At Kubrick, research forms part of how our data, AI, and engineering teams explore practical ways to solve complex operational problems. It helps us test new methods, build deeper technical understanding, and translate emerging approaches into capabilities that clients can use in real-world environments.
Our team’s research on physically informed anomaly detection for wastewater network asset management was accepted for presentation at the 2nd International Conference on Engineering, Technology and Management 2026 and is pending publication.
The work demonstrates a practical capability that matters for water utilities: using data, AI, and domain reasoning together to distinguish between expected wastewater network behavior and events that may require operational attention.
Making sense of wastewater signals with domain context
Wastewater networks are difficult environments for analytics. Flow, pressure, level, rainfall, and pump telemetry are noisy and highly variable. A rise in flow may be entirely normal during rainfall but concerning during dry weather. A pump signal may explain a sudden change in measured flow. Daily usage patterns can create predictable peaks that should not be mistaken for faults.
This is why anomaly detection in wastewater systems cannot be treated as a simple outlier problem. Detecting that something is unusual is not enough. Operators need to know whether it is unusual given the physical context of the network.
From network signals to operationally useful alerts
The paper addresses this problem by introducing a three-stage framework.
- The model estimates expected wastewater flow using known drivers such as hour of day, recent rainfall, and pump activity.
- Anomaly detection is applied to the residuals: the part of the signal not explained by those drivers.
- Detected events are classified into wet-weather or dry-weather contexts to support operational triage.
This approach reflects how experienced engineers and operators reason about sewer systems. They do not look at a flow spike in isolation. They ask whether it rained, whether pumps were running, whether the pattern fits normal daily demand, and whether the behavior is consistent with the asset’s expected response. Our method brings that reasoning into the analytics layer.
Using UK sewage pumping station data, the research showed that even a deliberately simple physically informed model could explain a meaningful share of flow variability. Daily timing, rainfall effects, and pump operation were all statistically and operationally important. After accounting for these drivers, the anomaly detector produced a more useful set of alerts than a raw-flow detector, with fewer rainfall-confounded alerts and a higher proportion of dry-weather anomalies.
Turning telemetry into actionable decisions
The result is a sharper operational signal: fewer rainfall-confounded alarms, clearer prioritization of dry-weather anomalies, and faster routes to investigation. For operators, this means less noise and more confidence in where to focus attention.
For utilities, the value is practical. Physically informed AI can help turn telemetry into decisions that support proactive maintenance, targeted intervention, and stronger evidence for environmental and regulatory performance.
This research also demonstrates how we approach data and AI consulting in infrastructure: start with the operating reality, encode the physical drivers, and build analytics that engineers and decision-makers can trust.
The framework provides a scalable foundation for richer diagnostics across flow, pressure, level, pump performance, asset hierarchy, network topology, and maintenance history. In practice, that means moving from generic anomaly alerts toward explainable signals of blockage, pump underperformance, abnormal inflow, infiltration, or sensor fault.
For water companies managing aging assets, environmental scrutiny, and growing telemetry volumes, this is the step change: analytics that are not only technically sound, but operationally insightful and actionable.
