Abstract: A machine learning method has been adopted to train models for H-mode prediction from the HL-2A experimental data. Stored energy and deuterium alpha signals from 206 discharges are selected for training, 15 for validation and 474 for testing. H-mode period of all the data is marked previously. Instances of 2-channel time sequence are generated in the way of time window. Finally, a 21 layers convolutional neural network was developed. It scored 98.17% of accuracy in test set and the recognized H-mode beginning and ending time error were lower than 10ms, which can be applied to the high confinement mode research.

Key words: HL-2A tokamak, High confinement mode, Convolutional neural network, Pattern recognition

摘要： 基于HL-2A装置的放电实验数据，利用卷积神经网络和时间窗口算法开发了高约束(H)模时段的识别算法，得到了可靠的高成功率的高约束模时段识别结果。算法中，选取206次放电实验数据中等离子体储能及氘a通道信号作为双通道原始数据进行学习，得到一个深度为21层的二分类卷积神经网络。该网络模型经过其他474次放电数据的测试集检验，高约束模识别的正确率达到了98.17%。

关键词: HL-2A装置, 高约束模, 卷积神经网络, 模式识别