Research On Wireless Sensing Technology Using Deep Learning And Channel State Information

Now,wireless sensing technique has received much attentions.Location based services and activity recognition have been widely used in many aspects.However,due to the complex environment,existing sensing technologies are difficult to obtain better estimation performance.In order to solve above problem,in this paper,the deep learning based wireless sensing algorithm using channel state information(CSI)is studied.The main contributions are described as follows:(1)The related theories and methods of wireless sensing are studied.The existing sensing technologies and methods are described first.Then the wireless sensing methods based on CSI are introduced.At last,the wireless sensing system is studied in which the measurement parameters,the hardware and software platforms and used machine learning techniques are described in detail.(2)A CSI based wireless sensing algorithm using convolutional neural network(CNN)and support vector machine(SVM)is proposed.In the proposed algorithm,two processes of the activity recognition and localization are realized in parallel but independently.In the off-line learning phase,the amplitude difference preprocessing is first conducted to reduce the measurement noise.Then,the CSI measurement information in the temporal,spatial and frequency domains are exploited to construct the CSI image.Next,the CNN is utilized to perform classification learning to obtain the optimal weight parameters for activity recognition.For another,a CNN based deep learning network is designed to extract the optimized deep image features of the CSI image firstly.Then the position regression models of the X-axis position and Y-axis position are obtained by SVM based regression learning.In the on-line phase,with the constructed CSI image,the activity and the position are estimated by the activity classification and position regression models,respectively.Extensive experimental results show that the proposed algorithm has better performance than other existing approaches.(3)A CSI based wireless sensing algorithm using multi-task CNN is proposed.In the proposed algorithm,two learning processes of the activity recognition and localization are joint each other.In the off-line phase,the amplitude based CSI image is constructed at first.Then,assigning three different labels,activity label,X-axis coordinate label and Y-axis coordinate label to CSI image,the training data is formed.The multi-task CNN is used for multiple-task learning and obtain activity classification and position regression models at last.In the on-line phase,with constructed CSI image,the activity and position can be estimated with activity classification and position regression models.Experimental results show the efficiency of the proposed algorithm.