Metallic Surfaces Binary Reconstruction using Eddy Current Sensors and Convolutional Neural Networks

A new Convolutional Neural Network (CNN) based algorithm is proposed for reconstructing metallic layers geometries from eddy current sensors data. The developed algorithm aims at the online quality control on metal additive manufacturing processes by enabling the real-time, layer-by-layer detection of the manufactured part geometry deviations. A simple eddy current sensor was used together with a dedicated instrument to gather real signals. A training set was prepared from different metal patterns offering a total of 332820 inputs to train the CNN. Diversity was achieved using randomly generated geometries and applying them with different scale factors. The Keras tuner optimizer was used to boost accuracy to levels around 98% on the reconstruction of features as small as 2 mm. Comparison with simple Threshold (TH) algorithm showed notable improvement specially for low signal-to-noise ratio conditions.