改进双向LSTM的地震震相拾取算法

Seismic Phase Picking Algorithm Based on Improved Bi-LSTM

为了更好地提升识别精度和算法的适应性,针对传统波形检测算法存在依赖人工设定阈值以及震相拾取的精度偏低等问题,提出了一种基于改进双向LSTM的震相拾取算法。该方法利用滤波、归一化、加噪等方式对原始波形数据进行预处理,通过双向LSTM模型对地震信号数据进行自适应深层特征提取;同时为解决双向LSTM模型存在的过拟合问题,引入Spatial-Dropout机制在随机区域对双向LSTM模型进行稀疏性约束;最后引入Time-Distributed机制从时域维度针对事件-噪声的二分类问题自动判定P波到时,并在相关地震数据集上进行对比实验。结果表明:模型对P波拾取的准确率达90%,P波拾取的精确度达80%,与LSTM和GRU等传统RNN网络模型相比,精确度分别提高了6%和5%.同时,该模型不需要人工设定阈值,并且对复杂波形数据的P波拾取问题具有较强的鲁棒性。

In order to solve the problem of traditional waveform detection algorithms such as relying on manual setting of thresholds and low precision of phase picking,a seismic phase picking algorithm based on improved Bi-LSTM was proposed.This method uses filtering,normalization,and noise addition to preprocess the original waveform data,and uses the Bi-LSTM model to perform adaptive deep feature extraction of the seismic signal.Meanwhile,in order to solve the overfitting of the Bi-LSTM model,Spatial-Dropout mechanism is introduced to restrict the sparsity of the Bi-LSTM model in random area,and finally the Time-Distributed mechanism is introducted to automatically determine the arrival time of P-wave from the time domain dimension for the event-noise binary classification problem.Comparative experiments on relevant seismic data sets show that the precision and accuracy of the P-wave pickup of this model is 90%,80%.Compared with traditional RNN network models such as LSTM and GRU,present algorithm gives a precision increased by 6%and 5%respectively.At the same time,the model does not need to manually set the threshold,and has strong robustness to the problem of P-wave pickup of complex waveform data.