ANL, Lemont, Illinois, USA
We explore the possibility of using a Machine Learning (ML) algorithm to identify the source of occasional poor performance of the Particle Accumulator Ring (PAR) and the Linac-To-PAR (LTP) transport line, which are parts of the injector complex of the Advanced Photon Source (APS) at Argonne National Lab. The cause of reduced injection or extraction efficiencies may be as simple as one parameter being out of range. Still, it may take an expert considerable time to notice it, whereas a well-trained ML model can point at it instantly. In addition, a machine expert might not be immediately available when a problem occurs. Therefore, we began by focusing on such single-parameter anomalies. The training data were generated by creating controlled perturbations of several parameters of PAR and LTP one-by-one, while continuously logging all available process variables. Then, several ML classifiers were trained to recognize certain signatures in the logged data and link them to the sources of poor machine performance. Possible applications of autoencoders and variational autoencoders for unsupervised anomaly detection and for anomaly clustering were considered as well.
ANL, Lemont, Illinois, USA
Particle accelerators require continuous adjustment to maintain beam quality. At the Advanced Photon Source (APS) this is accomplished using a mix of operator-controlled and automated tools. To improve the latter, we explored the use of machine learning (ML) at the APS injector complex. The core approach we chose was Bayesian optimization (BO), which is well suited for sparse data tasks. To enable long-term online use, we modified BO into adaptive Bayesian optimization (ABO) though auxiliary models of device drift, physics-informed quality and constraint weights, time-biased data subsampling, digital twin retraining, and other approaches. ABO allowed for compensation of changes in inputs and objectives without discarding previous data. Benchmarks showed better ABO performance in several simulated and experimental cases. To integrate ABO into the operational workflow, we developed a Python command line utility, pysddsoptimize, that is compatible with existing Tcl/Tk tools and the SDDS data format. This allowed for fast implementation, debugging, and benchmarking. Our results are an encouraging step for the wider adoption of ML at APS.