Causal Learning for Fault and Anomaly Detection in Unmanned Aerial Systems

Abstract

Unmanned Aerial Systems (UAS) are vital to tactical autonomy applications, where they operate in contested, communication denied, and high stakes environments with limited human oversight. In such contexts, robust fault detection is not merely a safety concern but a mission critical capability where engine failures or control surface malfunctions can result in asset loss, mission failure, or compromised operational security. Traditional anomaly detection methods rely on correlation-based approaches that identify statistical deviations but fail to reveal the underlying causal mechanisms driving failures. This opacity is particularly problematic for tactical autonomy, where operators and commanders require explainable diagnostics to maintain situational awareness and trust in autonomous systems. We argue that causal learning provides a more principled and interpretable framework for understanding UAS faults. By utilizing causal learning to identify the cause-and-effect relations from telemetry data we can answer not only whether a fault occurred, but what caused it, enabling root cause analysis essential for rapid field assessment and informed decision making. We present preliminary proof-of-concept results using the ALFA (AirLab Failure and Anomaly) dataset to learn causal structures from UAS telemetry data. Our causal graphs reveal directional relationships between sensor readings, control inputs, and failure states, offering actionable insights for predictive maintenance, real time fault diagnosis, and post mission analysis that correlation-based methods cannot provide. We position causal reasoning as a foundational capability for trustworthy tactical autonomy.