial filtering on channel signals obtained after weighting to extract feature vectors; and S3, monitoring the output confidence coefficient of the classifier on the feature vector and the comprehensive quality factor in real time, judging that the physiological state of the user drifts and triggering the incremental updating of the model parameters when the output confidence coefficient shows a continuous descending trend during the period that the signal quality is continuously higher than a first preset threshold value, and triggering a model protection mechanism to pause updating when the signal quality is lower than a second preset threshold value.
2. The method according to claim 1, wherein in step S2, the covariance matrix calculation process of coupling the weight vector to the spatial filtering algorithm specifically comprises: Weighting the original signal of each channel based on the weight vector to obtain a weighted channel signal; calculating an average covariance matrix of each task by using the weighted channel signals; solving a spatial projection matrix based on the average covariance matrix; and projecting the signals by using the space projection matrix, and calculating variance characteristics of the projected signals to form the characteristic vector.
3. The adaptive brain-computer interface feedback calibration method according to claim 2, wherein step S2 further comprises: After the weight vector is obtained by distributing initial weights based on the comprehensive quality factors, introducing a joint optimization target comprising an L2 regularization term and a Laplace smoothness constraint term constructed based on a channel physical position adjacency relation, optimizing the weight vector to obtain an optimized weight vector which is smooth in spatial distribution and avoids excessive concentration on individual channels, and using the optimized weight vector for calculating the covariance matrix.
4. The method according to claim 1, wherein in step S1, the multi-dimensional evaluation index on which the comprehensive quality factor is calculated further comprises an inter-channel spatial consistency analysis index based on pearson correlation coefficient and/or molan index.
5. The method according to claim 1, wherein in step S3, the triggering model parameter increment update is specifically: And (3) adopting a gradient descent method or a variant thereof, taking the feature vector corresponding to the triggering period as training data, calculating the gradient by a preset loss function, and fine-tuning model parameters by a preset small learning rate.
6. The adaptive brain-computer interface feedback calibration method according to claim 1, wherein step S3 further comprises a calibration consistency evaluation step of: periodically calculating a weighted integrated bias value including a spatial mismatch rate, a target bias, and an execution hysteresis coefficient, and calculating a mean square error thereof in one evaluation period; When the mean square error of a plurality of continuous evaluation periods exceeds the preset tolerance upper limit, an early warning or protection mechanism is triggered.
7. The adaptive brain-computer interface feedback calibration method according to claim 1, further comprising a silence calibration step of: and in a background thread parallel to the main interaction task of the user, periodically executing the signal quality real-time evaluation step and the condition triggering type self-adaptive calibration step to realize automatic calibration without user perception.
8. The adaptive brain-computer interface feedback calibration method according to claim 7, wherein the silence calibration step further comprises a reference model evolution strategy: Storing a reference model derived from the initial stable calibration of the user; Periodically calculating the normalized deviation degree of the current working model parameter and the reference model parameter; and when the deviation exceeds a preset evolution threshold and the online recognition rate of the system is kept stable, updating the parameter state of the current working model into a new reference model.
9. The adaptive brain-computer interface feedback calibration method according to claim 1, wherein the method is implemented using a distributed processing architecture, wherein: in the step S1, basic quality index calculation and artifact preliminary labeling with high real-time requirements are performed on a front-end sensing layer integrated in portable electroencephalogram acquisition equipment; The step S1 involves comprehensive quality factor calculation of complex spectrum analysis and global fusion, and the step S2 and the step S3 are executed at a back-end calculation layer outside the front-end perception layer; and the front-end sensing layer sends the calculated basic quality index and the preprocessed signal to the back-end calculating layer.
