Dual-Layered Unsupervised Anomaly Detection for Regulatory Vehicle Emissions Enforcement: A Physics-Constrained Synthetic Approach

Abstract

The Labeled Data Bottleneck The enforcement of vehicle emissions regulations - critical for urban air quality and global climate compliance - relies heavily on certified, decentralized testing centers. However, this infrastructure is highly vulnerable to systemic fraud, including hardware tampering (e.g., EGR deletes, defeat devices) and institutional corruption (e.g., Clean Scanning, Odometer Rollbacks). A primary bottleneck in applying Machine Learning to this regulatory domain is the strict absence of publicly available, labeled fraudulent datasets due to privacy laws and government liability. Proposed Architecture In this paper, we present a novel, dual-layered unsupervised machine learning architecture to detect emissions fraud without requiring historical labeled data. First, we introduce a physics-constrained synthetic injection engine that maps 7 distinct real-world fraud vectors into a harmonized 10-dimensional physical baseline. Second, we implement a Dual-Layered Isolation Forest architecture. Layer 1 evaluates thermodynamic outliers in vehicle hyperspace, while Layer 2 evaluates operational metadata (Throughput, Variance, Poisson-distributed temporal anomalies) to flag institutionally corrupt testing centers. Results Evaluated against a 48,000-row synthetic US Environmental Protection Agency (EPA) baseline, the model achieved perfect recall in isolating institutionally corrupt facilities with zero false positives. This paper demonstrates that tree-based unsupervised learning can provide the mathematical interpretability required for legally defensible regulatory enforcement.