Thread Subject: System identification - preprocessing data

Hello to all,

I'm wondering about significance of preprocessing data before
identification. It is generally recommended to detrend data (remove
mean and trends) before identifications. I have read that detrend is
not good in cases where "there is an integrator in the system, and the
input and output levels are essential to describe the effect of the
integration".

Now, I'm trying to apply this on my particular problem. I have data
from integrating process (level in vessel which is under vacuum). It's
from control level loop. Level is controled by manipulting control
valve (output flow). I have recorded a position of control valve
(0-100%) and level in the vessel. Controller is manipulating control
valve so it can be used as an input to the process.
I have made step test (open valve from 30 to 40% and then back to 30%)
and recorded output.
The real question here is should I use detrend data?
I have noticed that if I use detrend data (detrend function with
parameter 1, say detrend(x, 1)) I get much better model fits to
process model of first order plus dead time model. That is good, but I
don't know if I made mistake when detrending data. Percent of data
fitting to a model is much lower if data is not preprocessed with
detrend function.
Vesel is generally and object of integrating type, so the question is
should detrending be used at all.

What is your experience?

Thanks