In the present social and economic climate, Exploration and Production (E&P) companies are under pressure to embrace low-cost digital technology that offers a dual benefit of lowering well construction capital expenditures, as presented in Fig. 1, by reducing Non-Productive Time (NPT) and Invisible Lost Time (ILT), while also aiding in meeting Environmental, Social and Governance (ESG) objectives.
One of the main sources of NPT in well construction is stuck pipe, as such incidents represent one of the most challenging and costly problems in drilling operations. In these cases, the drillstring becomes immovable in the borehole caused by different mechanisms such as differential or mechanical sticking and packing off. Such incidents can lead to not only NPT but also potential loss of equipment, hence necessitating expensive technical detours that prioritize projects over authorization for expenditure (AFE). Furthermore, stuck pipe events can significantly affect the safety and wellbore integrity. While approaches to identify stuck pipe scenarios retrospectively are well documented, this paper describes a data-driven approach to identify and predict such events at the edge in real time using machine learning.
Nowadays, the constant development of digital technologies, such as digital twins modeling behavior of the wellbore and equipment in real-time (Arévalo et al. 2021), drilling automation applications aiding in the monitoring of wellbore condition (Ait Ali et al. 2023, Arévalo et al. 2023, Arévalo et al. 2022), among others, have prepared the ground for a more accurate detection of downhole hazards, such as conditions leading to stuck pipe. This paper presents the utilization of machine learning (ML) to analyze surface data, detect stuck pipe events and classify such events into categories. The approach consists of a fuzzy logic algorithm to categorize acquired surface data and a real-time classifier to be trained. The score calculated by the fuzzy algorithm is mapped into five categories, which serve as labels for training validation and test data. A Long Short-Term Memory (LSTM) and a decision tree are trained by a sequence-to-label prediction approach to predict whether the drillstring will operate unimpeded or whether higher drag or even a stuck pipe event is expected.
