Automation of SQUlD nondestructive evaluation of steel plates by neural networks

This paper presents a method for automation of SQUID nondestructive evaluation (NDE) using neural networks, exempting the need for a trained technician, necessary to most of the usual NDE methods. An LTS-SQUID susceptometer, with a 16 mm diameter planar concentric gradiometer, was used to image flaws in steel samples from the bottom of an oil storage tank. Natural and artificial corrosion pits of various sizes were present in the samples, and a vertical magnetic field of 0.5 mT was applied by a superconducting magnet concentric with the gradiometer coils. A finite element model was used to simulate the magnetic signals due to the flaws, yielding training sets for the artificial neural networks. A neural system composed of two cascaded networks was developed to preprocess and analyze the magnetic signals. The first network removes a distortion that occurs in the experimental magnetic signal, and the second network detects the presence of flaws, and also assesses their severity. The trained neural networks were successfully tested with the experimental data obtained with the SQUID system.