Identification of Solder Joint Failure Modes Using Machine Learning
Kyung‐Eun Min, Jaewon Jang, Soobin Kang, Cheolhee Kim, Sung Yi
Abstract
The reliability of solder joints is one of the most critical factors that determine the lifecycle of electronic devices, and the identification of solder joint failure modes is necessary to enhance the performance and durability of electronic devices. In this study, solder joint failure modes were identified using the fine-tuned visual geometry group 19 (VGG 19) pretrained model. Raw images (57 images) were augmented into 428 images by sectioning to classify the solder joint failure mode into two classes (good or not-good mode) for the binary classification model, and 265 not-good data points obtained from the binary classification were employed as input to classify solder joint failure mode into six classes (failure modes 1–6) for the multiclass classification model. The binary and multiclass classification models were trained and validated, achieving 99% accuracy. The binary model classified shadows and small voids as defects, identifying the failure mode as “not-good.” The multiclass model occasionally misclassified the failure modes due to the multiple modes or difficulty in classification. The trained binary and multiclass classification models were further verified using 102 and 64 third-party experimental data points, respectively, confirming 100% accuracy. The results demonstrated the successful automated classifications of solder joint failure modes using the convolutional neural network (CNN) model, indicating the potential for its use in the industry to improve the reliability and quality of solder joints.