Error Detection and Error Localization

Modern radio telescopes are super sensitive, but the scale and complexity of these systems make it increasingly difficult to find and resolve system malfunctions. The LOFAR radio telescope for example produces almost ten thousand spectrogram images per observation that provide information about the quality of the systems. It is not practical to manually go through this large amount of images to detect system errors, which is why research has been conducted into whether artificial intelligence (AI) could offer a solution.

In the ASTRON-NleSC AI project, a neural network has been built, a so-called 'autoencoder', that is able to automatically find different types of system errors in the spectrogram images and cluster them. This approach is particularly suitable when no labeled data (example data) is available, such as at the start of this project. Each cluster corresponds to a certain type of pattern or error visible in the images. Because similar errors are clustered, finding system malfunctions has become much easier: instead of thousands of images, only about twenty clusters now need to be viewed.

A prototype web application has been made available to the LOFAR telescope, and has already led to the discovery of several subtle system errors, which were subsequently quickly resolved.

In addition to the above, the project has also implemented a method to suppress radio interference in time-variable signals from radio telescopes. This was done by efficiently implementing the relevant algorithms in computer hardware for the Time Domain Pipeline (TDP) of the Square Kilometer Array (SKA) radio telescope. This work contributes, for example, to follow-up research into low-frequency gravitational waves recently discovered using pulsar timing arrays.