Convolutional neural networks for predicting the state of electrical switches

The aim of this master thesis is to expose the work done during my five months of internship at the Blue Reply company in Turin where I worked on a commission project from one of the largest energy companies in the world. It has consisted in the development of an API based on a machine learning algorithm that is able to receive as input an image of an eletrical switch, to detect the QR codes which are present and to extract the information on them, to crop the area of interest and to query the appropiate convolutional neural network (CNN) previously trained to give as output a prediction about the state of the switch. The CNNs are one of the most useful tool of the deep learning nowadays and they are applied in several tasks, indeed they cover a main role in topics like image and video recognition, image classification, neural language processing and others. The problem I dealt with can be classified as a computer vision problem, more precisely as an image classification problem, and its purpose is to increase occupational safety for operators who need to carry out maintenance and repair operations. The key role of the project is covered by neural networks and for this reason the first part of the thesis serves to provide their theoretical knowledge, in fact it aims to explain what they are and how they work, focusing the main attention on the convolutional architecture adopted for the problem and known as ResNet. After that, it is explain into details the work done. It concerns the development of the machine learning algorithm, the creations of different datasets, the object detection implementation which is another important tool of deep learning, the training of neural networks and the results evaluation. The performances are not only evaluated through more classic measures like accuracy and confidence, but also through the implementation of an explanatory model that allows to understand why a certain prediction is done and if it is reliable.