The issue of stuck pipe can significantly increase non-productive time, even leading to accidents such as drill string failure, which causes a sharp rise in drilling costs. Therefore, timely and accurate monitoring for signs of stuck pipe is crucial. This study establishes a hybrid physics-data model comprising a drag-torque model, hydraulic model, and unsupervised learning algorithm. The first model component enables real-time calibration of physical model parameters through Bayesian optimization using streaming data, achieving accurate dynamic calculations for three sticking-type characteristic parameters: theoretical hook load, theoretical torque, and theoretical pump pressure. The second component employs an unsupervised learning algorithm to monitor anomalous trends in characteristic parameters across different sticking types. Prior to real-time deployment, the model was trained on a 31-sample normal drilling dataset, with validation set sticking incidents subsequently guiding optimal threshold selection. Two field test cases demonstrate that the model’s reconstruction error effectively characterizes sticking progression, triggering alerts 4 and 30 min before friction-reduction operations respectively. This methodology addresses three critical limitations in existing approaches: (1) oversight of mechanistic distinctions among sticking types, (2) ineffective utilization of physics-based models, and (3) insufficient stuck pipe data availability for small-sample learning scenarios. The proposed framework establishes a novel paradigm for stuck pipe prediction research, enabling timely implementation of field-proven prevention and control strategies in oilfield operations.