Anomaly Detection for Residential HVAC System Using LSTM Autoencoder

Abstract

Heating, Ventilation, and Air Conditioning (HVAC) systems are extensively utilized in commercial and residential buildings. Specifically, in the case of residential HVAC systems, they not only inconvenience users in the event of malfunctions but also contribute to increased power consumption, energy usage, and carbon emissions. Ongoing research focuses on the Prognostics and Health Management (PHM) of residential HVAC systems, aiming to simplify repair procedures and assist service engineers in decision-making. Obtaining post-installation failure data poses a significant challenge in residential HVAC system. Hence, this study proposes a data-driven anomaly detection method utilizing a reconstruction autoencoder. The method involves training long short-term memory (LSTM) autoencoder model with data collected during normal HVAC system operation. The model is trained to reconstruct normal data patterns. Subsequently, the reconstruction error, which quantifies the dissimilarity between the input data and the reconstructed data, is transformed into an anomaly score based on the Mahalanobis distance. The effectiveness of the proposed method is verified by detecting anomalies in vacuum operation resulting from valve blockage. Furthermore, a comparative analysis is conducted to evaluate outlier detection performance, differences in reconstruction error distribution measured by Kullback-Leibler divergence. The findings demonstrate the applicability of autoencoder-based outlier detection for managing the operational reliability of residential HVAC systems, taking into account their characteristics.