A Calibrated Hybrid Edge Framework for Elevator Safety Warning with Statistical Anomaly Detection

Abstract

This paper presents a calibrated hybrid edge framework for elevator safety warning. The framework combines interpretable safety rules, robust statistical anomaly scoring, one-class anomaly-detection baselines, temporal persistence, and structured event logging on a Raspberry Pi platform. The prototype integrates acceleration, acoustic, environmental, and vision-based door-state sensing. To avoid treating fixed thresholds as arbitrary constants, severe safety constraints are kept as explicit rules, while broader motion and acoustic deviations are interpreted as calibrated statistical evidence before alarm escalation. On the available factory-test motion-noise recordings, the deployed full rule set detects all 10 abnormal or top-impact events but triggers nuisance alarms in 3 of 5 normal events, yielding precision, recall, F1 score, specificity, and accuracy of 0.769, 1.000, 0.870, 0.400, and 0.800, respectively. To assess robustness under controlled data expansion, an augmented motion-noise batch is constructed from annotated factory intervals using class-preserving resampling and signal perturbation. On this augmented robustness batch, the hybrid detector obtains precision, recall, F1 score, specificity, and accuracy of 0.904, 0.966, 0.934, 0.906, and 0.935, respectively, outperforming fixed-rule, robust-MAD, Isolation-Forest, and one-class-SVM baselines. A standalone door-state clip validation reports 0.943 accuracy and 0.946 macro-F1, while synchronized door-motion event validation remains a limitation for future multi-site testing.