The radio frequency (RF) spectrum has become increasingly congested. In this environment, future cognitive radar systems must be capable of real-time reconfiguration utilizing fast search algorithms to insure maximum power-added efficiency (PAE) at each operating frequency while meeting spectral constraints. Tunable amplifiers have been built based on micro-electrical mechanical systems (MEMS) technology using a genetic search algorithm , but MEMS technology is limited in its power-handling capability and genetic search algorithms are inherently slow in searches where criterion characteristic attributes are partially known. Particle swarm optimizations have been attempted, but hundreds of experimental queries can be required . In the present paper, we utilize a 90 W evanescent-mode cavity (EVA) impedance tuner designed by Semnani for radar applications . The impedance presented by the tuner to the amplifier is based on the variation of the two cavity piezoelectric disc positions, given by position numbers and. In a previous paper, we have demonstrated a modified gradient search for frequency agility .
This paper presents an addition to our previous algorithm  that reduces the time necessary for reconfiguration by using previous search results at each operating frequency using a look-up table (suggested by Sun ). Our objective is to use an EVA tuner  as the power amplifier load matching network in a cognitive radio that can perform target tracking while avoiding radio-frequency interference (RFI) (Fig. 1). In the eventual application, ten 10-MHz sub-bands between 3.25 GHz and 3.35 GHz will be used. The radar bands will be chosen in real time based on a Markov Decision Process predictive algorithm with reinforcement learning . In such a scenario, the radar will need to quickly perform multiple relocations. The ability to predict frequency will further extend previous work from the literature on spectrum scanning [7–8]. To develop and demonstrate fast frequency agility, we consider relocations directed by a random selection of one of the ten sub-bands.
The gradient-based search implemented in our previous paper  tunes n1 and
The initial search to populate the look-up table can either be performed with the system online or with a separate, off-line initial search. For the initial search used to populate the look-up table at each frequency, search parameters of Dn=10, Ds=100, and Dr=5 were used. For each subsequent time a frequency is visited, the lookup table starting location is used for the gradient part of the search, removing the five initial measurements required by “Sarvin’s Method” . The search parameters for subsequent searches utilizing the lookup table are changed for finer tuning and faster convergence to Dn=10, Ds=40, and Dr=2. The look-up table is updated if a higher constrained optimum PAE is achieved.
For all the results presented in this section, the input radar signal was limited to a bandwidth of 1 MHz. This allows the frequency to be changed as in the scenario of Fig. 1 without incurring effects of tuner impedance variation over the frequency range of the input signal spectrum. Future efforts will address a wider bandwidth approach.
A. Single Frequency
Measurements were first performed to examine the ability to reduce search time and maintain or improve performance in repeated measurements at a single frequency. For this test, the real-time search algorithm was repeated at a center frequency of 3.305 GHz to produce maximum use of the lookup table. Table I shows the search end values of n1 and n2, PAE, and Sm, along with the number of measurements and time required. PAE shows a slightly improving trend over the succession of searches, and the end cavity position numbers also shift for later searches. While time is shown, the number of measurements may be a better metric for use here, as a traditional load-pull bench is used and not the final cognitive radar platform, resulting in significant equipment communication overhead time. The experiment was repeated without allowing the lookup table to be used, and a comparison is presented in Fig. 3.
In this experiment, one of the ten center frequencies between 3.255 GHz and 3.345 GHz was randomly chosen for each iteration to emulate the frequency agility shown in Fig. 1. The search was repeated 100 times in total (this means that each frequency receives an average of approximately 10 iterations). Fig. 4 shows the number of measurements required for each iteration, with colored markers used to designate the operating frequency of each iteration. The average number of measurements required to reconfigure tends to decrease over time in general, as expected. By iteration 49, all frequencies have been used at least once. Figs. 5(a) and 5(b) show generally decreased measurements and consistent PAE performance for repeated searches at individual frequencies. The higher optimum PAE for higher frequencies (Fig. 5(b)) coincides with measured load-pull data.
A fast search algorithm has been demonstrated in a frequency-agile coexistence scenario using a high-power impedance tuner. The use of a look-up table to provide a good search starting point based on past search results provides for tuning with fewer measurements while maintaining PAE search results. The improvement on impedance tuning time will have a significant impact on radar performance metrics such as output power and detection range, as well as power-added efficiency. This will allow radar performance to be maintained in a dynamic cognitive radar spectrum sharing environment.
This work has been funded by the Army Research Laboratory (Grant No. W911NF-16-2-0054). The views and opinions expressed do not necessarily represent the opinions of the U.S. Government. The authors are grateful to John Clark of the Army Research Laboratory for assistance in development of this paper.