A neural network approach for the detection of the locking position in RFX

In the FFX (reversed field experiment), one of the most important reversed field pinch (RFP) devices in the fusion community, wall locked modes have always been present. Recently, a new technique has demonstrated the possibility of inducing a continuous rotation of the modes with respect to the wall. The non-linear coupling of the m=0 and m=1 modes has been used to decouple the modes themselves, and the mode rotation has been induced by means of a pre-programmed waveform of a toroidal magnetic field rotating ripple. Consequently, a feedback system for detecting the locked mode position along the toroidal co-ordinate and able to create a continuous rotation with variable speed has been envisaged. Neural networks (NNs) represent a promising approach for rapid detection of the locked mode angular position in such a system, and in this paper the performances of different NNs trained to identify the locked mode position are compared and discussed. In particular, their robustness to noise is analyzed, and it is shown that NNs provide reliable results, sometimes better than those computed with fourier analysis.