Following work to define the project’s scope in 2018, as well as an initial feasibility study, the main project got underway in 2019. For the first stages of development within the project, data has been used from the Linac 4 accelerator at CERN. In particular, we have used data from the 2 MHz radio-frequency source that is used to create the plasma; this presents periods of ‘jitters’ that influence the beam’s quality.
By nature, these data sets are extremely noisy and volatile, leading to difficulties in interpretation and labelling. Therefore, the first research objective was to establish an appropriate data-labelling technique that would make it possible to identify ‘jittering’ periods. This has led to the creation of an anomaly detection system that recognises early symptoms in order to make preventive maintenance possible. Several approaches based on statistics and neural-network technologies were used to solve the problem. These approaches are now being combined in order to offer a system that can be adapted to different sources.
The data has been shown to be extremely difficult for neural networks to categorise. Rather than using neural networks to detect anomalies themselves, we have therefore made use of them to define appropriate parameters for a statistical treatment of the data source. This will, in turn, lead to detection of anomalies.
A first solution is already trained to function in the radio-frequency source environment of Linac 4. Therefore, the first objective of 2020 is to start its on-site implementation and to set up continuous field tests. The next challenge will then be to consolidate our parameter-selection model and to test the technique on multiple data sources.