基于高斯混合模型的飞机进近着陆阶段运行异常检测

Abnormal operation detection for approach and landing phase based on Gaussian mixture model

为检测与分析飞机进近着陆阶段可能引发安全风险的运行异常现象,基于高斯混合模型提出了一种进近着陆阶段飞机运行异常检测方法。首先,将进近着陆阶段飞行数据输入期望最大算法构建高斯混合模型,研究各高斯分量随时间的变化特性,结合所输入的数据在各高斯分量概率密度函数中的计算结果,识别进近着陆阶段存在的飞行数据异常航段。然后,通过复核数据异常航段的原始飞行数据,检测与分析飞机进近着陆阶段可能引发安全风险的运行异常现象。最后,利用该方法检测与分析了462个实际运行航段中进近着陆阶段的运行异常。结果表明:当文中模型对应的BIC数值最小时,高斯分量个数为21,其中有11个高斯分量的从属度在接地后迅速衰减至0;在划定3个不同的检测阈值后,分别识别出27、15、8个飞行数据异常航段,最终通过复核检测出3个进近着陆阶段的运行异常现象。

Flight safety plays an important role in the development of the aviation industry.A large amount of non-exceeding flight data contains important information about flight safety and operational risk management.But the value of non-exceeding flight data has not received enough attention.To detect and analyze abnormal operations that may induce risks during aircrafts’approach and landing phase,this paper proposed an abnormal operation detection method for approach and landing phase based on the Gaussian mixture model.Firstly,it was assumed that flight data from real airlines obeyed the multi-dimensional Gaussian distribution.And by combining the Bayesian information criterion and K-means algorithm,the non-exceeding data of approach and landing phase was submitted to the expectation-maximization algorithm to construct the Gaussian mixture model.Then the temporal distribution of each Gaussian component was studied to characterize each Gaussian component’s appropriateness with time.And the calculation results of the probability density functions of Gaussian components were obtained.After that,a function was defined to judge whether a flight’s data was normal or not at a certain moment.And by comparing the function value of every second with the preset detection threshold,it was able to identify flights with abnormal flight data during the approach and landing phase.Finally,experts in the field of flight safety would review the original flight data of abnormal flights to detect and analyze abnormal operations.This method was used to detect and analyze the risks of 462 commercial flights that came from a specific airline.By comparing the decision value of each flight with three different detection thresholds,27,15,and 8 flights with abnormal data were identified respectively.After that,3 experts in the field of flight safety were invited to review 11 flights with abnormal flight data.Finally,3 kinds of abnormal operations of the approach and landing phase were detected.