Prototype of an early warning system based on deep learning for the CSNS accelerator

Abstract

To send out early warnings before some failures of the China Spallation Neutron Source (CSNS) accelerator, the feature models of the CSNS accelerator vacuums and drift tube linac (DTL) temperatures have been established based on deep learning, and a prototype of an early warning system has been developed. This prototype of an early warning system was built based on the experimental physics and industrial control system (EPICS) architecture, and it is mainly composed of three parts: training, recognition and information release. Python was adopted for program design and development, and functions such as training samples acquisition, deep learning networks design and training, online recognition and information release have been realized. The test results show that the accuracy of this prototype can reach 98.4% for the test set generated based on the historical data of the CSNS accelerator vacuums and DTL temperatures, and the anomalies of the CSNS accelerator vacuums and DTL temperatures can be recognized based on the real-time data, and the early warnings can be sent out, which proves its feasibility and effectiveness.