文献类型：期刊文献

中文题名：基于LSTM-CRF的躯干肌肉行为仿真校正

英文题名：Calibration of Trunk Muscle Behavior Simulation Based on LSTM-CRF

作者：王琦[1];王庆明[2];周志勇[1];杨杰[1]

机构：[1]上海电机学院设计与艺术学院,上海200240;[2]华东理工大学机械与动力工程学院,上海200237

年份：2022

卷号：39

期号：10

起止页码：304

中文期刊名：计算机仿真

外文期刊名：Computer Simulation

收录：CSTPCD;;北大核心:【北大核心2020】；

语种：中文

中文关键词：行为分类;椎旁肌;长短时记忆神经网络;条件向量场;两步聚类

外文关键词：Behavior classification;Paravertebral muscles;Long short-term memory network(LSTM);CRF;Two-step clustering

摘要：弯伸腰过程中,躯干肌肉的发力状态可迅速在拮抗和放松间转换,基于有限元分析的预测方法很难识别肌肉所处状态。提出了一种能记忆行为特征,用于识别发力状态的方法,校正现有人体有限元模型的预测结果。先同步采集弯伸腰动作中的肌电数据和运动数据,用肌电信号的聚类分组结果代替肌电信号,作为LSTM-CRF构架的输入,实现对一组躯干肌肉行为的编码和标注,标注步骤的F1值的均值为91.37%。方法能够在长时间内记忆个体行为特征,还可校正其它躯干人体组织受力状态的预测,预测结果对个体特征和手部工作任务差异敏感。
During the flexion-extension process, the exertion state of the trunk muscles may switch between antagonism and flexion-relaxation state rapidly. The method based on finite element analysis has difficulty recognizing the state adopted. Therefore, a method to record individual behavior characteristics was proposed which could recognize the exertion state and calibrate the predictions based on existing finite element analysis. The EMG and motion data during the flexion-extension process were collected synchronously. The clustering results of EMG signals took place of the EMG to be used as the input signals of LSTM-CRF architecture which was established to code and label the behavior of a group of trunk muscles. The averaged F1 value of the labeling steps is 91.37%. The method can memorize the individual features. At the same time, it’s able to calibrate the predictions on the exertion state of other trunk tissues. The prediction results are sensitive to individual features and diversified manual tasks.