Multivariate Air Pollution Anomaly Detection via LSTM Autoencoders on Beijing Multi-Site Sensor Data

Abstract

The current paper investigates the application of a Long Short-Term Memory Autoencoder (LSTM-AE) to identifying anomalies in the multi-variate time-series of air-pollution measurements. The algorithm was used to the sensor data of PM2.5, CO, O3, NO2, TEMP, and WSPM of the Beijing Multi-Site Air-Quality Data Set described on Kaggle. The test set included fifty synthetic anomalies that were used to assess performance. The anomalies were identified using the reconstruction error calculated using Mean Squared Error (MSE) as a dynamic threshold value of 0.002957. The presented model produced the Precision, Recall, F1-Score, and ROC-AUC of 77.00%, 100.00%, 87.01%, and 99.86%, respectively, proving its effectiveness in detecting minor and drastic changes in the pattern of air quality.