We finally found one of the possible causes of the problem.
Analyzing the SDA and SCL signals with an oscilloscope we noticed that the rise time of both signals was near to 600 nsec.
Changing the pullup resistors from 4.7K to 1.0K we have better signals (very deep borders) with a raise time below 200 nsec.
The communication is more stable and the very few bus frozen situations can be recovered with the I2C reset procedure.
