基于双向长短期记忆模型的起重机智能操控方法  

Intelligent manipulation method of crane based on BiLSTM model

作　　者：倪涛 刘海强 王林林 邹少元 张红彦 黄玲涛 NI Tao;LIU Hai-qiang;WANG Lin-lin;ZOU Shao-yuan;ZHANG Hong-yan;HUANG Ling-tao(School of Mechanical and Aerospace Engifieering,Jilin University,Changchun 130022,China)

机构地区：[1]吉林大学机械与航空航天工程学院,长春130022

出　　处：《吉林大学学报（工学版）》2020年第2期445-453,共9页Journal of Jilin University:Engineering and Technology Edition

基　　金：国家自然科学基金项目(51575219).

摘　　要：为解决地面指挥人员和司机协同操控起重机难度大的问题,提出一种基于双向长短期记忆(BiLSTM)模型的起重机智能操控方法,实现了对起重机的单人控制,降低了人力成本。首先,使用Kinect从指挥人员的一段动作指令中采集出人体关节点的坐标序列;然后,利用这些坐标构造人体关节向量,通过计算向量间的夹角和模比值构造出两种区分不同指令的特征矩阵;之后,将夹角特征矩阵输入到基于BiLSTM模型的指令识别网络,并与支持向量机(SVM)和反向传播(BP)神经网络的识别结果作比较;最后,将夹角和模比值特征矩阵进行融合识别,以进一步提升准确率。实验结果表明:本文指令识别网络具有较高的识别率;提出的融合识别方法有效地利用多种特征的信息,对训练集的识别准确率达99.13%,对测试集的识别准确率达96.75%。To solve the problem that it is difficult for commander on the ground and driver to operate cranes cooperatively,an intelligent control method of the crane based on BiLSTM model was proposed,which realizes single-person control of crane,thus,reducing manpower cost. Firstly,Kinect sensors were used to collect the coordinates of human joints from the commander′s instructions. Then,the vectors of human joints were constructed using the coordinates,and two kinds of characteristic matrices that distinguish different instructions were constructed by calculating the angles and modulus ratios between the vectors.After that,the feature matrix of angle was input into the instruction recognition network based on BiLSTM model,and the recognition results were compared with those of support vector machine(SVM)and back propagation(BP)neural networks. Finally,the feature matrices of angle and modulus ratio were fused to improve the accuracy. The experimental results show that the instruction recognition network designed in this paper achieves the highest accuracy rate and the proposed fusion recognition method utilizes the information of multiple features effectively. For the training set,the recognition rate can reach 99.13%,and for the test set,this rate can reach 96.75%.

关 键 词：起重机智能操控 吊运指令识别 特征矩阵 双向长短期记忆模型 多特征融合识别 

分 类 号：TH218[机械工程—机械制造及自动化]