Neural networks for back wall geometry reconstruction using the active thermography

In this work we propose a 3D reconstruction of a defect that exists in the inaccessible side of a material by an artificial neural network. We used a neural network that contains a single hidden layer with the sigmoid activation function to treat the standard thermal contrast deduced from the pulsed infrared thermography data. Three different data sets were used in the learning phase. The use of standard thermal contrast for training the neural network has significantly reduced the processed data number. This represented a considerable gain in data processing time and memory space. The 3D reconstructions were made by estimating the depth of each point of the defect geometry in the controlled material face. The differences between real and network estimated geometries are very low and are not exceeding 3%. The experimental results and the simulations by finite elements confirmed the performance of the presented method.