نوع مقاله : علمی - پژوهشی
1 دانشجوی دکترای اکتشاف معدن دانشگاه تهران عضو هیئت علمی دانشگاه صنعتی اراک
2 عضو هیئت علمی دانشگاه صنعتی اصفهان
3 عضو هیئت علمی دانشکده مهندسی معدن دانشگاه تهران
عنوان مقاله [English]
In current study, the application of artificial neural network (ANN) and support vector machine (SVM) methods in automatic detection of hidden cylindrical targets in GPR images is presented. In order to determine the predictor variables as the input to the ANN and SVM methods, through relationships between geometrical parameters of cylindrical targets and characteristics of the associated GPR hyperbolic responses, the forward responses of 194 simulated synthetic models resembling cylindrical structures were computed using an improved 2D time domain finite difference method. Through transforming the B-scan images into eigen value- eigen vector space using singular value decomposition algorithm, the minimum number of vector space bases capable of representing GPR forward responses were extracted. This approach could effectively reduce the number of input predictor variables to only five to eight normalized eigenvector subspaces to be fed into ANN and SVM algorithms. In addition, the geometrical parameters of the cylindrical object including radius, burial depth and horizontal location were set as the output target parameters for both ANN and SVM methods. The proposed approach was then tested on some GPR forward responses that were not used in training the ANN and SVM models. The predicted geometrical parameters of the cylindrical models were amazingly close to their actual values. Encouraged by preliminary results obtained by synthetic models, the same methodology were then evaluated on both forward responses added by 5 percent white noise and a real GPR profile data collected over a known sewage pipeline. The performance of both ANN and SVR were evaluated resulting in acceptable relative misfit error of 5 and 9 percent respectively. Finally, it is concluded that both methods are capable of predicting subsurface geometrical parameters with acceptable precision (less than 10 percent relative error), however the ANN method could produce more accurate estimations compared with equivalent SVM method.