Eddy currents and ultrasonic testing data fusion for Wire + Arc Additive Manufacturing

Advanced manufacturing methods such as high-deposition rate Wire + Arc Additive Manufacturing (WAAM) has seen an increasing uptake in high-value manufacturing of aerospace components owing to their capability of producing large components in a short time. The process is automated involving industrial robots that enable layer-by-layer deposition on substrates through arc-based welding processes. The components are prone to the typical welding defects such as lack of fusion, keyhole and porosities, and therefore Non-Destructive Evaluation (NDE) are crucial to ensure their fitness for service. Traditionally, are inspected post-manufacturing however, the fully automated WAAM deposition process unlocks the opportunity to also integrate and deliver the NDE during the manufacturing after deposition of every few layers, leading to process cost-effectiveness and opportunity for early process intervention and remedial rework.
A dual sensor concept for in-process inspection of WAAM components is presented in this work. The two sensor modalities deployed here are: a) a high temperature phased array Ultrasound Testing (UT) roller-probe, and b) a high-temperature flexible Eddy Currents (EC) array. The design and fabrication are presented for the robotically deployable dual-sensor head. To test the performance of the dual-sensor, titanium calibration blocks with artificial defects and geometries similar to WAAM were manufactured. A range of side-drilled holes and flat-bottom holes located at different depth were fabricated. The phased array UT and array EC scans were performed at once using the dual head through a fully integrated software encoding the data from both arrays with robot pose information.
C-scan images from both UT and EC arrays were produced during the scan. The images went through: Image registration, pre-processing, and fusion stages. Different pre-processing methodologies were devised and applied to the EC C-scan images to overcome the lift-off noise bands caused by the poor probe contact to the surface close to the corner fillets of the calibration sample. Two distinct data fusion approaches; namely ramping method, and novelty detection statistical method, followed by pixel-level summation, multiplication, and max pooling were applied to the filtered C-scans, and their performance were compared. For the statistical approach, distributions of a clean section of UT noise and EC differentials per channel across the scan were fitted. The resulting cumulative distribution function was then used with an acceptable criterion for false indication (e.g 0.1%) to give thresholded images. These images can then be combined with either max pooling or pixel summation. The results of data fusion showcased the complementary nature of the two inspection methods through successful detection of all the artificial defects in the calibration blocks.