The increasing complexity of industrial systems necessitates accurate and timely fault detection mechanisms to maintain reliability and avoid unplanned downtimes. This paper presents a comparative study on fault detection methods applied to a classifier drive system in a cement manufacturing plant. Four models: One-Class Support Vector Machine (OC-SVM), Random Forest (RF), Gated Recurrent Unit (GRU), and Long Short-Term Memory (LSTM) were evaluated across three time-series segmentation strategies: overlapping, non-overlapping, and expert knowledge driven. Current and speed data collected over a one-year period were used and expert-annotated fault dates serving as ground truth. Thirteen statistical and frequency-domain features were extracted from each segmented window and used to train both traditional machine learning (ML) and deep learning (DL) models. Results show that RF achieved the highest accuracy of 98.90%, successfully detecting 11 faults with a relatively low average computational time, ranging from 0.33 to 0.94 s. In terms of computational time, ML (OC-SVM and RF) models demonstrated far lower processing times compared to DL models, but expert knowledge-driven segmentation has significantly enhanced the DL (GRU and LSTM) models’ ability to align predictions with actual fault occurrences by achieving an average temporal error of just 0.2 days which is crucial in industrial settings for rapid and accurate fault localization. This study highlights the importance of selecting appropriate segmentation strategies and tuning model parameters for achieving accurate, interpretable, and efficient fault detection in complex industrial environments.