Enabling Automatic Detection of Anomalies in Wastewater: A Highly Simplified Approach to Defining “Normal” in Complex Chemical Mixtures

Abstract

Wastewater treatment demands management of influent conditions to stabilize biological processes. Generally wastewater collection systems lack advance warning of approaching water parcels with anomalous characteristics, which could then be diverted for testing or pre-treatment. A major challenge in achieving this goal is identifying anomalies against the complex chemical background of wastewaters. This work evaluates unsupervised clustering methods to characterize “normal” wastewater characteristics, using >17 months of 10-min resolution absorbance spectrometry data collected at an operating wastewater treatment facility. Comparison of results using K-means, GMM, Hierarchical, and DBSCAN clustering shows minimal intra-cluster variability achieved using K-means. The four K-means clusters include three representing 99% of samples, with the remaining cluster (<0.3% of samples) representing atypical measurements, demonstrating utility in identifying both underlying modalities of wastewater characteristics and outliers. K-means clustering provides a better separation than grouping based on factors such as month, precipitation, or flow (with 25% overlap at 1-σ level, compared to 93, 93, and 83%, respectively) and enables identification of patterns that are not visible in factor-driven grouping, e.g., shows that summer and November months have a characteristic type of behavior. When evaluated with respect to wastewater influent changes occurring during the SARS-CoV-2 pandemic, the K-means approach shows a distinct change in strength of diurnal patterns when compared to non-pandemic periods during the same season. This method may therefore be useful both as a tool for fast anomaly detection in wastewaters, contributing to improved infrastructure resilience, as well for providing overall analysis of temporal patterns in wastewater characteristics.