Abstract During industrial offshore deep-water drilling process, gas kick event occurs frequently due to extremely narrow Mud Weight (MW) window (minimum 0.01sg) and negligible safety margins for the well control purposes. Further, traditional gas kick detection methods in such environments have significant time-lag and can often lead to severe well control issues, and occasionally to well blowouts or borehole abandonment. In this study, firstly, the raw field data is processed through data collection, data cleaning, feature scaling, outlier detection, data labeling and dataset splitting. Additionally, a novel data labeling criterion for gas kick risks is proposed where five kick risks (Indicated by different colors in this study) are defined based on three key indicators: differential flow out (DFO), kick gain volume (Vol), and kick duration time (Time). Kick risk status represents one of the following cases: Case 0 - No indicators are activated (Green), Case 1 - Multi-drilling parameters deviation or DFO is activated (Orange), Case 2 - DFO and Vol are simultaneously activated (Light Red), Case 3 - DFO and Time are simultaneously activated (Light Red), Case 4 - DFO, Vol and Time alarms are simultaneously activated (Dark Red). Then, a novel data mining method using Long Short-Term Memory (LSTM) Recurrent Neural Network (RNN) is presented for early detection of gas kick events by analyzing time series data from field drilling process. The network parameters such as number of hidden layers and number of neurons are initialized to build the LSTM network. The learned LSTM model is evaluated using the testing set, and the best LSTM model (six (6)-layers eighty (80)-nodes (6 L*80 N)) is optimally selected and deployed. The accuracy of deployed LSTM model is 87 % in the testing dataset, which is reliable enough to identify the kick fault during the deep-water drilling field operation. Lastly, the LSTM model detected the gas kick events earlier than the “Tank Volume” detection method in several representative case studies to conclude that the application of LSTM model can potentially improve well control safety in the deep-water wells with narrow MW windows.

中文翻译：

---

基于监督学习算法的工业深水钻井过程气爆现场数据分析及风险评估

摘要 在工业海上深水钻井过程中，由于极窄的泥浆重量（MW）窗口（最小0.01sg）和井控安全裕度可忽略不计，气涌事件频繁发生。此外，在这种环境中，传统的气涌检测方法具有明显的时滞性，通常会导致严重的井控问题，偶尔会导致井喷或钻孔废弃。在本研究中，首先通过数据采集、数据清洗、特征缩放、异常值检测、数据标记和数据集拆分对原始现场数据进行处理。此外，还提出了一种新的气涌风险数据标记标准，其中基于三个关键指标定义了五个井涌风险（在本研究中用不同颜色表示）：差异流出 (DFO)、井涌增益量 (Vol)、和踢球持续时间（时间）。踢腿风险状态代表以下情况之一：案例 0 - 未激活任何指标（绿色），案例 1 - 多钻孔参数偏差或 DFO 已激活（橙色），案例 2 - DFO 和 Vol 同时激活（浅红色） , Case 3 - DFO 和 Time 同时激活（浅红色），Case 4 - DFO、Vol 和 Time 警报同时激活（深红色）。然后，提出了一种使用长短期记忆 (LSTM) 循环神经网络 (RNN) 的新型数据挖掘方法，通过分析现场钻井过程中的时间序列数据来早期检测气涌事件。初始化隐藏层数和神经元数等网络参数以构建 LSTM 网络。使用测试集评估学习的 LSTM 模型，并优化选择和部署最佳 LSTM 模型（六 (6) 层八十 (80) 个节点 (6 L*80 N)）。部署的 LSTM 模型在测试数据集中的准确率为 87%，足以可靠地识别深水钻井现场作业期间的井涌故障。最后，在几个代表性案例研究中，LSTM 模型比“储罐容积”检测方法更早地检测到气涌事件，得出结论，LSTM 模型的应用可以潜在地提高具有窄 MW 窗口的深水井的井控安全性。