TECHNICAL OVERVIEW
Purpose-built models. Not prompted.
Three domain-specific models trained from scratch on EV charging telemetry. No fine-tuned foundation models. No prompt engineering. Raw physics.
detection_pipeline.py
Why custom models?
- Custom Transformer Architectures — Purpose-built encoder-decoder transformers with rotary position embeddings. Not fine-tuned foundation models — trained from scratch on EV charging telemetry.
- Unsupervised Reconstruction — Models learn normal charging physics through reconstruction. Anomalies are sessions the model cannot faithfully reproduce — no labeled failure data needed.
- Physics-Aware Feature Engineering — Input features encode domain physics: power curves, energy gradients, SoC trajectories, per-feature activations matched to physical constraints.
- Per-Timestep Precision — Multi-scale temporal convolutions classify anomalies at individual timestep resolution. Pinpoints the exact moment and duration of each failure.
- Sub-5ms Inference — Full 3-layer pipeline < 5ms per session. Millions of sessions per day on a single GPU.
- Transfer Learning Pipeline — Layer 1 encoder is frozen and reused as backbone for Layer 2. Only the classification head trains — zero catastrophic forgetting.
- Per-Client Training — Small model footprint (~1.4M params) means each client can have models trained exclusively on their own network data. Dedicated models, not shared weights.
Rule-based monitoring catches errors. We catch failures.
OCPP error codes tell you what the station reported. Our models tell you what actually happened — including the failures the station never noticed.
Rule-Based / OCPP Solidstudio AI
Detection Error codes only Learned reconstruction
Coverage Known failure modes Silent failures included
Granularity Session-level Per-timestep
Speed Reactive Real-time (<5ms)
Learning Static rules Per-station baselines
Maintenance Manual rule updates Self-updating
SESSION · LAYER 1
CSAR v2
Charging Session Anomaly RecognitionTransformer autoencoder trained to reconstruct healthy charging sessions from engineered physics features. Uses rotary position embeddings (RoPE) for variable-length sequence handling. Anomaly score is derived from reconstruction error — sessions the model cannot faithfully reproduce are flagged.
╭──────────────────────────────╮
│ │
│ ┌╮ ╭──╮ ╭─╮ │
in ──┤│ │╰──╯ │ ╭╮ │ ╰──╮ ├── score
│ │ ╰──╯╰──╯ │ │
│ │ ▓▓▓▓ │
│ ╰────────────────────╯ │
│ reconstruct │
╰──────────────────────────────╯ - › Detects silent charging failures invisible to OCPP error codes
- › RoPE-based positional encoding — handles variable-length sessions natively, no padding
- › Per-feature output activations matched to physical constraints (rates, gradients, absolutes)
- › Encoder embeddings serve as transfer-learning backbone for Layer 2
- › Calibrated threshold derived from reconstruction error distribution on known-normal data
· · ·
// 14:22 — Berlin, connector_2
// Driver charges BMW iX, session looks normal
// OCPP reports: status=Charging, no errors
// ───────────────────────────────────
Energy delivery dropped 40% at min 35
OCPP never reported this.
// → session flagged, forwarded to CSAR-C v2
SESSION · LAYER 2
CSAR-C v2
Per-Timestep Anomaly ClassificationMulti-scale temporal convolution head trained on frozen CSAR v2 encoder representations. Consumes encoder embeddings, per-feature reconstruction error, and bypass features to produce per-timestep multi-label classification. The CSAR v2 backbone is completely frozen — zero catastrophic forgetting.
╭────────────────────────────────╮ │ t= 0···142━━━━168···T │ │ │ │ ep ····▓▓▓▓▓▓▓···· │ │ kd ··········▓▓··· │ │ gd ··················· │ │ sp ··················· │ │ │ │ → 6 labels × T timesteps │ ╰────────────────────────────────╯
- › Multi-label classification across 6 anomaly categories at every timestep
- › Multi-scale convolutions capture patterns across different temporal windows simultaneously
- › Per-feature reconstruction error as explicit input — the model knows which features CSAR v2 struggled with
- › Bypass features preserve signed values the autoencoder clamps (critical for reverse energy flow detection)
- › Post-processing: temporal segment merging and confidence-based filtering
Anomaly categories: ENERGY_PIT · KWH_DECREASE · GHOST_DRAW · STUCK_POWER · CHARGE_DRIFT · CHARGING_EXPECTED
· · ·
// 14:22 — same session, escalated from Layer 1
// Scanned 482 timesteps in 1.2ms
// ───────────────────────────────────
Found: energy_pit (6.5 min) + kwh_decrease
Inverter throttling at t=142
Counter rollback at t=165
Driver lost ~3.2 kWh without knowing
// → webhook → ops dashboard → ticket created
STATION · LAYER 3
SHADE
Station Health Anomaly DetectionCompact transformer autoencoder operating at the station-day level. Reconstructs 24-hour operational profiles from aggregated health metrics with RoPE temporal encoding. Uses directional scoring — an asymmetric error function that penalizes the right kind of deviation.
╭────────────────────────────────╮ │ │ │ ▁▂▃▅▇█▇▅▃▂▁ 24h profile │ │ ···········▼··· │ │ 08:30 │ │ │ │ score: 0.41 DEGRADED │ │ ▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔▔ │ ╰────────────────────────────────╯
- › Reconstructs full 24h station profile from aggregated health, utilization, and power metrics
- › Directional scoring: penalizes under-delivery of energy but not over-delivery
- › Activity masking forces reconstruction from health signals alone — prevents trivial shortcuts
- › Per-bucket, per-feature error breakdown enables precise temporal localization of degradation
- › Tracks multi-day score trends for predictive maintenance alerting
· · ·
// End of day — Berlin A7 daily health check
// 47 sessions processed
// ───────────────────────────────────
8 anomalies found
Morning peak (08:30) worst — utilization cratered
3rd consecutive day of decline
connector_2 degrading, connector_1 fine
// → field team dispatched for inspection
Start monitoring your network today.
Free plan. No credit card. 100 sessions included.