Combined ML- and Physics-Based Techniques Achieve Real-Time Stick/Slip Mitigation

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper IPTC 24509, “Real-Time Stick/Slip Mitigation Using Combined Machine-Learning and Physics-Based Techniques,” by Hana Yahia, SPE, Mines Paris Tech and Helmerich and Payne, and Thomas Romary and Laurent Gerbaud, Helmerich and Payne, et al. The paper has not been peer reviewed. Copyright 2024 International Petroleum Technology Conference. _ In recent years, artificial intelligence (AI) -based detection algorithms increasingly have been proposed in the literature for diagnosis of torsional stick/slip vibration. One drawback of AI methods, however, is their limited applicability to wells in the same field with similar geological formations and bottomhole assemblies (BHAs). The complete paper presents trained and tested machine-learning (ML) models capable of predicting the stick/slip severity index (SSI) using sequences of surface measurements and physical features. The paper also proposes real-time auto-updated stability heat maps to help drillers choose optimal drilling parameters to avoid stick/slip vibrations. Modeling Approach In this section of the complete paper, the authors introduce a regression model aimed at predicting a continuous nondiscrete SSI. The model takes as input time-series data of drilling surface measurements, including surface torque, surface weight on bit (WOB), rate of penetration (ROP), flow rate, and total rotation speed. These measurements were selected based on their relevance for addressing the stick/slip problem and their widespread availability in most wells. The data were sampled at a consistent frequency of 1 Hz for practical reasons. To label the data, downhole information was collected with a measurement subsystem installed along the BHA. This downhole data included bit-rotation speed and torque on bit (TOB). These data were collected from wells with different sections belonging to distinct drilling fields. To align the downhole data with the surface data, a time shift was estimated using cross-correlation, a method that measures the similarity between two signals by assessing their alignment over time. The resulting continuous data set is divided into 60-second windows. Each window contains multichannel surface data and serves as an input sample for the regression model. To label the corresponding surface data for each window, the downhole bit rotation speed is inspected. For each sequence of windows, an SSI is calculated. Model Characteristics Long Short-Term Memory (LSTM) Model. In this work, an ML model architecture is used to predict continuous SSI values from surface sequence measurements. This architecture comprises a series of LSTM layers. LSTM operates in a way similar to a recurrent neural network, where the output from a previous step is recycled as input for the subsequent step. Notably, LSTM possesses an internal state that acts as a memory space, enabling the storage and retrieval of information across multiple timesteps. This means that the input value, previous output, and internal state collectively contribute to predicting the current value.