Long‑horizon predictive maintenance requires models to separate slowly evolving degradation from normal operating variation over planning windows measured in days rather than hours. This work investigates whether an explicit conditional‑quantile representation can serve as an informative classifier interface for this task. The proposed TQRNN30d framework couples a dual‑stage quantile‑regression neural network (QRNN) feature extractor with a multi‑stream temporal‑fusion classifier. Each hourly snapshot of 81‑channel machine behaviour is mapped to a 324‑dimensional quantile‑state vector, and 720 ordered hourly snapshots form a 30‑day document fed to the long‑horizon model. The classifier fuses quantile states with dynamic covariates, channel‑level static metadata, and a 168‑hour latent‑history stream using gated residual processing, causal recurrent encoding, and metadata‑conditioned cross‑modal attention. A bounded instability‑aware signal derived from sustained one‑step‑ahead prediction‑error divergence provides auxiliary memory modulation at the longest horizon. Evaluation uses a machine‑disjoint 43/14/15 train/validation/test split across 72 machines in nine manufacturing facilities. At the 30‑day horizon TQRNN30d achieves 79.97% F1, 80.18% recall, 81.82% precision, 82.39% accuracy and 0.820 ROC‑AUC, outperforming all 18 baselines at the 7‑, 14‑, and 30‑day fixed‑threshold comparisons, with the largest F1 gain at 14 days. The results demonstrate strong held‑out‑machine performance within the homogeneous nine‑facility fleet, but do not establish generalisation to unseen sites, equipment types, or sectors.
Review