Each post streams voltage, current and internal temperature. Isolation Forest and autoencoder models both reduce that stream to one number: how far this reading sits from everything the model has learned is "normal" charging behaviour.
score = sqrt( ((V-V0)/σV)² + ((I-I0)/σI)² + ((T-T0)/σT)² )
flag for maintenance if score > sensitivity
- Stations — size of the monitored charging network.
- Fault injection rate — how often a post starts drifting away from its healthy baseline (a failing sensor, degrading fan, worn connector).
- Detection sensitivity — the anomaly-score threshold; lower catches drift earlier (more maintenance callouts) but a threshold set too high lets a station reach hard failure and drop offline before anyone is dispatched.
Green posts are healthy, amber posts have crossed the anomaly threshold and are queued for a technician, red posts have failed outright and gone offline — exactly the gap between "caught early" and "driver turned away" the article describes.