PhD in Mechanical Engineering · C-MAST, UBI · 2026

A sensor that knows when to defrost

Fridges and cold rooms defrost on a timer, whether there is frost or not. My PhD was about a cheap sensor that notices the frost instead.

The same evaporator four times: clean, then with low, moderate and heavy frost
The same evaporator, from no frost to a lot of it.
Thesis
Frost Detection in Refrigeration Systems through Sensor Technology Integration.
Supervisors
Pedro Dinis Gaspar and Pedro Dinho da Silva.
Patents
PT 116073 and PT 119851 came out of this work.
Licence
The thesis is open, CC BY 4.0.

First, a place to make frost

To test a frost sensor you need frost on demand. So I built a small wind tunnel with its own evaporator, where humidity, air speed and temperature could be set and repeated.

The test rig: an acrylic wind tunnel with the evaporator, coolant lines and electronics
The test rig.

A bit of cotton

The sensor measures resistance between two electrodes. Water and ice change it. After air gaps and ceramics, the design that behaved best had its electrodes wrapped in loose cotton, which holds the water where the sensor can feel it.

Microscope view of cotton fibres around the sensor electrodes
The cotton, under the microscope.
The fabric sensor beside a millimetre ruler: a thin cable with a cotton-wrapped tip
The whole sensor. About 30 mm.
The sensor between the evaporator fins, first dry, then covered in frost
In place between the fins: dry, then frosted.

Then a printed one

A second version is printed in conductive TPU, so it can be made in one go and shaped to the fins. It became its own patent application.

The printed sensor between the fins, with light condensation and then heavy frost
The printed sensor, with condensation and then with frost.

Checking it with a camera

To know whether the sensor was right, I needed an honest referee. A camera and some computer vision split the evaporator into tiles and measured the frost in each one. The same maps showed where frost forms first, which is where the sensor should go.

Black-and-white image of the evaporator with a 12 by 12 grid of tiles
Frost, tile by tile.
Stacked grey tile maps over time, flattened into one frost map
Many moments flattened into one map.

No batteries

The cold that makes the frost can also power the sensor. A thermoelectric heat exchanger with four Peltier cells gave between 0.20 W and 0.93 W in the tests, more than the sensor needs.

The thermoelectric heat exchanger connected to the rig
The thermoelectric heat exchanger in the rig.
Exploded render of the four Peltier cells between machined aluminium plates
And taken apart.

Did it work?

Mostly. It caught the start of frost every time. Telling when defrost had finished was harder: right 58% of the time, and every time when the sensor was in the right place. Which is why the camera maps mattered.

Where I would take it next

On the home page I put myself between research and industry. From there, my honest view: the sensor was the research, but for real fridges and cold rooms I would push the camera. It sees the whole evaporator instead of one spot, it does not get in the way of the airflow, it can tell when maintenance is due, and it leaves room for much more capable AI. The computer vision was part of this PhD too. That takes nothing away from the sensor; the camera simply has more room to grow.

None of it was done alone: thanks to my supervisors and to the people at C-MAST.

Back to all work