A particle detector at the Large Hadron Collider runs for years without stopping. Its sensors produce mountains of data every second, and among it all, you need to spot a fault quickly, like overheating or a broken channel. In the past, simple rules did this, but they often made mistakes. Now, a neural network called DepthViT steps in — light and fast, like a vigilant technician who knows what normal operation looks like.
The catch is that conditions inside the detector change over time: parts wear out, and the particle flux shifts. A neural network trained on old data starts to get confused and raises false alarms. The authors came up with a solution: they combine several models trained on different time periods. Like an experienced mechanic who remembers both the old and new quirks of a machine, this system stays accurate. On data from the CMS calorimeter, anomaly detection accuracy stays above 98.8% even under strong changes.
This approach is useful beyond physics. In factories, machines also operate under changing conditions: vibration, temperature, wear. A neural network that can adapt could help predict breakdowns before they happen. A detector that never tires is a step toward smart systems that can work for years without supervision.
The DepthViT neural network processes calorimeter occupancy maps: these are like heat maps showing where particles left energy. An anomaly looks like an unexpected spot on such a map.