Desenvolvimento de redes neurais artificiais para classificação de ensaios não destrutivos de ultrassom em PVDF

Abstract

The use of nondestructive (NDT) ultrasound (UT) test is widely diffused in industry. The application of this technique to polymers is, however, almost nonexistent due to the difficulty in interpreting the results. Polyvinylidene polyfluoride (PVDF) has, among many applications, been used as a sealing layer in risers in the petroleum industry. One of the problems is precisely the inspection by nondestructive testing of this material. A certain polymers fenomenon is called whitening, which follows the deformation of the material, and thus can be used as a fault indicator. Aiming for a solution to this problem, it is proposed in this work the use of artificial intelligence in the classification and detection of whitening in PVDF specimen inspected by UT. The obtained data were analyzed and processed, choosing to use only the backscattered portion of the signal. The Fourier and discrete cosine transforms were used, in addition to the Savitzky-Golay filter in the processing of the signals before the feeding the network. The technique used is that of artificial neural networks as a multilayer perceptron. The training was done using the Python language and its machine learning libraries, such as TensorFlow and Keras, resulting in a classification network with more than 95% of accuracy.