Vision Field Notes

Power, Cooling and Storage for a Reliable Edge AI Box

A model that works on your desk can still fail in the field. When an edge AI device goes down, the cause is rarely the neural network. It is usually a weak power supply, heat, a worn-out SD card or a process that nobody restarted. This article covers the basics that make an edge box run for months without a visit.

The examples use Raspberry Pi hardware because it is common, but the principles apply to any small computer.

Power

Use a proper supply

Small boards are sensitive to supply quality. A cheap charger or thin cable can drop voltage under load, especially when the board is busy with inference and USB devices. The result may be random crashes, corrupted files or throttled performance.

Practical rules:

On Raspberry Pi OS you can check throttling and under-voltage flags with vcgencmd get_throttled. Look this up in the current documentation to interpret the bit flags, and log the value over time in your health checks.

Power loss

Sudden power loss is the most common cause of file system corruption. Options to reduce risk:

Power over Ethernet

For IP cameras, PoE removes the need for a separate power cable at the camera. A PoE switch supplies power and data on one cable. Check the switch's total power budget against the number of cameras, because exceeding it causes cameras to drop out.

Cooling

Why heat matters

When a processor gets too hot, it lowers its clock speed to protect itself. This is thermal throttling. Your detection frame rate then falls without any error message. A device may seem fast in a short test and slow after ten minutes.

How to check

On Raspberry Pi OS, vcgencmd measure_temp reports the temperature, and vcgencmd get_throttled shows whether throttling has occurred. On other platforms, look for thermal zone files or vendor tools. Do this:

  1. Run your real workload, not an idle system.
  2. Let it run for a long time, at least until the temperature levels off.
  3. Log temperature and frame rate together.
  4. If frame rate falls as temperature rises, you have a cooling problem.
Two stacked line charts over 30 minutes of load. Top: temperature of a bare board climbs to a dashed throttle line at 80 degrees and stays there, while a board with heatsink and fan levels off near 60 degrees. Bottom: frame rate of the cooled board stays at 100 percent, while the bare board drops from about minute 12 to around 63 percent.
Simulated, illustrative data. A one-minute test looks identical for both boards; only a long run with temperature and frame rate logged together shows the throttling.

What to do

Storage

SD card wear

MicroSD cards are convenient, but they are the weakest part of many small deployments. Flash storage wears with writes, and cheap cards can fail or corrupt data, especially with frequent small writes such as logs and databases.

Reduce the risk:

SSD instead of SD

For heavier or long-lived use, an SSD is usually more robust and faster. With the Raspberry Pi 5, you can attach an NVMe drive through a HAT or base on the PCIe interface. Benefits: better sustained write performance and typically better endurance for write-heavy work. Costs: extra hardware, more power draw and a bit more setup. Check that the HAT and drive are compatible, and follow the current instructions for booting from NVMe.

Option Good for Watch out for
Standard SD card Prototypes, light use Wear, corruption, fakes
High-endurance SD card Light logging, appliance-style devices Still limited write endurance
NVMe SSD Video clips, databases, heavy logging Power, cost, compatibility

What to store

Decide what the device keeps and for how long.

Uptime practices

Hardware is half the story. Software must also recover by itself.

Run as a service

Use a process supervisor such as systemd or a container restart policy, so the application restarts after a crash or reboot. In Docker, a restart policy like unless-stopped does this. See Dockerize a computer vision app for container patterns.

Use a watchdog

A hardware or software watchdog reboots the device if it hangs. Many boards have a hardware watchdog. Combine it with an application-level heartbeat: if your main loop stops producing results, restart it.

Health checks and monitoring

Track a few signals and send them somewhere you will see them:

Alert on the absence of data, not just on errors. A device that is silent is often a device that has failed. Camera reconnection logic is covered in handling RTSP dropouts.

Two-column diagram. Left, what goes wrong: app crashes, main loop hangs while the process is alive, whole device hangs, power cut, silent for any reason. Right, what recovers it: process supervisor such as systemd or a Docker restart policy, an app heartbeat check, a hardware watchdog, a UPS with few writes and a read-only root, and remote monitoring that alerts when data stops.
Each failure needs its own recovery layer. The last layer, alerting when data stops arriving, catches whatever the others miss.

Time and logs

Keep the clock correct with network time, since wrong timestamps make events hard to trust. Rotate logs and cap their size.

Updates

Automatic updates can break a working system, and no updates can leave security holes. Pick a policy: test updates on a spare device first, pin versions and keep a way to roll back. Remote access should use keys, not default passwords.

A burn-in test

Before you ship a box, run it for several days under realistic load in the enclosure it will live in. Include a power-cut test and a network-cut test. Confirm that it recovers without a person touching it. The pre-deployment testing checklist lists more checks.

Key points

Next steps

Continue with Dockerize a Computer Vision App for Edge and Cloud, then A Pre-Deployment Testing Checklist for Vision Models.