Research On Recognition And Forecast Method Of User Daily Activity In Smart Home

In recent years,the increasing popularity of smart home equipment and the continuous development of smart activity analysis technology have greatly promoted the rapid development of smart home technology.The two major tasks in activity analysis research are activity recognition and activity forecast,which are of great significance in computer theory research and application practice.At present,the research of sensor-based activity recognition and forecast in the smart home field has made great progress in theory and application.However,there are still many issues in single-user and multi-user activity recognition and forecast,especially low performance of multi-user recognition and forecast that requires continuous further research.Focusing on these problems above,this paper researchs the daily activities of single-user and multi-user in the smart home environment,and proposes corresponding activity recognition and forecast methods.The main research contents and achievements include:(1)The current single-user activity recognition usually adopts the step-by-step execution mode of “cutting dataset into segments first,then extracting and classifying features”.The steps are tedious.And the irregularities at the beginning and end of behaviors lead to problems such as inaccurate division of activities.This paper proposes a single-user daily activity recognition method based on self-reinforcement learning.The method first uses sensor events as the activity category to construct the feature space of daily activities,then uses unsupervised sentence embedding learning method to convert the sparse feature matrix into dense feature vector representation,and finally establishes a self-reinforcement learning model to extend the activity recognition model.(2)In view of the lack of feature representation ability generated by the feature extraction method based on sensor frequency,which is commonly used in multi-user activity recognition,and the lack of mining potential rules between activity features and the loss of a large amount of timing information,a activity feature extraction method based on time series of sensor events is proposed for these problems.First,time tuples are extracted from sensor events to form a time series.Subsequently,several common-used statistic formulas are used to form the space of daily activity features.Finally,a feature selection algorithm is employed to generate final daily activity features.(3)Aimed at the problem that it’s easy to lose relevant task-related information in the asynchronous training of daily activity,this paper proposes a multi-task forecast method of daily activity based on information interaction from the perspective of recent sensor events.Firstly,most recent raw sensor events are pre-processed to form a feature space of daily activity.Secondly,an interactive multi-task learning model which combines the correlation between the forecast tasks are developed as the forecast model to ensure their co-training.And then the two forecast tasks are output effectively in parallel.In addition,in order to solve the problem of the difference of loss functions of heterogeneous tasks,this paper introduces an information interactive mechanism to provide more supervisory signals for the forecast tasks to balance such orders of magnitude difference.Finally,in order to verify the effectiveness of the method proposed in this paper,the algorithm is implemented in Python language,and several groups of experiments are conducted based on CASAS datasets provided by Washington State University.The experimental results show that the proposed methods are effective and practical to a certain extent.