Study and Development of Neural Network Architectures on Rad-Hard FPGAs

This thesis investigates the use of radiation-hardened FPGAs for neural network implementation in radiation-prone environments. For this purpose, the European-based NG-Medium board is examined and compared to the non-radiation-hardened Pynq-Z2. A UART communication system and a neural network for color classification were designed to function on both boards without significant alterations to ensure equal compatibility for a fair comparison. The difference between the utilization and performance of the NG-Medium and the Pynq-Z2 FPGA is interesting. The Pynq-Z2 demands more resources for code execution, whereas the NG-Medium struggles with proper code functioning, possibly contributing to its seemingly lower consumption. The Pynq-Z2 also performs better in speed due to the faster clock frequency. The classifier works well with Pynq-Z2 but does not function properly on the NG-Medium. The cause of the improper functionality is still unknown. Integrating neural networks on radiation-hardened FPGAs is a complex task. While the NG-Medium's feasibility for neural networks remains unproven, this study can still be helpful due to its insights into the utilization and architecture between the two boards. The thesis helps bring forward the potential benefits of this combination despite being unable to validate neural networks on the NG-Medium. While working on this project, new challenges and insights have come forward, which can help future researchers studying this topic. The use of the neural network on radiation-hardened FPGAs needs further investigation, exploring different neural network algorithms and training methods, addressing code issues on the NG-Medium, and performing radiation testing on the entire system. While the study could not determine if the NG-Medium is a good candidate for neural network implementation on board spacecraft, the study lays the groundwork for advancing intelligent systems in the future. In conclusion, neural network implementation on radiation-hardened FPGAs requires more research. However, despite the challenges and lack of proper results in this thesis, the research sets the stage for further exploring this topic of deep learning for space application.