Electroencephalogram emotion recognition method based on distributed progressive domain adaptive diffusion modeling

Abstract

The invention discloses an EEG (electroencephalogram) emotion recognition method based on distribution progressive domain adaptive diffusion modeling, which comprises the following steps of: firstly, acquiring an EEG (electroencephalogram) signal, and forming a source domain feature and a target domain feature; secondly, constructing a feature extraction model, and obtaining corresponding source domain potential representation and target potential representation according to the source domain features and the target domain features; then, based on the potential representation of the source domain and the potential representation of the target, the distribution from the source domain to the target domain is aligned through DDEEG, and target domain classifier training is carried out; and finally, using the trained target domain classifier in a test stage, and based on the input EEG signal, identifying the emotional state of the target domain. According to the method, more stable cross-domain feature alignment is realized, mode collapse is effectively avoided, a category structure is kept, and the emotion recognition accuracy under cross-subject and cross-session conditions is remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of EEG emotion recognition and brain-computer interface technology, and particularly to an EEG emotion recognition method based on distributed progressive domain adaptation diffusion modeling. Specifically, it is an EEG domain adaptation method used to address the significant differences in EEG signal distribution and insufficient generalization ability of emotion recognition under cross-subject and cross-conversation conditions. Background Technology

[0002] EEG, as a non-invasive method for recording brain neural activity, has important applications in fields such as emotion recognition and cognitive monitoring. However, EEG is inherently characterized by high noise, strong individual variability, and significant non-stationarity. The feature distributions of different subjects and different conversations show significant shifts, leading to a substantial decrease in the accuracy of deep learning-based emotion recognition models under cross-domain conditions, thus limiting their widespread application in real-world scenarios.

[0003] In recent years, deep learning has made significant progress in EEG emotion recognition, with convolutional neural networks (CNNs), recurrent neural networks (RNNs), and attention-based models all demonstrating excellent performance in time-frequency feature modeling. However, these methods generally rely on large-scale labeled data and are highly dependent on specific subjects or sessions, resulting in a significant performance drop when applied to target domain data that was not used in the training. To alleviate these issues, researchers have introduced domain adaptation techniques, which align source and target domain features at the distribution level to improve the model's cross-domain generalization ability.

[0004] Existing EEG domain adaptation methods are mainly divided into two categories: one is moment matching methods based on statistical distributions, such as Maximum Mean Difference (MMD) and Covariance Alignment (CORAL), which achieve global matching of feature distributions by aligning higher-order statistics; the other is adversarial domain adaptation methods, such as Domain Adversarial Neural Networks (DANN) and Multi-Adversarial Domain Adaptation (MADA), which introduce a domain discriminator to enable the feature extractor to learn domain-invariant features. Although these methods improve cross-domain performance to some extent, they still have limitations in dealing with the large-scale distributional differences prevalent in EEG: moment matching methods cannot guarantee semantic consistency, and adversarial learning methods suffer from pattern collapse and training instability. Furthermore, most existing methods attempt... Figure 1 Performing global distribution alignment in a single step is insufficient to handle the complex and wide-ranging domain differences in EEG across different scenarios.

[0005] As a rapidly developing generative modeling framework in recent years, diffusion models have the ability to decompose complex distribution transfers into multiple fine-grained steps, enabling stable and controllable distribution transformations. Diffusion models demonstrate powerful distribution representation capabilities in tasks such as image generation and signal modeling, providing a new solution for cross-domain feature alignment in EEG.

[0006] However, there is currently a lack of methods to combine diffusion models with EEG domain adaptation depth, especially a lack of research on maintaining semantic consistency of features and steadily improving cross-subject emotion classification performance during multi-step distribution transfer. Therefore, how to utilize diffusion-based distribution transformation to achieve progressive alignment between the source and target domains, and improve the discriminative ability of transferred features through semantic constraints, has become a key technical problem that urgently needs to be solved in the field of EEG emotion recognition. Summary of the Invention

[0007] This invention addresses the common problems faced by deep learning models in the field of EEG emotion recognition, such as large differences in feature distribution, difficulties in domain alignment, and semantic instability under cross-subject and cross-conversation conditions. In particular, existing moment matching methods and adversarial domain adaptation methods often suffer from coarse one-time distribution alignment, unstable training, and semantic degradation. This invention proposes an EEG emotion recognition method based on distribution-progressive domain adaptation diffusion modeling. By decomposing the large distribution difference between the source and target domains into multiple fine-grained transition steps and maintaining semantic consistency in each step, this invention achieves a smoother, more stable, and reliable cross-domain transfer process.

