Abstract:
Anomaly detection involves identifying data points or patterns that significantly diverge from expected behavior within a system, often due to external disruptions, faults, or uncertainties during operation. In smart buildings, where data such as electricity and air conditioning usage is represented as time series data, the early identification of these anomalies is essential for maintaining efficiency, reducing costs, and ensuring safety. Unlike traditional models that consider features separately, AB-LOF combines attention mechanisms with LOF to capture both temporal dependencies and interactions among features. The attention maps reveal interactions among features and dependencies over time, which are subsequently transformed into fixed-size window vectors via an ANN.The vectors are processed using LOF to calculate anomaly scores, helping the model to effectively detect un usual patterns, even when outliers are present. In the context of actual building data, the attention maps generated offer a way to interpret feature correlations and temporal dynamics, facilitating a deeper understanding of anomalies and the as sociated energy patterns. Furthermore, visualizations of the attention maps and anomaly scores enable comparisons across different features, window, and time points, enhancing the interpretation of the results and aiding in the identification of significant anomalies.
