Abstract Drilling complications, which are usually hard to be discovered in time using the traditional surface detecting methods, result in much time and money wasted in handling these problems. Restricted to data transmission speed with the measurement while drilling (MWD), downhole measured data is usually ignored in downhole complications detection. And the surface detection methods with some pressure and rate of flow sensors always demand much professional knowledge and contain detection delay. In this paper, we used the measured downhole parameters to discover the drilling complications combined with deep leaning methods. Firstly, we described the difficulties of applying deep learning methods into the exploring drilling data. Then we used wavelet decomposition and reconstruction method to reduce the influence of the data trend with well depth and remove the high frequency noise. The fluctuation items coupling analysis method, consisted with rock breaking theory and transient fluctuating pressure theory, was established to make sure whether the wavelet reconstruction results contain the information to do detection. We applied a deep learning method called Bidirectional Generative Adversarial Network (BiGAN) in complications detection. BiGAN can distinguish whether the data belongs to normal working condition data or not. An end to end deep neural network mainly composed with one dimensional convolutional neural network was established to determine the specific kind of normal working condition. Then, large numbers of real field drilling data collected by the measuring tool were used to test the detection method. The testing results indicated that BiGAN indeed learned the normal working condition data distribution and the end to end network performed high accuracy in the normal working conditions classification. Therefore, we chose the combination of BiGAN and the supervised neural network to detect drilling complications with six field cases. The experiment results showed that the detection method can detect the complications much earlier than the surface detection results except for nozzle clogging case.

中文翻译：

---

基于深度学习的井下工况分析及钻井并发症检测方法

摘要 使用传统的地面检测方法通常难以及时发现钻井并发症，导致在处理这些问题时浪费了大量的时间和金钱。受限于随钻测量（MWD）的数据传输速度，井下测量数据通常在井下并发症检测中被忽略。并且一些压力和流量传感器的表面检测方法总是需要大量的专业知识，并且存在检测延迟。在本文中，我们使用测量的井下参数结合深倾斜方法来发现钻井复杂性。首先，我们描述了将深度学习方法应用于勘探钻井数据的困难。然后我们使用小波分解和重构方法来减少数据趋势随井深的影响并去除高频噪声。建立了结合破岩理论和瞬态脉动压力理论的波动项耦合分析方法，以确定小波重构结果是否包含检测信息。我们在并发症检测中应用了一种称为双向生成对抗网络 (BiGAN) 的深度学习方法。BiGAN 可以区分数据是否属于正常工况数据。建立了以一维卷积神经网络为主的端到端深度神经网络，以确定正常工作状态的具体种类。然后，测量仪采集的大量现场钻井数据用于测试检测方法。测试结果表明，BiGAN确实学习到了正常工况数据分布，端到端网络在正常工况分类中表现出较高的准确率。因此，我们选择了 BiGAN 和监督神经网络的结合来检测六个现场案例的钻井并发症。实验结果表明，除了喷嘴堵塞情况外，该检测方法可以比表面检测结果更早地检测出并发症。测试结果表明，BiGAN确实学习到了正常工况数据分布，端到端网络在正常工况分类中表现出较高的准确率。因此，我们选择了 BiGAN 和监督神经网络的结合来检测六个现场案例的钻井并发症。实验结果表明，除了喷嘴堵塞情况外，该检测方法可以比表面检测结果更早地检测出并发症。测试结果表明，BiGAN确实学习到了正常工况数据分布，端到端网络在正常工况分类中表现出较高的准确率。因此，我们选择了 BiGAN 和监督神经网络的结合来检测六个现场案例的钻井并发症。实验结果表明，除了喷嘴堵塞情况外，该检测方法可以比表面检测结果更早地检测出并发症。