Hydraulic drilling operation state recognition under complex coal–rock structures based on mechanism–data fusion

Abstract Under complex coal–rock formations and variable load conditions, accurate recognition of drilling operation states is essential for the safety and efficiency of hydraulic drilling. To overcome the limited robustness of purely mechanistic or purely data-driven approaches in strongly nonlinear drilling environments, this study proposes a hybrid framework that integrates a physics-based mechanistic drilling dynamics model with a Transformer–long short-term memory data-driven model under a Bayesian inference scheme. The mechanistic model is established through analyses of drill bit–coal–rock interactions and drill string dynamics, while key parameters are estimated using an improved INIR-Harris hawks optimization algorithm, supporting reliable characterization of drilling instability states. The hybrid model is validated on a multivariate time-series dataset under different drilling types and geological conditions, achieving over 98% recognition accuracy with a false alarm rate (FAR) below 1.5% across multiple drilling types, and over 97% accuracy with a FAR below 2% under varying geological conditions. These results demonstrate the effectiveness, robustness, and engineering applicability of the proposed method for hydraulic drilling state monitoring.