CN-122020295-A - Electroencephalogram emotion recognition method based on distribution progressive domain adaptive diffusion modeling

CN122020295ACN 122020295 ACN122020295 ACN 122020295ACN-122020295-A

Abstract

The invention discloses an electroencephalogram emotion recognition method based on distribution progressive domain adaptive diffusion modeling, the method first acquires EEG signals to form source domain features and target domain features. And secondly, constructing a feature extraction model, and aiming at the source domain features and the target domain features, obtaining corresponding source domain potential representations and target potential representations. The target domain classifier training is then performed by aligning the source domain to target domain distribution via DDEEG based on the source domain potential representation and the target potential representation. And finally, using a trained target domain classifier in a test stage, and identifying the emotion state of the target domain based on the input EEG signals. The invention realizes more stable cross-domain characteristic alignment, effectively avoids mode collapse, maintains a category structure, and remarkably improves emotion recognition accuracy under the cross-tested and cross-session conditions.

Inventors

MENG MING
CHEN YUQI
MA YULIANG
WANG TING
XI XUGANG

Assignees

杭州电子科技大学

Dates

Publication Date: 20260512
Application Date: 20260123

Claims (7)

1. The electroencephalogram emotion recognition method based on the distribution progressive domain adaptive diffusion modeling is characterized by comprising the following steps of: Step 1, acquiring EEG signals to form source domain features and target domain features; Step 2, constructing a feature extraction model, and aiming at the source domain features and the target domain features, obtaining corresponding source domain potential representations and target potential representations; Step 3, aligning the distribution of the source domain to the target domain through DDEEG based on the source domain potential representation and the target potential representation; Step 4, training a target domain classifier after the source domain is aligned to the target domain; And 5, using a trained target domain classifier in a test stage, and identifying the emotion state of the target domain based on the input EEG signals.
2. The electroencephalogram emotion recognition method based on the distributed progressive domain adaptive diffusion modeling according to claim 1 is characterized in that step 1 is specifically implemented by acquiring EEG signals of different tested or different sessions by using an electroencephalogram acquisition device, performing band-pass filtering and segmentation processing on the EEG signals, and calculating differential entropy DE characteristics of a plurality of frequency bands to form source domain characteristics and target domain characteristics.
3. The electroencephalogram emotion recognition method based on the distribution progressive domain adaptive diffusion modeling according to claim 2, wherein the step 2 is specifically implemented by constructing a feature extraction model based on an MLP encoder, and obtaining corresponding source domain potential representations and target potential representations for source domain features and target domain features as initial source domain features and initial target features of a subsequent diffusion domain migration process.
4. The brain electrical emotion recognition method based on distribution progressive domain adaptive diffusion modeling of claim 3, wherein in step 3, a diffusion domain migration process is performed based on a source domain potential representation and a target potential representation, DDEEG models the distribution alignment of the source domain to the target domain as a series of continuous, controllable feature diffusion-back diffusion steps, and explicitly constrains the diffusion direction through a semantic reinforcement learning mechanism.
5. The brain electrical emotion recognition method based on distribution progressive domain adaptive diffusion modeling of claim 4, wherein said DDEEG implements distribution alignment of source domain to target domain as DDEEG generates semantically continuous intermediate domain features after each diffusion-back diffusion step And at the same time, the intermediate domain features are used for reversely training the target domain classifier, so that the classifier synchronously expands the judgment boundary in the process of gradually changing the distribution, and the gradual domain adaptation from easy to difficult is realized.
6. The distribution progressive domain adaptive diffusion modeling-based electroencephalogram emotion recognition method of claim 5, wherein the semantic reinforcement learning mechanism comprises: The classifier guides the semantic diffusion learning of the diffusion module in DDEEG, which is to restrict the diffusion-back diffusion process by using the classifier with discrimination capability; the intermediate domain characteristic drives the generalized generation learning of the classifier, namely the classifier is trained step by step through continuously generated intermediate domain characteristics, and the classifier is gradually adapted to the distribution of the target domain under the condition of not directly crossing large domain differences.
7. The brain electricity emotion recognition method based on distribution progressive domain adaptive diffusion modeling of claim 6, wherein the step 4 is specifically implemented by integrating all intermediate domain features to train a final target domain classifier after source domain to target domain distribution pairs, and integrating the intermediate domain features generated by the diffusion module As a combined training set, training is carried out by utilizing a dynamic confidence weighting mechanism, so that the target domain classifier obtains optimal recognition performance on a plurality of progressive distributions.

