Abstract: Based on the deep learning method, a set of fishbone mode (FB) recognition algorithm was developed on the HL-2A tokamak. The algorithm uses 858 shot discharge data (780 for training and 78 for verification) and about 463800 data slices for training and verification, and obtains a convolutional neural network mainly composed of convolutional layer, residual connection layer and fully connected layer. In order to measure the recognition ability of the algorithm, the algorithm was used to scan and identify the 780 discharge data, and a total of 86820 FB slices were identified, of which 4327 were misidentified, and the false alarm rate was 4.75%. In the actual 97145 FB slices, 10325 slices were missed, and the corresponding false negative rate was 10.63%, and the overall recognition accuracy rate reached 94.26%. The false positive rate and false negative rate can meet the accuracy requirements of FB recognition.

Key words: Fishbone, Neural network, Deep learning, HL-2A tokamak

摘要： 基于深度学习的方法，在HL-2A装置上开发出了一套鱼骨模(FB)识别算法。算法使用858(780次放电训练，78次放电验证)次放电数据、约46.38万数据切片进行训练与验证，得到了一个主要由卷积层、残差连接层、全连接层组成的卷积神经网络。为衡量算法的识别能力，该算法被用来扫描式地识别了HL-2A装置的780次放电数据，共识别出86820次FB区间，其中误识别4327次，误报率为4.75%。在实际的97145次FB区间中，漏识别10325次，对应的漏报率为10.63%，总整体的识别正确率达到了94.26%。该误报率和漏报率可以满足FB识别的精度要求。

关键词: 鱼骨模, 神经网络, 深度学习, HL-2A装置