@Alfred I used the 2x2 binning. And the deviation was soft with a very good RMS around 0,25 arcsec plus or minus 0,05 arcsec on each axes.
I still use the 2x2 binning but the result is so worst. In fact the amplification of the deviations values are not really very important if they are equally distributed around 0.
The value of deviation is mostly contained in plus or minus 1.5 arcsec. The main problem before reducing the deviation are those huge shots of more than 3 arcsec that didn't exists on previous autoguiding processes.
For calibration I had to increase the pulse duration to 5000 to have a decent move of the mount overall in AR (2000 with previous version).
Under this value, the calibration curve looks to nothing good.
This didn't changed anything on the guiding process. It remains as bad as with a less calibration pulse of 2000.
For the moment I am fully lost in adjusting the values of the parameters for calibration and guiding.
As I don't have the chance to benefit of good weather conditions very often, I guess that will take a while to find the good values.
Maybe I will get back temporarily to PHD2 to improve the autoguiding process.