可拓门控网络驱动的下肢外骨骼智能控制方法研究

PDF下载HTML

Title:

Extension gating network-driven intelligent control method for lower-limb exoskeletons

作者:

余昊阳^1,2, 李兴森^1,2, 颜伊庆^1,2

1. 广东工业大学 可拓学与创新方法研究所, 广东 广州 510006;
2. 广东工业大学 机电工程学院, 广东 广州 510006

Author(s):

YU Haoyang^1,2, LI Xingsen^1,2, YAN Yiqing^1,2

1. Institute of Extenics and Innovation Methods, Guangdong University of Technology, Guangzhou 510006, China;
2. School of Electromechanical Engineering, Guangdong University of Technology, Guangzhou 510006, China

关键词:

可拓神经网络; 时间卷积网络; 混合专家模型; 可拓智能; 可拓距; 下肢外骨骼; 关节力矩估计; 智能控制系统

Keywords:

extension neural network; temporal convolutional network; mixture-of-experts; extension intelligence; extension distance; lower-limb exoskeleton; joint moment estimation; intelligent control system

分类号:

TP18

DOI:

10.11992/tis.202510022

摘要:

下肢外骨骼技术在康复医疗与日常助行领域展现出广阔应用前景，但其在复杂地形下的自适应控制智能性不足。针对现有下肢外骨骼在复杂地形中控制策略迁移性差、动态适应性弱等问题，提出一种融合可拓门控网络与时间卷积专家的混合专家模型。该模型利用可拓距函数动态量化地形特征与专家模块的匹配度，实现多源步态信息的智能路由与自适应决策，并通过时间卷积专家系统的时序建模能力，生成符合人体运动特征的精准辅助力矩。在包含平地、斜坡等多种地形的实验中，模型髋关节力矩预测平均误差较LSTM(long short-term memory)降低18.71%，平均推理时间仅为37.09 ms。在场景转换实验中，模型平均误差较LSTM降低17.75%。该模型有效提升了外骨骼系统的跨场景适应能力，其可拓门控网络路由机制与信息融合架构对多模态机器人决策系统、神经网络路由优化等研究领域均有一定参考价值。

Abstract:

Lower-limb exoskeleton technology holds substantial promise for applications in rehabilitation medicine and daily mobility assistance. However, achieving adaptive control across complex terrains remains a critical challenge. To address the limited transferability of control strategies and inadequate dynamic adaptability in current systems, this study introduces a mixture-of-experts model that integrates extension gating network with temporal convolutional experts. The model utilizes an extension distance function to dynamically quantify the compatibility between terrain features and expert modules, thereby enabling intelligent routing and adaptive decision-making based on multi-source gait information. By leveraging the temporal modeling capabilities of temporal convolutional experts, the system generates precise assistive torques aligned with human motion patterns. In experiments involving multiple terrains including flat ground and slopes, the proposed model reduced the average hip joint torque prediction error by 18.71% compared to conventional LSTM, with an average inference latency of merely 37.09 ms. In scene transition experiments, the model demonstrates a 17.75% reduction in average error compared to LSTM. This model effectively enhances the cross-terrain adaptability of exoskeleton systems, and its extension gating network routing mechanism and information fusion architecture provide valuable insights for multimodal robotic decision-making and dynamic neural network routing optimization.