The occurrence of downhole high-frequency torsional oscillations (HFTO) can lead to the significant damage of drilling tools and can adversely affect drilling efficiency. Therefore, establishing a reliable HFTO identification model is crucial. This paper proposes an improved whale algorithm optimization support vector machine (TSWOA-SVM) for accurate HFTO identification. Initially, the population is initialized using Fuch chaotic mapping and a reverse learning strategy to enhance population quality and accelerate the whale optimization algorithm (WOA) convergence. Subsequently, the hyperbolic tangent function is introduced to dynamically adjust the inertia weight coefficient, balancing the global search and local exploration capabilities of WOA. A simulated annealing strategy is incorporated to guide the population in accepting suboptimal solutions with a certain probability, based on the Metropolis criterion and temperature, ensuring the algorithm can escape local optima. Finally, the optimized whale optimization algorithm is applied to enhance the support vector machine, leading to the establishment of the HFTO identification model. Experimental results demonstrate that the TSWOA-SVM model significantly outperforms the genetic algorithm-SVM (GA-SVM), gray wolf algorithm-SVM (GWO-SVM), and whale optimization algorithm-SVM (WOA-SVM) models in HFTO identification, achieving a classification accuracy exceeding 97%. And the 5-fold crossover experiment showed that the TSWOA-SVM model had the highest average accuracy and the smallest accuracy variance. Overall, the non-parametric TSWOA-SVM algorithm effectively mitigates uncertainties introduced by modeling errors and enhances the accuracy and speed of HFTO identification. By integrating advanced optimization techniques, this method minimizes the influence of initial parameter values and balances global exploration with local exploitation. The findings of this study can serve as a practical guide for managing near-bit states and optimizing drilling parameters.

Within the spectrum of complex downhole operational challenges, pipe sticking incidents emerge as one of the most prevalent and costly drilling complications. These incidents characteristically develop through progressive deterioration rather than abrupt failure, with discernible precursor signals typically manifesting as anomalous patterns in critical drilling parameters (torque fluctuations, drag anomalies, deviations in standpipe pressure). Consequently, early detection of these signals plays a pivotal role in mitigating pipe sticking occurrences. To systematically investigate the characteristic signatures pipe sticking events, this study employs two modal decomposition methods to extract salient features from near-bit downhole data. Conventional pipe sticking prediction methodologies exhibit three predominant limitations: rule-based systems suffer from poor generalizability, physics-based models demonstrate low computational efficiency, and data-driven techniques lack physical interpretability. To overcome these constraints, this study innovatively proposes a physically constrained prediction framework that integrates Variational Mode Decomposition (VMD) with near-bit measurement data. Experimental results demonstrate the superior predictive capability of the proposed VMD-based, near-bit data-physical constraint model. Based on a comprehensive evaluation using six benchmark models, the proposed approach achieves optimal performance, with an R² metric of approximately 0.9, significantly outperforming existing algorithms. When deployed in actual drilling operations, this model exhibits robust early detection of pipe sticking precursors, enabling proactive intervention. The practical implementation of this framework facilitates timely corrective actions, thereby substantially reducing the incidence of downhole pipe sticking events and enhancing operational safety.