Predicting Drill Stuck Risk Using Conditional Boundary Equilibrium GAN in Mining Drilling Tools

Drill sticking incidents pose a serious threat to coal mine safety and production efficiency. The intelligence level of drilling rigs is low, and the complex coal-rock structure makes drill tools prone to becoming stuck. Traditional methods predict drill sticking through the manual design of risk factors, but a single empirical formula fails to accommodate diverse drill tool types and dynamic processes, thus limiting its practical applicability. While data-driven methods leverage operational signals for qualitative diagnosis, they fail to develop quantitative metrics for dynamic risk assessment. Although data-driven methods leverage operational signals for qualitative diagnosis, they lack dynamic risk quantification metrics. This paper introduces the KI-CBEGAN model, built upon the Conditional Boundary Equilibrium Generative Adversarial Network (CBEGAN), to predict the drill sticking factor (KI). The generator network employs a conditional input mechanism, utilizing historical signal sequences as constraints to predict future $t_{p}$-step operational signal vectors of drilling tools. By mapping the multi-step prediction vectors to the historical feature space, it effectively reduces the complexity of temporal prediction. The discriminator network simultaneously processes historical signals and the generator's multi-step predicted signal sequences. Leveraging deep networks' adaptive temporal feature extraction capabilities, it constructs a dynamic risk quantification model. The two components operate collaboratively: the generator outputs predicted values of operational signals for future $t_{\mathrm{p}}$ steps, while the discriminator integrates historical and predicted signals. Through the boundary equilibrium adversarial mechanism, it calculates the drill sticking factor (KI) representing the holistic risk evaluation over the entire $t_{p}$-step prediction interval, thereby achieving quantitative risk assessment driven by multi-step signal predictions. The proposed method is validated on a simulated drill sticking dataset and benchmarked against other approaches, confirming its feasibility.