Fast Recovery Of Signal Estimates using Neural Networks (FROSENN)

Maritime domain situational awareness is becoming an increasingly difficult problem, largely due to the pace of  Technological advancement and the rapid proliferation of low-cost communications equipment and small radars. Current automated RF signal acquisition and analysis systems struggle to keep up as the signal density continues to increase and new signals of interest (SOI) spread across the globe. The US Navy’s surface ships, aircraft, and submarines that operate in this environment require a more robust approach in order to properly characterize and exploit the RF spectrum. For example, a communications link can be affected by multi-path scattering and therefore the link’s Bit-Error-Rate (BER) can fluctuate over time with motion or other variation channel with power deviations above and below nominal. Signal acquisition in dense signal environments is especially challenging where strong and weak signals mingle close to each other in frequency. This condition is exacerbated as the bandwidth of the receiver becomes wider and bit-depth is reduced, making it vulnerable to saturation and clipping. In applications where a single front-end provides coverage over a wide range of frequencies, a very wide-band system provides significant system simplicity, but strong out-of-band signals can cause non-linear effects--which must be solved to improve wideband system viability. The Fast Recovery of Signal Estimates using Neural Networks (FROSENN) effort develops machine learning techniques to address the recovery of corrupted signals due to saturation and digitization effects resulting from very wideband signal collection. The key result is a neural network architecture—the Desaturation Network—able to significantly recover data that has been corrupted by analog to digital conversion saturation effects. The Desaturation Network reduces the effects of clipping by estimating the samples where the signal is distorted. The results of the Phase I Base effort show that a degraded signal passed through the FROSENN network can recover a significant percentage of the performance gap between the distorted and undistorted signal. This represents a significant improvement in demodulation quality. The current results are based on using a simple communications signal as our signal of interest and has been shown to work across the usable Nyquist sampling region of the analog to digital converter (ADC) in the presence of zero or multiple interference signals. These results will be extended by expanding the capability of the neural networks and extending the signal environment complexity. Additionally, two training methodologies will be evaluated that will give us insight into whether, for a specific signal, knowledge of the signal can be used to increase the performance improvement against that type of signal. This becomes important when the signal under test is an intentional signal that we can process with foreknowledge rather than an unknown signal.