This thesis addresses one of the key challenges in the field of lower-limb exoskeleton robots—namely, the classification of locomotion conditions an‎d the prediction of future joint trajectories. These tasks play a crucial role in enhancing the safety, efficiency, an‎d intelligence of such robotic systems. To classify various locomotion conditions, including level-ground walking, ramp ascent/descent, an‎d stair ascent/descent, a hybrid model based on convolutional an‎d recurrent neural networks with an attention mechanism was designed an‎d trained. The model achieved an accuracy of over 95% across all conditions. Furthermore, to accurately predict the future trajectories of lower-limb joints, a multi-scale hybrid network was developed, capable of capturing complex spatiotemporal patterns. The dataset used in this study consisted of signals from inertial sensors an‎d goniometers, which were preprocessed before being fed into the models. The results demonstrated that the proposed methods achieved outstan‎ding performance in both classification an‎d prediction tasks, representing a significant step toward improving exoskeleton control an‎d assistive applications. In addition, the transformer-based network proposed in this work representing a novel approach to time series prediction not only exhibits high generalization capability but also effectively identifies influential an‎d relevant input features, making it a powerful tool for developing the next generation of wearable robotic systems.

حامد جلالي بيدگلي

محسن مجيري فروشاني , مريم ذكري