CONICET | Buscador de Institutos y Recursos Humanos

Ground Penetrating radar (GPR) is a method widely used to study the near-surface subsoil. Many GPR applications require the acquisition of large volumes of data. In these cases, the processing and analysis of the data involve considerable amounts of time and human effort, and the possibility of errors increases. Considering this, the implementation of dependablemethods for the automatic detection of GPR response-patterns of the targeted structures becomes clear, because they can contribute to the efficiency and reliability of the interpretation.In this work, we present three methods for automatic detection of pipe-flange signals in constant-offset reflection-GPR images. These methods were obtained using well-known supervised machine learning techniques, and data acquired during a previous study of an extensive section of a pipeline. The first two methods are based on support vector machines (SVM), combined with the image descriptors local binary patterns (LBP) and histogram of oriented gradients (HOG), and the third, on artificial neural networks (ANN). The training and validation of these types of algorithms require large numbers of positive and negative samples. From the mentioned study, we had only 16 experimental flange-patterns. Then, in this work, they were taken as references, together with available documentation on the geometry and materials of the pipe and flanges, for building a broad database of synthetic patterns corresponding to different depths of the pipe and characteristics of the environment. These patterns constitute the set of positive samples used for training and validation. They were also used for the final test of the algorithms. The negative samples for the three stages were directly extracted from the profiles.The results obtained indicate the usefulness of the proposed methodologies to identify the flanges. The best performance corresponded to the ANN, closely followed by SVM combined with HOG, and finally SVM with LBP. In particular, the ANN provided rates of false positive (FP) predictions for the validation and test samples of about 3%, and rates of false negative (FN) predictions of 1.67% for the validation samples and 18.75% for the test samples. Greater FN rates for the test experimental samples, in comparison to those obtained for the validation synthetic samples, were also observed for both SVM algorithms. The detection failures mainly originated in that some complex features of the experimental flange responses could not be appropriately reproduced through the performed numerical simulations, and therefore, some of the patterns were not satisfactorily represented in the sets of positive samples used for training and validation. A first option to improve the results is to obtain a significant number and variety of experimental samples of flange responses and use themto train and validate the algorithms. Other alternatives are to usemore sophisticated numerical simulation environments and to find more efficient attributes of the data.

enviar mensaje