Industrial AI and edge AI electronics for predictive maintenance
Sensor electronics, edge AI hardware and firmware that run analytics on site, designed and built in Vaasa.
In short: Comtec Labs designs and manufactures the electronics behind industrial AI: sensor front-ends, edge AI processor boards and the firmware that runs the model on the device, so monitoring and predictive maintenance work on site without relying on the cloud. We integrate and deploy models trained by you or your data-science partner, help design the data collection, and take the hardware from feasibility prototype to series production in Vaasa, Finland.
Who this service is for
The service fits companies that know what they want to detect or predict and need the hardware, firmware and production to make it work in the field. Typical customers are:
- machine builders and manufacturing companies monitoring production lines, motors, pumps and bearings
- energy and infrastructure operators with assets on power grids, in wind power or in oil and gas
- automation companies that want to optimise a process with on-site analytics instead of raw data streams to the cloud
- product and data-science teams that have a model and need a reliable, manufacturable device to run it
How an industrial AI project runs
- Problem and data. Tell us the machines, the failure modes or events you want to catch, the installation constraints and the system that should receive the results. We help you plan what data to collect.
- Feasibility. The idea is first tested on an evaluation board with your own data, before any custom hardware is designed.
- Architecture. We choose the sensors, the sampling, the compute tier and the connectivity that fit the problem, the power budget and the environment.
- Custom PCB and firmware. We design the board with the right processor tier and the firmware with the inference runtime, feature extraction and connectivity stack.
- Model deployment. Your trained model is integrated on the device, together with a deployment and update pipeline so it can be replaced in the field.
- Series production. Boards are assembled and tested in Vaasa, with programming and calibration on the production line.
What we deliver
| Area | What we offer |
|---|---|
| Hardware | Design and manufacture of AI-capable PCBs and PCBAs, integration of edge AI processors and accelerators |
| Sensor electronics | Measurement electronics and analog front-ends for vibration, current, temperature and other signals |
| Firmware | ARM Cortex-M (STM32, NXP, Nordic, Espressif) on bare metal, FreeRTOS or Zephyr; embedded Linux on Cortex-A and i.MX |
| Connectivity | CAN, Modbus, EtherCAT, OPC UA; BLE, Wi-Fi, LTE-M, NB-IoT and LoRa |
| Production | High-reliability PCBA for industrial environments, prototype to series, with programming and calibration on the line |
Edge AI hardware generally falls into three tiers. A microcontroller runs small quantised models for anomaly detection and simple classification on milliwatts. An application processor with a neural processing unit handles camera inspection and sensor fusion on a few watts. A dedicated accelerator runs several camera streams or large models, and then thermal and power design become the main board task. We design the board for the tier your problem needs.
Lead times
For a predictive maintenance sensor node we propose a sensor set, a node architecture and a feasibility prototype, usually within four weeks of receiving your description. Larger projects can start with a fixed-price concept study of one to three weeks under NDA, which gives the architecture, the risks, a development estimate and a unit-cost forecast. Development to a verified prototype typically takes three to six months, depending on scope. Series production then runs on the same line as our other assembly work.
Once the design is released and all components are in house, prototype boards are assembled in 3–5 working days, small series of up to 500 boards in 1–2 weeks and larger series in 2–3 weeks.
Development, test and production
Design, prototyping and production happen in one place in Vaasa. EMC pre-compliance testing is done in-house, so radiated and conducted emission problems are found before a formal test. We build test fixtures and test software ourselves, devices are programmed and provisioned on the production line, and sensor nodes are calibrated on the line before they ship. Assembly runs in an ESD-protected environment with AOI on every board.
See the full equipment list, capabilities and lead times
Quality, compliance and ownership
- Boards are built to IPC-A-610 class 2 as standard or class 3 on request, with AOI on every board and test results stored per board.
- Formal EMC, safety and radio tests are run with accredited laboratories, or radio approval is simplified with pre-certified modules.
- Connected devices get a security plan aligned with the EU Cyber Resilience Act.
- You own the schematics, layout, firmware source and documentation, and there are no licence fees.
Edge or cloud?
Running the model on the device is a constraint decision, not a fashion. These are the usual criteria in industrial systems:
| Criterion | Favours edge AI | Favours cloud AI |
|---|---|---|
| Latency | Control or safety reactions within milliseconds | Seconds are acceptable |
| Connectivity | Remote sites, mobile machines, metered links | Reliable wired network |
| Data | High-rate sensors such as vibration or camera, data that must stay on site | Low-rate telemetry |
| Model | Stable, small, quantised | Large, retrained often |
Most industrial systems end up hybrid: inference on the device, training and fleet-wide analysis in the cloud.
Frequently asked questions
Do you train the AI models?
We integrate and deploy models on the device. Training is usually done by you or a data-science partner, using data from the prototypes we build, and we help design the data collection.
Can a microcontroller really run AI?
Yes, for small quantised models such as anomaly detection and simple classification. Vision and larger networks need an application processor with a neural processing unit or a dedicated accelerator.
How quickly can we test a predictive maintenance idea?
We usually propose a sensor set, a node architecture and a feasibility prototype within four weeks of receiving a description of the machines, failure modes and installation.
Do we need a cloud platform?
Not necessarily. A node can report to a local PLC or gateway over Modbus or OPC UA. A cloud platform is useful for comparing a whole fleet and for training models.
Who owns the design and the firmware?
You do. Schematics, layout, firmware source and documentation belong to you, there are no licence fees, and you are not obliged to manufacture with us.
Related reading
- When Does Edge AI Make Sense on an Embedded Device, and What Hardware Does It Need?
- How to Build a Predictive Maintenance Sensor Node: Sensors, Sampling, Edge Processing and Electronics
- How to Design an Industrial IoT Device: Connectivity, Power, Security and Manufacturability
- How Is Firmware Developed for an Integrated Device? Hardware–Software Co-Design Explained
Related services
- PCB design services
- PCB prototype assembly
- PCB functional testing
- PCB series and low-volume production
- Robotics control electronics
Discuss your industrial AI project
Describe the machines, what you want to detect or predict and the installation. We propose a sensor set, an architecture and a feasibility prototype plan, with a named engineer as your contact.
