A Multitask Deep Learning Approach for sEMG-Based Human Motion Intention and Muscle Fatigue Levels Recognition

The natural interaction between the lower limb prosthetic and amputee directly affects the rehabilitation effect. Human motion intention and muscle fatigue are the most relevant factors affecting the rehabilitation training of amputees. Previous methods focus on structuring independent models for each task, neglecting the problem of motion intention recognition (MIR) under muscle fatigue constraints. To remedy this deficiency, this article proposes a multitask deep learning approach to tackle MIR and muscle fatigue classification (MFC) tasks simultaneously. First, the surface electromyography (sEMG) samples dataset of three limb motions with different muscle fatigue levels are collected and preprocessed. Second, we create a MIR recognition model based on neural networks and then utilize its knowledge-sharing backbone feature network to learn a new MFC task with a limited dataset. Next, a multitask loss function with uncertainty is leveraged to weigh different losses between two tasks and improve model performance. Finally, we compare the accuracy of the proposed model against other models train individually on each task, demonstrating the proposed model effectiveness in terms of data efficiency. Simultaneously, the proposed approach achieves 96.8% accuracy in MIR task under fatigue constraints, while achieves 93.2% accuracy in MFC task which a limited dataset, respectively.