Multimodal emotion recognition has demonstrated considerable potential for advancing the understanding of human affective states by integrating heterogeneous sources of information. Nevertheless, existing methods often encounter difficulties in effectively unifying modality-specific representations and in dynamically adapting to variable input quality. To address these challenges, this paper introduces FAME-Graph (Fidelity-Aware Multimodal Emotion Recognition with Graph Networks), a novel framework that combines modality-specific feature extractors, Mixture-of-Experts (MoE) transformers, and Dynamic Graph Convolutional Neural Networks (DGCNNs). The proposed approach first projects features derived from multiple modalities, including physiological signals, audio, video, and text, into a unified latent space. Each modality is subsequently processed by a specialized transformer encoder functioning as a domain expert within the MoE architecture. A fidelity-aware fusion mechanism, grounded in learned Beta distributions, adaptively assigns confidence-based weights to capture modality reliability. The fused representation is then refined using DGCNN-based feature integration, incorporating k-nearest neighbor graph construction, edge feature extraction, and residual connections. Extensive experiments are conducted on three benchmark datasets, including DEAP for physiological signal-based emotion recognition and IEMOCAP and MELD for conversational multimodal emotion recognition. Experimental results demonstrate that FAME-Graph achieves substantial improvements over state-of-the-art methods across different modality settings and challenging emotional contexts. These findings underscore that the synergy of modality-specific experts, fidelity-aware weighting, and graph-based representation learning yields a more accurate, robust, and generalizable solution for multimodal emotion recognition.

This study presents a memory sampled-data control design for the Takagi–Sugeno (T–S) fuzzy-based permanent magnet synchronous generator (PMSG) via an improved looped functional (ILF). To do this, first, the nonlinear PMSG is modeled as a T–S fuzzy system. Then, to derive the sufficient criteria, a novel ILF is proposed, which includes the available information about the sampling pattern characteristics from$\mathbf {x}(t_{k})$to$\mathbf {x}(t)$and$\mathbf {x}(t)$to$\mathbf {x}(t_{k+1})$together with the signal transmission delay. In addition, the ILF introduces a fractional parameter$0 < \hat {\beta } < 1$that gives more information of splited sampling intervals. Also, the external disturbances of the proposed system are attenuated by using the$H_{\infty }$performance. Furthermore, the stabilization conditions expressed as linear matrix inequalities (LMIs), the proposed closed-loop system’s asymptotic stability is ensured under the designed controller. Finally, comparison and simulation results demonstrate the effectiveness and feasibility of the proposed method and the designed control scheme.