edfas.org 5 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 26 NO. 3 Neerula et al.[18] proposed a deep-learning approach. The U-Net architecture is used for void detection. Wankerl et al.[19] used fully convolutional networks for 2D x-ray void segmentation. The encoder of the network is inspired from the VGG network. Cheng et al.[20] proposed a novel effective fusion network for BGA bubble segmentation with blurred boundaries. A combination of the transformer and CNN network is used and to deal with limited data situations data-efficient image transformers (DeiT) are used. Though 2D x-ray methods are widely used, these methods face challenges in providing accurate information on the 3D shape, size, and location of voids within the solder joint. Moreover, details of overlapping voids are challenging to accurately obtain from 2D methods.[11,21] To address the challenges with 2D x-ray methods, technology that provides high-resolution 3D information, such as x-ray microscopy (XRM), is gaining popularity.[22,23] XRM provides precise visualization of void morphology and spatial distribution, delivering valuable 3D information.[21,22] While XRM offers significant advantages over 2D x-ray, a major challenge associated with this technique is the generation of large amounts of data. The analysis of this data can be very time-consuming and requires significant intervention from field experts.[22–25] Eom et al.[26] proposed hybrid two-stage algorithms for defect detection on occluded BGA balls. The first stage uses image processing techniques, and the second stage uses the 3D x-ray technique oblique computed tomography. Zhang et al.[27] proposed an algorithm for common defect detection which also includes voids in 3D x-ray images of BGA solder joints. The method involves analyzing crosssection images of the solder joints, segmenting the BGA ball and background, extracting features of solder joint defects, and applying decision rules for inspection. Despite the methods and algorithms proposed, accurate automated segmentation of the BGA solder voids in 3D XRM data is still challenging. Therefore, the objective of this research is to develop a deep-learning approach for an automated segmentation of the BGA solder voids. EXPERIMENTAL PROCEDURES EVALUATION AND DESCRIPTION OF THE DATASET The dataset used in this study comprises 3D x-ray microscopy scans of two different BGA samples. Thus, the dataset consists of a total of two 3D scans with each scan containing more than 1000 2D slices with a resolution of voxel size of 3.7 mm3, ensuring detailed coverage of the sample geometry. Out of these two scans, one scan is used for the training set and one is used as a test set. Data acquisition was performed using a Zeiss Xradia Versa x-ray microscope.[28] The majority of XRM scanning parameters, such as projections, binning, and voxel size, were kept identical in both scans. However, parameters like voltage, current, and exposure time were adjusted by the image acquisition expert, based on the specific characteristics of each BGA sample to optimize image quality. Third-party software Dragonfly[29] is used for 3D data visualization, data pre-processing, and data labeling for segmentation model training. As part of the pre-processing, both scans were cropped to remove unnecessary information from the 3D scan in order to optimize the computational resources. The final remaining dataset contains the void region. The labeling of this cropped dataset was done manually with three labels: voids, solder balls, and background. Further, cleaning and adjustment of labels were done using connected component analysis. The labels were validated by subject experts to ensure accuracy and reliability. However, it is important to note that even after expert validation, the labels may contain some degree of human labeling error. The data set from the training data is visualized in Figs. 1-3. Figure 1 is a visualization of XRM scans of multiple solder balls. Figures 2 and 3 visualize the sample dataset that was used for training. Fig. 1 3D XRM scans of multiple solder balls; the visualization is extracted from Dragonfly. (a) Visualization of 3D XRM volume where voids and solder balls are visible and marked. (b) 3D visualization of solder balls by removing air from actual 3D volume for better visualization of solder balls. (b) (a)
RkJQdWJsaXNoZXIy MTYyMzk3NQ==