Radar (radio detection and ranging) is a detection system that uses radio waves to determine the distance (ranging), angle, and radial velocity of objects relative to the site. A radar system consists of a transmitter producing electromagnetic waves in the radio or microwave domain, a transmitting antenna, a receiving antenna (often the same antenna is used for transmitting and receiving) and a receiver and processor to determine the properties of the object(s). Radio waves (pulsed or continuous) from the transmitter reflect off the object and return to the receiver, giving information about the object’s location and speed.
Today specialized radars measure range as well as azimuth and elevation angles, enabling target detection and localization. Through Synthetic Aperture Radar (SAR), Inverse SAR (ISAR), or Interferometric SAR (InSAR), a 3D image of an object can be obtained
Radar systems have seen many technology improvements in apertures (antennas) and associated hardware and software since the nascent operational versions in World War II. What hasn’t changed significantly over the decades, however, is that radars still use linear signal processing between the aperture and the detector. In the 1940s linear radar signal processing used vacuum tubes and analog circuits, while current radars accomplish linear signal processing digitally with microchips and software. Future radar systems should effectively support a massive variety of applications with novel hardware solutions and innovative signal processing techniques.
DARPA launched Beyond Linear Processing (BLiP) program in Oct 2022, with goal to improve radar performance by applying innovative signal processing methods. BLiP will leverage high-power computer processing to explore new, non-linear and iterative signal processing techniques that could lead to lighter, smaller, and less expensive – but equally capable – radar systems. If successful, BLiP would enable the same radar performance achieved on large platforms today on much smaller sea, air, and ground platforms.
“A lot of radar improvements over the past 30 years have focused on growing the size of the aperture for greater sensitivity or increasing transmitter power,” said Frank Robey, BLiP program manager in DARPA’s Strategic Technology Office. “Those are important, but if we want to shrink aperture size by 50% and still get the same radar performance then we need to disrupt the linear signal processing paradigm. With the tremendous increases in computer processing power available today, we can take a fresh look at radar signal processing and explore iterative, leap-ahead techniques.”
BLiP will address the current immaturity of non-linear and iterative signal processing methods. Over the course of the two-year program, end-to-end radar signal processing chains will be developed, analyzed, implemented and tested – initially through non-real-time laboratory testing and culminating in real-time implementation and full-scale field testing using an operational National Weather Service radar. Key technical challenges for BLiP will be the development, understanding, and optimization of the signal processing chain, and the practical aspects of implementing BLiP algorithms using real-time, high-performance processing.
The theoretical foundation for radar signal processing is the optimal detection of a single, known
signal in Gaussian noise. This results in linear signal processing up to the detection and track
stages. In practice, these foundational assumptions are not true and, as a result, ad-hoc modification of the optimal processing are used to address multiple targets, the lack of knowledge of the parameters of the single desired signal, and non-Gaussian noise, clutter and interference.
Linear signal processing is attractive because the theory is well-developed and understood and because linear processing can be factored into a chain of semi-independent processing blocks with relatively straightforward implementation. The development of extremely high performance Graphical-Processing-Unit (GPU) commercial computing enables increasing the processing performance by several orders of magnitude over what is required to perform linear processing and justifies a fresh examination of the end-to-end radar signal processing.
BLiP is based on the premise that when more physically realistic target and environment models
are used as the basis for the radar estimation and detection problem, then solutions require nonlinear and iterative signal processing. When non-linear methods are implemented for individual processing stages then the output of that processing block no longer provides statistics consistent with linear processing and that are expected by the following processing stages. For example, a sparse-representation Doppler processing block would not provide data to a constant false-alarm rate (CFAR) detection stage with the same statistical properties that a linear processor provides and that most CFAR detectors are designed for. The incompatibility with interfaces results in an inability to evolve current radar signal processing by upgrading or replacing one processing block or stage at a time; end-to-end processing must be addressed to go beyond the current linear processing.
BLiP will address the current immaturity of non-linear and iterative signal processing methods
by developing and testing an end-to-end processing chain in real-time. Over the course of this
program, the processing chains will be developed, analyzed, implemented and tested initially
through non-real-time laboratory testing and culminating in real-time implementation and fullscale field testing.
