Intelligent Micro-Kick Detection Using a Multi-Head Self-Attention Network

Accurate micro-kick detection is crucial for blowout accident preventions. The more drilling parameters that change due to kicks, the more accurate the warning results become. However, when the micro-kick occurs, there is a significant time lag between these parameter changes. Dominant kick detection methods based on long short-term memory (LSTM) forget early parameter trends when dealing with long time series. To improve the recognition accuracy of micro-kicks and avoid potential blowout accidents by memorizing the early or long-term trends in drilling parameters, an intelligent micro-kick detection method based on a multi-head self-attention network is proposed. First, a novel multi-head structure is designed to separate various types of features due to different monitoring parameter changes at different speeds or trends. Second, a self-attention mechanism is employed to focus on parameter changes in separated monitoring data sequences. Then, a feed-forward network with parallel computation capability is utilized to analyze long-range correlations, thus avoiding the loss of early or long-term trend information. Finally, an artificial neural network is used to establish nonlinear relationship models between the trend features of each monitoring parameter and kick accidents. The experiment results demonstrate that the recognition accuracy of the proposed micro-kick detection method is 7.9% higher than that of the LSTM-based method.