Anomaly Detection Anywhere: Edge AI Diagnostics Meets Satellite IoT

Webinar Summary

The following summary is auto-generated from the webinar recording.

In this webinar, Blues and Edge Impulse teamed up to show how Edge AI and Satellite IoT combine to make anomaly detection truly "anywhere." We built an OBD-II anomaly detector with Edge Impulse, ran it on a Raspberry Pi, and used Blues Notecard + Notehub to route meaningful events to the cloud — with Starnote for Skylo providing satellite fallback when cellular coverage disappears.

Why Combine Edge AI with Satellite Connectivity?

Edge AI gives you the power to analyze raw sensor data locally — so instead of streaming everything to the cloud, your device only transmits high-value events. But those inferences are only useful if operators receive them, even when vehicles travel out of cellular range.

The result: local models that detect anomalies (misfires, leaks, overheating, etc.) + resilient connectivity that gets those events through no matter where the vehicle is.

Edge Impulse: The "Easy Button" for Edge AI

In short: if you need a custom model (sensor fusion, audio events, or image/object detection) that runs on constrained devices, Edge Impulse gets you from data collection to deployment with minimal friction.

• Collect data (CSV, sensor logs, audio, images)
• Build an Impulse (preprocessing + learning blocks)
• Train and test the model
• Deploy to targets ranging from MCUs to Raspberry Pi and Jetson

"When you need to train something unique, that's what Edge Impulse helps you to do." — David Tischler

Building an Anomaly Detector using OBD-II Data

We focused on a realistic automotive use case: detecting an air leak that causes abnormal NOx and RPM patterns, and ultimately a diagnostic trouble code (DTC). The workflow looked like this:

1. Tap into the vehicle's OBD-II port (CAN bus) using an ELM327-style adapter or direct CAN interface.
2. Record key sensor streams (MAF, RPM/throttle, NOx, temps, etc.) into CSV for labeling.
3. Label sequences as "healthy" or "air leak" based on a DTC or observed fault.
4. Upload labeled CSVs to Edge Impulse, add spectral/feature blocks and a classifier (or anomaly block), then train.
5. Deploy the trained .eim model to the runtime device (Pi, MCU, or Linux runner).

Key point: OBD streams are rich but massive. Pick the signals that matter for the failure mode you want to detect, and consider anomaly detection (k-means, clustering) for rare/unbalanced events.

From Model to Action: Pi + Model + Notecard

TJ showed how the trained Edge Impulse model is embedded into a small Python firmware loop. For demo purposes the script reads pre-recorded CSVs at a fixed cadence (0.5s) to emulate OBD sampling. Each sample is sent into the model and we look for the top label and its probability.

When the model predicts a fault (e.g., "leak" with >80% confidence), firmware can take immediate local action (indicators, safety logic) and queue an event for upload.

Blues Notecard: Connectivity without the SIM Headaches

Notecard is Blues' connectivity module: not just a modem, but a full managed wireless connectivity system that talks securely to Notehub in the cloud. Highlights:

• Global cellular reach (prepaid 500MB, up to 10 years of service with no monthly SIM fees for the cellular Notecard).
• Low power (µA idle currents) and an M.2-style swappable radio ecosystem.
• Simple JSON-based API so your firmware stays tiny while Notehub handles routing and integrations.

Notehub enriches incoming events with timestamps and location. If the Notecard has Wi‑Fi and is indoors, Notehub can triangulate using Wi‑Fi; outdoors you get GPS coordinates and/or cell tower triangulation.

Starnote for Skylo: Satellite Fallback

Cellular isn't always available. For truly critical alerts that must go through even when LTE access is gone, we presented Starnote for Skylo — a $49 accessory that pairs with a Notecard to provide satellite failover (via the geostationary Skylo network).

• Starnote is a Notecard accessory — it requires a paired Notecard to operate.
• Ships prepaid with 18 KB of satellite data (enough for one ~50-byte payload per day for a year).
• No monthly minimums from Blues; extra satellite KBs billed at usage ($0.75/KB).
• Supports NTN (Non-Terrestrial Networking) — a message-based, non-IP protocol optimized for tiny packets.

Typical pattern: operate on Wi‑Fi or cellular as primary transport, and configure the Notecard to fall back to NTN (satellite) when terrestrial networks fail. Use the note.template API to predefine compact payload shapes so the Notecard can compress as much as possible before sending via satellite.

Putting it Together

1. Device reads OBD-II data and runs on-device inference via Edge Impulse.
2. On fault detection, firmware uses note.add API to queue a Note/event on Notecard.
3. Notecard attempts a sync with Notehub. If unavailable, it uses Starnote/NTN for satellite transmission.
4. Notehub receives the event, enriches it (timestamp, location), and routes it (dashboard, webhook, SMS via Twilio, archives, etc.).

Edge AI and IoT Best Practices

• Start simple: capture a few critical signals for the failure mode you care about.
• Use Edge Impulse spectral features + classification or an anomaly block for rare events.
• Keep firmware logic minimal: detect and queue events; let Notehub handle formatting, enrichment, and routing.
• Design payload templates for satellite: compress and minimize fields sent via NTN.
• Test with CSV playback before wiring to live vehicles — it speeds iteration and debugging.

Conclusion

Edge AI and satellite IoT are a powerful pairing: models running at the edge let you sift through raw telemetry and only transmit what matters, while Notecard + Starnote give you the connectivity resilience to get those events to operators no matter where the vehicle goes. The end result is faster, more efficient maintenance workflows and fleets that stay visible even when cellular fades.

Let's make anomaly detection truly anywhere.