Research On Time Series Data Analysis And Storage Of CAFe Control System

Climate change is a significant global challenge.As the largest energy-consuming country,the transformation of green and low-carbon energy structure is not only a challenge in the field but also an important opportunity for the development of new energyrelated technology.Nuclear,as an energy source,is clean,reliable,and sustainable.Nuclear fuel has a high energy density and does not cause air pollution.However,nuclear power generation also faces the problems of spent fuel disposal and thermal pollution.Accelerator Drive Sub-Critical System(ADS)reduces the radioactivity of nuclear waste and the volume of nuclear waste by transmutation reaction to processing spent fuel.ADS system is composed of a superconducting proton linac accelerator,a spallation target,and a reactor.Based on ADS injector II,the Institute of Modern Physics of the Chinese Academy of Sciences has built a superconducting linac accelerator,named Chinese ADS Front-end Demo Linac(CAFe),which has completed the technical accumulation for a higher power device,China initiative Accelerator Driven System(Ci ADS).The control system,as an important part of the accelerator,will generate time-series data related to the state of the accelerator in the operation process.This dissertation carried out a series of research work on the time-series data generated by the accelerator control system.The accelerator alarm system monitors the state of the process variables in realtime then issues alarm information,and the time-series alarm data generated can be used for hidden danger investigation and fault diagnosis for the accelerator.In this dissertation,the concept of alarm events is proposed.In addition to alarm time,alarm events also include fields such as device type and alarm severity.Multi-dimensional column vectors are used to represent alarm events.Distance correlation is introduced to determine the sliding window width and transform time-series data into transactional datasets.Combined with the characteristics of accelerator alarm data,a parallel association rule CApriori algorithm,based on the big data computing engine Spark,is proposed with high correlation and low support as conditions.CApriori algorithm uses distance correlation in the second stage to filter candidate binomial sets with high frequency and low correlation.This algorithm performs partitioned parallel mining of frequent itemsets and improves algorithm performance by using Bloom filters and using transactions directly to generate candidate binomial sets.Finally,the time complexity of the algorithm is analyzed in this dissertation,and the effectiveness of the algorithm is verified by applying the proposed algorithm to the alarm data of CAFe and comparing it with the YAFIM algorithm.The machine protection system prevents the damage of the beam current to the equipment by preventing the occurrence of faults.The abnormal value of the actual signal causes the machine protection system to generate false alarms,which seriously affects the robust running of the accelerator.To solve this problem,this dissertation proposes an alarm discrimination algorithm based on deep learning to abstract the problem into the multivariate time-series outlier detection problem.The alarm discrimination algorithm consists of two modules: prediction and detection.The prediction module is composed of a two-layer LSTM model based on the attention mechanism.After the sample data is denoised by wavelet transform,it is used to capture the change law of analog quantity under normal conditions.The detection module uses the predicted value and the actual value to calculate the mean square error,then detects abnormal signals by comparing with the mean square error threshold and changes the false alarm rate by adjusting the mean square error threshold.Finally,experiments are carried out on BPM temperature dataset.The experimental results show that the alarm discrimination algorithm can effectively identify abnormal signals and reduce false alarms.With the increasing scale of accelerators,the storage and query of historical data face new challenges.This dissertation researches and implements a new EPICS data archive engine,named Archive Engine,based on distributed database HBase.According to the characteristics of different types of Process Variable,three sampling modes,which are Monitor,Monitor with Threshold and Scan,are implemented.Metadata is stored in the My SQL database.Archive Engine prevents reconnection of abandoned Process Variables by adding a pause flag to each Process Variable.Archive Engine is also an HBase client,so all Process Variables share the same buffer.According to the underlying physical structure of HBase and user data query habits,the HBase table structure is designed.This dissertation focuses on the system architecture and expansion method of Phoebus and realizes the retrieval of historical data through the Databrowser of Phoebus.Through the comparative experiment with EPICS Archiver Appliance,the retrieval performance of Archive Engine is verified.Aiming at time series data generated in the operation process of the CAFe device,different methods and technologies are used in the research.The work described above has important engineering application value for the intelligent diagnosis of accelerator faults,improvement of operation stability,and improvement of data archiving engine functions.Also,technical reserves for the construction of Ci ADS are provided.