Anomaly Detection In Civil Aviation Safety Based On Multiple Kernel Learning And Cluster Analysis

The safety level of civil aviation has aroused more and more concern, as the proportion of civil aviation transportation in the whole transportation industry grows steadily. We need to take the safety monitoring measures more actively, in order to improve the existing safety level of aviation. Flight Operations Quality Assurance(FOQA) program is one of the important means to ensure flight safety which is well known internationally. It has provided a large number of basic data sets for the purpose of studying different phases of aircrafts’ flight, and detecting the incident of fatal crashes through collecting different data types of the flight by the digital flight data recorder or quick access recorder. The data collected could be used in the laboratory scale models’ construction, and also for the analysis of the related algorithms.In this paper, we use the kernel function and cluster analysis technique in machine learning as our basic model construction algorithms, and take the heterogeneous data collected by the FOQA program as the sample of input. We established a multiple kernel learning anomaly detection model, presenting the model structure and the algorithm innovatively. We build the discrete kernel and the continuous kernel respectively, both of which has been used in our model to deal with the large scale of the heterogeneous flight data systematically. We take the Longest Common Subsequences(LCS) as our kernel over discrete sequences, whereas we use symbolic representation method to discretize the continuous variables in our model for further processing. Meanwhile, feature vectors are extracted from the input DFT coefficients, and clustered in the real FOQA data set. After detecting anomalies in the sample input, we identified the accident statistically. Then we analysis the flight of abnormal samples using aviation domain prior knowledge in depth. We have verified the robustness and effectiveness our algorithm, by operating on both the simulated data sets and real flight data sets. The outcomes has shown that our algorithm is better than the classical algorithm on the time cost and correct detection rate.