I looked at the article you referenced. They appear to be feeding four parameters into a neural network: voltage@f1, current@f1, voltage@f2, and current@f2 (where f1 is the 'natural lag frequency' of their 'resolver', and f2 is #poles*freq/2).
I know you probably are already doing this, but it bears checking: Make sure that you are using the correct voltage differential for your calculations that is appropriate for the 3-phase configuration you are using (e.g. delta or star). The last time I touched on this stuff was ages ago in school, so I won't try to get to much into details that I'd probably get at least partly wrong.
I also noticed that your data is a 1000x3 matrix. From the paper you referenced, I would have expected it to be 1000x6 (3 voltages and 3 currrents). It also wasn't clear from your data what the sampling rate was used. The paper you referenced was sampling at 100kHz, but it seems like your data was probably sampled at something much much lower (which would prevent you from distinguishing the frequencies specified in the paper).
TO SUMMARIZE: 1) Make sure you are sampling at a high enough rate for your metric of interest. 2) Make sure that you are measuring everything you need to calculate your metric your metric. 3) Make sure you are interpretting your data correctly (i.e. correct voltage differential/reference, correct filters for your sampling rate, etc.)
Good luck, good cheer, god bless, and may success smile on your efforts.
