Research And Application Of Family User Portrait Based On Deep Learning

With the continuous enrichment of household products and the popularization of applications,people’s home life is becoming more comfortable and convenient.At the same time,the model transformation from smart home based on active control mode to smart home based on artificial intelligence non-sense experience has gradually become a research hotspot in the field of smart home.Based on this background,how to analyze and understand users in smart home scenarios has gradually become a significant and indispensable task.With the rapid development of deep learning theory,the use of deep learning algorithms to analyze user behavior habits in home scenarios and then construct user behavior portraits in the physical world has only been dabbled in some large enterprises and has not been explored in depth.Through user portrait technology,researchers can dig out rich user tags,analyze users and understand users with clear data,and then provide users with personalized and customized services,making home devices truly intelligent.This thesis studies the problem of how to construct user portraits in smart home scenarios.In view of the characteristics of numerous household data sources and complex data formats,this thesis proposes an overall process framework for constructing user portraits.Starting from the use records of wearable devices and household devices,analyzing user behavior habits and device usage preferences,combined with the predefined label system,finally Visually display user portraits in the form of word clouds in the prototype system.First of all,for the problem of action recognition in family scenes,this thesis proposes a wearable device-based action recognition model MCACN,which processes multi-source sensor data separately in a multi-channel manner.At the same time,the proposed adaptive convolution module can integrate the features of different time spans,so that the processed high-dimensional vector contains more abundant information,thereby improving the classification accuracy of the action recognition model on complex similar actions.Secondly,aiming at the open set scenario encountered in the actual use process,this thesis continues to improve the action recognition model,and proposes an action recognition framework under the open set scenario.The model first extracts highdimensional features by the TC-Encoder module,combined with the proposed labelbased multivariate variational autoencoder structure and unknown detectors,so that the model can focus on known classes and reject unknown classes.Thirdly,for the cold start problem when new devices join,this thesis proposes a small-sample label classification model combined with meta-learning algorithms.The basic model uses a dual-stream form to extract the timing features and channel features of the device usage records,and then divides the entire process into two stages: metatraining and meta-testing,and uses MAML to learn a set of initialization parameters with strong generalization performance.When new devices are added,even with only a small number of samples,the model can quickly iteratively converge on this task and achieve excellent classification results.Finally,combined with the above research,this thesis builds a prototype system of family user portraits.Users can manage home devices in the system.At the same time,the data analysis module displays the status of each device in the family in a visual way,and comprehensively describes the user portrait.