An industry 4.0 deep learning neural network machine vision system classifier trained with synthetic computer-aided design training data

Abstract

Industry 4.0 describes the digitalization of industrial manufacturing processes, the integration of information and communication technologies, and the use of smart products and machines. Two of the most impactful aims of Industry 4.0 are to facilitate rapid adaptation in the manufacturing process to design changes and the reduction of batch sizes to as little as a batch of one item while maintaining quality and profitability. The implication being that manufacturing systems would need to deal with very fast turnaround times, possibly having many different products, variations of products, and constant addition of new products on the same assembly line. This would require quality control systems, especially Machine Vision Systems, which are capable of rapid reconfiguration. Current Machine Vision Systems are not suited for such an environment. This is largely due to the requirement of an Image Processing Engineer to handcraft the feature extractor of the Machine Vision Systems and then train the artificial neural network that performs the classification. This process normally requires several iterations to ensure accuracy and reliability and can be significantly prolonged with larger amounts of features or products that the system needs to process. Deep Learning has proven to be very successful in various image processing domains and holds great potential in developing more flexible Machine Vision Systems that can be retrained as required. With Deep Learning feature extraction and classification are integrated into the same structure, and features are automatically learned. Deep Learning requires significantly more data and processing power for learning. While these problems can be overcome in many general use cases by making use of techniques like transfer learning, image augmentation, and modern GPUs to train the Deep Learning network, there are many specialized cases, such as automated visual inspection of a product, where there is simply not enough data available. The development of a Deep Learning Machine Vision Systems Classifier would be of great advantage to any Industry 4.0 manufacturing system. However, obtaining the large amounts of training data required for training such a system can be labor-intensive, expensive, and, in some cases, not possible. This, however, raises a question: Would it be possible to create synthetic training data using computer-aided design and modeling tools and to train a Deep Learning system with the synthetic data? This study aims to create a synthetic image dataset of basic electronic components by using computer-aided design and modeling tools to automatically generate large amounts of synthetic images, use image augmentation to expand the synthetic dataset, and then train a Deep Learning classifier model on the resulting synthetic data. To achieve this, a literature study was done on the role of Deep Learning in Industry 4.0, Deep Learning Classifiers, training data requirements, and Blender as a platform for creating synthetic training data. Three-dimensional models of the electronic components were created and textured; an automated workflow was developed to generate and augment the synthetic data; a Deep Learning classification model was trained using the synthetic data; and lastly, the performance of the Deep Learning model was evaluated using a set of real-world testing images. The results show that it is feasible to use synthetic data to train a Deep Learning model, provided that the synthetic data accurately represents the real-world counterpart. However, it also shows that synthetic data that does not accurately represent the real-world counterpart is of no use and will reduce the overall performance of the classifier.