Fusing Long Short-Term Memory and Autoencoder Models for Robust Anomaly Detection in Indoor Air Quality Time-Series Data

Abstract

People spend most of their time indoors by choice or by need. Carbon dioxide (CO2) accumulation can cause various adverse health effects, including vertigo, headache, and fatigue. Therefore, monitoring indoor air quality(IAQ) is necessary for various health reasons. The market is flooded with air quality monitoring devices. However, the ordinary public does not make use of them because they are expensive and difficult to obtain. Several research studies have been carried out to monitor indoor air quality with the help of the Internet of Things(IoT), which has greatly simplified the method for monitoring IAQ. In this research, we offer an improved IoT based IAQ monitoring system with AI-powered recommendations. Our suggested system relies on the Message Queuing Telemetry Transport(MQTT) protocol for communication between IoT devices. In addition, the gathered CO2 occupancy data is used together with the deep learning approach of Long Short-Term Memory and Autoencoder (LSTM-AE) to detect anomalies or outliers in CO2 concentrations. Due to a close connection between air quality and human health and well-being, the detection of anomalies in the data of IAQ has emerged as an essential topic of study. Anomalies requiring the observation of correlations spanning numerous data points (i.e., often referred to as long-term dependencies) were not detectable by conventional statistical and basic machine learning (ML) related techniques in the sector of IAQ. Hence this research uses the LSTM-AE model to address this issue. In comparison to previous similar models, our experimental results on a generated CO2 occupancy time series reveal a robust and powerful accuracy of 99.49%.