ML Based Intelligent Filter with Real Time Calibration of Transmitter (MIR-CAT)

In response to the Army’s solicitation for an emitter system-agnostic solution for threat representative waveform generation and validation, Alphacore and the proposing team propose to develop an ML Based Intelligent Filter with Real Time Calibration of Transmitter (MIR-CAT). In Phase I, a prototype for assessing the efficacy of applying novel artificial intelligence & machine learning (AI/ML) algorithms to enhance the capability of both legacy and state-of-the art low-cost RF emitters will be developed. In Phase II, a carefully selected combination of hardware, including reconfigurable analog front-end (AFE), analog-to-digital converter (ADC) and FPGA, will be built to meet the low cost goal of the program while able to work with carrier frequencies up to K-band (27GHz, or beyond if needed). This proposed work will integrate advanced algorithms in an open software framework capable of real-time adaptation of pre-distortion techniques on commercial off-the-shelf (COTS) radio hardware. While there are numerous ML algorithms under investigation for intelligent control of RF transmitter front-ends (RFFE) systems, many of these do not consider the specific low-SWaP-C and real-time needs and capabilities of Army transmitters, nor the harsh radio frequency environments under which they operate. Digital preconditioning is a form of applying adaptive filters and non-linear functions to compensate for the distortions caused by the RFFE, particularly compression artifacts of the up-converter and power amplification stages of the transmitter. These artifacts reduce the quality of the transmitted signal, reducing the communication capacity of data links or degrading the performance of radar systems. While digital pre-conditioning techniques have the theoretical ability to correct for these impairments, in practice calibrating these effects is extremely difficult as the characteristics of the transmit chain vary over time, temperature, frequency band of operation, and output power. As such, applying a real-time feedback mechanism is considered optimal; however with the large number of configurable parameters comes the need for intelligent control of the digital preconditioning stage. Deep-learning models are a tractable solution to this problem as they allow for online learning, removing the reliance on expensive calibration efforts that are susceptible to unforeseen environmental conditions that would otherwise adversely affect the front-end performance.