從線積分信號中重建電漿參數（包括密度、溫度等）的空間分佈的傳統方法是假定電漿為宏觀靜態的狀態。 然而，在大多數情況下，電漿（尤其是在核聚變電漿實驗中）在空間和時間上是動態演化的。 因此，我們嘗試使用深度神經網絡 (DNN) 從干涉儀測量的線積分密度 n_e L(x,t) 中建立電漿密度截面輪廓在時空演化的新重建方法，其中 x 和 t 分別表示探測波路徑的位置和時間。這個想法是將重建過程視為圖像重建，這是 DNN 的專業領域。 也就是說，將線積分密度 n_e L(x,t) （等效於功率頻譜 (n_e L) ̃(x,f) 和相位頻譜 ϕ(x,f)）的時空演化作為輸入圖像到 DNN，相應的輸出是橫截面密度分佈的動態圖像。作為第一步，我們使用由干涉儀測量的數值模擬產生的數據集開發和訓練 DNN，以製備剪切旋轉等離子體的各種剖面，其密度剖面表示如下：n_e (r,t)=∑_r〖∑_m〖Amp(r,m)cos⁡(mθ-ω(r,m)〗 t+θ_0 〗) ，其中 r,Amp(r,m), m, θ, ω(r,m), …分別是徑向位置、幅度、極向模數、極向角、角頻率……。 m 範圍為 0 到 4。
在數值模擬中準備了以下三個數據集：（A組）在所有層中只允許單一且同一種模式存在，其中對於滿足m=m_single 的位置則在其所有的徑向層中Amp(r,m)≠0，而對於m≠m_single 的位置則 Amp(r,m)=0，其中 m_single 是作為單一模式的選定模式編號。 （B組）每個徑向層中只允許單一模式，而在不同徑向層之間可以是不同模式，並且各個模式可以有不同的值，其中Amp(r,m_single (r))≠0僅當m=m_single時成立，而對於m≠m_single 的位置則 Amp(r,m)=0。 （C組）允許多種模式存在於每個徑向層。 由數據集（A 組）訓練的 DNN 在三種情況中表現最好，平均絕對誤差（MSE）低於 1%，均方誤差MSE 低於 500 Hz（10% 的密度擾動的振盪頻率分佈的最大頻率）。 其他兩種情況（B 組）和（C 組）無法顯示具有可接受誤差範圍的重建結果。

Conventional reconstruction method of spatial profiles of plasma parameters including density, temperature, etc. from line-integrated signals assume macroscopically static plasma states. In most cases, however, plasmas (especially in fusion plasma experiments) evolve dynamically in space and time. Therefore, we attempt at establishment of a novel reconstruction method of spatio temporal evolution of cross-sectional profiles of plasma density with the use of deep neural network (DNN) from line-integrated density n_e L(x,t) measured by an interferometer, where x and t denote the position of the cords and time, respectively. The idea is to regard the reconstruction process as an image reconstruction, which is a DNN’s area of expertise. That is, a spatio temporal evolution of the line-integrated density n_e L(x,t) (equivalently power spectra (n_e L) ̃(x,f) and the phase spectra ϕ(x,f) ) is taken as an input image to DNN and the corresponding output is a dynamic image of the cross-sectional density distribution. As a first step, we develop and train DNNs using datasets produced by a numerical simulation of interferometer measurement for prepared various profiles of sheared rotating plasmas, whose density profiles are expressed as the following from: n_e (r,t)=∑_r〖∑_m〖Amp(r,m)cos⁡(mθ-ω(r,m)〗 t+θ_0 〗), where r,Amp(r,m), m, θ, ω(r,m), … are radial position, amplitude, the poloidal mode number, the poloidal angle, the angular frequency, …, respectively. m ranges 0 to 4. The following three data sets are prepared: (Group A) only a single m allows at all layers, in which Amp(r,m)≠0 in all radial layers for m=m_single, and Amp(r,m)=0 for m≠m_single, where m_single is a selected mode number as the single mode. (Group B) only a single m is allowed at each radial layer, and respective m can have different values in which Amp=Amp(r,m_single (r))≠0 only for m=m_single, (Group C) multiple m is allowed at each radial layer. The DNN trained by the data set (Group A) shows the best performance among the three cases, mean absolute error (MSE) lower than 1%, and the profiles of oscillation frequency of the density perturbation with MSE lower than 500 Hz (10% of the maximum frequency). The other two cases (Group B) and (Group C) could not show reconstruction results with acceptable error ranges.

[1] A.E. Costley 2016 Nucl. Fusion 56 066003
[2] ITER Disruption Mitigation Workshop, ITER HQ, 8 – 10 March 2017
[3] J.A. Casey, E. Sevillano, J.H. Irby, and B.G. Lane, “A Pseudo-Tomographic Fitting Algorithm for Density Profile Reconstruction from a Sparse 1-D Interferometer Array”, Review of Scientific Instruments, (1987).
[4] Yoshio Nagayama, “Tomography of m=1 mode structure in tokamak plasma using least-square-fitting method and Fourier-Bessel expansion”, Journal of Applied Physics 62, 2702 (1987).
[5] A logical calculus of the ideas immanent in nervous activity. Warren S. McCulloch & Walter Pitts. The bulletin of mathematical biophysics volume 5, pages115–133 (1943)
[6] Gradient Theory of Optimal Flight Paths, HENRY J. KELLEY, Published Online:6 Jun 2012
[7] Diogo R. Ferreira, ORCID Icon, Pedro J. Carvalho ORCID Icon, Horácio Fernandes ORCID Icon & JET Contributors, Pages 47-56 | Received 28 Jun 2017, Accepted 25 Sep 2017, Published online: 02 Feb 2018
[8] J. Santos, F. Nunes, M. Manso ,and I. Nunes, Review of Scientific Instruments 70, 521 (1999)