I'd define the stable states T1 and T2 by a moving box with an equilibration period and a delta. When the box was entered on the right side and left on the left side, the graph retrospectively is defined as stable on entering. When unstable, the box is exited on top or bottom. The delta could be a relative standard deviation for noise and maybe also a linear regression for drift.

The classic criterion for transition is the second derivative, but that heavily suffers from noise.