Neural network technique for improving accuracy, reliability and robustness of beam position monitor system

The beam position monitor (BPM) is a crucial instrumentation system for the commissioning and operation of the accelerator. Its accuracy and robustness are essential for ensuring the stability of the accelerator. Currently, the beam position is calculated by fitting a polynomial to the four voltage signals obtained from the BPM electrodes in BEPCII and HEPS. To improve the system’s robustness, a formula is provided that expresses the relationship between the three voltage signals and the position. The average fitting error is 40 𝜇m, but the error of the three-electrode calculation is not high. Therefore, we propose using neural networks for beam position calculation to improve the system’s robustness while guaranteeing its accuracy. This will ensure that the beam position can be provided stably, even in the case of one single electrode error. In our experiments, we use BPM calibration data from HEPS. The trained neural network’s performance on the test set meets the accuracy requirements, with an error of less than 15 𝜇m in both four-electrode and three-electrode predictions, and an average value of fitting error is 1 𝜇m. Furthermore, we validate the neural network’s generalization ability by using data measured by BPM on HEPS.