Similarity Augmented Incremental Learning-Based Flight Data Anomaly Detection Method for UAV Dynamic Conditions

Abstract

In response to the high accident rate associated with large-scale unmanned aerial vehicle (UAV) deployment, the prediction-based anomaly detection method has emerged as an important research field. This method uses airborne sensor data to predict the expected operational status of UAVs to determine whether their actual status deviates from normal. When the anomaly detection model is exposed to diverse flight conditions and missions, continuously changing sensor data characteristics may, however, lead to model mismatch, resulting in reduced detection accuracy (ACC). This article, therefore, proposes a similarity augmented incremental learning-based flight data anomaly detection method for UAV dynamic conditions. The core goal of this method is to maintain the ACC of anomaly detection in new conditions while preventing issues related to model mismatch. First, an long short-term memory (LSTM)-based anomaly detection model and the feature buffer are established using historical sensor data. The feature buffer is used to update the detection model when new operating conditions arise. Second, a Jensen-Shannon divergence (JS divergence)-based similarity measurement is proposed to monitor changes in the distribution of sensor data and identify whether new operating conditions that differ significantly from historical data occur. Finally, a similarity augmented replay-based incremental learning mechanism is designed for model updates when new operating conditions are detected. During the model update process, similarity measurements are employed to augment the replay-based update strategy, thereby improving the adaptability of the model. Experiments with simulated and real airborne sensor data demonstrate that the proposed method improves anomaly detection ACC, increasing ACC by 0.6%–1.8% and reducing false positive rate (FPR) by 16.5%–95.2%.