Air quality analysis based on threshold values and anomaly detection using machine learning models

This research addresses critical challenges in local air quality monitoring, specifically focusing on threshold-based condition classification and anomaly detection within micro-batch sensor data. Air quality degradation is characterized by dynamic fluctuations influenced by particulate matter (PM2.5, PM10), carbon dioxide (CO₂), temperature, and relative humidity. Data acquisition involved SCD41, BME688, and SPS30 sensors, generating micro-batches of ten sequential observations, which were subsequently stored in a DuckDB database. A dataset comprising 33,190 measurements was processed into 3,319 micro-batches, each characterized by 60 statistical features. For threshold-based classification, logistic regression, decision tree, random forest, and a rule-based baseline were evaluated against predefined thresholds for PM2.5, PM10, CO₂, temperature, and humidity. Anomaly detection was

assessed using Isolation Forest, Local Outlier Factor (with and without PCA), and One-Class SVM, benchmarked against rule-based anomaly labels and micro-batches exhibiting extreme feature values. The decision tree model effectively replicated the labeling logic, while the random forest achieved F1 scores of 0.857 under time-series cross-validation and 0.980 on the test set. Isolation Forest demonstrated superior F1 scores for rule-based anomaly detection. Local Outlier Factor exhibited enhanced sensitivity to outliers in the feature space. The computational efficiency was notable, with training times under 500 ms and inference times below 109 ms, confirming the practical implementability of both classification and anomaly detection tasks.

Message: Hi FE, I need help with “Air quality analysis based on threshold values and anomaly detection using machine learning models” in “Environmental Engineering”