Research On Accelerator Alarm System Optimization And Beam Stability Technology

Particle accelerators play an important role in high energy physics,material science and biomedicine,the high energy particle beams of which are widely used in the research on the fundamental properties of matter,damage mechanisms and irradiation repair of materials,tumour therapy and other areas.However,particle accelerators are affected by a variety of factors such as the external environment,malfunctions in the operating equipment and human error,which can lead to abnormal beam interruptions,beam position shifts and divergence,affecting beam stability and the accuracy of experimental data.Therefore,it is a hot research topic in the field of accelerator technology to study how to reduce the repair time of equipment failures,the number of accidental interruptions of beam and improve the self-correction level of beam position.To solve the above problems,technologies such as high throughput streaming,data analysis and visualization,mobility and standardized system packaging need to be introduced to achieve an efficient alarm information flow processing mechanism for the accelerator alarm system,improve the availability of the alarm system and provide an accurate alarm basis for equipment maintenance.Machine learning algorithms are also required to filter and process abnormal fault data and reduce the frequency of false interlocks.In contrast to traditional mathematical expression-based approaches to accelerator beam tuning,deep learning algorithms can automatically build network models to express highly non-linear functional relationships for predicting or identifying data.This thesis investigates the beam stability of particle accelerators based on the Chinese ADS Front-end Demo Linac(CAFe)and the High Energy Electron Radiography(HEER)platform to improve beam stability by improving the response ability and maintainability of the system at the software level,and reducing the false interlock and beam offset at the system level.The specific contributions and results of this thesis are summarized as follows:1.Methods to improve the availability of the accelerator alarm system are studied,solving the problems of low throughput,low visualization of logs,poor mobility and standardization in traditional alarm systems.Firstly,the original alarm system streaming platform is optimized based on Kafka to ensure that data is not blocked and lost when a large number of alarm messages are concurrent,which improves the real-time performance of the accelerator alarm data analysis and fault diagnosis.Secondly,the end-to-end migration of alarm data and multi-dimensional alarm log visualization techniques are studied to add a more effective means of data analysis to the alarm system.Then,improved methods for alarm system mobility are studied,enabling mobile access to the alarm system and enhancing the remote operation and maintenance ability of the accelerator.Finally,the alarm service is standardized and encapsulated to improve the cross-platform compatibility,operation and maintenance capabilities of the alarm system.The above research optimizes the availability of the accelerator alarm system in terms of overall architecture.The optimized alarm system has been run on the CAFe accelerator platform with good results,providing an accurate basis for fault maintenance and optimizing the maintainability of the equipment,thus improving the operational stability of the beam.2.An anomaly detection method of accelerator state based on isolation forest(i Forest)is proposed.The method is based on a machine learning algorithm to monitor the anomalous state of the Beam Position Monitor(BPM)temperature and reduce the number of false interlocks in the accelerator machine protection system.The method is used to characterize the degree of BPM anomalies by dividing the temperature data space to obtain different path lengths of data,and to construct a BPM anomaly detection model.The model achieves the detection of abnormal BPM temperature data,with the ability of filtering out abnormal datas with different characteristics,and achieves a good abnormality classification effect while meeting the engineering time requirement.The i Forest-based accelerator state anomaly detection method improves the stability of beam operation and provides a reference for alarm threshold setting.3.An automatic beam adjustment method based on Surrogate Model Assisted Optimization(SMAO)is proposed.This method studies the accelerator beam prediction model based on neural network,and trains the model in a data-driven manner to establish the mapping relationship between machine parameters and beam parameters with fast calculation speed.Based on this,the purpose of optimizing machine parameters through Bayesian optimizer is further studied.The black box model generated by the neural network is automatically searched for machine parameters to achieve automatic beam calibration and improve beam stability and beam adjustment efficiency.It lays an important technical reserve for the future accelerator construction.