Interpretable Anomaly Diagnosis and Root-Cause Analysis of LNG Storage Tanks Using Signed-Deviation Calibration and Mechanism-Consistent Evidence Reasoning

Large LNG storage tanks are key shore-based facilities linking maritime transport with receiving-terminal storage and send-out. Their continuous monitoring needs anomaly identification, preservation of physically meaningful deviation directions, and association of abnormal states with potential failure modes. This study proposes an interpretable anomaly-diagnosis and root-cause-analysis framework integrating a signed-deviation-calibrated multi-branch probabilistic autoencoder (SDC-MPAE) with mechanism-consistent evidence reasoning (MCER). SDC-MPAE learns continuous signed states from normal monitoring data through channel-adaptive signed perturbations, normal-manifold-anchored reconstruction, and signed-deviation calibration. MCER combines the mechanism-consistent relation matter-element model (MCREM) with distribution-aware uncertainty-retained Dempster–Shafer fusion (DUR-DS) to map signed states and fuse absolute mechanism support, candidate distributions, and inter-source conflict. A 200,000 m3 full-containment LNG storage tank in Tangshan, China, is the engineering case; its actual temperature, pressure, and liquid-level monitoring data support model training and testing, while parameterized abnormal test scenarios are independently constructed from representative failure mechanisms. SDC-MPAE achieved an accuracy of 0.9704 and a PR AUC of 0.9850, while MCER achieved Top-1 and Top-3 root-cause accuracies of 0.946 and 0.985, respectively, with an MRR of 0.971. Ablation and robustness analyses confirmed improved directional representation, candidate discrimination, and uncertainty expression, providing a feasible pathway for interpretable diagnosis of LNG receiving and storage infrastructure.