An Interpretable Composite Index for Real-Time RFID Anomaly Detection in Predictive Maintenance

Abstract

Radio-frequency identification (RFID) systems are widely deployed for asset and inventory tracking in industrial environments, yet their reliability degrades under dynamic conditions where early, subtle tag anomalies remain difficult to detect. The objective of this work is to develop an interpretable, lightweight, and real-time index for detecting anomalous tag behavior at the point of reading, addressing the limitation that most existing methods rely on a single indicator such as received signal strength or raw read counts and therefore lack sensitivity to incipient instability. The core idea is to fuse two dimensionally consistent sub-metrics into a single composite score: a read-speed deviation coefficient that quantifies instability in tag-read cadence, and a communication-frequency concealment coefficient that captures temporal and communication-rate irregularities, with each component mapping to a physically identifiable failure mode so the score is interpretable by construction. The contribution is this anomaly evaluation index together with a single-pass, constant per-event computation suitable for commodity reader hardware. Validation spanned three tiers: a simulated dataset of 500 virtual tags and 30,284 events with controlled anomaly injection, a semi-synthetic dataset built from industrial warehouse logs, and a public benchmark of approximately 1,100,000 reads across multiple tag manufacturers. Key results: on the simulated test set the index achieved 95.7% accuracy, an AUC of 0.924, a 6.0% false-positive rate, and a mean detection latency of 4.8 ms on an embedded ARM-class processor; ablation confirmed the complementary contribution of both sub-metrics (accuracy dropped when either was removed), and the index ran approximately 40× faster than a deep-learning baseline of comparable detection quality. The novelty lies in combining two dimensionally consistent, physically interpretable sub-metrics into one constant-cost score, delivering deep-learning-comparable accuracy with a 40× speedup, making it well suited to embedded, real-time anomaly detection in resource-constrained industrial deployments.