Abstract:
This work employs X-ray computed tomography (XCT), coupled with machine and deep learning techniques to non-destructively evaluate the quality (detecting defects and differentiating of materials) of dissimilar friction stir welded (DFSW) metallic joints.
Since the model will quantify the material mixing taking place within the dissimilar welds. A large number of pure and alloyed metals was scanned, and HU measurements and their standard deviations recorded for these metals. Since this work depends on XCT scanning to identify HU values of metals of interest and since not much is available in the published literature, much effort was spent on collecting such values experimentally. It was found that measured HU values depend strongly on the XCT scanning parameters including tube current and voltage, voxel size (function of the section width, table speed), etc... Therefore, before such an HU database can be used, it was necessary to delineate such dependence in a reliable fashion including using the broader16-bit HU scale to resolve the issue of saturation.
Given that mixing in FSW joints results in a multi-metal component alloy, a method had to be determined to relate the alloy HU value to those of its elemental constituents. A developed weighted average mixture model utilizes the individual HU values of each metal component and their relative weight percentages to predict the HU value for the alloy. The model is then compared against machine learning (ML) and neural network (NN) based methods which were created to also model the HU values- elemental composition relation.
After scanning of the raw materials was concluded and the relations between HU measurements, scanning conditions, and elemental composition well documented, FSW joints of dissimilar aluminum-based (AA6061) and magnesium-based (AZ31b) are fabricated under different sets of welding parameters (mainly the welding tool as well its feed and rotational speed). The FSW parameters are optimized to obtain relatively good-quality welds.
The welded joints are then XCT scanned at optimal scanning parameters. The collected XCT data is then used to train, validate, and test deep-learning-based neural networks, which will be able to detect the presence of defects, identify their locations, quantify their volumes, and differentiate the two metals in the weld zone. Allowing for the fast and reliable evacuation of dissimilar FSW welds.