模拟机飞行训练质量智能评估技术研究

The flight training on simulator has been an important means for pilots to improve flying skills, while automatic assessment of quality of training has become a research focus in flight training industry around the world. But due to many factors affecting the flight, its standard has been different and variable which makes the automatic assessment become a challenging research topic. Computational intelligence has the characters of self-learning, self-organizing and applicability etc., the project would model and learn the instructors’ experiences based on its theories and methods. In this project, the flight was expressed as a multivariate time series while the standard flight maneuvers were expressed as multiple time series fragments, thus the assessment process was divided into two stages: flight identification and automatic assessment. After that, the flight identification would be solved as a multivariate time series by using PSO and bee colony algorithm; the automatic assessment would be solved as a complex classification problem by using multi-modal neural network ensembles. Through this project, a mature and reliable automatic assessment technology for flight training on simulator was expected. Instead of manual assessment, this technology could avoid making large differences and different flight standards, improving the level of modernization of flight training.

模拟机飞行训练已经成为飞行员提高飞行技能的重要手段，对其训练质量进行自动评估成为全世界飞行训练行业的研究热点，但由于飞行动作影响因素较多，飞行标准本身即存在差异性、变异性等问题，使得飞行训练质量自动评估成为极具挑战性的研究课题。本项目以计算智能的理论和方法为基础，利用其自学习、自组织、自适应的特性，对飞行教员的评分经验进行建模与学习。将飞行过程表达为多元时间序列，各个飞行科目的标准飞行动作抽象为多元时间序列片段，评估过程分为飞行科目识别和科目质量自动评估两个阶段。科目识别问题转化为多元时间序列比对问题，采用基于群体智能的粒子群算法和蜂群优化算法进行求解；飞行科目的训练质量评估问题转化为复杂分类问题，采用多模态神经网络集成的方法对进行学习建模。期望形成成熟可靠的模拟机飞行训练自动评估技术，避免手工评分主观评估性过大而导致的评分出现较大差异、放飞标准不统一的弊端，提高飞行训练的现代化水平。

模拟机训练是目前飞行训练的重要手段，但由于技术封锁等诸多因素，现有模拟机的训练仅限于简单的人机交互，却难于对训练人员进行动作规范，对训练参数进行自动评估。针对这些问题，我们建立了模拟机训练质量评估系统，并研究实现了关键算法。针对模拟机训练参数，采用模式识别技术对关键参数进行提取。应用投影不变量命名的特征数的方法和共形粘合理论对训练人员规范动作进行监测。在科目识别过程中，基于传统蜂群算法的优缺点，我们提了多个基于蜂群算法的的优化算法，以提高科目识别的效率和准确度。对于自动评估机制，搭建评估模型，研究应用L1范数低秩矩阵和多视角非负矩阵分解进行聚类分析，并针对评估的合理性和准确性提出了基于iDBMM的优化算法。最后应用可视化技术对分析结果进行丰富的展示。