甘肃塌陷地震震相识别研究

选取甘肃省测震台网测定的甘肃平凉地区塌陷地震、甘东南地区天然地震各50个,同地区的30个地震事件为待测事件。采用近年来计算机领域里较先进的图像识别方法——卷积神经网络识别两种地震事件类型,设定波形通道总数的80%为地震事件分类阈值,超过分类阈值的设定为“0”或“1”,即可判定该待测事件为对应的事件类型。将地震事件作为原始图像,提取图像中最具代表性特征点;接着提取图像特征进入池化层,池化层会对该图像特征点进行归类压缩,提取最具代表性的图像特征,最后输出识别图像。结果表明:塌陷地震最终分类准确识别率为86.7%,天然地震准确识别率为93.3%,总识别率为90%,为今后平凉地区塌陷地震事件类别识别工作提供了可靠的参考价值。

Seismic Phase Recognition of Collapse Earthquakes in Gansu Area

Earthquake events can be divided into natural earthquakes and non-natural earthquakes according to the generating mechanism. Collapse earthquakes belong to the common earthquake events in non-natural earthquakes. In this paper, we selected 50 collapse earthquakes in Pingliang area of Gansu Province and 50 natural earthquakes in southeastern Gansu Province, and 30 earthquakes to be measured in the same area. The convolution neural network, an advanced image recognition method in computer field in recent years, was selected to identify two types of seismic events. 80% of the total number of waveform channels was set as the classification threshold of seismic events, that is, more than 80% of the waveform channels were "0" or "1". Seismic events were taken as the original images, and the most representative feature points in the images were extracted. Then the feature points were classified and compressed by the pooling layer, and the most representative image features were extracted to output the recognition image. The results showed that final classification accuracy of collapse earthquakes is 86.7% and that of natural earthquakes is 93.3%, and the total recognition rate is 90%. The results can provide a reliable reference value for future classification of collapse earthquakes in Pingliang area.