The literature on cointegration in large datasets or panels is really the only place where I've seen this sort of issue discussed. Breitung and Pesaran, among other places, talks about it.
I would recommend applying the PCA to the rate changes (perhaps with some kind of zero lower bound adjustment). Then, take the cumulative sum of each of the factors. This is effectively like the cumulative factor you would be using in a cointegration test or factor error correction model. Next, regress the rates (in levels) against the cumulative factors. At this point, you can test the residuals of this regression for stationarity (I would expect the residuals to be mean-reverting, there are a lot of nuances to panel unit root testing that I'm ignoring).
You have a number of options when creating an investment strategy based on this information. The most obvious is to just buy (sell) the bonds where the residuals are most positive (negative). Highly positive residuals indicates that the yields are more than what the model would suggest and should fall. Obviously, there are a lot of variations you can make on this.
Another alternative is to estimate a factor error correction model. A factor error correction model bears a lot of similarity to the approach described in the second paragraph, but just extends it a little. Once you estimate the model, you can forecast rates, and use that to determine your bond positioning.