Wind turbines undergo dynamic loads along all the phases of transformation of wind kinetic energy into power output to be fed into the grid. Gearbox breakdowns are one of the most common and most severe causes of energy losses and it is therefore crucial to prevent and forecast them. Straightforward vibration analysis is very demanding by the point of view of technology, costs and complexity of signal denoising. A considerable keystone in fault diagnosis is the analysis of Supervisory Control And Data Acquisition (SCADA) systems. In particular, thermal behaviour of wind turbines fits well with the common time scale of SCADA data; heating trends are fairly responsive as a consequence of rotor vibration. Machine learning techniques applied to SCADA data are very powerful in reconstructing inputs – output dependency. On these grounds, in this work an Artificial Neural Network approach is proposed for early diagnosis of gearbox faults. The method is validated on the data of a wind farm operating in Italy. It is shown that the method is capable in recognizing incoming faults with a very manageable advance also with data on short time scales.