Incident Detection Adapting to the Drilling Depth for Geological Drilling Processes Based on Domain Adversarial Dual Graph Convolutional Network

In geological drilling processes, it is of great importance to detect drilling incidents, so as to prevent serious consequences and improve the operational safety. However, prompt and accurate detection of drilling incidents is quite challenging, since the difference between data samples under normal and faulty states are usually inappreciable in the early stage of a drilling incident. Meanwhile, the formation commonly changes with the depth, making a model trained based on data from a certain depth hardly adapt to incident detection at different depths. Accordingly, this article proposes a new incident detection method adapting to the drilling depth for geological drilling processes based on the domain adversarial dual graph convolutional network (DADGCN). The contributions are threefold: 1) a temporal-spatial multifeature graph (TSMFG) construction method is designed to excavate the difference of adjacent samples under multiple states; 2) a dual graph convolution network (DGCN)-based incident detection method is proposed to mine the deep features in the graphs; 3) a DADGCN framework is designed to generalize the well-trained incident detection model to different drilling depths. The effectiveness and superiority of the proposed method are demonstrated by case studies involving real data. According to the results, the incident detection accuracy of DGCN reaches 99% under the same drilling depth, and the accuracy of DADGCN exceeds 94% when adapting to different drilling depths, which are much better compared to other state-of-the-art methods.