Anomaly detection finds readings or patterns in water data that differ from normal behaviour, such as sensor faults, leaks, bursts or pollution events.
Anomaly detection is the task of finding data points or patterns that do not fit expected behaviour. In water data, an anomaly can mean a problem with the sensor or a real event in the system, and telling the two apart is a large part of the job. Simple methods use rules: values outside a physical range, sudden jumps, flat lines where a signal should vary, or gaps. Statistical methods compare each value with a baseline, such as a moving average and standard deviation, or the typical value for that hour and day of the week. Model-based methods predict what the reading should be, for example with a hydraulic model, a regression or a neural network, and flag large differences between predicted and observed values (the residuals). Unsupervised machine learning methods, such as isolation forests or autoencoders, learn normal patterns from history and score how unusual new data look. Water examples: a rise in minimum night flow into a district metered area can signal a new leak or burst. A sudden fall in pressure across several loggers can point to a main break. A step change in conductivity or a dip in dissolved oxygen at a river monitoring station can indicate pollution. A stuck level sensor in a sewer reads the same value for hours. In groundwater networks, a sudden jump in a water level record may be a logger reset rather than a real change. What it needs: a long enough record of normal behaviour, timestamps and metadata (sensor changes, maintenance, known events), and ideally some labelled past incidents to test against. Limits: there is a constant trade-off between missed events and false alarms. With too many false alarms, staff stop looking. Seasonal patterns, holidays and rare but normal events, such as floods, can look anomalous. A flag is a prompt to check, not proof of a fault.