基于长短期记忆神经网络的深水钻井工况实时智能判别模型

深水钻井具有高投入、高风险等特点，其工况实时判别是提高钻井时效、减少复杂事故的基础和前提。传统深水钻井作业中，钻井工况主要通过基于编程方式的物理模型与经验模型进行判别，难以保证时效性和正确率。为此，创新性地将机器学习引入深水钻井工况判别全流程，考虑综合录井数据的长时间序列特征，基于长短期记忆神经网络建立了深水钻井工况实时智能判别机器学习模型。通过对29 856 140行深水综合录井数据预处理，选取钻头深度、井深、大钩高度、钻压、悬重、扭矩、转速、立管压力，共计8个综合录井参数作为输入特征，建立了20隐藏层×70节点的长短期记忆神经网络模型，实现了旋转钻进、滑动钻进、接单根、静止、循环、向下洗井、划眼、向上洗井、倒划眼、起钻、下钻及“其他”，共计12种常见深水钻井工况的实时智能判别，测试集上正确率高达94.09%，满足深水现场作业需求。该模型可实时智能地判别钻井工况，充分验证了长短期记忆神经网络用于钻井工况实时智能判别的可行性与时效性，为钻井时效分析和复杂事故预警提供了机器学习模型基础，并将进一步拓展机器学习在石油工程领域的应用范围。

Real-time intelligent rig activities classification model of deep-water drilling using Long Short-Term Memory (LSTM) network

Deepwater drilling has difficulties such as high investment and high risks. Real-time intelligent rig activities classification of deep-water drilling is the basis and premise of improving drilling efficiency and reducing complex accidents. In traditional deep-water drilling field operation, the rig activities are mainly classified by physical models and empirical models based on programming, which is difficult to ensure the timeliness and accuracy. Therefore, the method of machine learning is innovatively introduced into the whole process of deep-water drilling rig activities classification in this paper. Considering the long time series characteristics of comprehensive mud-logging data, a real-time intelligent machine learning model for deep-water drilling rig activities classification was established based on the Long Short-Term Memory (LSTM) network. After the preprocessing of 29,856,140 lines of deep-water comprehensive mud-logging data, there were eight comprehensive mud-logging parameters selected as input features, including Depth of drill bit in real-time (DBTM), Measured depth of hole (DMEA), Hook Height (HKH), Weight on Bit (WOB), Weight on Hook (WOH), TORQUE, Rate per Minute (RPM) and Standpipe Pressure (SPP). Then, a LSTM network model with 20 hidden layers ×70 nodes was established. It has realized the real-time intelligent classification of 12 common deep-water drilling rig activities, including rotary drilling, slide drilling, stand connection, static, circulate, wash down, ream, wash up, backream, trip out, trip in, and other. The accuracy of machine learning model in the testing dataset is as high as 94.09%, which meets the requirements of deep-water field operations. The rig activities were intelligently classified in real-time, which fully verifies the feasibility and timeliness of the LSTM network for the real-time intelligent classification of rig activities. Furthermore, the real-time intelligent method provides the machine learning models basis for the drilling efficiency analysis and complex accidents warning, which will expand the application range of machine learning in the petroleum engineering.