Research On Wi-Fi-based Cross-scene Behavior Perception Technology

With the development of the Internet of Things(Io T)and the popularization of commercial Wi Fi,behavioral perception technology has been widely used in various fields of production and life.Due to the wide distribution,low cost,and easy deployment of Wi Fi devices,researchers have begun to use commercial Wi Fi for human behavior perception in the past decade.Although numerous research work has proved that Wi Fi has great potential as a sensing technology,there is an obvious shortcoming at present.The Wi Fi behavior recognition system does not perform well in cross-scene recognition.The Wi Fi behavior recognition model trained in one scene is directly used in another scene,and the recognition performance will drop rapidly.In order to enable the Wi Fi perception model to be used across scenarios and enhance the robustness of the system,it is necessary to study cross-scene recognition technology.In order to solve the problems in the background,this thesis will research on Wi Fi-based crossscene behavior perception technology.This thesis uses commercial Wi Fi equipment to collect channel state information(CSI)data,and designs a series of preprocessing processes(such as filtering,normalization)to process the original CSI data stream,thereby improving the system robustness and training speed.After that,this thesis constructed two Wi Fi-based human behavior recognition models based on Bi LSTM and Conv1 D networks to realize behavior perception in the source scene.Before cross-scene behavior recognition,this thesis designs a behavior recognition model scene adaptation method based on fine-tuning.This method loads the parameters of the pre-trained model before training,and uses the target scene data set to fine-tune the model.After the fine-tuning is completed,the model can be used in the target scene.In order to make the fine-tuning process more efficient,this thesis also proposes a source model selection method based on the DBA and DTW algorithms.This method uses the DBA algorithm to merge the data set sequence,uses the DTW algorithm to calculate the similarity between the data sets,and based on the similarity give guidance on the source model selection before fine-tuning.The evaluation results show that our system is accurate and reliable.When training the source scene model from scratch,for the Bi LSTM-based system,the recognition accuracy of the two experimental scenes reached 96.55% and 91.22%,respectively.For the Conv1D-based system,the recognition accuracy of 5 experimental scenes is very high,of which 3 scenes exceed 99%.When using fine-tuning methods to adapt to new scenes,the accuracy of the Bi LSTM-based system is reduced by less than 5%,but the training time is reduced by 60%.For the Conv1D-based system,it can be maintained or even improved accuracy,and training time can also be reduced.For the source model selection method,the experimental results show that this method can further improve the efficiency of fine-tuning and make it easier to send positive migration during the fine-tuning process.