Antonin Sulc, Raimund Kammering, Tim Wilksen (DESY, Hamburg)
Interpretation of data from beam position monitors is a crucial part of the reliable operation of European XFEL. The interpretation of beam positions is often handled by a physical model, which can be prone to modeling errors or can lead to the high complexity of the computational model. In this paper, we show two data-driven approaches that provide insights into the operation of the SASE beamlines at European XFEL. We handle the analysis as a data-driven problem, separate it from physical peculiarities and experiment with available data based only on our empirical evidence and the data.
Note (20/08/2022): As we mention in the paper, the hypersphere center c isn't kept fixed and is updated through the optimization, batching should cause slight variations of the c and prevent the collapse of all points projected on the c. In the original paper, they particularly stress that this may lead to the projection of all points onto the c by zeroing all weights and keeping bias c = b. In future experiments, we plan to stick to the suggested technique (i.e. remove biases and keep c fixed), since this explains why many interations cause diminishing scores.
The files are compressed into pickles with gzip package with 8-bit float precission (for storage purposes). The pickle contains dictioary with first_bunch_x and first_bunch_y for horizontal and vertical BPMs respectively. Each element is a pandas (pandas 1.3.1) data frame. Each row is position of first bunch in a single indidivual bunch train.
Two issues shown in the paper.
SASE1 undulator server crashed after an unusual selection of colours for individual cells (2022-03-07 20:45:00)
zip.
zip.
To understand how to read data, check get_X_from_file in model.py. The function loads a list of files and normalizes it at the same time.
The model requires following libraries:
- pytorch '1.10.2+cu113'
- numpy '1.20.3'
- pandas '1.3.1'