Research On Wireless Sensing System For Commercial Mobile Devices

In recent years,wireless sensing has aroused widespread concern in academic circles,but the existing research is still conducted in a relatively ideal environment.There are still many challenges to using the wireless sensing system on a large scale.The main problems are as follows:(1)The lack of quality evaluation indexes of wireless sensing data makes it challenging to use it effectively;(2)The wireless signal is easily disturbed by the environment,which leads to the failure of the sensing model aiming at detection or recognition;(3)Dynamic change of sensing environment leads to increased training cost,and the robustness of model becomes worse;(4)The resource allocation and hardware conditions of commercial mobile devices lead to poor universality of the sensing model.Thus,this thesis studies the evaluation index of wireless sensing data quality and the sensing model’s effectiveness,robustness,and universality.The specific contents are as follows:(1)For the lack of quality evaluation method of wireless sensing data,a quantitative analysis method of sensing data based on deep learning and variability and confidence is proposed.By analyzing the behavior information of each wireless sensing data in each iterative training,the indicators for analyzing the usability of wireless sensing data are extracted.According to this index,the variability and confidence scalars are quantified,and a two-dimensional information data map is constructed.The quality of wireless sensing data is analyzed and quantified according to its distribution area in the data map.We verify the performance of the proposed method on data sets collected in different environments and public data sets,respectively.The results show that this method can quantify and analyze sensing data quality,distinguish valuable data from noise data,and be used as an evaluation index of wireless sensing data quality to guide the selection and use of data sets.(2)For the problem of model failure caused by complex interference in the wireless sensing environment,this thesis proposes a method to detect the effectiveness of the wireless sensing model by combining probabilistic and statistical assessments.According to the probabilistic and statistical assessments in the model’s prediction process,the model’s credibility and reliability are quantified.Using the anomaly detection method combined with the credibility and reliability information,we can judge whether the prediction results are consistent with the real results and detect the effectiveness of the sensing model.We have carried out largescale experiments on 11 representative wireless sensing models.The results show that this method can successfully reject 92.3% of false predictions and then successfully detect the effectiveness of the sensing models.(3)For the problem of unstable sensing model caused by dynamic changes in the sensing environment,a robustness enhancement method based on incremental learning and ensemble learning is proposed.According to information entropy and Euclidean distance,the data selection algorithm and model selection algorithm are designed for incremental learning and ensemble learning,respectively,which can reduce the overhead of incremental learning,improve the performance of ensemble learning and enhance the robustness of the sensing model with the minimum cost.We have conducted large-scale experiments on 11 representative wireless sensing models.The results show that this method can ensure the stability of the model after the dynamic changes in the sensing environment and enhance the robustness of the original sensing model.(4)For the problem of poor universality of wireless sensing model caused by limited hardware conditions of commercial mobile devices,this thesis proposes a method of improving the universality of sensing model based on classification with rejection and incremental learning.Firstly,by analyzing the basic principle of ultrasonic sensing,we use a convolution neural network model and support vector machine classifier to extract features and recognize gestures and design an ultrasonic sensing model that can be applied in different scenes.Then,the model is optimized by combining the classification with rejection and incremental learning to improve its recognition performance under different external environment changes.We have carried out large-scale experiments with different users,different locations,and other conditions.The results show that the system can achieve 98%recognition accuracy by using a pair of speakers and microphones built-in commercial mobile devices,and it is universal in different application scenarios and external environment changes.