Challenges of data classification for Machine Learning Algorithms: Case Studies in Drinking Water Distribution Systems

Online monitoring of water quality in drinking water distribution systems (DWDS) is critical to ensure public health and regulatory compliance. Online sensor systems can provide real-time insights into water quality variations and enable prompt detection of anomalies. This paper presents two case studies demonstrating a multi-sensor monitoring approach for detecting and diagnosing water quality issues in DWDS and evaluates challenges of water quality classification tasks for ML applications. In the first case study, an online sensor system with pH, turbidity, total organic carbon, electrical conductivity, and oxidation-reduction potential (ORP) sensors was deployed in the middle of a distribution network. Investigation revealed the presence of a partially closed valve, which increased water age. This case highlights the role of sensor networks in identifying hydraulic anomalies and their impact on water quality. The second case study involved monitoring water quality at iron removal water treatment plant. The sensor system detected recurring increases in turbidity, reaching 10-20 NTU every 2–3 days. Detailed analysis traced the issue to an eroded butterfly valve within the de-ironing plant, which failed to properly separate flushing water from treated water. Utilizing sensor data for classification in machine learning algorithms remains challenging. Data interpretation of DWDS processes require site-specific information and data is often highly variable. The findings highlight the necessity of integrating multi-sensor platforms with hydraulic and operational assessments to enhance the reliability and safety of DWDS. This paper was presented at the 21st Computing and Control in the Water Industry Conference (CCWI 2025) at the University of Sheffield (1st - 3rd September 2025).