Development of an Optimized Neural Network for the Detection of Pipe Defects Using a Microwave Signal

Abstract

Neural network technology is applied to the detection of a pipe wall thinning (PWT) in a pipe using a microwave signal reflection as an input. The location, depth, length, and profile geometry of the PWT are predicted by the neural network from input parameters taken from the resonance frequency plots for training data generated through computer simulation. The network is optimized using an evolutionary optimization routine, using the 108 training data samples to minimize the errors produced by the neural network model. The optimizer specified not only the optimal weights for the network links but also the optimal topology for the network itself. The results demonstrate the potential of the approach in that when data files were input that were not part of the training data set, fairly accurate predictions were made by the network. The results from the initial network models can be utilized to improve the future performance of the network.