Quantile-Led Feature Extraction for Industrial Predictive Maintenance

A dual-stage MLP-QRNN hierarchy framework improves multi-horizon predictive maintenance by learning ten-quantile conditional distributions for sensor channels. Using four retained mid-tail quantiles with attention achieves a 75.92% F1-score for 30-minute forecasting regimes across 72 industrial machines.