#13 🚀 A True Story of an Artificial Intelligence in Power Electronics
The Newsletter of High Frequency folks
👋 Hello! Dr. Molina here! 👨🔧
🤗 Thank you for reading and welcome to the show my Newsletter. I’m Chema Molina, Founder, and CEO at Frenetic, where we train and create Great Magnetic Engineers🚀.
Today I won´t talk. I have invited our Artificial Intelligence team to tell you how we started our Artificial Intelligence. The whole team of AI has been involved, but the main author has been Noelia Jiménez.
👉 Since the A.I. is driven by data, as much quality data you have as better results you can provide. 😍
✅ The price of our service is like the entropy, it only increasing because our A.I. software and our team are better every month.
💵 The sooner you are IN, the better magnetics you will have.
Frenetic first-born AI model: Leakage Inductance
By Noelia Jiménez (aka Noe)
This is the short story of the first AI model hosted by Frenetic Online. It was September 2020 and after several experiments, we had the intuition that breaking the magnetics design process in several small modules and applying AI could be the way to attack and defeat this problem. So, we gave ourselves three months to get the first successful proof of concept of one of these AI modules. Leakage Inductance was chosen first due to the difficulties in modeling this magnitude with complex behavior with equations.
This done, the first and usually one of the most painful questions when approaching an AI model came on stage: we needed a training dataset so our model could learn from reality. This took us to join forces as a team to get the first dataset in our lab (thousands of data rows gathered with sweat and tears) to train-test and validate our models. In a few weeks, we had our first dataset. This data acquisition was expensive in terms of time and money. In addition, after doing some data cleaning, around 10% of the rows had to be discarded due to human errors in data recording or measuring. In a relatively small dataset, this amount of measures would have had enough weight to spoil our models.
First models were generated, having as inputs plain constructive parameters of the magnetics, that is number of turns, wire widths… This first version had no stratification (all magnetics types were together in one single model), we were still a bit lost in feature engineering issues (what are the inputs that could give us the information we need?) and basic things such as several parallels or multi-winding were still far from being covered. Unsurprisingly, the results were not that good, even if we tried to do some feature engineering on our own.
Next step for the AI team was asking our magnetic experts to refine the approach we were taking. This way our models started to be separated into one and two windows, which was the first successful stratification. Then some feature processing was done, so information such as interleaving could be included, using variables such as maximum horizontal magnetomotive force (MMF). This was followed by models separated by type of wire and other considerations. During this first process of tests, we also tried several algorithms that could overcome the limitations of simple regression. Finally, our beloved XGBoost Regressor won by far, since its many configuration possibilities (some will be mentioned later on), speed, and accuracy found no competitor.
Then, more measurements were taken in our lab to include magnetics with several parallels as part of the space covered by our models.
By December 2020, we had achieved the first functional version of Leakage Inductance AI models (multi-winding not included). We had made it!
In 2021 we kept improving this module in terms of accuracy and coverage.
Accuracy was improved following three strategies. The first one was including new constructive parameters that affected Leakage Inductance that had not been taken into account so far: insulation distance between windings and wire extra length. The second strategy attacked one of our big problems: lack of data in certain regions of our inputs. We solved it by including analytical data in the recipe (the way to get to a good enough analytical model would be another article itself, credits to our CTO Alfonso Martínez). These data were included giving them weighted importance inside the model (the chosen algorithm, fortunately, offers this option). The mixed-data models allowed us to have good results in the whole covered space, avoiding crazy results due to gaps in the space covered. The third improvement included was, again due to the possibilities offered by our algorithm, forcing some dynamics that were known in the behavior of Leakage Inductance. That is, for instance, if the distance between windows grows, Leakage must grow if we keep the rest of the constructive parameters. Our models were trained to take these dynamic known rules into account.
So, with all this, we have done several iterations in train-test and validation and arrived at the current models. We present here an example of the results yielded by our models for some two windows, Litz wire with several parallels samples:
These models’ predictions have as inputs:
number of turns
number of parallels
wire widths
wire heights
number of strands
wire type
number of layers
winding width
winding height
mean length turn
wire packing factor
number of stacks
filling factor
distance between winding windows
maximum horizontal MMF
wire length
extra length
insulation distance.
Magnetics with one or two winding windows, with litz, round, or foil, and windings with several parallels are already included. Accuracy decreases in extreme cases though.
In 2021 multi-winding models were developed following a similar process (with all the previous knowledge). These models give Leakage Inductance for each secondary winding versus the primary winding. For this, new input variables such as the distance between the evaluated secondary and primary were taken into account.
Here are shown some of the predictions for multi-winding, one window with litz wiring case:
Other achievements of our 2021 were making our data acquisition more professional to avoid as much as possible human errors in the process. This includes automating measuring and data recording as much as possible, as well as making protocols to ensure replicability. Improving the quality of our data will allow us to have better models.
The story of this first-born AI model is still to be continued. We will keep looking for ways of improving our accuracy and start covering some options still missing, such as Leakage Inductance for toroids.
Top finds🧑🏽🌾
Book: Sapiens
Articles: DeepMind’s last publication
Thank you again for reading and don´t forget to subscribe.
Sincerely,
Chema 👋