CN-122004893-A - Electroencephalogram signal identification method based on trial characteristic reconstruction and self-attention

CN122004893ACN 122004893 ACN122004893 ACN 122004893ACN-122004893-A

Abstract

The invention discloses an electroencephalogram signal identification method based on trial characteristic reconstruction and self-attention, and belongs to the technical field of electroencephalogram signal decoding and stereoscopic vision function assessment. The method comprises the steps of collecting multichannel electroencephalogram signals when tested to watch DRDS stimulation, retaining an effective channel and aligning labels according to test times after bandpass filtering and independent component analysis, generating multi-segment samples by adopting an annular slicing strategy to carry out non-overlapping segmentation on single-test electroencephalogram signals, extracting segment characterization through a local module comprising multi-scale space-time convolution and channel recalibration, decoupling and splicing through a test secondary characteristic reconstruction module to relieve segment-test semantic mismatch, respectively modeling long-range dependence and cross-test global interaction in test times by utilizing a long-term and short-term self-attention module to obtain enhanced characterization, and outputting identification results through a full-connection layer and Softmax. The method can realize segment-trial alignment learning and global space-time modeling under a small sample, improves identification accuracy and robustness, and is suitable for objective stereoscopic vision function evaluation.

Inventors

SHEN LILI
YANG ZIHAO
WU YIMING

Assignees

天津大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (8)

1. An electroencephalogram signal identification method based on trial characteristic reconstruction and self-attention is characterized by comprising the following steps: S1, acquiring multichannel electroencephalogram signals and corresponding test time labels acquired during the stimulation of a tested dynamic random point stereogram; s2, preprocessing the electroencephalogram signals, including band-pass filtering, baseline correction and artifact removal, and screening out a reference channel and an interference channel to obtain effective channel electroencephalogram signals; S3, annular slice data enhancement is carried out on the brain electrical signals of the effective channels of each test time, the test time is connected end to form a closed loop, a plurality of starting positions are selected to carry out non-overlapping segmentation on the closed loop signals, and a plurality of segment samples are obtained; S4, inputting the segment samples into a local segment feature extraction module to extract segment-level space-time features; S5, inputting a plurality of fragment level features of the same test time into a test secondary feature reconstruction module, decoupling according to a channel, and splicing along a time axis to obtain a test secondary reconstruction feature; s6, inputting the secondary reconstruction characteristics into a long-term and short-term self-attention module, modeling time dependence in the test time through short-term self-attention, and outputting the secondary characterization with enhanced attention through long-term self-attention modeling global interaction; And S7, inputting the test secondary characterization with enhanced attention to a classification module, and obtaining a recognition result of the dynamic random point stereogram through a full-connection layer and Softmax.
2. The electroencephalogram signal identification method based on trial characteristic reconstruction and self-attention according to claim 1, wherein the band-pass filtering in S2 is 1-80 Hz, the artifact removal adopts independent component analysis to remove eye movement artifacts and myoelectric artifacts, and effective channels except for a reference channel are reserved after baseline correction.
3. The method of claim 1, wherein the annular slice data enhancement in S3 is specifically implemented by connecting a start sampling point and an end sampling point of each test signal to form a closed loop, selecting M different start positions along the closed loop for segmentation, and dividing the closed loop signal into N non-overlapping segments at each start position, thereby amplifying each test into M×N segment samples.
4. The electroencephalogram signal identification method based on trial feature reconstruction and self-attention according to claim 1, wherein the local segment feature extraction module in S4 comprises a multi-scale time convolution branch, a channel recalibration module and a multi-scale space convolution branch, and the convolution kernel length of the multi-scale time convolution branch is set in proportion to the sampling rate so as to extract local time sequence features of different time scales at the same time.
5. The electroencephalogram signal identification method based on trial feature reconstruction and self-attention according to claim 4, wherein the channel recalibration module is a Squeeze-and-specification module for adaptively weighting the feature channel response to highlight feature channels with larger contributions.
6. The electroencephalogram signal identification method based on trial-and-secondary feature reconstruction and self-attention according to claim 1, wherein the trial-and-secondary feature reconstruction module in S5 comprises: and decomposing each fragment level characteristic into a plurality of channel primitives in the channel dimension, sequentially splicing the characteristics of the same channel in different fragments along a time axis to obtain a test secondary characteristic sequence of each channel, and combining the test secondary characteristic sequences of the channels to form a test secondary reconstruction characteristic.
7. The electroencephalogram signal identification method based on trial feature reconstruction and self-attention according to claim 1, wherein the long-short-term self-attention module in S6 comprises two stages of short-term self-attention and long-term self-attention, wherein the short-term self-attention calculates time dimension attention weights for query, key and value as trial secondary reconstruction features to model long-range dependence in trial, and the long-term self-attention performs channel dimension self-attention calculation on short-term self-attention output to model cross-trial or global interaction.
8. The method for recognizing brain electrical signals based on trial feature reconstruction and self-attention according to claim 1, wherein S7 said classification module trains the network with cross entropy loss and introduces L1 regularization term to suppress overfitting.

