Use 3D Convolutional Neural Network to Inspect Solder Ball Defects

Abstract

Head-In-Pillow (HIP) is a solder ball defect. The defect can be caused by surface of solder ball oxidation, poor wetting of the solder, or by distortion of the Printed Circuit Board (PCB) by the heat of the soldering process. The current diagnosis of the HIP defects is difficult to find out the problems, and some destructive tests are not recommended for use. In this paper, we use different angles of 2D X-Rays images to reconstruct the 3D PCB volumetric data. We crop the 3D solder balls volumetric data to 46 × 46 × 46 pixels from the 3D PCB model. Because HIP problems do not happen often, we use the data augmentation method to expand our solder ball data. We propose a new 3D Convolutional Neural Network (CNN) to inspect the HIP problems. Our network uses convolutional blocks that consist of different convolutional paths and the dense connectivity method to connect the blocks. The network can learn various features through these convolutional blocks with different convolutional paths. Moreover, the features of each layer will be fully utilized in the following layers by the dense connectivity method, and also avoid some features lost in a deep convolutional path. In the last layer of our network, the global average pooling can let our network process more different sizes of solder ball data, and the normalized same size vector can be used to do the end-to-end learning. Compared with other classic models, our network not only has fewer parameters but also has faster training time.