Description

Electroencephalogram emotion recognition method based on distribution progressive domain adaptive diffusion modeling Technical Field The invention relates to the technical field of electroencephalogram emotion recognition and brain-computer interfaces, in particular to an electroencephalogram emotion recognition method based on distribution progressive domain adaptive diffusion modeling, and specifically relates to an electroencephalogram (electroencephalogram, EEG) domain adaptation method for solving the problems of significant difference of electroencephalogram distribution and insufficient emotion recognition generalization capability under a cross-test and cross-session condition. Background EEG has important applications in the fields of emotion recognition, cognitive monitoring, etc., as a non-invasive means of recording brain neural activity. However, EEG is inherently high in noise, strong in individual difference and remarkable in non-stationarity, and characteristic distribution among different tested and different sessions is remarkably offset, so that the accuracy of the emotion recognition model based on deep learning is greatly reduced under the cross-domain condition, and popularization and application of the emotion recognition model in an actual scene are limited. In recent years, deep learning has made remarkable progress in EEG emotion recognition, and Convolutional Neural Networks (CNNs), recurrent Neural Networks (RNNs), and attention-mechanism-based models all perform excellently in time-frequency feature modeling. However, this type of approach generally relies on large scale annotation data and has a strong dependence on specific subjects or sessions, with significant performance degradation when applied to target domain data that is not involved in training. To alleviate the above problems, researchers have introduced domain adaptation (domain adaptation) techniques to improve the cross-domain generalization ability of the model by aligning source domain and target domain features at the distribution level. The existing EEG domain adaptation methods are mainly divided into two types, namely a moment matching method based on statistical distribution, such as Maximum Mean Difference (MMD) and covariance alignment (CORAL), which realizes global matching of feature distribution by aligning high-order statistics, and an antagonistic domain adaptation method, such as Domain Antagonistic Neural Network (DANN) and multi-antagonistic domain adaptation (MADA), which enables a feature extractor to learn domain invariant features by introducing a domain discriminator. Although these approaches improve cross-domain performance to some extent, there are limitations in dealing with the large-scale distribution differences that are common in EEG, in that the moment matching approach does not guarantee semantic consistency, and the anti-learning approach suffers from pattern collapse and unstable training. Furthermore, most existing approaches attempt global distribution alignment at one time, and it is difficult to cope with complex and large span domain differences of EEG in different scenarios. The diffusion model, as a generative modeling framework that has been rapidly developed in recent years, has the ability to decompose complex distribution migration into multiple fine-grained steps, enabling stable, controllable distribution transformations. The diffusion model shows strong distribution expression capability on tasks such as image generation, signal modeling and the like, and provides a new solution for cross-domain feature alignment of EEG. However, there is currently no method of combining diffusion models with EEG domain adaptation depth, and in particular, there is no study that can maintain feature semantic consistency, stably boost the performance of classification across tested emotions during multi-step distribution migration. Therefore, how to use diffusion distribution transformation to realize progressive alignment of a source domain and a target domain, and promote the discrimination capability of migration features through semantic constraint becomes a key technical problem to be solved in the field of EEG emotion recognition. Disclosure of Invention Aiming at the problems of large feature distribution difference, difficult domain alignment, unstable semantics and the like commonly faced by a deep learning model under the conditions of cross-test and cross-session in the existing electroencephalogram emotion recognition field, particularly the defects of rough one-time distribution alignment, unstable training and semantic degradation of the existing moment matching method and the contrast domain self-adaptive method, the invention provides an electroencephalogram emotion recognition method based on distribution progressive domain adaptive diffusion modeling. According to the invention, larger distribution difference between the source domain and the target domain is disassem