Thanks for attaching three of your MD structures -- the minimized one looks amazing!
I've visualized the three structures in Jmol, and noticed quite significant distortions in the 50ns and 100ns structures. DSSR is performing as designed, even though the results are not what you expected. When a structure is distorted, the boundary between different helices are no longer clear. DSSR uses a geometric approach to identify stacking regions of base pairs. Using your 50ns structure as an example, helix#5 is a concatenation of two helices because of the stacking of DA240 on the DG95--DC242 pair. On the other hand, some DSSR-derived helices in 50ns are quite short. Overall, if you can visualize a fragment as helix-like in Jmol/PyMOL, DSSR should be able to find it.
Are the 12 helices a requirement during your MD simulations? Your 50ns and 100ns structures show that some pairs are broken or formed in the process. Moreover, the single-stranded connecting regions are not always 3 nucleotides.