Short answer: A predictive maintenance sensor node has five design decisions: measure the physical quantities that precede the failure (vibration for bearings and imbalance, motor current signature for electrical faults, temperature for lubrication and insulation, acoustic emission for cavitation and leaks); sample fast enough to see the fault frequencies (10–20 kHz for rolling bearings); compute features or run an anomaly model on the device so only exceptions are transmitted; choose connectivity and a power budget that match the installation (wired and machine-powered, or wireless with a multi-year battery); and ruggedise the electronics for the environment with sealing, conformal coating and an industrial temperature range. Comtec Labs designs and manufactures such nodes, including the sensor front-end, the edge processing firmware and the series production.
Which sensor detects which failure
| Failure mode | Sensor | Sampling and processing |
|---|---|---|
| Bearing wear, imbalance, misalignment | 3-axis MEMS or piezo accelerometer | 10–20 kHz; FFT, envelope spectrum, RMS, kurtosis |
| Electrical faults, rotor bars, load anomalies | Current transducer on motor phases | 5–10 kHz; motor current signature analysis |
| Lubrication, insulation, overload | Contact or IR temperature sensor | 1 Hz; trend and rate of change |
| Cavitation, leaks, valve faults | Acoustic emission / ultrasonic microphone | 50–200 kHz; band energy, on-device model |
| Process drift | Pressure, flow, humidity | 1 Hz; thresholds and trends |
Design the node around the data path
- Analog front-end: sensor bandwidth, anti-aliasing filter and ADC resolution decide whether the fault is visible at all; a 1 kHz accelerometer path cannot see a bearing defect.
- Edge processing: compute spectra and features on a Cortex-M microcontroller and transmit a few hundred bytes per interval instead of raw waveforms; add an anomaly model when the normal state is well defined.
- Connectivity: wired (Modbus RTU, CAN, Ethernet, IO-Link) when the machine has power and a cabinet; wireless (BLE, LoRa, LTE-M/NB-IoT, Wi-Fi) when cabling is impractical.
- Power: machine-powered 24 V with protection, or battery with duty cycling; a node that samples for 2 seconds every 10 minutes and sleeps in between can run 5+ years on a single cell.
- Ruggedisation: IP67 enclosure, conformal coating, -40 to +85 °C components, surge and ESD protection on every external line.
What we build
Comtec Labs designs the sensor electronics, the microcontroller board with the analog front-end, the firmware with feature extraction and anomaly detection, the connectivity stack and the production test, then manufactures the nodes in series with calibration on the line. This is the electronics behind our industrial AI and predictive intelligence offering for production lines and energy infrastructure.
Key facts
- Rolling bearing defects show at frequencies up to several kHz; sample at 10–20 kHz or you will not see them.
- On-device feature extraction typically cuts transmitted data by 99 % compared with raw waveforms.
- A duty-cycled wireless node can exceed five years on one lithium cell.
- Comtec Labs delivers the node, the firmware and the series production from one team in Vaasa, Finland.
How to start a predictive maintenance node project
Tell us the machines, the failure modes you have seen, the installation constraints (power, cabling, enclosure) and the platform that should receive the data. We propose a sensor set, a node architecture and a feasibility prototype, usually within four weeks, and take it through to series production.
Frequently asked questions
Can one node cover several failure modes?
Yes. A typical node combines a 3-axis accelerometer, a temperature sensor and optionally a current input, with the firmware computing separate features for each.
Wired or wireless?
Wired where the machine has power and a cabinet nearby: it is cheaper and more reliable. Wireless where cabling costs more than the node or the machine moves.
Do we need a cloud platform?
Not necessarily. The node can report to a local PLC or gateway over Modbus or OPC UA; a cloud platform helps for fleet-wide comparison and model training.
How accurate is on-device anomaly detection?
Good enough to flag deviations from a learned normal state; classification of the exact fault usually needs spectra analysed by an engineer or a trained model.
Ready to reduce PCB surprises?
Comtec Labs offers a full suite of services to streamline your workflow:
PCB design service
PCB prototyping service
PCB component sourcing
PCB component assembly
PCB testing service
PCB repair and modifications
Printed circuit board production
PCB mass production