Early Detection and Reconstruction of Abnormal Data Using Hybrid VAE-LSTM Framework

Abstract

Early failure detection and abnormal data reconstruction in sensor data provided by building ventilation control systems are critical for public health. Early detection of abnormal data can help prevent failures in crucial components of ventilation systems, which can result in a variety of issues, from energy wastage to catastrophic outcomes. However, conventional fault detection models ignore valuable features of dynamic fluctuations in indoor air quality (IAQ) measurements and early warning signals of faulty sensor data. This study introduces a hybrid framework for early failure detection and abnormal data reconstruction applying variance analysis and variational autoencoders (VAE) coupled with the long short-term memory network (VAE-LSTM). The periodicity and stable fluctuation of IAQ data are exploited by variance analysis to detect unusual variations before failure occurs. The IAQ dataset which is corrupted by introducing complete failure, bias failure and precision degradation fault is then used to verify the feasibility of the VAE-LSTM model. The results of variance analysis reveal that unusual behavior of the data can be detected as early as 12 hours before failure occurs. The reconstruction performance of the developed method is shown to be superior to other methods under different abnormal data scenarios