【作者】 李旭；

【导师】 尚家兴；

【作者基本信息】 重庆大学 ， 计算机科学与技术， 2021， 硕士

【摘要】 飞行安全一直是民航业关注的重点,其中着陆阶段是整个飞行过程中不安全事件最频发的阶段。重着陆作为一种着陆阶段的典型超限事件,尤其受到航空公司的重视。传统的重着陆相关研究大多依赖事故调查或专家人工分析,效率较低且缺乏客观的数据支撑。近年来,随着飞行参数记录仪(QAR)的广泛使用,产生了海量的飞行数据,全面记录了飞行过程中的实时动态信息。因此,如何借助先进的大数据、人工智能技术,基于QAR数据开展重着陆超限事件研究显得尤为必要。目前,国内外已有不少基于QAR数据的飞行安全相关研究。然而,现有研究存在以下三方面的问题:首先,大多数研究依赖于专家经验,无法提取出深层次的QAR数据特征;其次,现有研究的可解释性不强,简单的统计分析难以解释重着陆事件的成因;最后,大多数研究在进行重着陆预测时,其准确率普遍不高。针对上述问题,本文以可解释性为核心,从以下两方面开展研究:针对QAR数据特征提取和可解释性问题,本文提出了基于曲线聚类的重着陆成因自动识别方法。本文首先将QAR数据参数曲线化,根据专家经验将曲线特征与重着陆成因联系起来建立二级分类目录树。为了提高重着陆成因的识别率,本文通过插值和重采样提取出重着陆的曲线特征/模式;接着,融合专家经验,将聚类算法改进为半监督学习方法并应用在曲线特征上以自动识别重着陆模式和成因。最后,在少量重着陆航段样本上通过重着陆风险评估算法发现高风险航段。针对可解释性和预测准确率问题,依据QAR数据中的时序特性,将重着陆预测问题抽象成时间序列分类问题,并提出基于多分支时间卷积网络的重着陆可解释模型(IMTCN)。具体来说,为了充分学习时间序列中的特征以提高预测准确率,本文充分利用时间卷积网络(TCN)学习每个参数时间层面的特征。与此同时,借鉴类激活映射(CAM)可视化特征重要性的思想为重着陆事件提供较强的可解释性。本文工作的一大特色是在研究中加强了与飞行专家的合作与沟通,将领域知识深度融合到重着陆分析中。在37 943个A320航段样本上的实验结果显示曲线聚类的方法和IMTCN模型均能较好地识别和预测重着陆,其准确率分别达到了92.99%和95.20%,并且为重着陆事件提供较强的可解释性。与此同时,重着陆风险评估算法只需要少量的重着陆样本即可在海量的正常航段样本中发现高风险航段,对飞行安全预警意义显著。本文的工作为飞行安全研究提供了新的技术借鉴。更多还原

【Abstract】 Flight safety is the hot topic to civil aviation all the time,and landings are considered to be the most frequent phases of unsafe incidents during the flight.As a classic exceedance of the landing phase,hard landing incident is particularly valued by airlines.Traditional studies on hard landing mostly rely on accident investigation or expert manual analysis,which is inefficient and subjective without data support.Recently,as Quick Access Recorder(QAR)has been widely used,immense flight data is generated,which fully records the real-time dynamic information during the flight.Therefore,it is particularly necessary to study how to utilize advanced technology of big data and artificial intelligence to carry out research on hard landing exceedances based on QAR data.Currently,experts and scholars have studied flight safety issues based on QAR data.However,existing studies face the following three challenges: first of all,most studies depend on expert experience and fail to extract deep and hidden QAR data features;secondly,weak interpretability is the challenge faced by current studies,and it is difficult to explain the causes of hard landing incidents using simple statistical analysis;finally,most studies aim at predicting hard landing incidents with low prediction accuracy.To address the above issues,the thesis focuses on interpretability and conducts research from the following two aspects:In term of the problem of QAR data feature extraction and interpretability,this thesis proposes an automated recognition method for the causes of hard landing based on curve clustering.Specifically,we first extract QAR data parameters’ curve characteristics.According to expert experience,we link the curve characteristics with the causes of hard landing,establishing a two-level hierarchical classification of hard landing incidents.To improve the recognition rate of the causes of hard landing incidents,we extract hard landing curve-level features/patterns through interpolation and resampling.Subsequently,we turn the clustering into a semi-supervised algorithm by incorporating some expert experience and apply it on the curve-level features to automatically recognize the hard landing patterns and causes.Finally,we propose a risk evaluation algorithm on a handful of hard landing samples to discover high-risk flights from normal ones.In term of the problem of interpretability and prediction accuracy,based on the time series characteristics in QAR data,the thesis abstracts the hard landing prediction problem into time series classification problem,and provides Interpretable Multi-Temporal Convolutional Networks(IMTCN)to explain hard landing incidents.Specifically,to fully extract time series data features,we make full use of Temporal Convolutional Networks(TCN)to learn temporal features of each parameters.Meanwhile,based on the idea of Class Activation Mapping(CAM),visualizing feature importance provides strong interpretability for hard landing incidents.One of the contributions in this thesis is that we have strengthened the communication with flight experts,and deeply integrated the professional knowledge into the hard landing analysis.Experimental results on a dataset with 37 943 A320 aircraft flights show that both the curve clustering method and IMTCN model can exhibits good performance in recognizing and predicting hard landing incidents,whose accuracy reaches up to 92.99% and 95.20% respectively,and provides strong interpretability for hard landing incidents.Moreover,it only requires a handful of hard landing samples to discover high-risk flights from tremendous normal landing flights,which is critical for flight safety warnings.The work above provides a novel technical reference for flight safety.更多还原