FPGA based long short-term memory artificial neural network for pulse signal detection

In this thesis, an FPGA-based Long Short-Term Memory (LSTM) Artificial Neural Network (ANN) has been proposed for pulsed signal detection applications, which also has an important role in Electronic Warfare Support Systems. It is aimed to detect pulsed signals collected by the signal receiver at low signal-to-noise ratio (SNR) levels with this proposed FPGA-based LSTM, using only raw data without any processing on the pulsed signal samples that cannot be detected by the widely used Constant False Alarm Rate (CFAR) method. Previous studies show that LSTM neural networks can detect pulsed signals with a high success rate at low SNR. In addition to these studies, an FPGA-based LSTM is implemented in this thesis. According to simulation results, it is observed that pulsed signal detection could be performed with 94% accuracy at -5 dB SNR. In this thesis, MATLAB/Simulink models were designed to generate the datasets required for training and testing the FPGA-based LSTM. With these models, test datasets of the CFAR method were also generated. In this thesis, the performance criteria obtained from the synthesis process of the FPGA-based LSTM and the synthesis results of the CFAR method on FPGA is compared. The obtained results show that the FPGA-based LSTM, which proposed in this thesis, can be implemented on a hardware (FPGA) and can detect pulsed signal at low SNR where the traditional threshold method (CFAR) cannot obtain successful results, however, it shows that this proposed architecture cannot meet the real-time radar receiver requirements of an Electronic Warfare System.