Research On Device-Free Localization With Wireless Based On Deep Dictionary Learning

With the development of emerging technologies such as the Internet of Things,there is an increasing demand for wireless localization.Among them,Device-free Localization（DFL）technology utilizes widely available wireless signals for localization without requiring users to carry any devices or violating their privacy,and has good practicality.The technology is suitable for emergency or special scenarios such as smart homes,anti-intrusion detection,and health care for the elderly,and its development prospects are promising.However,due to the extensive background noise in the monitoring environment,the correlation of localization data is large,resulting in poor localization accuracy and robustness.To address these problems,the following research is conducted in this paper.（1）In this paper,a wireless sensor network-based localization system is built in the laboratory using Universal Software Radio Peripheral（USRP）,and the localization data is collected by arranging six receiving sensors and one transmitting sensor around the monitoring area.Since the Received Signal Strength（RSS）of wireless signal is distinguishable,the system uses RSS as the localization information to achieve DFL.（2）To address the widespread background noise problem in the localization data,the paper considers using more representative information as localization features,a deep dictionary learning DFL algorithm based on the?₁-norm is proposed.The algorithm enables potential enhancement of the main features and differentiation of the data through data expansion,extracts multi-layer features of the data through dictionaries with different descriptive capabilities and the multi-layer features are integrated into the sparse representation localization model to improve the localization performance.The algorithm has better robustness to noisy data,e.g.,the proposed DDL-DFL algorithm can achieve nearly 100%localization accuracy at a signal-to-noise ratio of 15 d B at a grid size of 50×50 cm²,while the existing algorithm requires a signal-to-noise ratio greater than 20 d B to achieve 100%localization accuracy.（3）To improve the localization accuracy,we propose the deep dictionary learning DFL algorithm DCDDL using the nonconvex sparsity constrained MCP（Minimax Concave Penalty）,which introduces a nonconvex MCP regular term in the dictionary learning step of each layer to obtain a more sparse solution than the?₁-norm and reduce the correlation of the data.For the nonconvex subproblem at each layer,the use of DCA（Difference of Convex Algorithm）is proposed to solve it efficiently.The algorithm is simulated on simulated data to verify the performance of the algorithm such as dictionary recovery rate.In addition,the validation is carried out on laboratory and public data sets,respectively.As the laboratory data experiments show,the algorithm has a large improvement in localization accuracy and robustness.For a noisy dictionary with signal-to-noise ratio equal to 5 d B at 50×50 cm²grid size,the traditional algorithm localization accuracy is 57%,DDL-DFL is about 93%,and DCDDL is about 99%.In this paper,we propose two DFL algorithms based on deep dictionary learning:DDL-DFL and DCDDL.Compared with the traditional DFL algorithm,both proposed algorithms have a greater improvement in localization performance,but the DCDDL algorithm can obtain better results due to the use of non-convex penalty terms.