Pipeline Posterior Scoring Module for out-of-distribution detection via attachable uncertainty quantification

Intelligent monitoring models effectively support pipeline structural health monitoring, yet closed-world training causes overconfident predictions on untrained samples, threatening structural safety and necessitating out-of-distribution detection. Existing uncertainty quantification methods require retraining or intervention in frozen production systems, limiting practical applicability. This paper proposes the Pipeline Posterior Scoring Module (PPSM), an attachable component designed specifically for deep neural network-based leak detection models, endowing them with out-of-distribution risk identification capability without modifying their parameters. PPSM fuses multi-level features spanning signal textures to fault semantics via self-attention, outputs Dirichlet parameters for Bayesian uncertainty quantification, and realigns the training objective from accuracy maximization toward risk perception under frozen-parameter deployment constraints, combined with a noisy validation strategy to select model configurations most sensitive to out-of-distribution samples. Experiments on ResNet101, VGG19, and MobileNetV2 across real out-of-distribution and synthetic fault scenarios showed that PPSM achieves AUROC exceeding 0.88 with missed alarm rates below 0.15 in the vast majority of configurations, consistently outperforming all comparison methods. With 263K parameters and 5ms inference time, PPSM provides an efficient architecture-agnostic solution for reliability enhancement in resource-constrained industrial pipeline systems.