BLiP is organized as an applied research program, intended to perform studies, design,
development, and prototyping that will improve radar performance. The BLiP program is
directed toward general radar needs with a view to demonstrating feasibility and practicality of
non-linear and iterative processing. Key technical challenges for BLiP will be the development,
understanding, and optimization of the signal processing chain, and the practical aspects of
implementing BLiP algorithms in real-time.
Focus Area One: Non-linear signal processing. Implement and test the iterative and/or
non-linear processing that transforms information from the aperture’s raw digitized Radio
Frequency (RF) streams to a format that can be used by a target parameter estimation and
Viable approaches to non-linear processing may include, but are not limited to: Maximum likelihood, Sparse representation, Least squares representation, L1, L2 decomposition, Non-linear noise excision, and Burg extrapolation
Focus Area Two: Non-repetitive waveforms and processing. This focus area is an
extension of Focus Area One with an addition being the use of non-repetitive waveforms.
The combination of non-repetitive waveforms and non-linear processing is expected to
overcome losses due to blind ranges or speeds and to mitigate or reduce the need for fill
pulses. Much as non-uniform sampling of time series and suitable processing are able to
determine waveform spectral content, non-uniform temporal sampling of the radar
environment and suitable processing should be able to determine range and Doppler of
the radar returns. This focus area includes both waveform design and the processing.
Waveform design will be pre-planned, that is, scripted, rather than being adaptive to the
Focus Area Three: Multi-hypothesis track-before-detect. Recent developments are able to
reduce the needed target Signal-to-Noise Ratio (SNR) by 7 or 8 dB relative to current
processing with detection followed by multiple-hypothesis tracking. Multi-hypothesis
track-before-detect potentially overcomes the poor performance in the presence of target
dynamics that is experienced by Hough-transform based track-before-detect algorithms.
Focus Area Four: Main-beam interference mitigation. With the increasing pressure on RF
spectrum usage, it is clear that radars and communications systems will be sharing the
same spectrum. Over the longer term this manifests as the need to jointly operate in the
same spectrum; radars must perform well even with communications emitter signals in
the main beam. Communications systems have been sharing the same channels for many
years. One approach to ensuring robust communications systems performance with
multiple signals is to estimate the undesired waveform and subtract it from the desired,
perhaps weaker signal. BLiP will perform research in mitigating main-beam interference.
The radar that will be used to generate the test data and that will be used for real-time testing of
the performer algorithms will be the National Oceanic and Atmospheric Administration (NOAA)
National Severe Storms Laboratory (NSSL) Phased Array Radar (PAR).
The software-defined PAR was built as a prototype for a next generation multi-function weather and air traffic control sensor. This civil radar has a modern architecture consisting of 24 overlapped dual-polarization receive subarrays, and a two-channel waveform generator/exciter to provide arbitrary polarization waveforms on transmit. Signals received by the 24 subarrays are each passed through software-defined digital receivers. The digital in-phase/quadrature (IQ) data is currently passed to a signal processing cluster where beamforming and weather information processing is accomplished. BLiP will record the IQ data stream prior to full array beamforming to provide data for the performers. Later in the program, a high-performance GPU-based signal processor will be installed at the PAR. The processor will access this IQ data stream and make the data available to performer’s software for real-time demonstration of beyond linear processing
The NOAA PAR in weather surveillance mode is typically operated with waveforms that have a
6 MHz bandwidth and with dwells separated by 16 MHz. The system supports nearly arbitrary
waveforms and can utilize much wider bandwidth waveforms. Performer development must be
flexible for radar parameters including operating bandwidth up to 30 MHz, pulse width, dwell
length and Pulse Repetition Frequency (PRF). Proposers should indicate any expected limitations
on radar operating parameters.
The signal processor that will be procured by the Government team will consist of a GPU
compute cluster along with a small number of servers. For planning purposes, proposers may assume that the GPU cluster capabilities are consistent with the recently announced NVIDIA DGX-H100. This will be combined with two or more additional servers containing high core-count processors. The operating system will be Red Hat® Enterprise Linux. Accessing the DGX-H100’s 7.7PFLOPS of performance requires use of tensor core programming and attention to the processing precision.