| In the Smart Environment,due to the changing factors caused by different requirements,the system adapts to these changes only through algorithms,resulting in poor adaptability of the system,which makes it difficult for the intelligent system to be widely applied and popularized in the current environment.Software-Defined Intelligence(SDI)separates the common elements that support intelligent implementation into an "intelligent layer" pre-installed on the node device and then summarizes the changing factors into rules.The intelligent layer solves the problem of poor adaptability caused by various changes in the way of updating rules.This thesis applies SDI to the sleep action recognition system in the Smart Environment,which mainly includes the following work:(1)Based on the hierarchical model of SDI,this thesis proposes a system framework for action recognition during sleep,which is mainly composed of data acquisition,data preprocessing,feature extraction,and work engine modules.The working engine is the core component of SDI for rule-based reasoning,and it deals with various changing factors based on the rule updating method,thereby improving the adaptability of the system for sleep action recognition.(2)A time queue is presented to extract the features of actions in real-time,including energy,frequency-domain entropy,root mean square,etc.By implementing the Classification and Regression Tree(CART)algorithm in training these features,the decision tree model for action recognition during sleep is obtained.Based on the feature importance parameters provided by the model,a feature selection algorithm is proposed,which can screen out the important features and retrain them to get a new decision tree model.Then the new model is pruned according to the random search algorithm and the Cost-Complexity Pruning(CCP)algorithm.Based on the pruned model,this thesis devises a rule extraction algorithm for extracting sleep action recognition rules required by the working engine.Depending on these rules,the working engine reads the user data in real-time and infers sleep action conclusions.(3)Through the contrast experiment,it is verified that the system framework can quickly adapt to changes in node position,number of nodes,and user requirements in sleep action recognition by updating rules.Experiments show that the system framework can make full use of the computing resources of multiple nodes and recognize nine regular types of sleep actions: the recognition precision of each type can exceed 96%;the total recognition accuracy can reach 98.9%.Notably,this system framework can facilitate the recognition of other-type actions.Importantly,it has more robust adaptability than other systems while guaranteeing the accuracy of action recognition.(4)A nursing system based on sleep action recognition is designed.Users can check the sleep action status of the people being nursed through the system,including action conclusion data graph and action statistics data graph,and can also obtain expert knowledge recommended by the system according to related diseases. |
