We use AI for leak detection
We are fully making the switch to natural refrigerants. Until that conversion is complete, we will minimise as much as possible the emissions from our existing installations that use synthetic materials. We use artificial intelligence to detect leaks at an early stage.
Detecting leaks
Across all our stores and central buildings, we manage around 2,200 individual refrigeration units. A big part of those currently still runs on synthetic materials. Every year, about 4% of these harmful substances leak out, which harms our environment. At 4%, we do better than others do in our sector, where leakage rates are sometimes as high as 20%.
Replacing all those installations with alternatives that work with natural refrigerants is no easy feat. Until then, we are doing everything we can to avoid leaks and detect them as quickly as possible. Easier said than done. Detecting leaks is traditionally very labour-intensive: a technician has to go round the entire installation with a detection device. There had to be a better way, and so our engineers decided to look for a faster way to detect leaks, which would allow us to send out technicians faster to locate and repair them.
Quick and targeted interventions
Our goal was to develop an automated leak alert system that would allow our technicians to intervene in a quick and targeted manner. Project engineers Margo Parmentier and Margot Van den Brande immediately encountered a first problem: you could not tell from the installation's operational parameters whether or not there was a leak. A failure only occurred when the buffer tank - a usually fairly large storage facility that absorbs fluctuations in the quantity of liquid refrigerant - was completely empty. At that moment, a significant amount of refrigerant had already escaped.
Margot Van den Brande: "A logical first step is to simply monitor the level of refrigerants in the buffer tank. That is why we equipped some installations with level gauges. Those tests showed that the procedure was too complicated to lead to reliable results. To monitor all our installations efficiently, we had to find a simpler method."
Working with sensors
The buffer tank contains two simple sensors that give a signal when the level of refrigerant is higher than the sensor: one at the bottom and one halfway down the tank. With the sensor halfway down, we can get an idea of the evolution of the amount of refrigerant, provided it contains the correct amount. After all, when it contains so much refrigerant that the level never drops below half, the sensor is of little use. After consulting with our technicians, we found that the amount of refrigerant could be adjusted accordingly.
Now, the team disposed of a series of signals that indicated whether the refrigerant level was above or below the halfway sensor. The next thing that we needed to figure out was what this said about the condition of the installation, and whether a leak could be identified from this. The latter is a crucial point. If the alarm is not sensitive enough, a lot of refrigerant will escape before anyone notices. However, if you set it too sensitive, false alarms will be generated.
Artificial intelligence
That is why we used artificial intelligence to train a model that predicts the expected level of refrigerant in the buffer tank. This took into account several parameters that influenced the operation of the installation, including the outside temperature, the opening hours of the store and the scheduled defrosting cycles (which are always carried out at fixed times).
We validated this model using the installations that were already equipped with a level gauge. Based on that set of parameters, the model predicts how the level in the buffer tank will fluctuate. If the measured values deviate too much from the predicted ones, an alarm is raised. Because each installation is different, we need to train a separate model for each installation. Only after a week, we will have data that is sufficiently varied.
From theory to practice
A promising model, which turned out to be a major success in practice as well. The detection system raised an initial alarm, and after inspection, the leakage rate was 3.66%. Well within the target of detecting leaks below 5%. This excellent start facilitated the roll-out of the system to all installations involved and gave confidence to our technical teams.
Project engineer Bram Neckebroeck: "It's important that we don't send out our technicians unnecessarily. After all, detecting a leak costs many man-hours. On that level, the system offers a lot of added value: whenever it sounds an alarm, there effectively is a leak. We have already detected around fifteen leaks at an early stage. Therefore, whenever we now receive a signal from the detection system, that intervention is immediately given top priority.
Internal knowledge and experience
A key factor in the success of this project was our large internal technical department. All knowledge and experience regarding the operation of our installations is available in-house. Margo Parmentier: "Feedback from the field enabled us to develop a system that was also workable. The refrigeration team makes it possible to test things, like the installation of level gauges in the buffer tanks."
In the end, this was not the solution, but it did point us in the right direction to find something that would work. Margo: "Their expertise was also important for adjusting the refrigerant content of the installations, so we could use the existing sensors. The input of the technicians was crucial in achieving an effective system. Meanwhile, we are on track to reach our goal: halving the leakage loss by the end of 2021."
