Data-driven Flight Status Anomaly Detection And Analysis

The aviation industry is currently going through an important stage from manual risk detection to big data analysis of anomalies.Correct and efficient detection of anomalous events,from which factors influencing overrun events can be found,is of dual value in ensuring aviation safety and improving user experience.However,the overrun detection methods currently applied in the aviation industry need to address the challenges of aircraft type diversity,environmental diversity and climate diversity,making it difficult to adjust thresholds to detect anomalies based on each situation.Therefore,it is of great significance to Accurately detect abnormal events,analyze the causes of abnormal events,and propose corresponding solutions to improve aviation safety.In this paper,by introducing the common anomalous instantaneous transient anomalies and flight-level anomalies in aviation data,we propose two algorithms to efficiently detect anomalies respectively,and finally achieve the unification of risk detection and risk interpretation.The details of the thesis are as follows.(1)By studying the development of aviation data analysis at home and abroad,the historical evolution and data format of Quick Access Recorder(QAR)are introduced.The following,the overall idea of combining and matching aviation data with anomalous parameter libraries to analyze anomalies is proposed.(2)For the characteristics of aviation transient anomalies,a scheme of running Long Short-Term Memory(LSTM)trained by historical data and dynamic threshold detection is proposed.The LSTM is first used to fit the original data,and then the dynamic threshold is used to compare the difference between the predicted data and the original data,with the points larger than the confidence interval considered as anomalies.It can not only effectively detect anomalous events,but also visualize the difference between predicted and true values,which can provide a more intuitive understanding for practitioners.(3)For aviation flight-level anomalies,this paper takes the most frequent long landing event in the approach landing phase as the research object.We firstly construct a landing anomaly dataset by quadrature method,then a hybrid feature extraction algorithm is used,which make advantage of Spearman’s Rank Correlation Coefficient(SRCC)to remove redundant parameters and Gradient Boosting Decision Tree-Recursive Feature Elimination-Cross-Validation(GBDT-REF-CV)to extract key features in turn.Next,the Condition Generative Adverse Network(CGAN)is used to generate new abnormal data to adjust the imbalance of data from the data level,followed by an improved e Xtreme Gradient Boosting Tree(XGBoost)algorithm to improve the anomaly detection performance of the model.the experimental results show that the proposed model can effectively detect long landing events in the test dataset.(4)After the long landing event is detected by the model,the factors affecting the occurrence of long landing events were analyzed using SHapley Additive explanations(SHAP),a machine learning interpretable framework.In-depth analysis of the model showed that altitude is the most important feature of land safety,especially when the altitude less than 200 feet.The angle of the slide at less than 3 degrees below 100 feet is prone to long-landing events,which coincides with the aircraft dynamic equation.