Using Risk Analysis for Anomaly Detection for Enhanced Reliability of Unmanned Autonomous Systems

Unmanned Autonomous Systems (UAS) are intended to improve the safety of offshore operations by residing on the seabed and monitoring and inspecting assets and the environment. The UAS can collect and analyze data in real-time through sensor measurements and video analysis, warning onshore operators of data anomalies that indicate potentially hazardous environmental events. Due to the rarity of hazardous events, data on them are scarce, resulting in a data imbalance between normal and anomalous occurrences. Consequently, it has become increasingly challenging for UAS to recognize potentially hazardous circumstances. Thus, the UAS may overlook early warning signs of hazardous events or may overwhelm operators with trivial, resource-intensive information in the form of false alarms. Recent research has addressed data imbalances by simulating underrepresented data, extrapolating it using causal knowledge, or adding parameters to data and methods as a form of semi-supervision. However, in this research, we examine risk analysis as a tool for providing a semi-supervised approach to anomaly detection. We emphasize the overlapping properties of risk analysis and anomaly detection within the objectives of a highly autonomous system. Finally, we apply the derived insights to anomaly detection in sensor data to lower the likelihood of false alarms or missed signals.