Drill string stuck pipe remains a pervasive and costly Non-Productive Time (NPT) challenge in petroleum operations. While the previous studies have often focused on binary stuck/non-stuck prediction using low-resolution daily reports, this research introduces two key innovations: (1) the utilization of high-resolution Mud Logging data to capture subtle pre-failure signatures, and (2) a novel unsupervised labeling methodology using K-Means clustering to identify the specific stuck mechanisms (Wellbore Geometry, Differential Pressure, and Hole Packing), a critical step given the absence of physical labels in field data. Daily drilling reports from 28 wells were initially screened; however, the predictive models were developed using high-resolution mud logging data from 8 wells selected for data completeness, providing a methodological framework for real-time classification. For the binary early warning task, the Coarse Gaussian Support Vector Machine (SVM) demonstrated superior reliability, achieving the lowest count of critical False Negatives (FN=1) on the test set. For the multi-class mechanism identification task, the Linear SVM (while the Differential Pressure mechanism remained unresolved with the available surface data) proved optimal, achieving perfect classificationon the distinguishable classes on all distinguishable stuck types (Wellbore Geometry and Hole Packing) and an overall accuracy of 98 percent. This framework validates that the combination of high-resolution data with unsupervised labeling provides a computationally efficient and reliable foundation for deployment on the driller's panel, enabling proactive operational adjustments and significantly reducing NPT.