LLM-Augmented BiGRU-MHA for SSD Health-State Classification Using SMART Telemetry

Abstract

Solid-state drives are routinely monitored through SMART telemetry, yet health-state classification remains difficult because vendor attributes are heterogeneous, failures are rare, and a single operational snapshot mixes normalized values with highly skewed raw counters. This study presents LLM-Augmented BiGRU-MHA, a compact sequence classifier that represents each SMART attribute with two numeric measurements and a fixed semantic description. A bidirectional gated recurrent unit models dependencies across the ordered attribute sequence, while multi-head attention emphasizes the most diagnostic indicators. Experiments were conducted locally on a 3,000-drive corpus organized according to the Alibaba SSD Open Data schemas, using a stratified 64/16/20 train/validation/test split. The proposed model achieved macro-F1 of 0.700, balanced accuracy of 0.726, weighted-F1 of 0.877, and one-vs-rest macro AUC of 0.952. It improved macro-F1 over the non-semantic BiGRU-MHA and detected 9 of 13 failed drives; the remaining four were assigned to the warning state rather than healthy. Ablation results show that semantic descriptions, bidirectional recurrence, and multi-head attention each contribute to minority-state sensitivity. Learned attention concentrates on wear-out degree, program-fail count, percentage used, and erase-fail count, providing an interpretable signal for storage-health triage.