[0008] This invention is the first to introduce a diffusion-based feature transfer mechanism into EEG emotion recognition. It constructs a series of intermediate feature distributions that progressively approximate the target domain through forward and backward diffusion, achieving a gradual transformation from the source domain to the target domain. Simultaneously, a semantic reinforcement learning (SRL) mechanism is proposed. This mechanism utilizes the semantic constraints of the classifier to guide the diffusion process and uses the intermediate features generated by diffusion to inversely enhance the classifier's cross-domain adaptability, ensuring the invariance of emotional semantics throughout the transfer process. This invention significantly improves the accuracy of EEG emotion recognition across subjects and conversations, and reduces the instability caused by distribution alignment.

[0009] According to the invention, the specific steps include: Step 1: Collect EEG data and extract frequency domain features. Use EEG acquisition equipment to acquire EEG signals from different subjects or different sessions. After bandpass filtering and segmentation, calculate the differential entropy (DE) features of multiple frequency bands to form source domain features and target domain features.

[0010] Step 2: Construct a feature extraction model to obtain the corresponding latent representations of the source domain and the target domain for the source domain features and target domain features, which serve as the initial source domain features for the subsequent diffusion domain transfer process.

[0011] Specifically, based on the MLP encoder, a feature extraction model is constructed to obtain the corresponding source domain latent representation and target latent representation for the source domain features and target domain features, which serve as the initial source domain features and initial target features for the subsequent diffusion domain transfer process.

[0012] Step 3: Based on the latent representations of the source and target domains, a diffusion-based domain transfer process is performed. DDEEG models the distribution alignment from the source domain to the target domain as a series of continuous and controllable feature diffusion-anti-diffusion steps, and explicitly constrains the diffusion direction through a semantic reinforcement learning mechanism.

[0013] Specifically, during the diffusion-based domain transfer process, DDEEG models the distribution alignment from the source domain to the target domain as a series of continuous and controllable feature diffusion-convergence steps. It also explicitly constrains the diffusion direction through a semantic reinforcement learning mechanism to avoid semantic drift. In detail, DDEEG generates semantically continuous intermediate domain features after each diffusion-de-diffusion step. The intermediate domain features are then input into the target domain classifier, and the diffusion results are supervised using the source domain labels, thus directly embedding semantic consistency into the distribution transfer process itself. Simultaneously, these intermediate domain features are used to back-train the target domain classifier, enabling the classifier to synchronously expand its discriminative boundary as the distribution gradually changes, achieving a progressive domain adaptation from easy to difficult.

[0014] This semantic reinforcement learning mechanism includes: The classifier guides the semantic diffusion learning of the diffusion module in DDEEG: it uses a classifier with discriminative ability to constrain the diffusion-anti-diffusion process, ensuring that each step of feature transfer only occurs at the distribution level without destroying the category semantics; Generative learning that drives classifier generalization through intermediate domain features: The classifier is trained step by step by continuously generated intermediate domain features, so that it gradually adapts to the target domain distribution without directly crossing large domain differences.

[0015] Step 4: Integrate all transitional features to train the final target domain classifier.

[0016] Specifically, after distributing the source domain to the target domain, all intermediate domain features are integrated to train the final target domain classifier, which then incorporates the intermediate domain features generated by the diffusion module. As a joint training set, a dynamic confidence weighting mechanism is used for training, enabling the target domain classifier to achieve optimal recognition performance on multiple asymptotic distributions.

