Multimodal fusion of Electroencephalography (EEG) and functional Near-Infrared Spectroscopy (fNIRS) has shown great promise in Brain-Computer Interface (BCI) tasks. However, due to differences in physical mechanisms, temporal dynamics, and semantic representations between the two modalities, the fusion process faces significant challenges such as heterogeneity and temporal misalignment. To address this, we propose a cross-modal decoupled deformable distillation (CMD$^{3}$) method, which aims to achieve flexible, efficient, and interpretable EEG-fNIRS fusion learning. CMD$^{3}$first decouples the feature representations of each modality into modality-independent and modality-specific spaces to separately model commonality and complementary information. A deformable feature extraction network is then designed to process shared and specific features individually, enabling cross-modal temporal alignment via predicted dynamic offsets, thereby mitigating response delays between modalities. Furthermore, to facilitate inter-modal knowledge transfer, we construct a dual-space graph distillation module to explicitly migrate semantic information across modalities, with learnable edge weights used to adaptively regulate the distillation strength. CMD$^{3}$is systematically evaluated on public datasets covering emotion recognition and motor imagery tasks. Experimental results demonstrate that CMD$^{3}$consistently outperforms existing fusion approaches in classification performance. Offset visualization further reveals physiologically meaningful temporal attention patterns learned by the model, validating the effectiveness and explainability of the proposed method.

Multimodal peripheral physiological signal fusion for emotion recognition seeks to perceive or recognize human emotions using peripheral modalities such as electromyography, electrodermal activity, and respiratory wave. Previous approaches to multimodal fusion primarily focus on emotion-sensitive signals such as electroencephalogram (EEG), often overlooking the potential value of peripheral physiological signals in emotion recognition. Moreover, the inherent heterogeneity among different modalities continues to pose challenges to fusion quality. In this article, we propose a multimodal decoupled multimodal fusion (DMF) to address these issues, enabling flexible feature decoupling, cross-modal feature interaction, and relational knowledge learning. Specifically, each modality’s representation is first decoupled into two components: common features and modality-specific features; second, the DMF employs progressive cross attention to facilitate the exchange of modality-specific features across different modalities; and finally, it uses relational knowledge to learn multimodal spliced features, embedding both inter-modal and intra-modal feature relationships. DMF offers a dynamic multimodal emotion recognition framework that leverages the emotional information contained in diverse modalities. Experimental results demonstrate that the DMF method consistently outperforms previous approaches and provides a viable solution for multimodal peripheral physiological signal fusion.

Domain generalization remains a key challenge in human activity recognition based on channel state information (CSI). Different domains correspond to distinct data distributions, deviating from the typical assumption of independent and identically distributed (i.i.d.) data, which leads to significant performance degradation when models are applied to unseen domains. To address this issue, we propose a novel domain generalization model that integrates meta-learning initialization and an adaptive channel grouping attention mechanism. First, a meta-learning strategy is employed to acquire well-initialized parameters from multiple source domain tasks, enabling the model to implicitly enhance its cross-domain generalization ability. Second, an adaptive grouping attention mechanism is designed in the feature extraction stage to effectively capture the sensitivity differences of different subcarriers to human activities. Meanwhile, a random masking training mechanism is introduced to simulate real-world domain variations and improve model robustness. In addition, a domain adversarial training framework based on the gradient reversal layer (GRL) is adopted to mitigate domain-specific feature dependency, further enhancing the model's generalization capability. We evaluate our proposed method on both a self-collected dataset, which includes human activity data from nine volunteers across six different environments, and a public CSI dataset. The experimental results demonstrate that our method significantly outperforms existing approaches in domain generalization performance, verifying its effectiveness and practical applicability.