Research On Daily Activity Prediction Method On Multi-task Interaction And Temporal-Spatial Convolutional Network

At present,many countries in the world have entered the aging society,and there are more and more elderly people suffering from different degrees of cognitive impairment,The tracking and monitoring of the cognitive function level of the elderly has become one of the main reasons for the high cost of the elderly care.With the continuous improvement and increasing popularity of smart home devices in recent years,automatic tracking and monitoring of the cognitive function level of the elderly is expected to become one of the main ways to reduce the cost of elderly care.Since the automatic monitoring of the cognitive function level of the elderly depends on the analysis of their daily behavioral characteristics,the automatic identification and prediction of daily behaviors has become a hot topic of current research at home and abroad.The research work in this thesis is aimed at two problems in the field of daily activity prediction.To address the problem that the single-task prediction approach ignores the connection between different tasks of daily activity prediction,this thesis proposes a multi-task daily activity prediction method based on interactive feedback.The method first extracts features from the original sensor events based on the association between most recent sensor events and daily activities,introduces the most recent activity category information as the initial features,and generates a sample feature space with this.Then a multi-task activity forecast model is constructed based on the association between daily activity tasks in smart homes,based on which a feedback mechanism is introduced to balance the differences between the loss functions of two heterogeneous tasks for daily activity prediction,solving the problem of inconsistent gradient descent rates of heterogeneous task loss functions during the training process,and also increasing the diversity of features.On the basis of the feedback mechanism,a multi-headed attention mechanism is introduced as a feature selector to give more attention to the features with high discriminative ability,so as to select the features that are more relevant to the task.To address the problem that the existing sensor event prediction research ignores the sensor spatial information,this thesis proposes a sensor event forecast method based on spatio-temporal convolutional network.The method firstly extracts features from the original sensor event stream to form a feature matrix;then abstracts the spatial layout of sensors in smart homes as an undirected graph to form a sensor adjacency matrix and establishes a spatio-temporal convolutional prediction model;Convolutional neural networks(Graph Convolution Networks,GCN)obtain sensor spatial dependencies,and gated recurrent units(Gate Recurrent Unit,GRU)are used to obtain sensor temporal dependencies.The spatio-temporal convolution prediction model is used to combine temporal and spatial information to realize the prediction of sensor events.Finally,this thesis uses the Python language to implement the above method in code and validates the effectiveness of the proposed method in this thesis on the publicly available CASAS dataset from Washington State University.The experimental results show that the methods proposed in this thesis have better prediction results.The average Precision,average Recall,average F-score,and average R~2of the multi-task daily behavior prediction method based on interactive feedback improved by 7.0%,6.49%,6.62%,and 3.81%,respectively,and the average MAE and average RMSE decreased by 8.54 and 27.61 compared with the three baseline models.Compared with the two baseline models,the sensor event forecast model based on spatio-temporal convolutional network improves the average Precision,average Accuracy,average Recall and average F-score by 4.08%,4.06%,3.97%and 4.05%,respectively.