In order to reduce the influence of signal transmission quality and distance estimation algorithm on the localization accuracy, deep learning was applied to the ultra wide band（UWB）indoor localization system, the gated recurrent unit（GRU） network was used to replace the traditional trilateral measurement, to improve the indoor localization accuracy of UWB. After get the distance from the localization tag to the base station, the distance information was input into the GRU network, and the final localization coordinates were output. As a variant of recurrent neural network（RNN）, GRU not only had the advantages of RNN in processing time series data, but also solved the long-term depenencies problem in RNN. The parameters of different learning rate, optimizer, batch size, network layer and the number of hidden neurons in the GRU networks model were adjusted and trained. The result showed the UWB indoor localization system based on the GRU networks model improved the localization accuracy significantly, the average localization error was 6.8 cm.