无监督学习型凿岩钻臂逆运动学求解方法  

Inverse kinematics solution of an unsupervised learning drilling boom

作　　者：吴江 许皓渊 闫昊琪 熊锋 曹星宇 高路路[1] 段京良 马飞[1] WU Jiang;XU Haoyuan;YAN Haoqi;XIONG Feng;CAO Xingyu;GAO Lulu;DUAN Jingliang;MA Fei(School of Mechanical Engineering,University of Science and Technology Beijing,Beijing 100083,China)

机构地区：[1]北京科技大学机械工程学院,北京100083

出　　处：《工程科学学报》2024年第8期1479-1488,共10页Chinese Journal of Engineering

基　　金：国家重点研发计划资助项目(2021YFB3401502);国家自然科学基金资助项目(52202487);中央高校基本科研业务费专项资金资助项目(FRF-OT-23-02)。

摘　　要：凿岩台车钻臂智能寻孔控制对提升凿岩钻孔作业精度和效率具有重要意义,逆运动学求解是实现钻臂精确快速寻孔控制的核心.现有解析法或数值法无法满足复杂钻臂逆运动学求解精度或时间效率要求,而传统神经网络方法依赖于标签数据,求解结果的可靠性较低.针对上述问题,本文提出一种考虑安全约束的无监督学习型神经网络逆运动学求解方法.区别于传统解法,该方法不依赖标签数据,直接将期望钻臂末端位姿作为网络输入,以实际末端位姿与期望末端位姿的差异作为优化目标,通过梯度下降驱动网络更新.同时,为确保关节位置的安全性,本文构造了安全碰撞惩罚,将罚项引入到求解目标函数中,促使网络输出的关节量满足特定环境的约束条件.上述的研究方法不仅提高了逆运动学求解的精度,而且显著降低了逆运动学解的碰撞率.实验结果表明,使用无监督学习型神经网络逆运动学求解方法所求得的寻孔误差均值在5~7 mm之间,相较于监督学习型方法,逆运动学求解精度提升约70.72%;引入约束后,该方法在不损失求解精度的前提下,逆运动学解的碰撞率降低了90.28%.Intelligent control of the hole-seeking process of a rock drilling rig boom is crucial for enhancing the accuracy and efficiency of rock drilling operations.The inverse kinematics solution(IKS)is the core of achieving precise and rapid hole-seeking control of the boom.However,existing analytical and numerical techniques are inadequate in fulfilling the accuracy and time efficiency requirements for IKS in complex drilling boom scenarios.Conventional neural network(NN)approaches heavily depend on labeled data comprising target borehole sets derived from the activities of each joint and forward kinematics.The distribution of drill endpoints generated by this data cannot cover the entire workspace,resulting in the low reliability of the solution.To overcome these challenges,this study introduces an unsupervised learning-based NN method for IKS emphasizing safety constraints.This innovative method differs from traditional approaches in that it does not depend on labeled data.Instead,it utilizes the desired end position of the drilling boom as the network input.The network generates an eight-dimensional joint vector,and the actual drill end pose is derived through forward kinematics calculations on this vector.Then,the difference between the actual and desired drill end poses is used as the optimization objective,driving the network updates through gradient descent.The advantage of this method lies in eliminating the need for complex joint label data required in supervised learning IKS and using the differences in drill end poses directly as optimization objectives,which helps improve the accuracy of IKS.Meanwhile,a critical innovation of this study is integrating a safety collision penalty into the objective function of the solution,ensuring that the network’s output for joint positions adheres to specific environmental limitations.If the actual distance falls below the safety threshold,penalty terms are incorporated into the objective function,allowing for the adjustment of weights to maintain a balance between the

关 键 词：凿岩钻臂 逆运动学求解 无监督学习 神经网络 碰撞惩罚 

分 类 号：TG142.71[一般工业技术—材料科学与工程]