基于卷积神经网络的压裂微震实时监测参数优选研究

Microseismic real-time monitoring is an effective tool for evaluating hydraulic fracturing. Due to kinds of factors, it is hard to accurately estimate the quality of monitoring record before fracturing working. For real-time processing of monitoring record, such as microseismic event recognition and microseismic source locating, it is necessary to select processing parameters according to the quality level of monitoring record. The existing methods are fail to realize automatic selection of processing parameters. For these methods, manual intervention is necessary in this situation. But it is not conducive to real-time monitoring. A method is proposed to solve this problem in this fund application. The proposed method is based on convolutional neural networks. Using existing monitoring record, different levels of microseismic events can be obtained as training samples of network weights. Then, the model can be created to predict microseismic event type in real-time record. According to predicting result, it is easy to select suitable processing parameters. So, the accurate level of event recognition and processing efficiency can be greatly improved during the real-time monitoring.

微震实时监测是评估压裂效果的有效方法。因监测记录品质受多种因素的影响，很难提前对实时记录的品质做出准确估计。对监测记录的实时处理，如微震事件识别与微震震源定位，应由实时记录的特征动态选择的处理参数。现有方法都很难实现这一点，往往需要人工对处理参数进行调整，这不利于压裂作业的实时监测。本项研究提出使用卷积神经网络来解决此问题。在已有的监测记录中，提取出大量的不同级别微震事件作为训练样本，对神经网络的权重进行训练，构建微震事件的识别和处理模型，使用该模型对实时监测数据进行预测分析，依据预测结果选配相应的处理参数，以此提高实时监测的识别准确度和处理效率。

微震实时监测是评估压裂效果的有效方法。现场采集到的数据品质会受多种因素的影响，很难提前对监测数据品质做出准确估计。相应的实时处理，特别是微震事件识别，需要根据实时记录的特征对处理参数进行动态调节，这不利于对压裂作业进行实时监测。本项研究提出使用卷积神经网络来解决此问题。首先，根据微震数据的特点，将前期采集到的微震监测数据样本分成高信噪比微震事件、中低信噪比微震事件和噪声干扰三类。根据不同信噪比水平，就可提前设定好对应的处理方法与参数。在以往监测数据的基础上，共收集到5209个高信噪比事件，1672个低信噪比事件和5246个干扰信号。并采用合成数据集的方法解决样本类别数据不均衡问题。在建立好的样本数据集中，随机抽取出30%的样本作为测试数据集。搭建的卷积神经网络模型包含一个输入层，两个卷积层和两个池化层以及一个全连接层和一个输出层。经过测试数据集的测试，相比于三种常规的事件识别算法，本项研究采用卷积神经网络算法具有更高的准确率和更高效率。同时，对微震监测的相关问题开展研究，例如相对到时的估算、速度模型的快速校正和微震监测解释，并取得了相应的成果。