摘要：AbstractIn this contribution we investigate the application of radial basis functions artificial neural networks embedded in hardware for real-time moving-horizon state estimation. The solution of the optimal moving-horizon state estimation problem may be faced as the mapping from the inputs and outputs of the system to the state estimates according to the system model. This mapping may be solved offline with the optimal formulation and then approximated by any higher order function approximation algorithm, such as the ones from machine learning. An approximate version with radial basis functions neural networks is developed and implemented in a Field Programmable Gate Array (FPGA) showing good results in terms of accuracy and computational time. We show that the state estimate using the approximate version of the moving-horizon algorithm can be run using a laboratory scale kit of approximately 500 kHz for an inverted pendulum at a clock rate of about 110 MHz. The latency to provide an estimate can be further reduced when FPGAs with higher clocks are used as the artificial neural network architecture is inherently parallel.
