E N A B L I N G T E C H N O L O G I E S

1 7 4 H I G H L I G H T S 2 0 2 3 I

New features of FASTVAC, the tool to detect and classify vacuum pressure events

FASTVAC, a fast-data vacuum pressure event detection and archiving system, was deployed in 2019 on the new EBS storage ring. Since then, analysing the data of thousands of vacuum pressure events has become challenging and time- consuming. The ESRF Vacuum group has recently developed automatic data-sorting using machine learning, with very promising results.

In 2016, the ESRF Vacuum group developed a system, FASTVAC, which allowed the detection and archiving of vacuum pressure events (sudden pressure changes in the storage ring) with a high time resolution compared to standard measurements [1]. The aim was to have a better understanding and analysis of vacuum behaviour. In particular, it assists experts in troubleshooting the precise location of the event, and provides qualitative and quantitative analysis and cross-correlation with other operation parameters.

Since the first beam injection in the new Extremely Brilliant Source (EBS) storage ring at the end of 2019, the average pressure has continuously improved, reaching 7x10-10 mbar today. At present, FASTVAC has recorded over 10 000 vacuum pressure events (from about 300 penning pressure gauges) with different signatures related to different causes. Every archived event consists of a pressure versus time record over a period of 5 seconds, with a time step of 5 ms. The 200 Hz acquisition rate signals, compared to the 1 Hz standard measurements, provide much more information and allow the identification of different types of pressure profile corresponding to different event causes, as illustrated in Figure 138. To simplify and automate the classification process, artificial intelligence, in the form of a simple neural network known as a perceptron, has been added to the data treatment process and has been running for the past year.

A perceptron is an algorithm for the supervised learning of binary classifiers [2,3]. In principle, the number of neurons needed in the perceptron would correspond to the number of possible vacuum event classes identified from all records. The learning and classification tasks are based on the use of the pressure plot black-and-white images as inputs. Considering this, each individual neuron aims to deliver its own independent answer: whether the

Fig. 138: a) Archived pressure versus time event recorded from standard measurement at 1Hz and from three consecutive gauges (named G2, G3 and G4). b) Same event recorded with FASTVAC at 200 Hz. c) Example of images created from

pressure versus time data and used as inputs for the artificial intelligence classification; R% is the automatic classification success rate for the different vacuum event classes.