<p>Continuous casters play a crucial role in the continuous casting process of steel production, with the rolls serving as core components. The reliability of these rolls is vital as roll failures can severely impact the quality of the produced slab. Despite the implementation of regular maintenance practices to ensure the reliability of drive rolls, the timely detection of roll abnormalities remains a crucial necessity. This paper introduces a novel method for detecting abnormality in drive rolls. The proposed method revolves around the analysis of multidimensional time series withdrawal force data obtained from a continuous caster in the plant. Initially, the data undergoes preprocessing, which involves classification and Piecewise Aggregate Approximation (PAA) smoothing. Subsequently, relevant features are extracted from the data using dimensionality reduction and feature selection methods, including sliding window and statistical analysis in the time frequency domain. Afterward, the extracted features are utilized to train and evaluate ten different machine learning models. Upon evaluation, it is revealed that the Random Forest Extreme Learning Machine (RFELM) exhibits superior performance compared to the other models. Therefore, RFELM is chosen as the final model for effectively detecting abnormalities in the equipment. Finally, the abnormal state of the equipment is determined by comparing the predicted outcomes with the actual results. The results demonstrate that the method can accurately and timely detect abnormal states of the continuous casting drive roll, thus providing a guarantee for the stable operation of the continuous caster.</p>