基于代理遗传优化的智能驾驶系统加速测试方法
Accelerated Test Method of Intelligent Driving System Based on Surrogate Genetic Optimization Model
朱冰 1汤瑞 1赵健 1张培兴 1李文旭1
作者信息
- 1. 吉林大学 汽车仿真与控制国家重点实验室,吉林 长春 130022
- 折叠
摘要
提出了一种基于代理遗传优化的智能驾驶系统加速测试方法.首先,通过场景要素层次分析权值与优解区域特征改进参数采样模块中的拉丁超立方采样区间,实现了采样效率与优化效果的协同提升;其次,利用参数采样结果和重复度筛选机制增加遗传寻优模块的种群多样性,克服了传统遗传算法的局部收敛难题;然后,利用基于循环更新机制的代理筛选模块对场景测试结果进行预测,平衡了加速算法与代理模型应用之间的效率与精度矛盾;最后,搭建仿真平台在高维时序分解的前车变速场景下对待测智能驾驶系统进行加速测试与验证.结果表明,本文提出的方法可有效搜寻大量关键场景并提升测试效率.
Abstract
This paper proposed an accelerated test method for intelligent driving system based on the surrogate genetic optimization model.First,the Latin hypercube sampling interval in the parameter sampling module was improved by using the weights and optimal solution region features of the scenario element hierarchical analysis method,achieving a synergistic improvement in sampling efficiency and optimization effect.Next,by utilizing parameter sampling results and repeatability screening mechanism,the population diversity of the genetic optimization module was increased,overcoming the local convergence problem of traditional genetic algorithms.Then,the surrogate filtering module based on cyclic update mechanism was used to predict the test results of the scenario,which balanced the contradiction between the efficiency and accuracy of the accelerated algorithm and the application of the surrogate model.Finally,a simulation platform was built to accelerate test process and verification of the intelligent driving system to be tested in the front vehicle speed change scenario of high-dimensional time series decomposition.The results indicate that the method proposed in this paper can effectively search for a large number of key scenarios and improve testing efficiency.
关键词
汽车工程/智能驾驶系统加速测试/代理模型/遗传算法/拉丁超立方采样Key words
automotive engineering/intelligent driving system accelerated testing/surrogate model/genetic algorithm/Latin hypercube sampling引用本文复制引用
基金项目
国家重点研发计划(2022YFB2503402)
国家自然科学基金(U22A20247)
国家自然科学基金(52172386)
吉林省科技发展计划(20220201023GX)
出版年
2024
国家科技期刊平台
NSTL
万方数据