Sensor-Drift-Aware Time-Series Anomaly Detection for Climate Stations

Sensor data collected from climate stations has been used in various scientific applications and environmental monitoring. Maintaining the data quality is essential to guarantee the reliability and accuracy of science outputs, potentially impacting many critical decision making processes. Existing sensor anomaly detection techniques are mostly designed for general purposes, and may not be suitable for climate sensors which require complex handling of seasonality, spatial relationship and sensor interdependency. Current quality control process is deficient in climate sensor drift detection, which is a slow degradation of sensor accuracy over time. Recent development of anomaly detection in climate sensor domain is limited, it’s often constrained to particular sensor types, and not focused on drift detection. In this paper, we present a new drift-aware time series anomaly detection framework which leverages the spatial-temporal correlation of the climate sensor network and significantly improves climate sensor drift detection capability. Moreover, the proposed semi- supervised learning approach helps to generalise the solution for various types of sensors and anomalies. Our experiments using real-world dataset have demonstrated promising and competitive performance in regards to sensitivity, false alarm control, and computational efficiency suitable for real-time or near-real-time applications.