Model Reduction - Export matrices from ANSYS to MATLAB

So my first question would be - Should I build my flexible body with deformation model for each part (say it is a table, should I describe every leg of my table and the plate using rigid body and deformation model)?

That would be a reasonable approach. You haven't told us much about what you want to do with your model in Simulink, but as a general principle, it's easier if you retain any positions where you want to apply things like loads or constraints to the model as master degrees of freedom, even if they are not geometrically actually on the "boundary."

That might mean you need to keep some master degrees of freedom on the table top, in addition to the interface to the legs, if you want to model what happens when you apply a load at those positions.

An alternative would be to model the whole table top plus legs as a single object, with the master degrees of freedom as the ends of the legs (presumably they will be constrained to be resting on the ground, in the Simulink analysis) and any other positions you are interested in.

Concerning model reduction I am really not sure what to do - should I use Component Mode Synthesis Method - Fixed-Interface Method (CB Method)?

I would recommend that for your first attempt, definitely. It is conceptually simpler than the alternatives, and it automatically avoids some "elephant traps" where load paths through stiff parts of the structure (in your model, axially along the legs, for example) are not captured as part of the low-frequency internal modes of the model, but they are guaranteed to be included in the "rigid body" part with the fixed-interface method. Leaving those stiff load paths out of the model may produce "plausible looking complete nonsense" - which is the worst type of error to have, when you are inexperienced in what you are doing!

Should I generate superelements for every part selecting master degrees of freedom on the interfaces connecting to the other parts of the structure?

Yes, that is implied by my answers to above. Alternatively, treat the complete table as a single superelement.

If yes, what is then my next step?

Basically, just set up the model and run it. You should be able to find a similar tutorial example somewhere.

An important decision you need to make is how many internal modes to include in each component. One strategy to answer that is to include all the internal modes below some cutoff frequency, which depends on what frequency range you want to use the final model for.

If you don't have any more problem-specific advice available, set the cutoff frequency for generating the component models to about 3 times the highest frequency you are interested in. That may mean you get many more modes in the table-top model than in the legs - but don't worry too much about that, unless the reduced model is still too big to work with when you assemble all the superelements.

Another issue is how to check your reduced model is correct. One fairly simple strategy is to do two vibration analyses in Ansys, on the unreduced and reduced models. Check the mode frequencies and mode shapes match up, and in particular check the reduced model contains all the modes of the unreduced model, in the frequency range you are interested in.

If you were doing this sort of modelling repeatedly, it is possible to automate some of the checking with software, but for a first attempt it's probably better to do it hand, comparing the mode shapes visually - though it may get tedious if you have a large number of modes.