Smart data driven defect detection method for surface quality control in manufacturing

Abstract

The use of machine vision systems for quality control of reflective metal surfaces has increased over the years with new systems that combine higher resolution cameras and better illumination of inspected objects. With the advances in artificial intelligence pattern recognition of images, the integration of machine vision systems in a manufacturing line for accurate automatic classification of defects would work towards the application of a smart manufacturing concept. To investigate the feasibility of such integration, a vision system that combines a 4K camera and chromatic confocal technology was employed to analyze surfaces of copper parts after the laser machining process. By the application of three machine-learning algorithms (decision trees, random forest and multi-layer perceptron) on features extracted from the Sobel edge detector, segmentation of defects has been performed using the Weka segmentation plugin. A simple convolutional neural network (CNN) was also applied for the classification of defects. Later on, using smart data rather than big data, transfer learning (TL) has been successfully performed with retraining the mobilenet-v1 model, which is based on CNN. This lean learning process can be implemented in devices that are limited by their computation resources. The maximum average of validation accuracy achieved using TL trained over 500 epochs was 90.5%. Whereas for the simple CNN classification models, the best validation accuracy was achieved by a model with a batch size equal to ten and with 40% of validation data with an average equal to 98.7% over 500 epochs.