부산대학교 도서관

Interest is increasing in the use of neural networks and deep-learning for on-board processing tasks in the space industry [1]. However development has lagged behind terrestrial applications for several reasons: space qualified computers have significantly less processing power than their terrestrial equivalents, reliability requirements are more stringent than the majority of applications deep-learning is being used for. The long requirements, design and qualification cycles in much of the space industry slows adoption of recent developments. GPUs are the first hardware choice for implementing neural networks on terrestrial computers, however no radiation hardened equivalent parts are currently available. Field Programmable Gate Array devices are capable of efficiently implementing neural networks and radiation hardened parts are available, however the process to deploy and validate an inference network is non-trivial and robust tools that automate the process are not available. We present an open source tool chain that can automatically deploy a trained inference network from the TensorFlow framework directly to the LEON 3, and an industrial case study of the design process used to train and optimise a deep-learning model for this processor. This does not directly change the three challenges described above however it greatly accelerates prototyping and analysis of neural network solutions, allowing these options to be more easily considered than is currently possible. Future improvements to the tools are identified along with a summary of some of the obstacles to using neural networks and potential solutions to these in the future.