Neural Networks for Defect Detection on Eddy-currents-based Non-destructive Testing

This paper focuses on the detection of hole-like defects in materials using non-destructive testing methods. The proposed approach utilizes perturbances in induced eddy currents, captured by an application-specific integrated circuit (ASIC) and signal acquisition system based on magnetoresistive sensors. The system provides the capability to detect micrometric defects. To enhance defect identification in noisy signals (SNR below 6 dB), an artificial neural networks (ANN) approach is employed. The ANN is trained on fully synthetic data and analyzes 2D scans obtained from the probes, column by column accurately pinpointing hole-like defects in a manner that is independent of defect size and shape. Experimental results on an aluminum mockup with drilled holes demonstrate the effectiveness of the proposed method, in clearly highlighting the defects even at depths of 500 mu m and a diameter of 100 mu m.