An Ensemble Neural Network Location Algorithm Based On Deep Abstraction

With the rapid development of wireless network technology,and the popularity of smart phones,Wi Fi hotspot of large-scale coverage.Indoor location using Wi Fi signal is becoming a hot research hotspot.This paper uses CSI as the position fingerprint,and proposes a new algorithm which integrated the deep learning algorithm and ensemble neural network,this algorithm can effectively improve the indoor positioning effect.Specific contents are as follows:1)At first,this paper studied the CSI,through experiment contrast CSI and the characteristics of RSSI.The results show that CSI has better stability and position sensitivity in the complex indoor environment,so it is more suitable for indoor location.Then we use the BP neural network algorithm for indoor positioning research.The results show that the location algorithm based on neural network is affected by the CSI data and the CSI sample selection,which has two problems with large fluctuations and accuracy is limited;2)For the two problems of neural network location algorithm,this paper proposes an innovative solution method.For the problem of future selecting can impact the positioning accuracy,we proposed a method based on the deep learning algorithm which can abstract the feature of the original CSI data,through the stacked auto encoder,the high-dimensional CSI data down to low dimensional data and extract the future which can characterize the essence of the data,this effectively improve the positioning accuracy.For different CSI sample set selection causes the problem of locating the problem,this paper proposes a weighted ensemble algorithm based on localization error.First,we trained a number of different neural network models with different number of CSI data.Then,according to the error of location,the weight distribution of each model is obtained;3)Based on the above conclusions,this paper proposes an ensemble neural network localization algorithm based on deep abstraction.In the offline phase,we use CSI data training stack auto encoder,then use trained stacked auto encoder to abstract all CSI data,then these abstract features of CSI as a sample set trained neural network model,and according to the location error we calculating the weight of each model.In real-time positioning stage,the collected raw CSI data will be abstracted future through the trained stacked auto encoder,and then we use neural network models carry on location respectively,and the final location result is integrated from different location results by weight.