The Huddersfield projectAnomaly detection on the River Colne
Three simple rules, thousands of alarms, and a careful look at which ones matter.
The learner asks the Environment Agency's data service for every 15-minute level reading at Longroyd Bridge in 2025. The gap rule fires first: a whole day, 1 January, is missing. It looks like a sensor outage, until the learner reads the service's notes. The filter they used, min-date, leaves out the starting date itself; mineq-date includes it. Downloaded again properly, the year has all 35,040 readings, none missing. The first anomaly was a bug in the learner's own request.
Our Python anomaly rules on Longroyd Bridge levels, 2025, 27 September 2026| Rule | Threshold | Alarms | What checking showed |
|---|
| Gap in the readings | Any missing 15-minute slot | 1 day, then 0 | Our download bug, not the sensor |
| Flatline | Same value for 2 hours | 723 runs | Mostly calm low water, 32.6% of readings |
| Flatline | Same value for 24 hours | 1 run | 0.223 m from 6 August, worth a question |
| Spike | Up and straight back by 0.05 m | 13 | Short bursts of fast water |
| Spike | Up and straight back by 0.1 m | 3 | All in April and September |
| Spike | Up and straight back by 0.2 m | 0 | Nothing that extreme |
The flatline rule is next: flag any value that repeats unchanged for two hours. It raises 723 alarms, and a third of the whole year sits inside them. That cannot all be faults. Sorting by level shows why: 517 of the runs are at or below the year's median level of 0.275 metres. In calm, low water a sensor that reads to the millimetre really can show the same number for hours. Lengthening the threshold to a full day leaves one run, at 0.223 metres from 6 August, which deserves a human look.
Every one of the year's readings carries the Agency's own quality label, and all 35,040 say Good. So the learner treats the rules as questions, not verdicts, and measures them: how many alarms per month, how many at low water, how many survive a stricter threshold. The final program flags only the day-long flatline and the three largest spikes, and writes a short report explaining each. Tests use a made-up series with a planted stuck sensor and a planted spike to prove the rules still catch real faults.
Ages 8 to 11
Take the temperature every hour for a day and circle any reading that looks odd, then explain why.
Ages 11 to 15
Write the gap and flatline rules in Python and count the alarms.
Ages 15 and up
Add the spike rule, tune thresholds by level, and test with planted faults.
Agency readings, our rules
The readings and quality labels come from the Environment Agency hydrology service, and the gauge description from the National River Flow Archive. The rules, thresholds and counts are ours.
