I have seen similar on my own line in the past. It looks like youve had a noise burst. The separation is something to do with bitswap and gain. There's usually a bit of tx power left in reserve for the bitswap process, which can increase gain to raise the SNR a little bit sufficient to cover the bit loading.
When one of the subchannels has a reduction in SNR then it's not unusual for one of the other subchannels to see an increase. The increase in U2 never seems to be as much as the decrease in U0 or U1 though.* When I had a line fault last year they would even cross over each other and then back again.
The first graph is the mean average of all the tones across all three subchannels. The second graph is the mean average for each of the subchannels.
*U2 carries far more tones (@810) than U1 (@334) than U0 (@25), so because the mean average is used, then it will take less tones in U0 or U1 than in U2 to skew the difference. I know what Im trying to say, but its late and Im not doing a very good job of explaining it, other than say it would only take a couple of tones to be below par in U0 to make it show a greater difference than it would for the corresponding increase same number of tones in U2 to affect the mean average.. which will in turn correct the total mean average.
ahhhhh.. sorry can some one explain it better re the mean averages and why it graphs like that, yet still be correct?