IMPLEMENTATION OF RECURRENT NEURAL NETWORK-BASED UNSCENTED KALMAN FILTER ON FIELD PROGRAMMABLE GATE ARRAY FOR REAL-TIME CONTROL APPLICATIONS

Abstract

In this research work, a real-time recurrent neural network-based unscented Kalman filter (RTRN-based UKF) has been formulated and implemented on a field programmable gate array (FPGA for real-time estimation, tracking and control of an NXT-robot with a validated model of the NXT-robot. The mathematical model of the NXT-Robot has been formulated and integrated with RTRN-Based UKF. In other to meet real-time constraint for real-time critical application, a high performance embedded computing platform which employs FPGA has been proposed for the tracking and control of the NXT-Robot using the RTRN-Based UKF model of the NXT-robot estimation. In other to synthesize and implement the RTRN-Based UKF algorithms on a digital hardware, the Xilinx AccelsDSP synthesis tool has been employed to generate the register transfer level (RTL) model. The development of the RTRN-Based UKF as an algorithmic co-processor model (Pcore) has been accomplished using the Xilinx system generator. In other to meet real-time constraint, the Xilinx embedded development kit (EDK) which incorporates the Xilinx platform studio (XPS) and the Xilinx software development kit (XDSK) has equally been employed for the embedded PowerPCTM440 hard processor design from where the RTRN-Based UKF is integrated as an algorithmic co-processor. The real-time recurrent neural network-based extended Kalman filter (RTRN-based EKF) of the so called extended Kalman filter (EKF) has equally been developed, implemented and compared with the proposed RTRN-based UKF. Comparing floating-point simulation results of the NXT-robot estimation, tracking and control using the MATLAB RTRN-based UKF algorithm with a total computation time of 118.3252 seconds, the RTRN-based UKF algorithm with a total computation time of 103.4173 seconds. The FPGA implementation with RTRN-based UKF appears to be a suitable implementation platform due to the significant reduction in computation time. And further study is still very necessary on the implementation of RTRN-based UKF on a dedicated platform (FPGA).