基于非线性分类网络的气体钻井风险智能识别方法

Intelligent Identification Method of Gas Drilling Risk Based on Nonlinear Classification Network

  • 摘要: 针对气体钻井领域现有的风险智能识别方法泛化性差,且在风险发生初期识别正确率偏低,导致风险正确识别时间点滞后的问题,对气体钻井随钻监测数据进行挖掘分析,并提出了一种适用于该场景下的改进卷积神经网络模型,用于钻井工况的风险识别。该模型采用卷积层提取多个随钻监测参数随时间变化的关联特征,采用具有非线性分类能力的RBF(radial basis function)网络替换全连接层进行特征分类,提升分类精度的同时,也将工况过渡阶段风险识别的时间点进一步提前。引入K均值算法对RBF聚类中心以及中心宽度值进行初始化,解决不合理初始值导致模型对工况样本训练无法收敛的问题。引入基于距离均方误差的损失函数分量,对整个网络模型进行训练来优化模型参数,以保证网络训练过程中能不断调整最佳聚类中心。制定了一种隐藏层节点数选取准则以提高模型的计算效率和泛化性。与传统卷积神经网络模型进行了现场应用对比,结果显示,该网络模型对产气、出水、卡钻三种常见风险识别时间点提前56 s、16 s、8 s,充分展现了非线性分类神经网络在油气田勘探开发领域的适用性。

     

    Abstract: The existing risk intelligent identification methods in the field of gas drilling have poor generalization, and the low recognition rate in the early stages of risk leads to a delay in the correct risk identification. In this study, we perform data mining analysis of gas drilling while drilling monitoring data and propose an improved convolutional neural network model to identify the risk of drilling conditions. In this model, we extract the time-varying correlation features of multiple monitoring parameters while drilling using the convolution layer. We also replace the full connection layer for feature classification by using a radial basis function (RBF) network with nonlinear classification ability. The use of RBF improves classification accuracy and reduces the time of risk identification in the transitional stage of working conditions. Since the model cannot converge to the working condition sample training due to an unreasonable initial value, we introduce the K-means algorithm to initialize the RBF clustering center and center width to resolve this issue. We introduce the loss function component based on the distance mean square error to train the whole network model. The loss function also allows the optimization of model parameters, ensuring the continuous adjustment of the optimal clustering center in the process of network training. We also formulate a criterion for selecting the number of hidden layer nodes so as to improve calculation efficiency and facilitate the generalization of the model. The network model is then compared with the traditional convolution neural network model for field application. Our experimental results show that the identification time points of three common risks, i.e., gas production, water production, and sticking drill, are advanced by 56 s, 16 s and 8 s when using our network model. This finding fully demonstrates the applicability of nonlinear classification neural network in the field of oil and gas field exploration and development.

     

/

返回文章
返回