Detecting lighthouse faults with machine learning
Since 2021 we have worked with the General Lighthouse Authorities of the UK and Ireland (GLA) on machine learning that helps spot faults in lighthouses and other aids to navigation early.
The challenge
Many lighthouses switch on and off by themselves, using a sensor that detects daylight. Over time the sensor can drift, so the light comes on late. For a ship, a light that has not come on is a missing warning. Lighthouses are often remote, so engineers cannot check them often, and faults may only be found at the next scheduled visit. There is also almost no record of past sensor failures to learn from.
What we did
We used past records from seven Trinity House lighthouses, together with weather and sunlight data, to teach a computer program what normal behaviour looks like at each one. Because real faults were not available, we simulated a drifting sensor and checked whether the program noticed that something was wrong.
It did: the program's performance fell steadily as the simulated fault grew, and a delay of 10 to 15 minutes was already visible. A drop like this could act as an early warning to engineers, before the light is noticeably late. This is one project in a programme with the GLA that is now continuing as a Knowledge Transfer Partnership.