Many factories want monitoring and safety alerts but cannot or will not stream data to the cloud. Modern edge hardware can run the whole pipeline locally, from sensor to alarm.
Why process on the edge
- Privacy and policy: video and process data never leave the site.
- Latency: alerts in well under a second, without a network round trip.
- Reliability: monitoring continues when the internet does not.
- Cost: no ongoing server or bandwidth bills.
Choosing a detection approach
| Approach | Good for | Notes |
|---|---|---|
| Thresholds and rules | Simple limits like temperature | Easy, but misses subtle patterns |
| Statistical baselines | Vibration, current, pressure trends | Learns normal behaviour per machine |
| Autoencoders and small neural models | Complex multi-sensor patterns | Needs representative normal data |
| Vision models on an NPU | People in restricted zones, missing guards | Needs camera placement and labelled data |
Building the pipeline
Collect normal data
Record weeks of normal operation from each machine before training anything.
Start simple
A statistical baseline per machine often catches most real problems.
Add vision where it pays
Use an NPU board for camera-based safety checks.
Act locally
Drive relays, lights and on-screen alerts directly from the device.
Design for 24/7
Watchdogs, read-only file systems and automatic recovery after power cuts.
Frequently asked questions
Do I need labelled failure data?
Not for anomaly detection. It learns what normal looks like and flags deviations.
Can it run on a microcontroller?
For a few sensor signals, yes. Video needs a Linux board with an NPU.
Can results still reach a dashboard?
Yes, on a local network dashboard, without any cloud connection.
Embedded systems consultant at TechProgeny, Bengaluru. 7+ years and 29+ shipped products across firmware, embedded Linux, edge AI, apps and IoT servers.