Small Uncrewed Aerial Systems (sUAS), commonly referred to as drones, have become ubiquitous across various domains and applications. Examples range from drones taking part in search-and-rescue operations to drones being used for delivering medical supplies or packages. As sUAS and their applications exhibit safety-critical behavior, ensuring their safe operation within operational boundaries has become a top priority. Thus, continuous and rigorous monitoring of sUAS at runtime, during flight operations, is essential.

However, sUAS generate vast amounts of data, for example, multi-variate time series which need to be analyzed in real-time to detect potential emerging issues. This poses a significant challenge, due to resource constraints imposed on the onboard computation capabilities of sUAS. To alleviate this problem, we introduce ADAM, a novel self-adaptation anomaly detection framework for sUAS that facilitates adaptive monitoring. ADAM selectively monitors a small subset of the time-series data streams, which serve as indicators of anomalous behavior. In the event of a raised alert, ADAM adjusts its monitoring strategy by enabling additional detectors and taking further actions to mitigate the issue.

We evaluated the effectiveness of ADAM through simulations in Gazebo, analysis of real flight logs taken from sUAS forums, and tests performed with real-world drones. The results demonstrate ADAM's capability to enhance the safety and efficiency of sUAS operations by dynamically managing anomaly detection processes and optimizing computational resources.