Abstract

Quality standards fulfilment is an essential task in manufacturing processes that involves high costs. One target is to avoid the presence of burrs in the edge of machine workpieces, which reduce the quality of the products. Furthermore, they are not easily removed since the part can even be damaged. In this paper, we propose an optimized Convolutional Neural Network, to detect the presence of burrs in images of milling parts. Its design is focused on the optimization of classification (accuracy) and performance metrics (training time and number of trainable parameters). The proposed architecture identifies burrs with a 91.16% accuracy in the test set, outperforming existing models as EfficientNetB0. It also reduces the number of trainable parameters from other models as AlexNet by 1.5 million. The prediction process just takes 48.39 milliseconds per image. Finally, in order to check if the model gets a high activation in the region of interest, a visual explanation of the model is also carried out by using Gradient-weighted Class Activation Mapping.