The intelligence layer for EV charging infrastructure.
Data Pigeon detects failures in real time and predicts them before they happen — so charging networks stay online.
Anomaly detected — CHG-042
Health score dropped 54 pts over 8 min
The problem
EV infrastructure is scaling.
The intelligence layer isn't.
0+
public EV chargers globally
Infrastructure is scaling fast.
The number of public charge points has tripled in four years. Grid operators, CPOs, and fleets are deploying hardware faster than they can maintain it.
—
Operators flying blind
The tooling doesn't exist yet.
Charge point operators are not negligent — they're working with what's available. Most fault detection still relies on driver complaints or manual inspection cycles.
0+
per charger per day, undetected downtime
Downtime is expensive and avoidable.
A failed charger costs operators in lost revenue, missed SLAs, and reputational damage. Most failures have detectable signals hours or days in advance.
How it works
Three steps from signal to action.
Detect
Real-time anomaly detection on live OCPP streams.
We ingest raw OCPP message traffic — boot notifications, meter values, status transitions, error codes — and score each charger continuously. Isolation Forest flags deviations from each charger's learned baseline the moment they appear.
OCPP health score — live charger stream
Under the hood
Built on real fleet data.
0M+
OCPP messages analyzed
Raw protocol streams spanning boot events, meter values, error codes, and status transitions.
0
Chargers in pilot fleet
Multi-site deployment spanning urban and highway corridors with diverse hardware.
0
Failure modes modeled
Gradual degradation and sudden hardware failure — treated as distinct model families.
Real‑time
Real-time latency
Anomaly scores updated on every incoming OCPP message, not on a polling schedule.
Methodology
Three model families. One unified signal.
Our models ingest raw OCPP message streams — boot events, meter values, error codes, status transitions — and learn what normal looks like for every charger in a fleet. When something drifts, we know. We combine unsupervised anomaly detection (Isolation Forest), gradient-boosted classifiers for fault prediction, and LSTM sequence models for temporal patterns. Class imbalance is handled with F1 and AUC-ROC as the source of truth, not accuracy.
Model architecture
Two failure modes. Two models.
Treating all failures as one problem leads to brittle models. We split the prediction problem into separate families from the start.
Gradual degradation
Slow drift in meter values, increasing error rates, subtle shifts in power delivery — these signals accumulate over days before a charger fails. Detectable from OCPP data alone using trend analysis and sequence modeling.
- Predictable from OCPP streams
- Days-to-failure time horizon
- LSTM + gradient boosting ensemble
- Supported in current pilot
Sudden hardware failure
Component failures — connector relays, cooling systems, power electronics — often appear with little OCPP warning. Catching these requires sensor fusion: temperature, vibration, electrical telemetry. We're building toward this with hardware partners.
- Requires hardware telemetry
- Hours-to-failure time horizon
- Multimodal sensor fusion approach
- In development with telemetry partners
Cross-operator intelligence
One model.
Every network.
Fleet-level models trained across multiple operators make every charger smarter. Failure signatures seen in one network inform predictions across all others — without sharing sensitive operational data.
- Cross-operator failure pattern library
- Fleet-normalized health baselines
- Privacy-preserving model federation (roadmap)
Research & recognition
Built with academic rigour and validated through real-world deployment partnerships.
“The Data Pigeon team is approaching EV charging reliability from first principles — grounding their models in real OCPP telemetry and maintaining methodological honesty about what the data can and cannot tell you.”
Research & writing
We show our work.
The team
Built by people who care about the grid.
Senara
Product & Strategy
Sina
Systems & Diagnostics
Suhani
Machine Learning
Get started
Ready to see what your fleet is telling you?
We work with a small number of operators at a time to ensure a high-quality integration. Request access and we'll be in touch within 48 hours.
We'll respond within 48 hours. No sales scripts.