| 摘 要: 针对膝关节外骨骼任务无关控制对实时力矩估计的需求，提出一种基于多尺度特征融合深度可分离时序卷积网络(MSFF-DSTCN)的力矩估计方法。该模型融合深度可分离卷积、多尺度特征提取与通道注意力机制，仅利用髋、膝关节角度估计膝关节力矩。在11种日常活动数据集上，平均R2=0.69，RMSE=0.149 N·m/kg，模型大小5714 KB，推理时间3.9 ms。相比标准TCN，精度提升5.7%，模型压缩8%。消融实验验证了各模块的有效性。该方法在简化传感器配置下兼顾精度与效率，为外骨骼轻量化控制提供算法基础。 |
| 关键词: 膝关节力矩估计 任务无关控制 时序卷积网络 多尺度特征 深度可分离卷积 |
|
中图分类号: TP183
文献标识码:
|
| 基金项目: 浙江省自然科学基金项目 |
|
| Research on knee joint moment estimation based on improved temporal convolutional network |
|
BianYuwei, WangChen
|
Zhejiang Sci-Tech University
|
| Abstract: To address the demand for real-time moment estimation in task-agnostic knee exoskeleton control, this paper proposes a moment estimation method based on a Multi-Scale Feature Fusion Depthwise Separable Temporal Convolutional Network(MSFF-DSTCN). The model integrates depthwise separable convolutions, multi-scale feature extraction, and channel attention mechanisms to estimate knee moment using only hip and knee joint angles. On a dataset containing 11 daily activities, the method achieves an average R2 of 0.69, RMSE of 0.149 N·m/kg, a model size of 5714 KB, and an inference time of 3.9 ms per sequence. Compared with standard TCN, accuracy improves by 5.7% and model size reduces by 8%. Ablation studies validate the effectiveness of each module. The method balances accuracy and efficiency, providing an algorithmic foundation for lightweight exoskeleton control. |
| Keywords: knee moment estimation task-agnostic control temporal convolutional network multi-scale features depthwise separable convolution |
