飞机维修保障资源配置问题的离散变量仿真优化方法研究

For airplane, it is of vital importance to keep high availability by optimizing its maintenance support policy. To achieve satisfactory maintenance support policy, a new idea is proposed, which evaluates the maintenance support system's effectiveness by multi-agent simulation (MAS), and substitutes neural network prediction for simulation output in order to improve the low efficiency of simulation optimization for complex systems. We will focus on three problems: MAS model construction, neural network prediction, and optimization algorithm. By means of depicting the working parts, spare parts, and repairman as different agents, the simulation of the airplane maintenance system is built by MAS model. Then, we focus on the key problems of constructing neural network prediction model, including the selection and online appending of the samples data set generated by simulations, as well as the structure and parameters design of the neural networks. At last, the heuristic algorithm for optimization is designed by analyzing the fitness landscape of the input-output relationship obtained by neural network model. The expected results of the project will provide technical support for airplane maintenance policy making, as well as new ideas for the discrete optimization via simulation (DOvS) of complex systems, which is meaningful in theory and has broad application prospect.

飞机维修保障方案优化设计是提升飞机维修保障效益，确保飞机完好率的重要途径。针对实现满意的维修保障方案问题，提出了利用多主体仿真对维修保障系统进行评估，同时为解决复杂系统仿真优化效率低的问题提出利用神经网络代替仿真模型进行优化的思想。本项目拟围绕多主体仿真、神经网络预测、优化算法三个问题展开研究。通过将工作部件、库存备件、维修人员视为Agent，构建维修保障资源配置的多主体仿真模型；围绕仿真模型生成样本集的选择和在线增补，以及神经网络结构和参数选择等关键问题，构造神经网络预测模型；最后，通过适应值曲面分析选取合适的启发式算法实现优化。本项目的预期研究成果将为设计飞机的最优化维修保障策略提供技术支撑，并为复杂系统的离散变量仿真优化提供一种新的解决思路，是一项既有理论意义又有广阔应用前景的研究课题。

项目提出了利用神经网络代替仿真模型解决离散变量仿真优化问题的思路，重点探索了基于多主体仿真的飞机维修保障仿真建模、基于神经网络的仿真优化算法、离散变量仿真函数的优化算法。首先，系统分析了飞机多级维修保障的流程，在此基础上构建了基于多主体仿真的离散事件仿真模型，设计了各修理级别的备件、维修人员Agent的状态机模型，以评估给定维修备件、人员条件下的飞机战备完好率指标。其次，提出了一种基于神经网络的仿真优化算法，研究了神经网络的选择问题：为实现以尽可能少的仿真样本实现全局最优解，重点研究了仿真模型初始样本集的生成策略和增补样本集的选择策略问题；为避免因神经网络拟合曲面与真实曲面存在的误差，可能导致将局部极小误判为全局极小的问题，提出了首先在拟合曲面获取全部局部极小，然后逐步生成样本的策略，以实现神经网络在局部极小附近回归曲面的精确拟合。再次，提出了基于适应值曲面特征分析的优化算法，以实现神经网络拟合曲面的局部极值寻优。最后，针对离散变量的仿真函数优化问题，设计了基于边际优化的离散变量优化算法和模式搜索优化算法。项目的研究成果可为飞机维修保障资源优化配置问题提供技术支持，同时为复杂系统的离散变量仿真优化问题提供一种新的解决思路，具有较强的理论意义和实践价值。