Increasing Spacecraft Autonomy through Embedded Neural Networks for Semantic Image Analysis

University essay from Luleå tekniska universitet/Institutionen för system- och rymdteknik

Abstract: In the scope of this thesis, a possible usage of embedded artificial neural networks for on-board image analysis is investigated. After an introduction which imparts basic knowledge about artificial neural network and autonomy in spaceflight, a possible system design is elaborated based on previously defined reference scenarios. The reference scenarios are based on two projects that are currently under development at the professorship of Space Technology at the University of Würzburg: ASMET and SONATE. As base for the neural network architecture, a convolutional neural network called SqueezeNet was chosen, since it was developed for similar input data and performs sufficiently well. In addition, the SqueezeNet architecture requires extremely little memory for the trained-in model compared to other architectures which makes in-orbit updates of the model feasible. The system concept in this thesis is designed for offline learning, i.e. the training phase will be done on an ordinary computer. The resulting, trained-in model is then transferred to the embedded system. On the embedded side, a FPGA-based approach was chosen since FPGAs allow to parallelize the neural network execution and therefore accelerate it significantly. Even though not all components of the designed concept could be implemented in the scope of this thesis, all key elements were implemented and tested, either on real hardware or by using testbenches. Especially the tests conducted with the convolution unit of the embedded system went well and allowed to make a quite promising assessment of the expected execution speed. In addition, a tool with graphical user interface was developed to guide a potential user of the system through the steps of training-in and setting up the system. For the training process, a neural network framework called Caffe was used within this tool. In summary, this thesis provides as intended a profound starting point for further research on artificial neural network for space applications at the professorship of Space Technology of the University of Würzburg.

  AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)