Spatially-Aware Autoencoders for Detecting Contextual Anomalies in Geo-Distributed Data

The huge amount of data generated by sensor networks enables many potential analyses. However, one important limiting factor for the analyses of sensor data is the possible presence of anomalies, which may affect the validity of any conclusion we could draw. This aspect motivates the adoption of a preliminary anomaly detection method. Existing methods usually do not consider the spatial nature of data generated by sensor networks. Properly modeling the spatial nature of the data, by explicitly considering spatial autocorrelation phenomena, has the potential to highlight the degree of agreement or disagreement of multiple sensor measurements located in different geographical positions. The intuition is that one could improve anomaly detection performance by considering the spatial context. In this paper, we propose a spatially-aware anomaly detection method based on a stacked auto-encoder architecture. Specifically, the proposed architecture includes a specific encoding stage that models the spatial autocorrelation in data observed at different locations. Finally, a distance-based approach leverages the embedding features returned by the auto-encoder to identify possible anomalies. Our experimental evaluation on real-world geo-distributed data collected from renewable energy plants shows the effectiveness of the proposed method, also when compared to state-of-the-art anomaly detection methods.