[0017] Step 5: In the testing phase, the trained target domain classifier is used to identify the emotional state of the target domain. The EEG features of the target domain are input into the feature extractor and the final classifier, and the corresponding emotion category is output to achieve emotion recognition across subjects and across conversations.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: Achieving more stable cross-domain feature alignment: This invention utilizes a diffusion model to decompose the large distributional differences between the source and target domains into multiple fine-grained steps, making feature transfer smoother and avoiding instability caused by one-time alignment.

[0019] Effectively avoids pattern collapse and maintains category structure: By generating multiple transition features through progressive diffusion, this invention can preserve the category pattern structure of the source domain during the migration process, avoiding the feature aliasing problem common in traditional adversarial methods.

[0020] Introducing semantic reinforcement learning to maintain semantic consistency: The SRL mechanism of this invention uses a classifier to impose semantic constraints on the diffusion process, while using generated features to inversely improve the classifier's capabilities, so that the label structure information is not lost during the entire domain adaptation process.

[0021] Two-way optimization of classifier and feature generator improves cross-domain recognition performance: Through dual-path training, the feature generator and classifier promote each other, significantly improving the accuracy of emotion recognition under cross-subject and cross-conversation conditions.

[0022] Improving model robustness and practical application capability: Experiments show that the present invention achieves better recognition performance than existing methods on both the SEED and SEED-IV datasets, and has better generalization and stability. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This diagram illustrates the impact of each module of the emotion EEG recognition method of the present invention on the classifier. Figure 2 This is a flowchart of the present invention; Figure 3 This is a flowchart of DDEEG. Detailed Implementation

[0025] The present invention will be further described below with reference to specific embodiments. The following description is for illustrative and explanatory purposes only and does not constitute any limitation on the present invention.

[0026] The EEG emotion recognition method based on distributed progressive domain adaptive diffusion modeling includes the following steps: Step 1: EEG signal acquisition and preprocessing.

[0027] EEG signals are characterized by strong noise, non-stationarity, and large individual variability, making it difficult to directly model stable patterns across subjects and sessions using the raw time-domain signals. Therefore, frequency domain decomposition combined with statistical feature modeling is needed to extract representations that are sensitive to emotions and relatively robust to individual differences. The raw EEG signal is downsampled to 200Hz and then filtered using a bandpass filter of 0.5-70Hz. This yields signals in five frequency bands: δ (1 Hz-3 Hz), θ (4 Hz-7 Hz), α (8 Hz-13 Hz), β (14 Hz-30 Hz), and γ (31 Hz-50 Hz). Differential entropy (DE) features are extracted for each band, and using DE features can effectively improve the performance of the emotion recognition model. Differential entropy (DE) is defined as follows: Wherein, the random variable X follows a mean of _____. The variance is normal distribution Step 2: Train the feature extractor and classifier on the source domain data.

[0028] Since the target domain is unlabeled, direct modeling can easily lead to semantic uncertainty and misalignment. Therefore, it is essential to first obtain a semantically stable source domain classifier with reliable discriminative ability. For example... Figure 1 As shown in the figure, the Source domain represents the source domain; the Target domain represents the target domain; Negative represents negative; Neutral represents neutral; and Positive represents positive. The left figure shows features after only preprocessing and feature extraction, the middle figure shows features generated using only DDEEG, and the right figure shows features using DDEEG and incorporating SRL. This invention uses a Multilayer Perceptron (MLP) as the feature extractor to map multi-band DE features to a low-dimensional latent space. The source domain data and target domain data share the same feature extractor to maintain feature space consistency, obtaining corresponding source domain latent representations and target latent representations, which serve as the initial source domain features and initial target features for the subsequent diffusion domain transfer process. The classifier is trained entirely on supervised training based on the source domain data, and its classification performance is optimized through cross-entropy loss for predicting different sentiment labels. The trained classifier will provide semantic constraints to the diffusion module in subsequent steps, thereby ensuring class consistency of features during cross-domain transformation.

[0029] Its loss function is defined as follows: in, Represents cross-entropy loss, Indicates the source domain tag. Represents source domain characteristics.

[0030] Step 3: Construct a progressive domain alignment process using a diffusion model.

