Robust and Explainable Semi-Supervised Deep Learning Model for Anomaly Detection in Aviation

Abstract

Identifying safety anomalies and vulnerabilities in the aviation domain is a very expensive and time-consuming task. Currently, it is accomplished via manual forensic reviews by subject matter experts (SMEs). However, with the increase in the amount of data produced in airspace operations, relying on such manual reviews is impractical. Automated approaches, such as exceedance detection, have been deployed to flag safety events which surpass a pre-defined safety threshold. These approaches, however, completely rely on domain knowledge and outcome of the SMEs’ reviews and can only identify purely threshold crossings safety vulnerabilities. Unsupervised and supervised machine learning approaches have been developed in the past to automate the process of anomaly detection and vulnerability discovery in the aviation data, with availability of the labeled data being their differentiator. Purely unsupervised approaches can be prone to high false alarm rates, while a completely supervised approach might not reach optimal performance and generalize well when the size of labeled data is small. This is one of the fundamental challenges in the aviation domain, where the process of obtaining safety labels for the data requires significant time and effort from SMEs and cannot be crowd-sourced to citizen scientists. As a result, the size of properly labeled and reviewed data is often very small in aviation safety and supervised approaches fall short of the optimum performance with such data. In this paper, we develop a Robust and Explainable Semi-supervised deep learning model for Anomaly Detection (RESAD) in aviation data. This approach takes advantage of both majority unlabeled and minority labeled data sets. We develop a case study of multi-class anomaly detection in the approach to landing of commercial aircraft in order to benchmark RESAD’s performance to baseline methods. Furthermore, we develop an optimization scheme where the model is optimized to not only reach maximum accuracy, but also a desired interpretability and robustness to adversarial perturbations.