Prediction of Unusual Emission of Visible Light in Plasma Experiments by using Machine Learning
In plasma experiments at the Large Helical Device (LHD), unusual emission of visible light inside the plasma vessel has been observed. This emission must be predicted to avoid unexpected damages on the plasma vessel. This research experimentally shows that the unusual emission of visible light inside the plasma vessel can be predicted using a Support Vector Machine, a machine learning method.
In plasma experiments at the Large Helical Device (LHD), an unusual emission of visible light inside the plasma vessel has been observed. This emission must be predicted to avoid unexpected damage to the plasma vessel. This research experimentally shows that the unusual emission of visible light inside the plasma vessel can be predicted using a Support Vector Machine, a machine learning method.
This research proposes a method for predicting an unusual emission of visible light inside the plasma vessel by using a Support Vector Machine (SVM), a machine learning method.
A video is a sequence of pictures, which are called frames. The luminance value of a frame is used as a feature value. The width and height of the original video frame are 352 and 240 pixels, respectively. The unique number of the video and the elapsed time are included at the top and bottom of the video, respectively. As these are not related to the plasma phenomenon itself, we removed them and set the frame to a size of 256x128 pixels. The frame was converted to a grayscale image and divided into 64x64 blocks. This means that one block comprises four pixels in width and two pixels in height. By adopting the mean value of the luminance values of the pixels in a block as a pixel value, a picture of 64x64 pixels was obtained. A vector of 4,096 luminance values was obtained from the 64x64 pixel picture. Each luminance value was normalized to zero to one by being divided by 255, the maximum luminance value. This was used as an input of a SVM.
The videos used were obtained from the plasma experiment in the 10th experimental campaign. This study used 199 videos with an unusual emission and 254 without. Moreover, six videos with an unusual emission and six without were prepared for evaluating the prediction performance. The duration of videos is within a range of a few seconds to several tens of seconds. A non-luminous frame was considered as inappropriate to be included in the dataset. Thus, frames with a mean luminance value less than 40 were excluded. A total of 27,681 frames were obtained as the dataset. Of them, 7,968 frames included an unusual emission (positive examples), while 19,713 frames did not (negative examples). For machine learning, the dataset was divided 8:2 for the training and test datasets.
The SVM was constructed using a Python machine learning library scikit-learn.
In order to examine the performance of the predictor, the probability of an unusual emission was calculated to judge whether a frame includes an unusual emission or not for 12 evaluation videos. As a result, it shows that an unusual emission of visible light is predicted before it occurs. An unusual emission is seldom reported for the plasma videos without such a phenomenon. More than 96% of accuracy, 93% precision, and 93% recall were attained, which means that the prediction performance was very high. These results proved that the SVM is effective in predicting the unusual emission.
As a few unusual emissions were not reported as such, an improvement of the prediction performance is required.
This research was conducted by Shota Nakagawa, Teruhisa Hochin, and Hiroki Nomiya from Kyoto Institute of Technology, and Hideya Nakanishi and Mamoru Shoji.
This research result was published on June 2021 in Fusion Engineering and Design, an international journal for fusion energy and technology.