Scholarlink

Abstract <p>During exoskeleton rehabilitation training, providing the appropriate torque for recovery can enhance both the effectiveness and comfort of the training. However, due to torque sensor failures and wireless transmission issues, data acquisition may suffer from a loss of 200–800 time steps, which limits the model’s predictive accuracy. Therefore, we designed a torque data reconstruction method based on a simplified neuromusculoskeletal model. This approach uses an interpretable model to generate estimated torques for reconstructing missing segments, thereby simulating the temporal characteristics of torque data. Experimental results demonstrate that training neural network models with reconstructed data yields an average 12.70% improvement in the <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> of predicted torque compared to training with non-reconstructed data. This performance surpasses that of reconstruction methods based on periodic shift interpolation, chained equations multiple imputation, and neural network-based imputation for missing time series data.</p> Graphical abstract <p></p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A reconstruction method of missing torque data using a simplified neuromusculoskeletal (NMS) model

  • 期刊论文

  • 发表日期：

    2026年8月5日

查看全文
  • Zhe Sun,
  • Xiaoyun Bi,
  • Lubin Hong,
  • Jinchuan Zheng,
  • Zhihong Man

摘要

Abstract

During exoskeleton rehabilitation training, providing the appropriate torque for recovery can enhance both the effectiveness and comfort of the training. However, due to torque sensor failures and wireless transmission issues, data acquisition may suffer from a loss of 200–800 time steps, which limits the model’s predictive accuracy. Therefore, we designed a torque data reconstruction method based on a simplified neuromusculoskeletal model. This approach uses an interpretable model to generate estimated torques for reconstructing missing segments, thereby simulating the temporal characteristics of torque data. Experimental results demonstrate that training neural network models with reconstructed data yields an average 12.70% improvement in the \(R^2\) of predicted torque compared to training with non-reconstructed data. This performance surpasses that of reconstruction methods based on periodic shift interpolation, chained equations multiple imputation, and neural network-based imputation for missing time series data.

Graphical abstract