Every DC fast charger streams telemetry — connector temperature, contactor cycle time, insulation resistance, cooling-fan draw — dozens of times a minute. Hardware faults rarely appear as a single dramatic spike; they show up as a slow drift away from a station's normal operating envelope. This lab renders that drift two ways: as physical chargers on a forecourt, and as points in a live 3D feature space where an isolation-forest-style detector scores how "isolated" each reading is from the dense, normal cluster.
Isolation Forest needs no labelled failure examples at all — it only assumes faults are "few and different," which makes it well suited to charging networks where catastrophic hardware failures are thankfully rare and expensive to label by hand.
A live EV charger forecourt paired with a 3D isolation-forest feature space: watch simulated sensor drift push a station's telemetry point away from the healthy cluster until an unsupervised model flags it before it actually fails.
Each charger's temperature, cycle-time jitter and power-draw readings become coordinates in a 3D feature space. An isolation-style score measures how far a point sits from the dense, normal core — the same idea Isolation Forest and autoencoder reconstruction error use to flag hardware degradation early.
Raise the fault injection rate to make stations start drifting, then tune detector sensitivity to see the decision-boundary shell tighten or loosen. Add sensor noise to see how it complicates detection, and toggle the autoencoder overlay to see reconstruction-error vectors.
Because catastrophic charger failures are rare and expensive to label, real fleet-monitoring systems lean on unsupervised methods like Isolation Forest and autoencoders that need no failure examples at all — only a model of what "normal" looks like.