INVESTIGADORES
BONOMO Nestor Eduardo
congresos y reuniones científicas
Título:
Identification of Pipe Flanges in GPR Images by Using Neural Networks
Autor/es:
BORDÓN, PABLO; MARTINELLI, PATRICIA; BONOMO, NÉSTOR
Lugar:
Malmö
Reunión:
Congreso; 23rd European Meeting of Environmental and Engineering Geophysics; 2017
Institución organizadora:
EAGE
Resumen:
Pipe flanges detection is arelatively recent application of GPR (Bonomo y otros, 2011). The simple-offset(SO) reflection methodology has demonstrated very good detection efficiency ofthis type of targets and it has the advantage, with respect to othergeophysical methods, of making possible to prospect large sections of thepipeline in relatively short times. Other relevant benefit of the SO-GPRmethodology is that it avoids removing the soil around the pipe and alteringits normal functioning, since this methodology works from the ground surface ina noninvasive way. On the other hand, processing andanalyzing SO-GPR data usually requires considerably longer time than dataacquisition. This can be a problematic feature when searching for flanges alonglarge portions of oil, gas, mineral and water pipelines, since the associatedhuman effort, probability of error and monetary cost significantly increase. Inthese cases, it is particularly relevant to implement efficient automatic orsemiautomatic procedures of detection and classification, so that theinterpreter can be relieved of exhausting tasks and concentrate on relevantdecisions. Artificial Neural Networks (ANNs)is one of the Supervised Learning techniques most frequently applied to GPR data(e.g. Núñez-Nieto et al., 2014; Kilic and Unluturk, 2014). The optimal size andstructure of an ANN for a given problem is determined empirically by monitoringthe network performance. Normally, various algorithms are trained by providinga number of training GPR patterns that are representative of the reflections atthe investigated targets and surrounding soil as input and their correspondentinterpretation as output. The efficiency of these algorithms is monitored byapplying them to validation patterns, not used in the training. Finally, oncethis process is completed and the best ANN structure is selected, its generalperformance is evaluated using unanalyzed test data (e.g. Bishop, 1995; Hastieet al., 2009). In this work, we explore the use ofANNs for identifying pipe flanges reflections in SO-GPR images. Since none or afew images of the flanges are usually available in this kind of applicationbefore the prospection, simulated patterns have to be created and used to trainthe network. We first describe the training and validation process; then, wecheck the results of the final ANN with simulated and experimental test data.