Anomaly detection in low-cost sensor data using machine learning

Abstract

In December 2021, a system was deployed at the weather station in Makerere University to evaluate low-cost sensors that are used for environmental monitoring. The system has four kinds that is, the SHT sensor, BME sensor, HTU sensor and the HDC sensor. All the sensors mentioned are low-cost sensors which are monitoring the temperature and humidity of the environment. In environmental monitoring a reliable sensor is very important since the system will constantly obtain readings from the environment. In our project, we aim to carryout anomaly detection in low-cost environmental monitoring sensor data using machine learning. This implies we will be using the data collected by the sensors on the environment monitoring system. The data will be analyzed in order to identify if the sensors exhibit anomalous behavior while monitoring the environment. The model we developed is an LSTM Autoencoder which basically is an Autoencoder based on the LSTM encoder-- bit decoder architecture. The autoencoder can be used to train a model without relying on label s which makes it suitable for this kind of unsupervised training approach. In comparison with the KNN means and isolation forest, the LSTM Autoencoder was the most suitable to use because of its ability to minimize the mean absolute error (MAE) and detect anomalies in the sensor data. Finally, we have designed a dashboard that’s accessible through a web browser in order to visualize anomalous points in the sensors' data (temperature and relative humidity).