The use of intelligent systems to improve manufacturing processes is the basis for the development of robotic solutions in Industry 4.0. Monitoring operators manipulating tools and objects is one of the key tasks. Deep learning methods are obtaining state-of-the-art results to solve this problem but large amounts of labelled data should be provided to these networks. However, no specific manufacturing tools datasets exist. For this purpose, we proposed a new dataset for this type of environment. An hybrid dataset of 29550 images has been proposed for network training that combines real and synthetic images of tools and components commonly used in manufacturing cells. This project is part of a set of proposed modules of a solution that allows us to evaluate in real-time the execution of assembly instructions of the operators throughout the production process.