Plugging and Abandonment (P and A) operations must be carried out at the end of the well s life cycle, to guarantee the preservation of the environment. An essential step in P and A operations is the evaluation of cement quality, done through acoustic logging techniques based on the propagation of sonic or ultrasonic waves, where the integrity of the cement layer and its sealing capacity are verified. Currently, this procedure requires prior removal of the production column, which increases the costs and complexity of these operations. Thus, it is of great interest to the oil and gas industry to develop tools and techniques capable of performing cement evaluation without removing the tubing. Therefore, this work aims to apply feature extraction methods to process signals obtained through acoustic logging experiments in oil wells with the presence of the production column, and then use the resulting features to train and test neural networks of the Multi-layered Perceptron (MLP) type. Promising results were obtained for the defect classification task, with an accuracy of almost all methods above 0,8 and an AUC above 0,9.