Multi-parameter correlation anomaly detection for petroleum product compliance testing
Аннотация
Routine petroleum product testing evaluates each parameter independently against its specification limit under a single-parameter compliance paradigm that implicitly assumes parameter independence. This study challenges that assumption, proposing that multi-parameter correlation structures carry diagnostic signals invisible to single-parameter review. An Isolation Forest (IF) model was fit to 10,120 full-test-panel samples (gasoline: 5432; diesel: 4688) from an ISO/IEC 17025-accredited testing laboratory in Guangdong, China (2018–2026), retaining 3,279 gasoline samples across 32 parameters and 3534 diesel samples across 22 parameters. Five alternative anomaly detection methods—MCD Mahalanobis distance, Local Outlier Factor (LOF), One-Class SVM (OC-SVM), PCA reconstruction error, and a shallow autoencoder—were compared at equivalent contamination settings (1.0 %). IF detected 33 gasoline (1.01 %) and 36 diesel (1.02 %) anomalies, with 100 % top-20 ranking stability across contamination rates spanning an order of magnitude. The six-method comparison revealed low pairwise agreement (maximum Jaccard similarity 0.320, IF–PCA; IF–LOF 0.000), establishing that method choice is consequential and that different algorithms operationalize fundamentally different notions of “anomaly.” A structural hierarchy emerged: methods capturing global correlation structure (IF, PCA, autoencoder) detected globally isolated deviation; density-based methods (LOF) identified an entirely distinct anomaly population. Feature importance rankings were physically interpretable: compositional parameters (anti-knock index, distillation residue, aromatics) and oxidation stability indicators dominated gasoline; cold-flow properties—cold filter plugging point (CFPP), cloud point, pour point—and acidity dominated diesel. Retrospective verification of eight flagged samples traced to raw instrument records revealed two distinct anomaly categories: (a) genuinely hidden multivariate anomalies invisible to single-parameter logic, and (b) objective specification violations missed by the item-by-item review workflow—both representing failures of the single-parameter review paradigm. The framework requires no new instrumentation, translating anomaly flags into parameter-specific retest recommendations for laboratory personnel.