基于神经网络和强化学习的车辆装配系统中的多载量小车实时调度方法

This project aims to study and solve a kind of real-time scheduling problems of multiple-load carriers of advanced vehicle assembly systems.With characteristics of the complexity, dynamicity, and combinatorial optimizations etc., a dynamic scheduling problem domain of material handling systems of multiple-load carriers is established with the system coupling. In the expanding and improving of alternative graphs, the constraint conditions and a formal description method are researched for a scheduling problem of multiple resource constraints of multi-load carriers. Analyzing multiple-objective functions and constraint sets, the multi-objective scheduling model is established with multiple constraints. An optimization selection method of real-time scheduling rules is studied based on artificial neural network and knowledge bases. The scheduling algorithms of four interrelated decisions sub problems (material handling tasks generated decision problem, dispatching decision making problems, handling task selection decisions and ordering decision problem) are designed. To solve the optimization goal of the scheduling problem for a long time, the scenario reinforcement learning algorithm based on relaxed semi-Markov average reward technique（R - SMART） is constructed. A prototype system of multiple-load carriers is developed and verified for the applications. Therefore, through this research project, the scheduling and modeling theories can be expanded, and design methods of scheduling decision heuristic algorithms can be enriched. The research results of this project have important theoretical significances and practical values for being used to improve production efficiency of the automobile manufacturing enterprises and production management levels.

本项目旨在研究并解决一类先进车辆装配系统中的多载量小车动态调度问题。针对具有多系统耦合的多载量小车搬运系统的复杂性、动态性、组合优化等特点，建立动态调度问题域。在拓展和完善替代图的基础上，研究具有多资源约束的多载量小车调度问题的约束条件及形式化描述方法。分析多目标函数及约束集，建立带多目标、多约束的调度模型。研究基于人工神经网络和知识库的多载量小车实时调度规则组合优化选取方法。设计调度问题中相互关联的四个决策子问题（物料搬运任务生成决策问题、派遣决策问题、搬运任务选择决策以及排序决策问题）的调度决策算法。构造基于R-SMART的场景式强化学习算法，解决利于长期优化目标的调度决策问题。开发多载量小车调度原型系统，并进行应用验证。因此，通过本项目的研究可拓展多资源的调度与建模理论，丰富动态调度决策的启发式算法设计方法，对提高汽车制造企业的生产率以及生产管理水平有重要的理论意义和实际应用价值。

本项目对先进车辆装配系统中的多载量小车静态和动态调度问题展开了研究。首先，针对具有多系统耦合的多载量小车搬运系统的复杂性、动态性、组合优化等特点，建立了静态和动态调度问题域。从分析线边库存、成套配送、JIT配送以及线边超市等配送模式出发，在拓展和完善替代图的基础上，研究具有多资源约束的多载量小车调度问题的约束条件及形式化描述方法。分析多目标函数及约束集，建立了多种配送模式下带多目标、多约束的调度模型。研究基于准启发式、人工神经网络以及知识库的调度问题中相互关联的四个决策子问题（物料搬运任务生成决策问题、派遣决策问题、搬运任务选择决策以及排序决策问题）的调度算法。构造了强化学习算法，解决利于长期优化目标的调度决策问题。最后，进行了多载量小车调度模型以及算法的实验分析与评价，并进行应用仿真验证。因此，通过本项目的研究拓展了多资源的调度与建模理论，丰富动态调度决策的启发式算法设计方法，对提高汽车制造企业的生产率以及生产管理水平有重要的理论意义和实际应用价值。