Research On Indoor Location Algorithm On Deep Learning

Indoor positioning technology has been a research field of great interest in recent years.As satellite positioning technology cannot be used in indoor environments,indoor positioning has always been a challenging problem.To solve this problem,researchers have proposed many indoor positioning solutions based on different technologies and algorithms,among which deep learning-based indoor positioning algorithms have demonstrated outstanding performance in terms of positioning accuracy and robustness.However,the mainstream indoor positioning algorithms still face issues such as low positioning accuracy and low recognition rate of Non Line of Sight signals.To address these issues,this paper designs and implements a deep learning-based indoor positioning algorithm that includes research on positioning algorithms,NLOS signal recognition,and trajectory prediction and correction,after analyzing the Ultra Wide Band indoor positioning technology.The main contents are as follows:(1)To solve the problem of low accuracy of the initial input dataset of traditional deep learning-based indoor positioning algorithms,an improved Chan-Taylor algorithm is proposed,which uses the solved position coordinates as the input dataset for the deep learning algorithm.This algorithm uses the improved analytical solution obtained by the Chan algorithm as the initial value for the Taylor algorithm’s iterative calculation.It takes advantage of the fast solving of the Chan algorithm while using it as the initial value for the Taylor algorithm to obtain more accurate final position points.Simulation experiments demonstrate that the proposed algorithm has higher positioning accuracy and better performance as the input dataset for deep learning algorithms.(2)To solve the low recognition rate of NLOS signals in indoor positioning,a Back Propagation neural network and a Convolutional Neural Network are proposed.First,the BP neural network is applied to the recognition of NLOS signals,but the effect is not ideal.Therefore,a series of architecture modifications are made based on the classical CNN Le Net-5 to better extract feature information.Adjustments are made to some parameters in the network structure,such as convolution kernels and learning rates,and real datasets are used to test and achieve good results.The overall recognition rate in various non-severe occlusion scenarios is above 88%.(3)To further improve the accuracy of indoor positioning in complex environments,a bidirectional long short-term memory network is proposed to perform sequence modeling on the coordinate positions obtained by the improved Chan-Taylor algorithm,and applied to the prediction and calibration of indoor positioning trajectories.By learning sequences of feature vectors,the temporal and spatial features of signal data are captured to achieve position prediction.An independently designed positioning system with hardware,firmware,and software components was used to demonstrate the high accuracy and superior performance of the proposed method for indoor positioning through real-world measurements,outperforming the performance of the long short-term memory network.