Industrial robotic station for general-purpose quality inspection driven by convolutional neural network classifier

In the last decade the exponential growth of computational power, together with the huge availability of data, eased the spread of artificial intelligence in a variety of applications, one of which is the manufacturing business. The following thesis work will outline an industrial machine prototype that can automatically recognise faulty objects moving on a conveyor belt by means of a camera and an AI-powered image classifier and, whenever a defect is detected, remove such object through a robotic manipulator. The project was entirely developed at Tera Automation S.R.L. which in the last few years is establishing itself as a leading company in the valuable metals manufacturing. In particular, according to the company’s needs, the objects being examined are Christian Dior’s clips, designed to join handbags to shoulder straps. The two main problems that will be tackled in this thesis are: the almost-standalone task of automatically recognising the manufacturing quality of a clip through the processing of a picture of the same object, and the complementary coordination task required to remove faulty objects from the conveyor belt by means of the robotic arm. The first task will be accomplished by training a Convolutional Neural Network (CNN) to discriminate between good and faulty clips. The CNN programming will be carried out using the Keras framework for Python. To build the dataset, around 3600 images are collected, equally distributed among bad and good quality. The impurity on damaged objects are hand-made in order to simulate the most common defects that could occur during the manufacturing of these pieces, such as dots, scratches and deformations. Also, data augmentation will be applied on the dataset, making the CNN more robust to slight variations of the input images and less prone to overfitting. Several neural network architectures will be tested in order to achieve the best results: from a custom-built neural network trained from scratch, up to more complex pre-trained architecture such as VGG, ResNet, Inception and MobileNet which are designed to recognise a thousand different classes of objects from the ImageNet dataset. In the latter case, the transfer-learning technique will be performed in order to export the recognition capabilities of such CNNs, obtained on the general ImageNet dataset, onto the recognition of the specific two classes of interest for the application. Finally, an exhaustive comparison of the performance of the different models will be outlined, with the best performing one achieving an accuracy on a balanced test dataset of 98% circa. To handle the second task, the more mechatronic part of the application is also addressed. A pure computer-vision algorithm will be designed to detect the clips on the conveyor belt and to derive their coordinates from the picture. The robotic arm routines will be programmed, as well as the supervisor PLC, which will take care of most of the coordination features of the whole system. Also, given the variety of different devices involved in the application, how the communication between each of them is achieved will be highlighted.