用于等离子体相干模式自动识别的谱聚类算法实现

Implementation of spectral clustering algorithm for automatic identification of plasma coherence patterns

高约束模式对改善等离子体约束有着重要意义,但目前主要依赖人工进行模式识别,其效率低、成本高,导致核聚变装置中大量的诊断数据没有得到充分分析.为了解决这个问题,本文将机器学习中的谱聚类算法应用到EAST托卡马克装置上的电子回旋辐射成像、一维诊断系统电子回旋辐射计、磁探针、软X射线和快辐射等不同诊断系统的数据上,在时域及频域上识别出了锯齿模,验证了谱聚类方法的迁移性及准确性,解决了监督学习在数据处理上迁移性差以及需要依赖大量标签数据的问题.此外,本文实现了特定模式的筛选;最后利用电子回旋辐射成像及磁探针数据发现了一种可能的新模式,为新模式探索提供了一种新思路.

The number of data accumulated by controllable nuclear fusion devices is too large,and a large number of data have not been fully exploited.In such big data processing machine learning can play an important role.Therefore,in this work the spectral clustering method is used to realize the automatic processing of data,which can easily and quickly find the pattern information contained in the data.The discovery of these patterns is of great significance in improving plasma confinement and understanding plasma physics.In addition,in this work the spectral clustering method is applied to the electron cyclotron emission imaging(ECEI),one-dimensional diagnostic system electron cyclotron emissiometer,magnetic probe,soft X-ray,fast radiation(fast bolometer)and other different diagnostic systems on the EAST tokamak device.The sawtooth pattern is identified,the migration of the spectral clustering method is verified,and the problems of poor data processing migration in supervised learning and the need to rely on a large number of labeled data are solved.Finally,in this work,the ECEI and magnetic probe data are used to discover a possible new mode in the time domain and frequency domain respectively,which provides a new idea for exploring new modes.