If the dropouts are truly time coincident can't you just apply a linear discriminator to the bottom plot there? Figure out where the threshold is, is it a vertical line at 70.355? Or a somewhat vertical line with a bit of a drift to the right at the bottom? Once you've got the equation for your line then anything to the left/bottom of that line is a dropout.
The data looks a bit noisy, but maybe you could also do some clustering with k-means (if you know how many dropouts and where your data is) or dbscan and reject the clusters with values that correspond to the dropouts.
You might want to normalize the data first, by the mean of each and the standard deviation, like a z-score.