中国组织工程研究 ›› 2026, Vol. 30 ›› Issue (16): 4125-4136.doi: 10.12307/2026.685

• 组织构建与生物力学 tissue construction and biomechanics • 上一篇    下一篇

基于Transformer-卷积神经网络模型实现单节点腰部康复训练动作识别任务

余圣涵1,2,成贤锴1,2,郑  跃1,2,杨  颖2,3   

  1. 1中国科学技术大学生物医学工程学院(苏州),江苏省苏州市  215163;2中国科学院苏州生物医学工程技术研究所,江苏省苏州市  215163;3苏州高新区人民医院,江苏省苏州市  215129


  • 收稿日期:2025-07-01 接受日期:2025-08-22 出版日期:2026-06-08 发布日期:2025-11-27
  • 通讯作者: 成贤锴,理学硕士,副研究员,中国科学技术大学生物医学工程学院(苏州),江苏省苏州市 215163;中国科学院苏州生物医学工程技术研究所,江苏省苏州市 215163
  • 作者简介:余圣涵,男,2001年生,福建省莆田市人,汉族,中国科学技术大学在读硕士,主要从事惯性传感器数据分析的研究。
  • 基金资助:
    国家重点研发计划项目(2023YFC3604804);江苏省重点研发计划项目(BE2022064-2),项目负责人:成贤锴

A Transformer-based convolutional neural network fusion approach for single inertial recognition of lumbar rehabilitation exercises

Yu Shenghan1, 2, Cheng Xiankai1, 2, Zheng Yue1, 2, Yang Ying2, 3   

  1. 1School of Biomedical Engineering (Suzhou), University of Science and Technology of China, Suzhou 215163, Jiangsu Province, China; 2Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215163, Jiangsu Province, China; 3The People’s Hospital of Suzhou New District, Suzhou 215129, Jiangsu Province, China
  • Received:2025-07-01 Accepted:2025-08-22 Online:2026-06-08 Published:2025-11-27
  • Contact: Cheng Xiankai, MS, Associate researcher, School of Biomedical Engineering (Suzhou), University of Science and Technology of China, Suzhou 215163, Jiangsu Province, China; Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215163, Jiangsu Province, China
  • About author:Yu Shenghan, MS candidate, School of Biomedical Engineering (Suzhou), University of Science and Technology of China, Suzhou 215163, Jiangsu Province, China; Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou 215163, Jiangsu Province, China
  • Supported by:
    National Key Research and Development Program of China, No. 2023YFC3604804; Key Research and Development Program Project of Jiangsu Province, No. BE2022064-2  (to CXK)

摘要:


文题释义:
深度学习:是机器学习的分支,无需人工设计特征,通过多层神经网络自动从大量数据中提取特征,其核心在于端到端的学习,能够处理图像、语音、文本等复杂数据,在计算机视觉、自然语言处理等领域表现卓越。
Transformer:是一种基于自注意力机制的深度学习模型,最初用于自然语言处理,通过并行计算和全局依赖建模,解决了传统神经网络难以捕捉长距离依赖的问题。Transformer剔除循环结构,仅通过注意力机制和前馈网络进行信息传递,结构简单,计算效率高,广泛应用于语言模型、图像处理和时间序列分析等任务中。

背景:惯性测量单元被广泛用于人体姿态感知与动态捕捉。深度学习已逐步替代传统规则与特征工程,广泛应用于动作识别任务。卷积神经网络在提取局部动态特征方面表现良好,Transformer则在建模长时序依赖方面展现出强大能力。
目的:通过基于Transformer-卷积神经网络融合模型识别方法,实现在单惯性传感器条件下的腰部康复训练动作识别任务。
方法:采集6名健康受试者佩戴单个惯性传感器条件下执行腰部康复动作的加速度与角速度数据,以动作类型为数据进行标注,制作腰部康复动作数据集。通过腰部康复动作数据集对Transformer-卷积神经网络融合模型进行训练,构建动作分类模型。通过留一交叉验证评估模型准确性,并与线性判别分析、支持向量机、多层感知、经典Transformer等模型进行性能对比。
结果与结论:在5类动作识别任务中,Transformer-卷积神经网络模型准确率达96.67%,F1-score为0.966 9。在单传感器输入的条件下,相较于传统模型,在识别精度与泛化能力方面具有明显优势。验证了基于单惯性测量单元数据的深度模型在腰部康复动作分类任务中的实用性,为轻量化、高部署性的居家腰部康复训练系统提供基础。

https://orcid.org/0009-0000-3591-4871(余圣涵)https://orcid.org/0009-0000-3591-4871(余圣涵)

中国组织工程研究杂志出版内容重点:干细胞;骨髓干细胞;造血干细胞;脂肪干细胞;肿瘤干细胞;胚胎干细胞;脐带脐血干细胞;干细胞诱导;干细胞分化;组织工程

关键词: 慢性腰痛, 康复训练, 深度学习, Transformer, 单节点惯性传感器, 动作分类

Abstract: BACKGROUND: Inertial measurement units are widely used for human posture perception and dynamic capture. Deep learning has gradually replaced traditional rules and feature engineering, and is commonly used in action recognition tasks. Convolutional neural networks perform well in extracting local dynamic features, while Transformer-based convolutional neural networks approach demonstrates strong capabilities in modeling long-term dependencies.
OBJECTIVE: To develop a recognition method based on a Transformer-convolutional neural networks-based fusion model to classify lumbar rehabilitation exercises using data from a single inertial measurement unit.
METHODS: A dataset was constructed by collecting tri-axial accelerometer and gyroscope signals from six healthy participants performing standardized lumbar rehabilitation movements with a single waist-mounted inertial measurement unit. Each trial was labeled according to the exercise type. Transformer-based convolutional neural network fusion model was trained using this dataset to construct a motion classification system. Model performance was evaluated using leave-one-out cross-validation and compared against baseline models, including  linear discriminant analysis, support vector machine, multi-layer perception, and the standard Transformer. 
RESULTS AND CONCLUSION: Experimental results show that the proposed Transformer-based convolutional neural network fusion model achieved a classification accuracy of 96.67% and an F1-score of 0.966 9 across five exercise categories. Compared with conventional algorithms, it demonstrated superior accuracy and generalizability under the single-sensor constraint. These findings validate the practicality of deep learning models using single inertial measurement unit data for lumbar rehabilitation monitoring and provide a foundation for developing lightweight, highly deployable home-based rehabilitation systems.


Key words: chronic back pain, rehabilitation training, deep learning, Transformer, single inertial sensor, action classification

中图分类号: