Work

Reliability-Aware Multivariate Temporal Learning for Early Risk Detection in Cellular Network Infrastructure

Year: 2026
Type: preprint
Abstract: Thermal and electrical degradation in cellular network infrastructure develops through coupled changes in equipment temperature, ambient conditions, direct-current (DC) voltage and current, power demand, cooling behaviour and network load. This study develops a reliability-aware multivariate temporal learning framework for forward identification of thermal or electrical abnormal states under sparse and irregular telemetry. A real operational dataset of 12,000 observations from 250 cellular sites and 1,499 assets is analysed with a forward target that excludes contemporaneous alarms and downstream consequences to prevent outcome leakage. The framework separates current-state evidence, recent temporal dynamics and history reliability, and evaluates these components with site-grouped cross-validation. At the 72-h horizon, temporal extreme gradient boosting (XGBoost) achieves a mean area under the receiver operating characteristic curve (ROC-AUC) of 0.5592 ± 0.0295 versus 0.4869 ± 0.0310 for the current-state baseline, with a paired site-bootstrap AUC gain of 0.0741 (95% confidence interval 0.0235–0.1243). A mask-only control reaches 0.5021 and a seven-day rolling-history baseline 0.4745, indicating that sampling availability or longer-window aggregation alone does not reproduce the temporal result. An exploratory learned reliability-gating ablation improves Brier score modestly but reduces ROC-AUC to 0.5220 and is therefore not promoted to the primary learner; instead, history completeness, elapsed gaps and temporal consistency are treated as evidence-quality variables for deployment interpretation. Chronological and leave-one-asset-class-out stress tests further show that generalization remains bounded by the single-environment, sparse-event dataset. The study therefore contributes a reliability-aware evaluation and deployment framework for temporal early-risk identification rather than claiming universal predictive-maintenance capability.
Language: English
Open Access status: green
OpenAlex ID: W7230749337
Cites: 20
Cited by:
Related to: 10