Resident Daily Activity Recognition And Prediction Methods Based On Improved Long Short Term Memory Model

The research goal of daily activity analysis in the smart home environment is to recognize(daily activity recognition)which daily activity residents are performing and predict(daily activity prediction)which daily activity residents will perform by analyzing the environmental data sensed by the sensors deployed in the smart home.The results of the daily activity analysis can provide health protection for residents,effectively saving on diagnosis and care costs.It can also provide personalized assistance and services for residents in their daily life.Although current Long Short Term Memory model has been proved to be an effective model for daily activity analysis,the effect of daily activity recognition and prediction is still unsatisfactory,and the accuracy of daily activity recognition and prediction still needs to be improved.In order to improve the effect of daily activity recognition,this thesis proposes a Bi LSTM-Attention daily activity recognition method based on data dimension upgrading.The method firstly uses the feature extraction algorithm to obtain the relevant feature vectors,and then transfers the feature vectors to the Bidirectional Long Short Term Memory model to learn the time-series relationship between features for activity recognition.Then,the Bi LSTM Neural Network and Attention Mechanism are combined to improve the ability to improve the discrimination ability of daily activity features.Finally,the original model is improved,and the input data of the Bi LSTM-Attention model is upgraded to improve the daily activity recognition effect.The proposed model was validated on two publicly available datasets and the proposed model improved the F1-score by 19% and 11%respectively over the benchmark approach.In addition,this thesis proposes a daily activity prediction method based on the CNN-Attention-LSTM model.This method uses sliding windows to extract features,and then uses Convolutional Neural Network and Long Short Term Memory to extract local neighborhood information and long-term dependency information,and then introduces Attention Mechanism on this basis to effectively address the problem of poor prediction performance due to the neglect of inter-task correlations,thereby improving the prediction effect of the model.The proposed model was validated on two publicly available datasets.For the prediction of daily activity categories,the F1-score of the proposed model are15.8% and 15.3% over the benchmark method.For the prediction of the occurrence time of daily activity,the Root Mean Squared Error of the proposed model are improved by 10.72 and 23.05 respectively compared with the benchmark method.