Research On Wireless Sensing Technology Based On Sharing Mechanism Under Multi-Task Learning Framework

Due to the wide coverage,security,and high-speed advantages of Wi-Fi networks,now an increasing number of researchers are beginning to utilizing Wi-Fi signals for wireless sensing research.By receiving Wi-Fi signals,we can obtain Channel State Information(CSI),which is a measurement parameter that estimates the channel state through orthogonal frequency division multiplexing,reflecting characteristic information about the surrounding environment.CSI represents a more granular physical layer information.In order to achieve high precision and reliability in wireless sensing,this paper utilizes a multi-task machine learning framework to carry out CSI-based research on the perception of personnel location and identity.The main contributions are as follows:(1)The research explores the theory of CSI-based wireless sensing and establishes an experimental platform.Firstly,the CSI measurement values are described,followed by an introduction to the framework structure,loss functions,and theoretical advantages of multi-task learning.The commonly used machine learning algorithms in wireless sensing are then investigated.Finally,both hardware and software platforms for the CSI-based wireless sensing system are built,laying a solid foundation for subsequent research work.(2)The algorithm proposed is based on a hard parameter sharing mechanism and Deep Residual Shrinkage Network(DRSN)for wireless sensing.In the offline phase,outliers and noise in the raw CSI amplitude measurements are eliminated using Hampel filtering and wavelet thresholding.Then,a CSI fingerprint database is constructed by combining location and identity labels.Subsequently,the DRSN network is employed to extract deep feature information from the CSI,enhancing task-related features by learning thresholds and soft thresholds while suppressing irrelevant features.Finally,a dual convolutional neural network is utilized for classification learning of position estimation and identity recognition,resulting in a multi-task prediction model.In the online phase,the received raw CSI measurements are preprocessed,and the position and identity information is estimated using the multi-task prediction model.Experimental results validate the effectiveness of the proposed algorithm.(3)The proposed algorithm is based on a hybrid sharing mechanism and NDDR-CNN for wireless sensing.In the offline phase,a multi-task training dataset is constructed using action labels,X-axis position coordinates,and Y-axis position coordinates.The DRSN network is employed for feature extraction,followed by the NDDR-CNN network for identity classification learning and position regression learning,resulting in a multi-task prediction model.This network automatically integrates features from different tasks at each layer,effectively capturing the correlation between tasks and reducing the risk of overfitting.The Grad Norm loss optimization algorithm is utilized to balance the losses between asymmetric tasks,dynamically adjusting gradient magnitudes and learning rates.In the online phase,the constructed CSI fingerprint samples are processed through the multi-task learning model to obtain estimates of position and identity.Experimental results demonstrate that the proposed algorithm achieves improved performance in terms of localization and identity recognition.