Description

Electroencephalogram signal identification method based on trial characteristic reconstruction and self-attention Technical Field The invention relates to the technical fields of electroencephalogram signal decoding, stereoscopic vision recognition and deep learning, in particular to an electroencephalogram signal recognition method based on trial characteristic reconstruction and self-attention. Background Stereoscopic vision (Stereopsis) relies on binocular parallax information integration, which is a core index reflecting advanced binocular vision processing capability, and in clinical and scientific research, binocular coordination and depth perception levels of subjects are often estimated through stereoscopic visual acuity or stereoscopic recognition capability. Traditional stereoscopic vision test relies on subjective feedback of a subject, is easily interfered by subjective factors, and is difficult to realize objective and automatic assessment. The dynamic random point stereogram (DRDS) can effectively eliminate monocular clues, has unique advantages in stereoscopic vision mechanism research and objective evaluation, and can capture the neural response of the visual cortex to the DRDS stimulation due to the characteristics of millisecond time resolution, non-invasive acquisition, low cost and the like of an electroencephalogram signal (EEG), thereby providing a reliable basis for objective decoding of stereoscopic vision functions. Existing EEG visual recognition methods fall into two categories, traditional machine learning and deep learning. The machine learning method is characterized in that the characteristics such as variance, differential entropy, power spectral density and the like are manually extracted from original EEG signals, and then the characteristics are input into a classifier such as linear discriminant analysis, random forest or support vector machine and the like to complete identification, but the mode is highly dependent on prior knowledge in the field, key characteristic information is easy to miss, and complex nonlinear modes in the EEG signals are difficult to capture due to the linear expression capability of a shallow model. Deep learning automatically learns high-discriminant abstract features through multi-stage nonlinear transformation, and end-to-end learning paradigms remarkably improve robustness and generalization capability of EEG signal decoding. Despite the significant advantages of deep learning, EEG data characteristics still present a serious challenge in that EEG acquisition requires a tightly controlled experimental environment, subject individual differences are significant, and existing data sets have a generally limited sample size, which can easily lead to complex model overfitting. The existing method usually adopts a sliding window to slice the test signal into a plurality of fragments to expand samples, but the assumption that the characteristic of the fragment level learning-test secondary annotation' paradigm implies that the fragment characterization is consistent with the test label, and the characteristic that the EEG signal is non-stable and the contribution degree of different time slices is uneven, is easy to cause the mismatch of the fragment level characteristics and the test secondary semantics. Meanwhile, the limited receptive field of the traditional convolutional neural network is difficult to capture the long Cheng Shikong dependence, and based on the inductive deviation of the local similarity and translational invariance, the receptive field is in intrinsic conflict with the non-stationary dynamic characteristics of the EEG signal, so that the dynamic importance of a key time window is ignored, the global characterization modeling capability is restricted, and therefore, an electroencephalogram identification method capable of explicitly completing segment-to-trial alignment reconstruction and synchronously modeling long-short-term space-time dependence is needed. Disclosure of Invention 1. The invention aims to solve the technical problems: The invention aims to solve the problems of feature-tag mismatch and difficulty in capturing test-level long-range space-time dependence caused by fragment level learning in the prior art, and specifically comprises the following steps: (1) The semantic mismatch problem of the segment level learning and trial secondary labels; Existing EEG identification methods often slice the test signal into multiple segments through a sliding window to increase the sample size, but this approach implies the assumption that the "segment characterization is consistent with the test label". The electroencephalogram signal has a non-stationary characteristic, the contribution of different time segments to the test time label is uneven, so that the segment level features and the test secondary semantics are mismatched, the traditional sliding window is easy to crack the global time structure, the feature alignmen