An FPGA-Based Architecture for Interpretable PM2.5 Anomaly Detection with Counterfactual Explanations
Abstract
Black-box machine learning models have a well-known limitation in safety-critical environmental monitoring: they produce no explanation for their decisions. This paper describes an FPGA-based PM2.5 anomaly detection system that addresses this gap by generating counterfactual explanations, reporting to the user exactly which threshold was violated and by how much.The architecture combines adaptive threshold learning, EPA PM2.5 categorization (Good through Hazardous), and cold-start protection. Five anomaly types are detected, safe operating values are computed, and confidence scores are assigned according to severity. A cold-start controller injects nominal reference values at initialization, preventing the threshold learner from being biased by polluted startup conditions before enough clean history has accumulated. Cycle-accurate verification targets an Artix-7 XC7A35T FPGA running at 50 MHz. The system achieves 40 ns processing latency, 6.55 mW total power consumption, and 3.6% LUT utilization, confirming that hardware-accelerated interpretable anomaly detection is feasible on resource-constrained edge devices.