How to choose an edge AI development kit for industrial sensing

An edge AI development kit is a bundle of compute hardware, sensors and software that lets an engineering team record data, train or adapt models and run them next to a machine before committing to a production design. For industrial sensing, most kits on the market are built around the compute module and treat sensors as an afterthought. That is usually the wrong way round.

This article sets out vendor-neutral criteria first. At the end, we describe our own kits factually, so you can apply the same criteria to them. Our related article on edge AI in manufacturing covers the broader architecture question of what to evaluate at the machine.

What an edge AI development kit for industrial sensing has to do

A general-purpose AI board is designed to run models. A development kit for industrial sensing has to do four jobs:

1. Acquire signals from the right sensors at the right rates, with known quality.
2. Record them reliably, with time stamps and integrity checks, for model development.
3. Run models next to the machine and present results to operators.
4. Transfer to production without redesigning hardware or software.

If a kit only does the third job well, your team will spend its time building the other three.

Seven criteria for choosing an edge AI development kit

1. Sensor interfaces and acquisition quality

Start from your sensors. Cameras usually connect through MIPI CSI or USB; check the number of camera connectors and lanes, and whether global-shutter sensors are supported by the camera stack. Vibration and acoustic channels need analogue front ends with suitable sample rates, resolution and anti-aliasing; a general-purpose board rarely has them. Ask for the exact sensor part numbers, not just "supports cameras".

2. Time stamps across sensors

Ask how each stream is time-stamped and how offsets between devices are measured and recorded. A kit that time-stamps data on arrival in an application, with no information about buffering or clock offsets, makes multimodal analysis much harder later.

3. Compute: CPU, GPU and NPU

Edge processors combine CPU cores, a GPU and often a neural processing unit (NPU). A headline TOPS figure (tera operations per second) tells you the NPU's peak throughput, not whether your model runs on it. Check which model formats and operators the vendor's toolchain supports today. Portable formats such as ONNX, described as an open format to represent machine learning models, and runtimes such as ONNX Runtime, which targets CPU, GPU and NPU across platforms, reduce lock-in. For many industrial signal models, which are small, a modern CPU is already sufficient.

4. Recording, storage and export

Model development needs raw data. Check storage type and capacity (NVMe is far more robust for continuous recording than SD cards), how recordings survive power loss, whether files are hashed for integrity and which formats you can export without proprietary tools.

5. Software support and operating system

Find out which operating systems and versions are supported, for how long, and who maintains board support packages. A kit tied to a single, unmaintained image will not survive a three-year project.

6. Security and regulatory readiness

Connected industrial devices are increasingly subject to cybersecurity regulation. The EU Cyber Resilience Act, Regulation (EU) 2024/2847, entered into force on 10 December 2024. According to the European Commission, its reporting obligations have applied since 11 September 2026, and its main obligations apply from 11 December 2027. A development kit is not the final product, but if the same platform goes into production, ask now about security updates, vulnerability handling and update mechanisms. Our article on the Cyber Resilience Act for connected sensors goes into detail.

7. Path to production

The best development kit uses the same board, sensors and software as the production system. Check availability and lifecycle of the key components, whether the vendor can deliver in production quantities and whether a custom configuration is possible without starting again.

A scorecard for comparing kits

Use a simple scorecard to compare candidates. Score each criterion 0 to 2 and weight it by importance for your project.

How to run a short kit evaluation

A scorecard filled in from datasheets is a start. A short hands-on evaluation on a real machine tells you far more. As a rough estimate, the following sequence fits into two to three weeks for most teams.

1. Pick one machine and two or three failure modes you already understand, so you can judge whether the data makes sense.
2. Install the sensors as you would in production, including mounting, cabling and line of sight, not on a desk.
3. Record continuously for several days, including start-ups, shutdowns and product changes, and deliberately pull a cable or cut power once to see how the kit handles gaps.
4. Export the data and open it in your own tools without the vendor's software. If that is painful now, it will be painful for the life of the project.
5. Run one simple model at the edge, even a threshold on a feature, and measure the effort from trained model to deployed model.
6. Write down every workaround your team needed. That list is the most honest comparison between kits.

If you are unsure which features to compute for that first model, our article on feature engineering for industrial signals is a practical starting point.

Common mistakes when choosing a kit

• Starting with the accelerator. Most industrial sensing failures trace back to sensor choice, mounting or data quality, not to compute.
• Ignoring recording. Teams often plan for inference and forget that the first months are spent collecting and labelling data.
• Buying for a demo. A kit that runs an impressive object detection demo may have no answer for an 18-bit vibration channel.
• Assuming the NPU is used. Check whether your model actually runs on the accelerator with the current software, or falls back to the CPU.

For more on what a QCS6490-class platform can do, see our article on edge AI hardware for sensing.

Where NeuroControls kits fit

Our kits are built around the Edge AI Hub on a RUBIK Pi 3. According to the RUBIK Pi 3 user manual, the board uses a Qualcomm QCS6490 with eight Kryo 670 cores, a Hexagon NPU rated at 12 TOPS and an M.2 Key M slot with two PCIe 3.0 lanes. Our hub adds NVMe storage and the sensors listed below.

The Multi-Camera Hub combines three OmniVision OV9281 global-shutter cameras (capture up to about 100 fps, recording rate set per order), a FLIR Lepton 3.5 radiometric thermal core, a 48 kHz USB microphone and 512 GB NVMe. The Vibration Node uses an ADXL1002 accelerometer with an 18-bit, 200 kSPS front end; the Acoustic Node uses a Microtech Gefell MK 301 E capsule. Every stream is time-stamped on the hub, recordings are hashed per segment and export to NWB, MKV/HEVC, JSONL and CSV.

On compute, we are precise about the status: custom models are trained for you by our engineers, on-device inference as ONNX on the QCS6490's CPU was first proven on our Vibration Node, and NPU deployment is coming soon. See the kit overview and the Edge AI Hub for details, or request a quote.