Evaluation of core loss and frequency information extraction

[hvater]

Hello everyone,

I have two questions:

1. In the README.md it is written that the core loss estimation of our models will be evaluated as one of the judging criteria. As far as I know we can only accurately estimate the core loss for complete (major) BH-loops. Does that mean that there will be specific test cases in the test set where there are full BH-loops so that this evaluation can be performed? From the task description I understand that the models we are building should work for incomplete loops as well. Are we are allowed to use the information of whether the loops are full in inference? Or should we assume that we cannot know about the "fullness" of the curve?

2. We can theoretically extract the frequency of the PWM from the excitation signal. Are we allowed to use that information? For the data that is given this is likely helpful, but I understand that the models should work for arbitrary excitation signals. Should we assume to not know about the excitation signal shape for the purposes of the challenge?

I guess for both questions, the underlying question is how it is judged if information that is present from the measurment setup is withheld from the model. Withholding information will likely lead to poorer accurary but I guess it would result in more general models that are less specific to the presented measurement setup.

Cheers,
Hendrik

[minjiechen]

Thanks for the question ~ Answers:

• The testing data will be similar to the training data, so you should not assume a full complete loop. In fact we are still debating how to formulate a "fair" core loss evaluation given the loops are incomplete, and we may completely drop the core loss evaluation and only count on time domain signals. We will host a webinar + Q&A to make the evaluation criteria clear in September.

• Feel free to extract the frequency information if that helps with your prediction (and if your algorithm learnt that from the data it is benefitial to do so). Although, extracting frequency information reduces the generality of your algorithm.