[0031] This invention constructs a progressive domain adaptation framework, DDEEG, based on a diffusion model, the structure of which is shown in the attached figure. Figure 2 As shown. This module borrows the idea of ​​a denoising diffusion model, decomposing the transfer process from source domain features to target domain features into T consecutive steps. In each time step, the forward diffusion operator first... Noise is continuously injected into the source domain features, causing them to gradually deviate from the original distribution, and then the back-diffusion operator is used. At the same time step, the perturbed features are recovered, allowing them to slowly converge towards the target domain distribution. Each diffusion-anti-diffusion operation generates a new intermediate feature, thus forming... This invention utilizes continuous feature sequences to decompose large-scale cross-domain distribution differences into multiple fine-grained intermediate features, enabling source domain features to gradually migrate to the target domain distribution in a stable and controllable manner. Through this stepwise feature transformation mechanism, large-scale domain differences can be effectively decomposed, allowing the feature distribution to approximate the target domain more smoothly. To train the backdiffusion operator, this invention constructs a back-learning loss on the target domain features. This is used to improve its ability to reconstruct the target domain distribution from noise.

[0032] Step 4: Maintain feature semantic consistency through semantic reinforcement learning.

[0033] To ensure that the intermediate features generated by DDEEG do not undergo semantic drift during the transformation process, this invention proposes a semantic reinforcement learning (SRL) strategy, the structure of which is shown in the attached figure. Figure 3 As shown, it includes the following two parts: (1) Semantic diffusion learning Using the classifier trained in the previous stage, label prediction is performed on the currently generated transitional features to maintain consistency with the source domain labels. The cross-entropy loss between the predicted results and the true labels is used to optimize the DDEEG module, guiding the diffused features to retain correct semantic information at the class level.

[0034] (2) Generative learning All generated intermediate features As new training samples, these are used to further optimize the target domain classifier. To avoid the impact of low-quality early features on training performance, this invention introduces a dynamic confidence weight mechanism, assigning different training weights to features at different stages, enabling the classifier to better adapt to the continuously "evolving" feature distribution. The classifier's update loss is: in Indicates the dynamic confidence weight. This indicates that the DDEEG module processes the t-th time step. This indicates that the classifier processes the data at time step t. The two processes are executed alternately, enabling the classifier and the diffusion module to form a bidirectional collaborative optimization, thereby improving the ability to preserve cross-domain semantics.

[0035] Step 5: Generate the final target domain features and output the classification results.

[0036] After T diffusion-anti-diffusion iterations, this invention obtains the feature that most closely approximates the distribution of the target domain. The target domain data is processed by the same feature extractor and DDEEG module and then input into the classifier to obtain the final emotion recognition result. Experimental results show that the present invention achieves superior cross-subject and cross-session classification performance on both the SEED and SEED-IV datasets compared to existing methods.

[0037] To verify the rationality of the method proposed in this invention, several methods were compared, as shown in Table 1: Table 1

[0038] As can be seen from the table, the proposed DDEEG and SRL collaborative framework achieves significantly better performance than existing methods in both cross-subject and cross-session experiments on SEED and SEED-IV datasets, demonstrating superior performance in both mean accuracy and stability (standard deviation). Compared to traditional methods based on statistical distribution alignment, this invention utilizes a diffusion model to decompose the distributional differences from the source domain to the target domain into multiple progressive steps, making the domain offset alignment process smoother and more stable. Experimental results show that this invention achieves mean accuracy of 91.05% and 91.40% in cross-subject and cross-session experiments on the SEED dataset, respectively, and 75.64% and 73.11% on the SEED-IV dataset, respectively. It shows varying degrees of improvement over representative baseline methods (such as MS-MDA, PR-PL, STGATE, DANN, DAN, etc.), verifying that this invention can still achieve stable and efficient cross-domain feature transfer capabilities under large-scale domain differences. Furthermore, by observing the attached... Figure 1The three figures in this paper demonstrate that, compared with adversarial methods such as DANN and MS-MDA, the semantic reinforcement learning strategy introduced in this invention effectively alleviates the instability and semantic drift problems of adversarial training. The classifier maintains good discriminative ability on intermediate features generated at each stage. Furthermore, after introducing the SRL strategy, the visualized features exhibit more compact intra-class clustering and clearer inter-class separation, with obvious classification boundaries between different emotions. Overall, the EEG emotion recognition method based on distributed progressive domain adaptation diffusion modeling proposed in this invention can more effectively capture cross-domain invariant features of EEG signals in different subjects and different conversations, and significantly improve the robustness and accuracy of cross-domain emotion recognition, providing a more stable and interpretable solution for subsequent EEG domain adaptation methods.

