Toward Reliability-Aware Hybrid Anomaly Decision Systems: A Multimodal Prediction–Detection Framework for Cyber–Physical Environments

Abstract

Reliable anomaly decisions require both discriminative evidence and a defensible rule for reallocating trust when branch quality changes. This paper presents a Reliability-Aware Hybrid Anomaly Decision (RAHAD) layer that calibrates heterogeneous branch scores and adjusts their contributions without retraining the upstream models. Two complementary experiments are used. First, the reliability-allocation principle is tested on M2DGR through 800 clean controls and 21,600 controlled RGB, LiDAR, and IMU degradations. Under leave-one-sequence-out evaluation, the three branch models achieve ROC-AUC values of 0.9624, 0.7382, and 0.7115. Their estimated reliabilities correlate with known controlled quality at Spearman coefficients of 0.9555, 0.6372, and 0.6349. Equal weighting reaches a sequence-macro ROC-AUC of 0.7481, compared with 0.7473 for nested-selected fixed weights; the sequence-level difference is unresolved. Reliability-aware weighting suppresses the degraded modality, but stronger adaptation also increases weight drift on clean controls. Second, a frozen prediction-detection layer is evaluated once on 54 locked ShanghaiTech sequences. Fixed prediction scoring obtains ROC-AUC 0.6485, whereas regularized RAHAD obtains 0.6475. Complete-sequence bootstrap analysis confirms a small ROC-AUC disadvantage and a small reduction in mean sequence variance for regularized RAHAD, while differences in PR-AUC, low-false-alarm sensitivity, and transition rate remain unresolved. A causal moving average provides the strongest stability result with comparable ranking. Thus, the experiments do not establish universal superiority of adaptive fusion. They support reliability-aware allocation as an explicit trust-reallocation mechanism, while equal weighting and causal smoothing remain strong baselines when reliability is uncertain or temporal stability is the main objective.