Application of autoencoders on multivariate anomaly detection in building automation systems with variable selection based on semantic metadata of the facility

Description

We propose a method for training multiple individual anomaly detectors using the metadata of the facility. Each anomaly detector is specified by querying the semantic representation of the facility and the accompanying IoT network using system, subsystem, connectedness or spatial criteria. Additional criteria can be used depending on the richness of the underlying representation and may include geometrical (core or perimeter zones, horizontal or vertical adjacency, space size), control strategy and similar information. List of used data points is further expanded with weather data such as temperature, humidity and wind speed, as well as synthetic data points obtained by lagging existing data points, measuring intervals between readings or calculating aggregations or derivations. Individual autoencoder models are trained on data belonging to different systems and of differing spatial characteristics. We show that described method scales well to facility-wide generation of anomaly detection models. Quality and normality of input data remains a requirement for training anomaly detection models based on unsupervised algorithms. Validation possibilities with existing facility data, and data sanitation techniques required to avoid saturation of output with false positives are evaluated. Approaches at incorporating operator feedback on detected anomalies are also discussed.

Period	2 Jul 2024
Event title	ECOS 2024: 7th International Conference on Efficiency, Cost, Optimization, Simulation, and Environmental Impact of Energy Systems
Event type	Conference