Study On Automatic Recognition Of Seepage Piping Based On The FDTD Method

Embankment pipe is one of the most common and dangerous progressive hidden danger with great damage.It is invisible before it happens,and destructive while it happens.There,in Guangdong,are many small reservoirs and embankments,most of them filled by Earth-Rock,operating over 50 years.Because of the poor building technique and management,they are series of structural problems that cannot be fixed still though most of them are reinforced.They are still the best part for the development of piping.So the detection technique is the key to prevent the piping happening.The GPR detection is efficient,non-destructive and visualization and widely used that promoted to water conservancy construction.However,the interpretion of GPR data has dependence on the expreinces of users that probably make misreading.Building a mechanism of recognition for piping based on BP neural network is the purpose of this paper.Using the pattern built by computer to realize the auto recognition of piping can prevent the mistake happening and promote the efficiency.The paper introduces the ANN into the GPR interpretion,and creatively uses the virtual data instead of real data for the training of ANN by using FDTD to make forward modeling of piping as the training samples for ANN.It is an efficient way to solve the problem of sampling,and it is worth to promote to other geophysics detectiong usings.Building a useful and efficient neural network needs lots of the sample,and it is nearly possible for real data sampling.But FDTD method is the most popular way to study the electromagnetism,so it is the best way to finish the job of sampling.Meanwhile,we use the traingdx way to improve the BP neural network to prevent the slow convergence.The paper used the FDTD method to make the samples for ANN training,including the A-scan data and B-scan data which are grouped into different types by its piping number and layers of background,and each of the samples were different of the piping position.After sampling,we build two different improved BP neural network models by the two kinds of the samples and made comparative analysis of the two network models.This experiment shows the application value and guiding significance of technical ideas of this paper.