RF Sensing for Internet of Things: When Machine Learning Meets Channel State Information
Type of DegreePhD Dissertation
DepartmentElectrical and Computer Engineering
MetadataShow full item record
With the rapid development of Internet of Things (IoT) techniques, RF sensing has found wide applications for, e.g., indoor localization, activity recognition, and healthcare. In this dissertation, we investigate the problem of RF sensing for IoT using channel state information (CSI) and machine learning techniques. In particular, our work mainly focuses on indoor localization using deep learning and vital sign monitoring for RF sensing. In this dissertation, we first study the problem of CSI based indoor localization. For first three works, we exploit deep learning for three different indoor localization systems using CSI amplitudes, CSI calibrated phases, and CSI bimodal data, respectively. Moreover, we study and analyze CSI data, which is stable for indoor localization. We consider deep autoencoder networks to train CSI data, and employ the weights of the deep network to represent fingerprints. A greedy learning algorithm is leveraged to train the weights layer-by-layer to reduce computational complexity, where a sub network between two consecutive layers forms a Restricted Boltzmann Machine (RBM). In the online stage, we use a probabilistic method for online location estimation. Then, we exploit deep convolutional neural networks (DCNN) for indoor localization. Since DCNN is a supervised method, it only requires to train one group of weights for all the training data with related labels, which is different with our prior works that requires training weights for every training location. Specially, we use estimated angle of arrival (AOA) images from CSI data as input to the DCNN. By executing four convolutional and subsampling layers, the system can automatically extract the features of the estimated AOA images, to obtain training weights. To improve indoor localization accuracy, we propose deep residual sharing learning for training two channels CSI tensor data. Moreover, we can stack many residual sharing blocks for adding the depth of the deep network, thus achieving higher learning and representation ability for CSI tensor data. The proposed system can achieve decimeter level location accuracy, which is better than other deep learning methods. This dissertation also focuses on vital sign monitoring using CSI and machine learning techniques. First, we consider CSI phase difference data to monitor breathing and heart beats with commodity WiFi device. We implement data preprocessing for the collected CSI phase difference data to obtain the denoised breathing signal and the restructured heart signal. Moreover, we leverage the peak detection method for breathing rate estimation and FFT based method for heart signal estimation. To estimate breathing rates for multiple persons with CSI data. We leverage the tensor decomposition technique to handle the CSI phase difference data. This work first uses CSI phase difference data to create CSI tensor data. Then Canonical Polyadic (CP) decomposition is applied to obtain the desired breathing signals. A stable signal matching algorithm is developed to find the decomposed signal pairs, and a peak detection method is applied to estimate the breathing rates for multiple persons. To improve the robustness of breathing signs monitoring, we exploit bimodal CSI data, including amplitude and phase difference, for realtime breathing monitoring. Then, we implement the data preprocessing, adaptive signal selection, and breathing signal monitoring modules, and employ peak detection to estimate breathing rates. The last work of this dissertation considers a phase based active sonar to monitor breathing rates with smartphones. We implement several signal processing algorithms, including signal generation, data extraction, received signal preprocessing, and breathing rate estimation. Specially, we propose an adaptive median filter approach to remove the static vector in the received signal, which allows to effectively extract the inaudible phase information. Our experimental results validate the superior performance in different indoor environment settings.