Cross-subject EEG cognitive state recognition method based on prototype clustering domain adaptation algorithm

Abstract

The invention discloses a cross-subject EEG cognitive state recognition method based on a prototype clustering domain adaptation algorithm. According to the method, the concept of category domains isintroduced, on one hand, on the basis of multi-source domain alignment of labels, feature distribution differences between different categories are considered, structural fine-grained alignment underthe category conditions between different source domains in a feature space is researched, and the problem of category imbalance in the multi-source domains is converted into a mode of the category domains; and prototype theoretical clustering alignment between the source domain and the target domain is carried out, i.e., clustering between similar source domains is carried out on the target domain by taking a dynamic adjustment prototype center as a constraint, and similar features and sparse heterogeneous features between the domains are realized, wherein the former realizes intra-domain class conditional structure feature alignment, and the latter realizes global fine-grained structure feature alignment. According to the invention, the method can be compatible with category balance andimbalance, effectively solves the problem of individual difference of electroencephalogram signals in the field of brain cognitive calculation, has high generalization ability, and can be well suitable for clinical diagnosis and practical application.

Description

technical field [0001] The present invention relates to the neural electrophysiological signal analysis technology in the field of brain cognitive computing, and the multi-source domain adaptive model construction method in the field of unsupervised learning. The method of recognizing the cognitive state is applicable to the situation of category imbalance in multi-source domains and between domains, and solves the problem of individual differences in EEG. Background technique [0002] At present, the fruitful results based on deep learning benefit from the supervised learning of a large amount of labeled data. But for unsupervised learning, the main hurdle in designing general-purpose network models is extending models trained on known labeled data to new unlabeled domains. For the target task that lacks labeled data, making full use of the feature information of the source domain labeled data is the key breakthrough point to solve this problem. However, when the trained ...

AI Technical Summary

Problems solved by technology

[0007] The present invention proposes a prototype-based clustering alignment algorithm that is suitable for the problem of category imbalance in multi-source domains. In view of the possible existence of quantitative mismatch, individual differences, and category imbalance in EEG data, the structural characteristics of label samples are fully learned, and the The feature distributions of the target and multi-source domains are aligned for efficient transfer of features between domains

Images