FRFNet: a fatigue-robust EEG–EMG fusion network for hand movement intention recognition

Abstract

Introduction Accurate movement intention recognition is essential for intelligent rehabilitation systems, wearable assistive devices, and human–machine interaction. Electroencephalography (EEG) reflects cortical motor intention and cognitive regulation, whereas surface electromyography (sEMG) provides peripheral muscle activation information related to movement execution. Although EEG–EMG fusion can provide complementary information from both the central and peripheral nervous systems, EMG signals are easily affected by muscle fatigue during long-duration or high-intensity rehabilitation tasks, resulting in amplitude attenuation, signal-to-noise ratio degradation, and modality reliability imbalance. Methods To address this problem, this study proposes a fatigue-robust brain–muscle fusion network, termed FRFNet. The proposed framework contains a multiscale adaptive temporal convolutional network for EEG encoding, a dual-path fatigue disentanglement encoder for EMG representation, and a fatigue-aware dynamic weighted adaptive gating fusion mechanism. The EEG branch captures relatively stable movement-related cortical patterns through multiscale temporal modeling and spatial–temporal attention. The EMG branch separates fatigue-invariant action features from fatigue-sensitive degradation information, enabling implicit fatigue estimation without requiring explicit fatigue labels. The dynamic weighted adaptive gating module adjusts the contributions of EEG and EMG according to modality quality and fatigue state. Results Experiments were conducted on a public multimodal EEG–EMG dataset involving intuitive upper-limb movement tasks. Synthetic fatigue-related EMG degradation levels from 10 to 90% were introduced for controlled robustness evaluation. Under the within-subject protocol, FRFNet achieved 60.0% accuracy at 90% degradation, exceeding E2FNet, DCA Fusion, and DMEFNet-adapted by 12.7, 15.2, and 11.7 percentage points, respectively. Under strict leave-one-subject-out evaluation, FRFNet achieved 57.0% accuracy at the same level and maintained the highest overall performance among the six evaluated methods. The deployment inference path contained 1.166 million parameters and achieved a mean single-trial forward latency of 5.321 ± 0.934 ms. Conclusion The results indicate that degradation-aware adaptive fusion can suppress unreliable EMG features and improve EEG–EMG movement decoding under progressive EMG degradation. FRFNet therefore provides a robust multimodal decoding framework for fatigue-aware rehabilitation and adaptive human–machine interaction under the evaluated synthetic EMG degradation conditions.