Hardware-Software-Data Framework for Intelligent and Cognitive Radars

Emerging applications of radars such as automotive imaging and security sensing are characterized by an increase in the volume of data collected and requires higher processing powers to convert data into information. In addition, the interest to add artificial intelligence capabilities to radars and/or fuse data with other sensors adds to computational requirements. Further, as the electromagnetic spectrum becomes congested, the need for spectral coexistence with other wireless systems also becomes important. Despite the additional complexity, radars will be expected to perform in real-time with high energy efficiency. These challenges present opportunities for innovations in intelligence and cognition in radars. Our group explores the following research topics to address the above:

Our group is interested in developing AI algorithms and new architectures that will enable low latency and high energy efficiency in radars. We are particularly interested in embedded-AI solutions for radars. 

AI can learn to master benchmarks but fail in the real world if the training datasets are not representative of real-world scenarios. Hence data collection and curation are important, and often neglected, components to the success of AI-based technological solution. Benchmark wireless signal datasets (communication and radar signals) for training and testing AI do not currently exist. A major challenge to acquiring shareable radar datasets is the the RF/microwave/millimeter wave front-end systems add distortion that are specific to the instrument used to make the measurement. Therefore, our group is investigating calibration techniques to create receiver-agnostic wireless signal datasets. 

Multi-mode sensors often provide more information about the scene under observation (see pactive sensor project below). Our group is interested in novel approaches to fuse radar data with data from other sensors to convert data to information with high efficiency. 


Towards Pactive (Passive + Active) Sensors for Security Applications

My PhD dissertation demonstrates a new approach to noninvasively extract the complex permittivity and thickness of dielectrics using a combination of passive and active microwave measurements using a pactive (passive+active) sensor. Pactive sensors show potential in security application since data from two sensors can provide mode information about concealed items and aid in better identification of the item. In my doctoral work the passive system is a prototype total power radiometer operating at 23.55 GHz and the active system is a prototype frequency modulated continuous wave (FMCW) radar operating within 18-26 GHz. 

Imperfections in the RF front-end as well as background clutter impact the performance of both the radar and radiometer. To mitigate these issues, the following calibration techniques are developed: 

(i) a calibration algorithm to correct RF front-end distortion in FMCW radars resulting in a measured range resolution of 2.5 cm

(ii) a radiometer system equation (RSE) for correcting mismatch and temperature-dependent insertion loss contributions in the RF front-end, 

(iii) a novel calibration technique to correct background noise for radiometers in indoor applications resulting in less than 0.5 K (kelvin) average error in measured radiometric temperatures. 

Electromagnetic models to predict the radar and radiometric responses of multi-layer dielectric targets are investigated and experimentally verified using dielectric stacks consisting of one or two relatively low loss dielectric substrates, representing clothing and an unknown material, separated from a simplified human phantom (warm water) by a Styrofoam + air spacer layer. The models are then used to extract the dielectric constant, loss tangent and thickness of an embedded AD450 layer within a dielectric stack consisting of LDPE, AD450, Styrofoam, air, and warm water. The extraction process is based on optimization of four objective functions – three of which are radar-based while the other is based on the radiometric error. For best results in the presence of measurement error, a two-step approach is recommended consisting of a coarse optimization to find a narrow range of values for all three parameters of interest followed by a fine optimization using a different radiometric data point to extract the values. Preliminary results yielded less than 5% overall error in the extracted parameters. These results demonstrate a methodology for using a combination of passive and active microwave measurements for dielectric characterization.

Wireless Signal Datasets for Spectrum Classification and Prediction 

With technology poised to enter the age of the Internet of Things, the world has seen developments in wireless communication standards to facilitate high-data transmissions in parallel with an exponential growth in the number of wireless devices. Managing the congested EM spectrum in the face of this exploding wireless market is a challenge that can be solved with AI. My Postdoctoral research focused on generating and curating datasets for wireless signal classification and prediction. Some projects I worked on are: