COMPARATIVE ANALYSIS OF MACHINE LEARNING MODELS FOR PREDICTIVE MAINTENANCE: CHALLENGES AND OPPORTUNITIES

Predictive maintenance of Hard Disk Drives (HDDs) using machine learning is vitalfor preventing data loss and operational downtime in data centres. SequentialSMART (Self-Monitoring, Analysis, and Reporting Technology) data provides a richsource for forecasting impending failures. However, a clear consensus on the mosteffective modelling approach is absent, constituting a knowledge gap caused bystudies that focus on a single paradigm without a comprehensive, comparativebenchmark across multiple task types. This research aims to systematicallyevaluate and rank the performance of over 20 models across regression,classification, time-series forecasting, and anomaly detection tasks to identify theoptimal strategies for predicting HDD failures. We conducted an extensiveempirical study, employing tree-based models, deep learning architectures,including LSTMs (Long Short-Term Memory networks), GRUs (Gated RecurrentUnits), and Temporal Fusion Transformers (TFT), as well as various autoencoders.A full hyperparameter sweep was performed for time-series models to ensurerobust comparisons. Tree-based models excelled in static analysis, while deeplearning was superior for temporal sequences. The TFT model achieved thehighest average accuracy (86.12%) and best generalisation. The GRU modelattained the highest failure recall (92.52% peak accuracy). Conversely, anomalydetection methods demonstrated only moderate performance (65-73% accuracy).The key contribution of this work is a definitive, evidence-based model selectionframework that addresses this gap. The TFT architecture is identified as the mostrobust and efficient model for analysing temporal SMART data. Practitionersshould prioritise tree-based models for static tabular data, TFT for accurate andreliable forecasting, and GRU for maximising failure detection sensitivity, therebysignificantly enhancing predictive maintenance frameworks.