Artificial neural network predicting optimal cooking program for UNOX ovens

Technological progress driven by the “mobile” market has overturned different types of market in a few decades, effectively transforming any device into an intelligent, intuitive and connected object. In this context, the catering equipment sector is no exception, of which UNOX, a man- ufacturer of professional ovens, is one of the world leaders. The introduction of increasingly cutting-edge digital displays for interfacing with this type of household appliance has allowed enormous progress in terms of energy efficiency, cooking performance and ease of use. The Deep Cooking project is part of this last area, with the aim of making the control panel of UNOX ovens able to receive feedback from the user at the end of a cooking process, to then propose any changes to the program itself, in order to obtain the desired cooking result. To this end, a combined “data-driven” and “physics-based” approach will be adopted. The ”data driven” approach involves the creation of a Deep neural network capable of propos- ing changes to the cooking program just concluded on the basis of two inputs: the sequence of cooking steps that make up the program originally set by the user; user feedback on the cooking result just obtained. The training of this neural network requires a very large data-set, made up of thousands of cooking tests associated with the relative judgment of a professional chef. The aforementioned tests would require an unsustainable amount of time and resources, requiring an alternative method for constructing the dataset. It is at this point in the project that the ”physics-based” approach is necessary. In fact, the ”physics-based” approach involves the implementation of a physical (one-dimensional) and parametric model, capable of modeling the phenomena of state change, mass transport and heat exchange that occur within the oven cavity during cooking. With the use of a second algo- rithm capable of transforming the evolution of temperature and humidity during cooking into a final feedback on the cooking itself, the simulator will make it possible to create a ”simulated” cooking dataset, arbitrarily large and varied, to be used for neural network training. The thesis work focuses on the implementation of the framework of the data driven ap- proach. Using a simulated dataset, the most effective way to carry out the training of the NN will be studied, starting from the management of the dataset and the definition of an input space, up to the NN architecture and the necessary hardware specifications.