GTST-HDD: A Graph-Temporal Survival Transformer for Hard Drive Failure Prediction

In this study, we used the famous Temporal Fusion Transformer (TFT) framework to envisage the presumable storage unit problems through time-snapshot information and disk-level features. There were two key units prepared namely a time-snapshot file, merged_20000_sample.csv with serial number and the dedicative disk-level feature file, per_drive_features_full_20k_ready.csv with 30-day mean, and reversion coefficients). The target was defined as a binary classification of "failure within a window H = 30 days" at each time-snapshot, and time sequences with a maximum length of 32 (max_encoder_length = 32) were constructed for each disk (with fill and tagging of missing values). The data were segmented into training/validation/test sets at the serial number level to prevent data leakage.
The TFT architecture included: a static variable encoder, a variable selection network for each time period, a local processor (LSTM) to capture short-term dependencies, and an interpretable multi-head attention layer to capture long-term relationships. We used BCEWithLogitsLoss binary loss with positive class weights to handle imbalances, and the model was trained with the Adam optimizer at a 1e-3 learning rate and dropout of 0.1, with an AUPRC-based early stop on the validation set. For assessment, we applied PR-AUC (AUPRC), ROC-AUC, Accuracy/Recall, and F1, with a verge analysis to assess suitable run point. The selected variable was illustrated with features which contributed to the forecast of choices, to support operative model’s adaptability within the disk to monitor the ecosystem.