Development of an AI environment for AOCS applications.

The present work is intended as the final product of the double master's degree program between Politecnico di Torino and ISAE-Supaero, result of the final internship at CNES, the French space agency. The objective of this research was first and foremost to provide the CNES AOCS service with a bibliographic reference tool regarding Neural Networks and all their possible applications by understanding their mathematics and theory. It is then demonstrated how it is possible to reproduce such networks and their different configurations in the Matlab/Simulink environment through the development of an entire AI library in which each type of layer is represented with its characteristics, having Python as a reference. Each layer of neural network studied in depth in the state of the art (MLP, RNN, CNN) is therefore reproduced in respective Simulink models in which it is possible to load weights from Python of the corresponding network already trained through specific functions able to read .h5 and .json files. Once obtained the structure of the net, the loading of the weights in the Simulink models is done through configuration files coded according the OCEANS simulation environment, that is the main simulator used in the service for AOCS control. The aim is therefore to have a structure in Simulink that can reproduce the networks and use them in feedforward to make predictions useful for AOCS analysis. The training phase is therefore carried out in Python and is not reproduced in Simulink. In order to make possible the integration of this library with the AOCS environment all the models are coded in the OCEANS environment too. Such a library thus gives the service the ability to use any type of network by knowing its structure via Python files. The final part lastly focuses on validating these models by comparing the predictions between Python and Simulink of concrete cases through the use of networks already validated internally in the CNES service and from previous case studies. The whole work thus allowed us to fully explore and understand how neural networks work and their possible applications, how it is possible to reproduce their architecture in an environment different from Python by applying their mathematics and respecting their structure.