Abstract: The one-step ultra-wideband (UWB) coordinate error correction models based on the generalized regression neural network (GRNN) and back-propagation neural network (BPNN) were proposed to address the difficulty of correcting the coordinate error existing in UWB positioning based on conventional means. The correction models took the UWB original positioning coordinates,the distance between it and different base stations as inputs,and the UWB relative high-precision reference value error as output. The correction models were trained with GNSS RTK point coordinates as the dynamic experimental reference values and total station point coordinates as the static experimental reference values,respectively. Besides,the correction models were employed to correct the UWB coordinates of the non-modeled sample points. Then a comparative analysis of the accuracies before and after correction and the accuracies of the different correction models was conducted. The results show that the method of using artificial neural networks to construct the one-step UWB coordinate error correction models is feasible and it is easier and faster without the need to solve the coordinates using the corrected distance. The correction models can effectively improve the dynamic and static positioning coordinate accuracy of UWB overall. Among them,the correction performance of the GRNN-based correction model is the most significant. Moreover,the GRNN-based correction model can correct the UWB coordinate error more effectively than the BPNN-based correction model. The accuracy of the corrected UWB dynamic positioning planar coordinate can reach the centimeter level,and the accuracy of the static positioning planar coordinate is as high as the millimeter level.

Key words: ultra-wideband positioning, coordinate error correction, generalized regression neural network, backpropagation neural network, directly correction