Deep Smooth Random Sampling and Association Attention for Air Quality Anomaly Detection

Abstract

Real-time monitoring and timely warning of air quality are vital components of building livable cities and implementing the “Healthy China” strategy. Real-time, efficient, and accurate detection of air quality anomalies holds great significance. However, almost all existing methods for air quality anomaly detection often overlook the imbalanced distribution of data. In addition, many traditional methods cannot learn both pointwise representation and pairwise association, so they cannot solve complex features. This study proposes an anomaly detection method for air quality monitoring based on Deep Smooth Random Sampling and Association Attention in Transformer (DSRS-AAT). Firstly, based on the third geographical law, the more similar the geographical environment, the closer the geographical target features are. We cluster sites according to the surrounding geographic features to fully explore latent feature associations. Then, we employ Deep Smooth Random Sampling to rebalance the air quality datasets. Meanwhile, the Transformer with association attention considers both prior associations and series associations to distinguish anomaly patterns. Experiments are carried out with real data from 95 monitoring stations in Haikou City, China. Final results demonstrate that the proposed DSRS-AAT improves the effectiveness of anomaly detection and provides interpretability analysis for traceability, owing to a significant improvement with the baselines (OmniAnomaly, THOC, etc.). The proposed method effectively enhances the effectiveness of air quality anomaly detection and provides a reference value for real-time monitoring and early warning of urban air quality.