The article proposes a methodology to determine appropriate descriptors for the design of a classifier based on neural networks that allow the detection of loss of material in metal pipes based on nondestructive testing signals of Magnetic Flux Leakage (MFL). For this it has been proposed a method which consists of two stages: the first, corresponding to the signal processing, the Discrete Wavelet Transform (DWT) transform is used to implement a nonlinear threshold filtering or Shrinkage and correction baseline which seeks to eliminate or mitigate the different types of noise or phenomena found in the signal that make difficult the process of extracting relevant information to the subsequent detection of loss of material. In the second, corresponding to the design of the classifier, it seeks to identify a window width and descriptors in the time domain and the Power Spectral Density (PSD) to characterize the signal and differentiate areas of metal loss or no metal loss.