MIE.09 – DREAM LAB

• Rosalie Beirne
• Anna Biskup
• Josh Bustos
• Imran Khan
• Heba Raafat Elsayed Salem

The project focused on improving the accuracy of data reading(code)that will be implemented in the future to be used to identify defective products, ranging from different structures, colors, and sizes. When considering a specific individual development, the effect is replicated to every defined detail, and this is one of the main objectives when it comes to a manufacturing company—having the minor percent error of a defective product when processing the products—alluding to the fact that Artificial Intelligence (AI) now can automate tasks, such as detecting defective products by analyzing copious amounts of images from a product. This is done by separating the data types into specific folders and, if plausible, dividing them into categories; in our case, we had two separate folders, “intact” and “damaged.” Each image in the “damaged” folder had a DE pixelated spot from the original “intact” data. When having multiple varieties of “damaged” data, it becomes easier for the model to distinguish the “intact” from the “damaged.” The team was provided with three data sets consisting of manufacturing parts. The first data we accounted for was with boxes, and the differentiation of augmentation was minimal but had an abundance of data. As for the second data set, which was metal nuts, there was a more significant differential augmentation so much that the data was categorical and had far fewer data than the boxes. In most instances, the more the data, the better, but narrowing the data to specifics will be far more beneficial. The last data set was composed of different surfaces; the data set was based on six categorical factors: rolled-in scale, patches, crazing, pitted surface, and inclusion. In machine learning, these can be defined as class modes – these categories then become permutable, and the model has eased in cycling through the data due to their unique extinction. Thus, having a more significant margin of validity compared to the first two data sets. The team improved the accuracy of an automated defect detection software on the three data sets using a deep learning methodology based on artificial neural networks. Two neural network models were used in the project such as a neural network using Principal Component Analysis (PCA) and a Convolutional Neural Network (CNN). Both designs aim to recognize patterns and classify images. However, each neural network processes image inputs in a different way. Therefore, the team trained both models using the three data sets and made predictions. The obtained accuracy, learning rate, and uncertainty were used to evaluate and compare both models’ performance for each data set. The project’s main objective was to increase the accuracy of the provided codes through training and testing multiple data sets; there was no intent in applying it to a manufacturing process, for that requires extensive implementation. Expansion of this project can be used to develop a prototype system that uses the codes and other elements to implement a “smart” quality control methodology in a manufacturing company to ensure the maintenance and improvement of its manufactured product’s quality.