Model Comparison — H₂ Anomaly Detector

or Simulate cycles

Model Overview

Model	Approach	Strength	Limitation	Recommended for
Isolation Forest	Tree isolation	O(1) per reading, low FPR	Conservative recall	Live stream
Local Outlier Factor	Density-based	Best F1, catches local anomalies	Slower per prediction	★ Batch
One-Class SVM	Kernel boundary	Tight cluster detection	Higher FPR on noisy data	–
Elliptic Envelope	Multivariate Gaussian	Captures sensor correlations	Assumes Gaussian distribution	–
Z-Score	Statistical baseline	Fully interpretable	Misses multi-feature anomalies	Baseline

Metrics Summary

Precision — share of alarms that were real anomalies Recall — share of anomalies that were detected F1 — harmonic mean of both; primary ranking metric FP Rate — false alarm rate during normal operation TP correct detections · FP false alarms · FN missed · TN correctly ignored

Model	Precision	Recall	F1	FP Rate	TP	FP	FN	TN

F1 Score Comparison

Precision vs Recall

Detection Timeline (anomaly scores — ground truth marked in red)

Scroll to zoom · drag to pan