Recently, in the wake of the 4th Industrial Revolution, a new concept called "smart factories" emerged in the manufacturing industry. This enables flexible process operations beyond existing factory automation and lays the foundation for collecting various process data. Accordingly, several studies on predicting process defects using the data collected in the process are being conducted, but most of them are limited to the environment in which image data is collected. Therefore, in this study, a defect prediction of a process in which only numerical data is collected was carried out. The data used to conduct this study are from the foaming process that produces polyurethane foam, and a total of six factors were collected. The collected data were preprocessed, such as by normalization, and the normalized numerical data was changed to an image with a gray scale for application to CNN. To address the problem of disparity between fine and defect data, SMOTE techniques were applied to ANN datasets and cutout techniques were applied to CNN datasets, as well as to augment data and construct different datasets to compare prediction accuracy based on the size of each augmented and constructed dataset. As a result of the progress, a high prediction accuracy of 53.57% to 95.54% was derived. Of these, 53.57% accuracy is achieved when an unbalanced dataset is applied to ANN, and 95.54% predictive accuracy is achieved when 5 times and 10 times the augmented dataset is applied to CNN. Based on these results, it was confirmed that defect prediction with high accuracy was possible by the method conducted in this study. Since this prediction technique relates to the relationship between process factors and quality, it is considered that it is not limited to the foaming process and can be applied to various processes.