基于样本增强的卷积神经网络震相拾取方法

为了快速、高效地从地震数据中识别地震事件和拾取震相，本文利用基于样本增强的卷积神经网络自动震相拾取方法，将西藏林芝地区L0230台站3个月数据作为训练集，该区内另外6个台站连续1个月的波形数据作为测试集，采用高斯噪声、随机噪声拼接、随机挑选噪声、随机截取地震事件等4种样本增强的方法扩增训练集，以提高自动震相拾取技术的准确率。结果显示:样本增强前模型在测试集上的地震事件识别准确率为80%，样本增强后提升至97%，表明样本增强有效地提高了模型的泛化性能和抗干扰能力；在0.5 s误差范围内，震相自动拾取准确率高于81%，在1.0 s误差范围内，准确率高于95%；利用基于样本增强的卷积神经网络震相拾取方法能够检测出人工拾取震相中误标和漏检的震相。 

Seismic phase identification using the convolutional neural networks based on sample enhancement

In order to quickly and efficiently identify earthquake events and pick up seismic phases from seismic data, this paper a small sample enhancement-based automatic phase picking method based on a convolutional neural network. A total of three months of data from the L0230 station in Linzhi, Tibet were used as a training set, and one-month continuous waveform data from the other six stations in this area were used as the test sets. Gaussian noise, random noise splicing, random selection of noise and random interception of seismic events were used to enhance the training set so as to improve the accuracy. The results show that the accuracy of the model on the test set was 80% before the samples were enhanced, and up to 97% after that, suggesting that the sample enhancement effectively improves the generalization perfor-mance and anti-interference ability of the model. Within the 0.5 s error range, the accuracy of the seismic phase automatic picking is higher than 81%, and within the 1.0 s error range, the accuracy is higher than 95%. Moreover, it is able to detect mislabeled and missing seismic phases in manually picking seismic phases by this method.