Research On CSI Indoor Positioning Technology Based On Data Preprocessing And Deep Learning

Since indoor localization plays an important role in our social life,it has become a research hotspot.High precise localization is faced with great challenges in complex indoor environments.In this paper,choosing the channel state information(CSI)as the measurements,the key technologies of indoor localization using data preprocessing and deep learning are studied.The main contributions are described as follows:(1)The related theories of CSI indoor localization using deep learning technique are studied.The principle of machine learning based indoor positioning system is introduced at first.Then,the theoretical knowledge and experimental platform of CSI localization are described.At last,the learning framework and classical network structure of deep learning are described.(2)A convolutional neural network(CNN)based localization algorithm using principal component analysis(PCA),wavelet transform and color feature extraction is proposed.The PCA and wavelet threshold transform are used to reduce the dimension and measurement noise of the amplitude and phase information of CSI measurement,respectively.Then the linear mapping method is used to construct the CSI amplitude and phase image.The color square equalization method is used to extract the color features and form the fingerprint.Next,the CNN is used for X axes and Y axes based classification learning.The position classification models of X and Y axes are obtained at last.Experimental results demonstrate the efficiency of the proposed algorithm.(3)A two-stream convolutional neural network based localization algorithm using CSI image feature is proposed.The CSI amplitude and phase images are constructed using CSI amplitude and phase data at first.Then the CSI image features are extracted to form the fingerprint.The wavelet transform is used to extract low frequency information and remove high frequency noise.For another,the Sobel operator method is used to detect edges and extract image features.At last,the two-stream convolutional neural network is used for position based classification learning and the position based classification model can be obtained.Experimental results show that the proposed algorithm has better localization performance than some existing algorithms.