[0039] The present invention has been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described herein. Various modifications, substitutions and variations made by those skilled in the art without departing from the core ideas of the present invention should fall within the protection scope of the present invention.

Claims

1. A brainwave emotion recognition method based on distributed progressive domain adaptive diffusion modeling, characterized in that, Includes the following steps: Step 1: Acquire EEG signals and form source domain features and target domain features; Step 2: Construct a feature extraction model to obtain the corresponding latent representations of the source domain and the target domain for the source domain features and the target domain features; Step 3: Based on the source domain latent representation and the target latent representation, align the distributions from the source domain to the target domain using DDEEG; Step 4: After aligning the source domain to the target domain distribution, train the target domain classifier; Step 5: In the testing phase, use the trained target domain classifier to identify the emotional state of the target domain based on the input EEG signal.

2. The EEG emotion recognition method based on distributed progressive domain adaptive diffusion modeling according to claim 1, characterized in that, Step 1 is specifically implemented as follows: using an EEG acquisition device to acquire EEG signals from different subjects or different sessions, performing bandpass filtering and segmentation processing on the signals, and calculating the differential entropy (DE) features of multiple frequency bands to form source domain features and target domain features.

3. The EEG emotion recognition method based on distributed progressive domain adaptive diffusion modeling according to claim 2, characterized in that, Step 2 is specifically implemented by constructing a feature extraction model based on an MLP encoder, and obtaining the corresponding source domain latent representation and target latent representation for the source domain features and target domain features, which serve as the initial source domain features and initial target features for the subsequent diffusion domain transfer process.

4. The EEG emotion recognition method based on distributed progressive domain adaptive diffusion modeling according to claim 3, characterized in that, In step 3, a diffusion-type domain transfer process is carried out based on the latent representation of the source domain and the latent representation of the target domain. DDEEG models the distribution alignment from the source domain to the target domain as a series of continuous and controllable feature diffusion-anti-diffusion steps, and explicitly constrains the diffusion direction through a semantic reinforcement learning mechanism.

5. The EEG emotion recognition method based on distributed progressive domain adaptive diffusion modeling according to claim 4, characterized in that, The DDEEG method specifically aligns the distribution from the source domain to the target domain by generating semantically continuous intermediate domain features after each diffusion-de-diffusion step. The intermediate domain features are input into the target domain classifier, and the source domain labels are used to supervise the diffusion results, directly embedding semantic consistency into the distribution migration process itself. At the same time, these intermediate domain features are used to back-train the target domain classifier, enabling the classifier to expand its discrimination boundary synchronously as the distribution changes, achieving progressive domain adaptation from easy to difficult.

6. The EEG emotion recognition method based on distributed progressive domain adaptive diffusion modeling according to claim 5, characterized in that, The semantic reinforcement learning mechanism includes: Classifier-guided semantic diffusion learning in the diffusion module of DDEEG: using a classifier with discriminative capabilities to constrain the diffusion-re-diffusion process; Generative learning that drives classifier generalization through intermediate domain features: The classifier is trained step by step by continuously generated intermediate domain features, gradually adapting to the target domain distribution without directly crossing large domain differences.

7. The EEG emotion recognition method based on distributed progressive domain adaptive diffusion modeling according to claim 6, characterized in that, Step 4 is specifically implemented as follows: after the source domain to target domain distribution pairs are obtained, all intermediate domain features are integrated to train the final target domain classifier, and the intermediate domain features generated by the diffusion module are used to... As a joint training set, a dynamic confidence weighting mechanism is used for training, enabling the target domain classifier to achieve optimal recognition performance on multiple asymptotic distributions.

Images