Hm, what about CG321 CG321 CG205 HGA2 improper?
Weird, but I didn't find the CG321 CG331 CG204 HGA2 improper for the aldehydes...
Do not introduce impropers for sp3 centers!
Not even all sp2 centers have impropers in CHARMM. You should have exactly 1 improper parameter in cyclohexanone, which should be explicitly defined both in the topology and in the parameter.
Make sure you thoroughly understand the meaning of the different terms in the force field's potential energy function. Additionally, it wouldn't hurt to read some of Alex MacKerell's force field papers to make yourself familiar with CHARMM's parameterization philosophy.
Also, what if to take some of the dihedrals from all35_ethers FF?
We did that already when building CGenFF. As a result, Alex and I together spent many weeks if not months eliminating side-effects. You see, in force field parameterization, dihedral parameters typically are the final term optimized, thereby accounting for the contribution of long range non-bond interactions to the PES. In this light, you may understand that you're just asking for errors in your force field. And good luck explaining to your papers' reviewers why you chose to play mix-and-match with force fields whilst having a force field at your disposal that has all the parameters you need and is expressly designed for the class of molecules you're studying...
What you could
do is copy the "outer" dihedral parameters from other general force field molecules. Specifically, CG2O5 CG321 CG321 CG321 and CG2O5 CG321 CG321 HGA2 can respectively be copied from the ester's CG2O2 CG321 CG321 CG321 and CG2O2 CG321 CG321 HGA2. It's not ideal, but it may do. You will probably even get away with copying CG321 CG2O5 CG321 HGA2 and OG2D3 CG2O5 CG321 HGA2 from acetone's CG331 CG2O5 CG331 HGA3 and OG2D3 CG2O5 CG331 HGA3. But you can hardly escape doing a potential energy scan on butanone for CG321 CG2O5 CG321 CG321 and OG2D3 CG2O5 CG321 CG321 ! This is actually not a bad deal. You don't need to optimize any charges, bonds, angles or impropers; all you have to do is one potential energy scan on a very simple model compound. Following the procedure and the example scripts in the tutorial, this should really be a breeze. If you want, I can even take a look at your toppar file once you're finished.