Repository navigation
Description
Split from #58, as agreed. That PR fixes how PoSER pairs, completes and rescales the modes of the single setups, but it deliberately leaves uncertainties as they are today. This issue tracks what to do with them.
Current state
PoSER (MultiSetup_PoSER.merge_results())
- Each single setup can compute uncertainties with SSI (
calc_unc=True). Aftermpe, every setup providesFn_std,Xi_stdandPhi_stdfor each selected mode. - The merge ignores all of them. The merged
Fn_std/Xi_stdare just the scatter of the values between setups (population standard deviation). So:- a very uncertain setup counts as much as a clean one in the mean;
- with Fix PoSER merging: mode pairing, missing modes and mode-shape scaling #58, a mode found in one setup only gets
Fn_std = 0, even when that setup has its own σ; - two setups that happen to agree get a small
Fn_std, whatever their σ.
- The merged result has no mode-shape uncertainty (
MsPoserResulthas noPhi_std).
PreGER (MultiSetup_PreGER + SSI_MS)
- Uncertainties are never computed.
SSI_multi_setupbuilds the Hankel matrix withcalc_unc=False(src/pyoma2/functions/ssi.py), and theSSI_MS.rundocstring states that the uncertainties areNonefor multi-setup.
Points to decide
-
PoSER frequencies and damping ratios. The first version of Fix PoSER merging: mode pairing, missing modes and mode-shape scaling #58 used the inverse-variance weighted mean when every setup that identified the mode provides σ:
$$\bar f = \frac{\sum_j f_j/\sigma_j^2}{\sum_j 1/\sigma_j^2}, \qquad \sigma_{\bar f} = \Big(\sum_j 1/\sigma_j^2\Big)^{-1/2}$$ Example: 10.00 ± 0.02 Hz and 10.20 ± 0.50 Hz give 10.0003 ± 0.020 Hz, instead of the current 10.10 Hz.
The weakness: σ of the weighted mean ignores disagreement between setups. With 10.00 ± 0.01 Hz and 11.00 ± 0.01 Hz it gives 10.50 ± 0.007 Hz, although the setups differ by 1 Hz. Options:
- weighted mean + σ of the weighted mean;
- weighted mean + weighted scatter between setups;
- weighted mean + σ of the weighted mean inflated when the scatter exceeds what the σ's explain (e.g. by the Birge ratio);
- something else you prefer.
-
Meaning of
MsPoserResult.Fn_std/Xi_std. Today they are the scatter between setups. If they become the uncertainty of the merged value, should the scatter be kept in separate fields, so neither meaning is lost? -
Mode shapes. Should
Phi_stdbe propagated to the merged mode shape? It would need to account for the uncertainty of the reference sensors, since they define the scale factor applied to the roving part of each setup. -
Incomplete information. What to do when some setups have no σ, or σ is 0 / NaN for a mode (for example after the CoV hard criterion)? The prototype falls back to the plain mean and scatter for that mode.
-
PreGER. Is multi-setup uncertainty for
SSI_MSin scope, or should it be a separate issue?
Starting point
A first implementation of option 1, with tests, is on the branch wip/poser-merge-uncertainty: it is the first version of #58, before the uncertainty part was removed.
Activity