Since 18 of December 2019 conferences.iaea.org uses Nucleus credentials. Visit our help pages for information on how to Register and Sign-in using Nucleus.
13-17 May 2019
Daejeon, Republic of Korea
Europe/Vienna timezone
Meeting Material is now available and accessible from the left-menu

Real-time ELM recognition system based on deep learning

15 May 2019, 15:25
2h 35m
Daejeon, Republic of Korea

Daejeon, Republic of Korea

Board: P/2-1
Poster Advanced Computing and Massive Data Analysis Poster

Speaker

Prof. Fan Xia (Southwestern Institute of Physics)

Description

Based on the deep learning method, this paper introduced a set of ELM real-time recognition system on HL-2A. The system uses 5200 shots data (about 241,900 data slices) for learning. After more than 70 adjustments, a 22-layer convolutional neural network is obtained. The network can recognize whether a 30ms data slice contains the ELM signal, After smoothing, the system can recognize the ELMy H-mode and output the start/end time of H-mode.
The system has recognized all historical data of the HL-2A since it achieved stable H-mode discharge in 2009. A total of 1665 shots of H-mode has been recognized, of which 35 shots were misidentified, with the false positive rate (FPR) was 2.10%. In the actual 1634 shots of H-mode, the system missed to recognize 4 of them, with the false negative rate (FNR) was 0.24%. In a total of 25,321 shots, the recognition accuracy rate was 99.85%. For the data of latest years, the system obtains good results with a FPR of 3.69% and a FNR of zero. In the correctly recognized shots, the error of the H-mode start/end time less than 20ms. The data quality of HL-2A in different discharge periods has a certain correlation. These results demonstrate that the system has good generalization ability and recognition precision to fulfill the requirements for applying to new data.
In this experiment, a total of 14,900 data slices were used to test the speed of the neural network in the simulated real-time environment. The average calculation time of a single slice is 0.46ms less than PCS control cycle which is 1ms. This result show that the system can satisfy the calculation speed requirements of the real-time ELM recognition.
Since the HL-2A experimental has not yet ended, the system has not been transplanted to PCS, but the recognition precision and the calculation speed of the system have satisfied the requirements of real-time ELM recognition and real-time ELMy H-mode control.
This research introduces the deep learning method into the fusion research of HL-2A, and obtains the results that can satisfy the actual production requirements. It proves the feasibility and advantages of the combination of artificial intelligence method and fusion research field represented.

Primary authors

Prof. Fan Xia (Southwestern Institute of Physics) yao huang Dr Zongyu Yang Prof. Wulyu Zhong

Presentation Materials