Combining multi-sensor data and machine learning for improved wastewater contamination detection in stormwater systems

Stormwater drainage generally flows untreated into receiving waters, such as rivers, groundwater, and the sea. Stormwater networks can serve as pathways for contamination to enter receiving waters. Misconnections, illicit connections and discharges, overflows, and leaks from damaged sewers are the primary causes of such contamination. These issues not only degrade water quality, posing public health and environmental risks, but also create a range of expensive operational challenges for water and wastewater companies.

Detecting wastewater contamination and tracing its entry points into stormwater systems remains a significant challenge predominantly due to various potential sources of incoming wastewater, dilution and dispersion of contaminants by tributary stormwater flows, and significant differences in consistency, regularity, and flow rate of inflows.

We conducted an investigation on an urban stormwater pipeline in the UK suspected of receiving wastewater from multiple misconnections. The aim of the investigation was to determine whether the stormwater system was being impacted so that the statutory undertaker could address the contamination issues and improve the quality of the receiving water environment. The source of the misconnections was uncertain prior to the investigation but it was suspected that they may have been inputs from domestic households, small to medium-sized businesses, or both. The study employed a comprehensive approach combining water sampling for microbiological indicators (total coliforms and E. coli) and an array of chemical analyses, including trace elements, organics, nutrients, petroleum hydrocarbons, and volatile and semi-volatile organic compounds. The collection of grab samples was complemented by the use of a Proteus multi-sensor sonde (Proteus Instruments, UK), which measured parameters such as tryptophan-like fluorescence (TLF), chromophoric / fluorescent dissolved organic matter (CDOM/fDOM), electrical conductivity, pH, ORP, turbidity, temperature, ammonium (NH4), and dissolved oxygen (DO). Moreover, data collected with the multi-sensor sonde was used to model microbial parameter concentrations over a period of approximately three weeks. Two modelling approaches were tested: one following the methodology recommended by Proteus Instruments, and another employing the machine learning Random Forest method. The latter approach offers potential advantages in addressing challenges commonly associated with fluorescence-based sensors. The findings demonstrate the potential for enhanced detection of wastewater misconnections, providing a more efficient and accurate method for identifying sources of contamination within stormwater systems.