Improvement of feature extraction and intelligent identification method for the edge coherent mode in EAST
Introduction
The high confinement mode (H-mode) is a promising plasma regime for tokamak devices, because of its relatively high energy confinement and particle level, as well as good stability and repeatability [1]. One of the most important plasma filament of H-mode operation is edge localized mode (ELM) eruption. However, the problem of ELMs leading to transient giant heat flux is difficult to solve. In [2] the authors demonstrate that the edge coherent mode (ECM) can provide sufficient particle and energy exhaust across the pedestal in H modes in the absence of ELMs. In other words, ECM can provide an alternative to plasma control. However, so far ECM identification is mostly performed manually by a researcher. Research on the statistical behavior and intelligent recognition of ECM that contribute to understanding the dynamics of ECM are scarce. In addition, due to its inherently complex properties, the intelligent recognition of ECM is challenging.
In recent years, machine learning and deep learning have become research topics in nuclear fusion. Deep learning, by means of special neural networks called autoencoders, is applied to two classification problems of the TJ-II fusion database [3]. After classification boosting algorithms can be used to build data-driven models to perform classification of five images types of the TJ-II Thomson Scattering diagnostic [4]. A Faster R-CNN as a filament detector for the Mega Amp Spherical Tokamak-Upgrade (MAST-U) is also proposed [5]. However, the development of rule-based machine learning classifiers has allowed the implementation of a series of tools for disruption prediction, which provide very competitive performance in terms of both success rates and false alarms [6]. Meanwhile, a genetic algorithm improved model of disruption prediction and a real-time disruption predictor has been implemented [7, 8]. However, the above methods are either not suitable for ECM recognition, or the recall is poor. Moreover, the decision function of such models is usually a "black box", preventing intuitive understanding of the extracted feature.
Generally, ECM is detected by radial edge Helium Beam Emission Spectroscopy (He-BES) diagnosis on the EAST, and the sample rate of the data is1 MHz. In previous work, the data source of spectrogram is a correlation coefficient matrix (m0 × n0), where m0 is the number of timepoints and n0 the number of frequencies, which both depend on the FFT settings. The spectrogram becomes a binary image by denoising. We define white pixels in the binary image as "information points", and we call the total number of information points in a single row as "row information points". The existence of ECM is identified by estimating the change of the row information points in the binary image [9]. The advantage of this approach is that it can identify most ECMs quickly. However, we found that the features of a few ECMs are different from those of the others, and that the use of a fixed threshold to remove turbulence and noise will cause a small number of weak ECMs to be missed, therefore, it is difficult to further improve the precision and recall of this approach. In this context, we propose an adaptive denoising method that can better retain the main features of the weak ECM. We prove that a double-channel convolution algorithm can distinguish ECM from broad-spectrum turbulence. The combination of the adaptive denoising method and the double-channel convolution algorithm provides competitive performance for ECM recognition.
Section snippets
Image pre-processing
In order to understand ECM intuitively, we present three spectrograms. Fig. 1(a)–(c) are respectively the spectrograms of shot #55,861, #54,800, and #55,560, whose data source is the correlation frequency matrix (m0 × n0), where m0 is the number of timepoints and n0 the number of frequencies. As one can see from Fig. 1(a), the ECM fluctuation frequency started at ~50 kHz, at a time of ~3.4 s, after the L-H transition, then the frequency of the ECM gradually decreased, and finally stabilized at
Adaptive denoising method
In the image processing filed, the denoising problem can be considered as the problem of how to distinguish between the foreground and background of an image. In our case, the goal is that the ECM will be retained as the foreground, and null signal, noise and broad-spectrum turbulence will be removed as the background.
The variance, which represents the degree of change of the pixel value in the image, is one of the important criteria that can be used to evaluate each pixel of image C (show in
ECM feature extraction model
Generally, broad spectrum turbulence is strongest at the bottom of the spectrogram, and gradually becomes weaker with increase of frequency. However, if there is an ECM, the turbulence weakens rapidly. This occurs because the former may obtain free energy from the turbulence and saturate [10]. This feature is represented by the gradient of the pixel values in the ECM spectrogram. After pre-processing, the edge of the ECM in the spectrogram becomes clearer and the gradient of the pixel values is
Performance evaluation
Clearly, because of the fused feature, ECM recognition becomes easier, as shown in Fig. 7. 1) If the fused feature area is greater than a threshold that we set, it may be ECM. The threshold is obtained from the precision-recall (P-R) curve, which will be described in detail in the following. 2) Moreover, the Δf/fpeak of the ECM spectrogram is normally < 50%, where Δf and fpeak denote the half-peak bandwidth and the characteristic frequency of the mode respectively [11]. Thus, if the Δf/fpeak of
Conclusions
ECM can drive particle transport in the pedestal regime in EAST, the intelligent recognition of ECM is a critical task. In this paper, firstly, we define an adaptive denoising inequality based on the variance, secondly, we propose a dual-channel convolution feature extraction model based on adaptive denoising to extract ECM feature, and finally we distinguish the extracted feature by threshold and evaluate the algorithm performance. Compared with the ECM recognition approach based on the
Credit author statement
Yuqian Yang:Conceptualization;Methodology;Coding; Writing
Ying Liu:Investigation;Analysis, or interpretation of data for the work, Reviewing and Editing
Jianjun Huang:Supervision
Ye Yang:Determine ECM identification criteria
Bin Long:Responsible for the experiment of deep learning part
Fulin Zeng:Test and validation
Zhongxuan Wu:Data Annotations
Declaration of Competing Interest
We declare that no conflict of interest exits in the submission of this manuscript, and manuscript is approved by all authors for publication. I would like to declare on behalf of my co-authors that the work described was original research that has not been published previously, and not under consideration for publication elsewhere, in whole or in part. All the authors listed have approved the manuscript that is enclosed.
Acknowledgment
This work is supported by the National Natural Science Foundation of China under contract nos. 11805132, and the Natural Science Foundation of Guangdong Province under contract nos. 2017A030310139. This work is partly supported by the Institute of Plasma Physics Chinese Academy of Sciences.
Reference (12)
- et al.
Automatic feature extraction in large fusion databases by using deep learning approach
Fusion Eng. Des.
(2016) - et al.
Image classification by using a reduced set of features in the TJ-II Thomson scattering diagnostic
Fusion Eng. Des.
(2018) - et al.
Towards an automatic filament detector with a Faster R-CNN on MAST-U
Fusion Eng. Des.
(2019) - et al.
On the potential of ruled-based machine learning for disruption prediction on JET
Fusion Eng. Des.
(2018) - et al.
A multidimensional linear model for disruption prediction in JET
Fusion Eng. Des.
(2019) - et al.
Real-time implementation with FPGA-based DAQ system of a probabilistic disruption predictor from scratch
Fusion Eng. Des.
(2018)
Cited by (3)
A detection method of Edge Coherent Mode based on improved SSD
2022, Fusion Engineering and DesignCitation Excerpt :DCCN uses traditional image processing methods to identify ECM, which improves the recognition accuracy more than the statistical information method. However, the low recognition speed still challenges the applications [10]. Different from recognition, the detection result must include the location of ECM in the spectrogram.
Intelligent recognition and location of the edge coherence mode in EAST
2023, Plasma Physics and Controlled FusionRecognition of Edge Coherent Mode in EAST with Convolutional Neural Network
2022, Fusion Science and Technology