Early Intelligent Warning of Stuck Pipe Events Based on the Fusion of Machine Learning and Rule-Based Models

The timely and effective early warning for stuck pipe events during drilling operations is crucial for reducing non-productive time and ensuring drilling progress. Existing machine learning methods mainly focus on identifying the risk of stuck pipe events, but do not involve early warning. Due to limited sample sizes, it is difficult to establish effective feature engineering methods, resulting in high false alarm rates for stuck pipe events. This paper proposes a method that combines machine learning with rule-based models. We collected 641 stuck pipe cases from 114 wells in Sinopec’s Northwest Field Area and performs correlation analysis between 42 drilling parameters and potential risks. Rule-based model is incorporated to enhance the extraction of input feature values, such as the frequency and amplitude of hook load, to improve both the quantity and quality of feature values. Sliding window method with dynamic window sizes and step sizes are applied to the features for the 5-min period before and after the risk event. In order to identify early signs before the occurrence of risks, a weighted method that uses the inverse of the time interval from the risk event as the weight is applied in each sliding window to enhance the sensitivity of the deep convolutional neural network to the timing of risk events. Compared to existing machine learning methods, the results show that this approach can automatically detect early warning signals 1–2 min before the stuck pipe event occurs, with an F1 score higher than 0.85 (It also has a lower false alarm rate and missed detection rate),and can automatically generate a three-level warning system.