A Tree-based Machine Learning Model for Go-around Detection and Prediction

The approach phase of a fight represents a safety-critical fight operation that requires close and timely coopera­tion between pilots and tower controllers leading to a smooth landing operation. Go-around or missed-approach procedures are in place to discontinue an unsafe landing. These procedures may generate further safety concerns due to their complex manoeuvre and time constraints. Due to the availability of high-fidelity air traffic data, such as ADS-B, new data-driven metrics can be derived in order to enhance the situational awareness of tower controllers and thus, increase the safety level of landing operations. This paper proposes a novel safety metric based on machine learning techniques that may assist tower controllers in detecting and predicting go-around events. First, a data-driven model is developed for labeling go-around events. Then, features are engineered for a tree-based learning model to predict go-around events. The model is trained, validated, and tested using ten months’ of ADS-B data for fights arriving at Philadelphia International Airport (PHL), comprising 132,118 fights with 662 go-around events. Results demonstrate that the best prediction results are found at 2 NM away from the runway threshold. For the down-sampling data, the model is able to predict 56% of the go-around with only a 10% false-positive alert rate, while for the full data set the model is able to detect 33% of the go-around with a 25% false alert rate. The proposed model outperforms state-of-the-art methods in terms of decreasing the false-positive alerts in the system by 60%. The proposed model achieves a generic solution that is not specific for a single runway, and therefore can be deployed at other runways without a need for extensive model training.