Deep Learning Based Anomaly Detection Approach for Air Pollution Assessment

Environmental air pollution has become a cause of global concern due to its adverse effects. Unusually high concentration of air pollutants can be regarded as an anomaly indicating certain air quality problems. This paper presents a deep learning based anomaly detection approach to identify anomalous concentrations of five different air pollutants: Carbon Monoxide (CO), Ozone (O-3), Nitogen Oxide (NOX) and Particulate Matters (PM2.5, PM10) in a real-life environmental dataset. The collected data is multivariate in nature containing hourly generated information about several air pollutants and atmospheric parameters from a non-polluted city of India. The proposed framework contains a Bidirectional Long Short Term Memory (Bi-LSTM) based predictor model with self-attention to capture the normal pollutant levels in the time series dataset. The predictor model is responsible for predicting the value at the next timestamp, corresponding to a given window of the time series data. A subsequent anomaly detector is utilized to identify the anomalous pollutant levels based on the predictions of predictor model. Anomalies detected by the proposed framework are utilized to analyze the correlation of temporal and atmospheric parameters with the anomalous concentration levels. Experimental results illustrate the predominance of proposed approach over existing approaches towards air pollution assessment.