MDF-Net: A Misclassified Data Fusion Network for Locomotion Mode Misclassification Detection and Correction in Lower-Limb Exoskeletons

Accurate recognition of human locomotion intent is essential for the cooperative control of exoskeletons in human-machine interaction. Threshold-based or result-driven strategies are widely used for misclassification detection and correction but overlook the intrinsic traits of misclassified data. This study proposes a misclassified data fusion network (MDF-Net). Firstly, a dual-label supervision mechanism is introduced, combining misclassification indicators with locomotion mode labels to enable multiple functions, including locomotion mode recognition, misclassification detection, and correction. Secondly, multi-class focal loss (MFL) replaces the conventional cross-entropy (CE) loss to address class imbalance between correctly and incorrectly classified samples. In addition, a hyperparameter is defined to adjust the loss weights for multi-class classification tasks. A motion dataset collected from a human-exoskeleton experiment, where subjects wore a rigid exoskeleton to perform locomotion tasks, was used to evaluate MDF-Net. MDF-Net achieved an average recognition accuracy of 99.99% across three locomotion modes and two slope conditions for five subjects, which was superior to the state-of-the-art models tested. Furthermore, MFL reduced misclassifications by 87.5% compared to CE. MDF-Net maintained an average recognition accuracy above 95% when testing on augmented data. On the exoskeleton control board, the model ran with an average inference time of 0.45 ms per sample. Results show that MDF-Net can effectively detect and correct misclassifications, significantly improving locomotion mode recognition accuracy, while demonstrating real-time performance and generalization capability, thereby ensuring safe human-exoskeleton interaction.