基于多粒子群零样本神经架构搜索的道路病害预测  

Road Disease Prediction Based on Multi-particle Swarm Zero Sample Neural Architecture Search

作　　者：章一颖 郑杰 贺春林[1] 谭睿 徐黎明[1] 李林波 何利蓉 刘昊 侯晓宁 ZHANG Yiying;ZHENG Jie;HE Chunlin;TAN Rui;XU Liming;LI Linbo;HE Lirong;LIU Hao;HOU Xiaoning(School of Computer Science,China West Normal University,Nanchong,Sichuan 637002;Department of Architectural Engineering,North China Institute of Aerospace Engineering,Langfang,Hebei 065000;China Merchants Roadway Information Technology(Chongqing)Co.,Ltd.,Chongqing,400067;Chongqing municipal facilities operation support center,Chongqing,400015;Chongqing Wukang Technology Co.,Ltd.,Chongqing,400067)

机构地区：[1]西华师范大学计算机学院,四川南充637002 [2]北华航天工业学院建筑工程学院,河北廊坊065000 [3]招商局公路信息技术(重庆)有限公司,重庆400067 [4]重庆市市政设施运行保障中心,重庆400015 [5]重庆物康科技有限公司,重庆400067

出　　处：《公路交通技术》2024年第6期21-29,共9页Technology of Highway and Transport

基　　金：重庆市自然科学基金面上项目(CSTB2022NSCQ-MSX1405);国家自然科学基金项目(62206224);中国博士后面上项目(2023M732428);四川省自然科学基金项目(2022NSFSC0866)。

摘　　要：为解决道路病害预测中手工设计网络面临的效率低,准确性差以及易陷入局部最优的问题,提出基于多粒子群零样本神经架构搜索方法,自动探索用于道路病害预测的最佳神经架构。先利用多粒子群策略在尺度自适应搜索空间初始化高质量架构;后采用粒子群动态自适应更新架构,防止陷入局部最优;再结合零样本学习、参数和浮点运算进行多目标优化,实现轻量化并提高预测精度。结果表明:1)尺度自适应搜索空间能有效捕捉多尺度道路病害信息;2)粒子群动态自适应更新避免了搜索过程陷入局部最优;3)多目标优化使得算法在分类准确率、F1分数、卡帕系数、AUC、指数平衡和搜索效率方面分别提升19.34%、23.37%、23.77%、4.28%、20.26%和91.30%。To solve the problems of low efficiency,poor accuracy and easy to fall into local optimal in road disease prediction,a multi-particle swarm zero sample neural architecture search method is proposed to automatically explore the optimal neural architecture for road disease prediction.Firstly,the multi-particle swarm strategy is used to initialize the high-quality architecture in the scale-adaptive search space.Then particle swarm dynamic adaptive update architecture is used to prevent local optimization.Finally,zerosample learning,parameter and floating-point operation are combined for multi-objective optimization to achieve lightweight and improve prediction accuracy.The results show that:1)Scale adaptive search space effectively captures multi-scale road disease information;2)PSO dynamic adaptive updating effectively prevents the search process from falling into local optimization;3)Multi-objective optimization improves the classification accuracy,F1 score,Kappa coefficient,AUC,exponential balance and search efficiency by 19.34%,23.37%,23.77%,4.28%,20.26%and 91.30%,respectively.

关 键 词：道路病害预测 粒子群 尺度自适应 多目标优化 神经架构搜索 

分 类 号：U418.5[交通运输工程—道路与铁道工程]