Abstract: In order to realize the rapid and efficient identification of earthquakes, blasting and collapse events, this paper applies the Convolutional Neural Network(CNN)in deep learning technology to design a deep learning training module based on single station waveform recording of single event and a real-time test module based on multiple stations waveform recording of single event.
On the basis of ensuring that the data is comprehensive, objective and original, the three-component waveforms of the first five stations that recorded the P-wave arrival time of each event are input, and the current mainstream convolutional neural network structures are used for learning test. The four main convolutional neural network structures of AlexNet, VGG16, VGG19 and GoogLeNet are used for learning training, and the learning effects of different network structures are compared and analyzed. The results show that in the training process of various convolutional neural network structures, the accuracy rate and the cost function curve of the training set and the test set of each network are basically the same. The accuracy rate increases gradually with the increase of the training times and exceeds 90%, and finally stabilizes around a certain value. The cost function curve decreases rapidly with the increase of the training times, and eventually the stability does not change near a relatively small value. At the same time, over-fitting occurred in all convolutional neural network structures during training, except for AlexNet. In the end, the cost function of each type of structural training set and test set is finally lower than 0.194, and the recognition accuracy of each type of structure for training sets and test sets is over 93%. Among them, the recognition accuracy of AlexNet network structure is the highest, the accuracy of the training set of AlexNet network structure is as high as 100%, the test set is 98.51%, and no overfitting occurred; the accuracy of VGG16 and VGG19 network structure comes second, and the recognition accuracy of GoogLeNet network structure is relatively low, and the trend curves of the accuracy and cost function in training and test set of each network in the training process are basically the same. Subsequently, in order to test the event discrimination efficiency of the CNN in deep learning in the real-time operation of the digital seismic network, we select the trained AlexNet convolutional neural network to perform event type determination test based on the waveform recording of multiple stations of a single event. The final result shows that the types of a total of 89 events are accurately identified in the 110 events with M ≥0.7 recorded by Shandong seismic network, and the accuracy rate is about 80.9%. Among them, the accuracy rate of natural earthquake is about 74.6%, that of explosion is about 90.9%, and that of collapse is 100%. The recognition accuracy of collapse and explosion events is relatively high, and it basically reaches or exceeds the recognition accuracy of manual determination in the daily work of the seismic network. The accuracy of natural earthquake identification is relatively low. Among the 18 misidentified natural earthquakes, up to 13 events were judged as blasting or difficult to identify due to distortion of waveforms recorded by some stations(They are determined to be explosion and earthquake each by the records of two of the five stations). If sloughing off the recognition type error events caused by waveform distortion due to the background noise interference that overwhelms the real event waveform or waveform drift, the recognition accuracy of earthquake will become 91.4%, and the recognition accuracy of all events will increase from 80.9%to 91.7%, which is basically equivalent to the recognition accuracy of manual judgment in the daily work of the seismic network. This indicates that deep learning can quickly and efficiently realize the type identification of earthquake, blasting and collapse events.

Key words: deep learning, non-natural earthquake, automatic identification

摘要： 为实现天然地震与爆破、 塌陷事件类型的快速高效识别, 文中应用深度学习技术中的卷积神经网络模型, 设计了基于单个事件单个台站波形记录的深度学习训练模块和基于单个事件多个台站波形记录的实时测试模块。 以每个事件P波到时最早的5个台站记录到的原始三分向波形为输入, 分别采用目前主流的AlexNet、 VGG16、 VGG19、 GoogLeNet 4种卷积神经网络结构进行学习训练, 结果显示各类卷积神经网络结构对训练集与测试集的识别准确率均达93%以上, 且各个网络在训练过程中的训练集与测试集的准确率及代价函数的走势曲线基本一致。 其中, AlexNet网络结构的识别准确率最高, 测试集为98.51%, 且未发生过拟合现象; VGG16、 VGG19网络结构的准确率次之; GoogLeNet网络结构的识别准确率相对较低。 为检验深度学习卷积神经网络在数字地震台网实时运行过程中的事件判别效能, 选取训练好的AlexNet卷积神经网络开展基于单个事件多个台站波形记录的事件类型判定检验。 最终结果显示, 在山东台网实时触发的110个M≥0.7事件中, 共有89个事件的类型被准确识别, 准确率约为80.9%。 具体到各个类型事件中, 天然地震的准确率约为74.6%; 爆破的准确率约为90.9%; 塌陷事件的准确率为100%。 若删除其中由于波形失真而造成的类型识别错误事件, 则天然地震的识别准确率将提高至91.4%, 而所有事件的整体识别准确率也将由80.9%提高至91.7%, 与目前地震台网日常工作中人工判定的识别准确率基本相当。 这表明, 深度学习技术可以快速高效地实现天然地震与爆破、 塌陷的事件类型识别。

关键词: 深度学习, 非天然地震, 自动识别