10. An adaptive brain-computer interface feedback calibration system for implementing the adaptive brain-computer interface feedback calibration method of any one of claims 1 to 9, comprising: The signal acquisition and quality evaluation module is configured to acquire multichannel electroencephalogram signals in real time and calculate the comprehensive quality factors of all channels; A dynamic weighting and feature extraction module configured to dynamically generate channel weight vectors based on the comprehensive quality factors, couple the weight vectors into covariance calculation of a spatial filter to perform adaptive spatial filtering and extract feature vectors; The condition triggering calibration and model management module is configured to monitor the confidence coefficient and the signal quality of the classifier, trigger the incremental update of the model when the physiological state is determined to drift, and trigger the protection of the model when the signal quality is too low; the dynamic weighting and feature extraction module and the condition triggering calibration and model management module form closed-loop linkage according to the comprehensive quality factors output by the signal acquisition and quality evaluation module.

Description

Self-adaptive brain-computer interface feedback calibration method and system Technical Field The application belongs to the field of brain-computer interfaces, and particularly relates to a self-adaptive brain-computer interface feedback calibration method and system. Background The brain-computer interface technology is used as a leading edge means for realizing direct information interaction between the brain and external equipment, has important application value in the fields of medical rehabilitation, auxiliary driving and nerve engineering, takes an electroencephalogram signal as a core information carrier, and provides basic data for analyzing the cognitive state and the movement intention of a user by capturing the electrophysiological activity of a central nervous system. However, the brain-computer interface faces a complex electrophysiological environment in practical application, and particularly in a long-time interaction task, the coupling of drift of a user physiological state and multi-source environmental noise puts higher demands on the calibration accuracy and real-time feedback capability of the system. However, the traditional calibration method generally depends on static training or fixed online updating step length before experiments, so that extremely strong non-stationarity of the electroencephalogram signals is difficult to effectively analyze, meanwhile, due to electrode impedance fluctuation, myoelectric interference generated by limb actions and the existence of environmental electromagnetic noise, an acquisition system often captures abnormal signals containing a large number of artifacts, the existing model updating logic lacks dynamic perceptibility of signal quality, so that classifier parameters are polluted and cliff type drop of identification accuracy is caused, in addition, the traditional scheme lacks a dynamic weighting mechanism for multi-source heterogeneous channels, characteristic contribution of each channel cannot be automatically adjusted according to real-time signal-to-noise ratio, and self-adaptive fine adjustment of parameters is difficult to realize on the premise of guaranteeing model purity, so that robustness of the system in a complex environment is insufficient. Therefore, the existing brain-computer interface system often has the problems of insufficient calibration precision, poor self-adaptive capacity, weak anti-interference performance and the like when facing a long-time and non-stable practical use field Jing Shi. It is necessary to propose an adaptive feedback calibration method capable of evaluating signal quality in real time, dynamically adjusting channel weights, and triggering model update under reliable conditions, so as to improve stability and recognition performance of the brain-computer interface system in a real environment. Disclosure of Invention In order to solve the technical problems, the invention provides a self-adaptive brain-computer interface feedback calibration method and a system, which adopt the following technical scheme. In a first aspect, a method for adaptively calibrating feedback of a brain-computer interface includes: s1, acquiring multichannel electroencephalogram signals in real time, and calculating to obtain comprehensive quality factors of all channels based on multi-dimensional evaluation indexes at least comprising time domain features, frequency band energy distribution features, artifact identification results based on a lightweight deep convolutional neural network and electrode contact impedance; S2, dynamically distributing initial weights of all channels based on the comprehensive quality factors to form weight vectors, coupling the weight vectors into a covariance matrix calculation process of a spatial filtering algorithm to generate a spatial projection matrix capable of adaptively shifting according to real-time signal quality, and performing spatial filtering on channel signals obtained after weighting to extract feature vectors; and S3, monitoring the output confidence coefficient of the classifier on the feature vector and the comprehensive quality factor in real time, judging that the physiological state of the user drifts and triggering the incremental updating of the model parameters when the output confidence coefficient shows a continuous descending trend during the period that the signal quality is continuously higher than a first preset threshold value, and triggering a model protection mechanism to pause updating when the signal quality is lower than a second preset threshold value. Preferably, in step S2, the covariance matrix calculation process of coupling the weight vector to the spatial filtering algorithm specifically includes: Weighting the original signal of each channel based on the weight vector to obtain a weighted channel signal; calculating an average covariance matrix of each task by using the weighted channel signals; solving a spatial projection matrix based on the a