基于深度学习控制障碍函数的无人机安全避障方法  

UAV Safety Obstacle Avoidance Method Based on Deep Learning Control Barrier Function

作　　者：刘晨晖 刘建敏 刘钦朋 潘泉[2] LIU Chenhui;LIU Jianmin;LIU Qinpeng;PAN Quan(School of Science,Xi'an Shiyou University,Xi'an 710000,China;College of Automation,Northwestern Polytechnical University,Xi'an 710000,China)

机构地区：[1]西安石油大学理学院,西安710000 [2]西北工业大学自动化学院,西安710000

出　　处：《无人系统技术》2025年第1期50-67,共18页Unmanned Systems Technology

摘　　要：在无人机应用日益广泛的背景下,其系统模型不确定性给规避控制带来挑战,针对此提出基于深度学习的安全学习控制框架。首先在已知环境中,明确避撞安全要求,以预设避撞算法在仿真环境中生成专家轨迹数据,并进行数据泛化、构建边界集等处理,结合人工数据集和安全数据滤波器构建综合数据集,用神经网络拟合控制障碍函数,通过拉格朗日乘子法求解相关约束优化问题,设计特定神经网络作为逼近器。其次,在未知环境下,从激光雷达传感器获取数据转换到世界坐标系,依观测模型区分安全与不安全区域收集样本,利用支持向量机拟合控制障碍函数,采用核函数及松弛系数优化,结合Sigmoid函数等得到逼近器。接着构建安全学习控制框架,离散化系统模型并添加不确定性参数,定义包含系统状态和控制指令序列的成本函数,设置硬、概率、软三种安全约束,建立相应算法模型。最后在PyCharm平台实验,验证已知和未知环境下相关算法及集成框架在不同障碍物环境的有效性,如已知环境中学习后的控制障碍函数能助力无人机避障并到达目标,未知环境下基于观测学习的算法可消除噪声实现避障。结果表明框架性能良好,有望用于实际提升无人机自主安全控制能力。With the increasing application of unmanned aerial vehicles(UAVs),the uncertainty of the UAV system model poses challenges to avoidance control.This paper proposes a deep learning-based safety control framework.Firstly,in the known environment,the safety requirements for collision avoidance are defined.Expert trajectory data generated by preset algorithms in the simulation environment or collected by operators are processed through data generalization,boundary set construction,etc.The combined artificial data set and safety data filter form a comprehensive data set.The control barrier function is fitted with a neural network,and the Lagrange multiplier method is used to solve the related constrained optimization problem.A specific neural network is designed as an approximator.Secondly,in the unknown environment,data is obtained from the LiDAR sensor and converted to the world coordinate system.Samples are collected by distinguishing safe and unsafe areas according to the observation model.The control barrier function is fitted using the support vector machine,optimized with the kernel function and relaxation coefficient,and combined with the Sigmoid function to obtain the approximator.Then,a safety learning control framework is constructed.The system model is discretized and uncertainty parameters are added.A cost function containing the system state and control instruction sequence is defined,and three types of safety constraints(hard,probabilistic,and soft)are set.The corresponding algorithm model is established.Finally,experiments are carried out on the PyCharm platform to verify the effectiveness of the relevant algorithms in the known and unknown environments and the integrated framework in different obstacle environments.For example,in the known environment,the learned control barrier function can help the UAV avoid obstacles and reach the target.In the unknown environment,the observation-based learning algorithm can eliminate noise and achieve obstacle avoidance.The results show that the framework has good

关 键 词：安全避障 深度学习 控制障碍函数 专家轨迹 观测学习 仿真验证 安全约束 

分 类 号：V249.1[航空宇航科学与技术—飞行器设计]