Research On Sensor Activity Recognition Based On Recurrent Neural Network

With the concept of ubiquitous computing filtering into human's mind,wearable sensors are increasingly appearing in people's daily life.The most import signal processing problem in Assisted Living(AL)systems is improving the accuracy and realtime of the redcongnition of human activity(HAR).Because of the high performance and abandoning extracting features manually,deep learning is becoming more and more popular in HAR.This paper mainly studies the application of the recurrent neural network(RNN)in the field of human activity recognition,applies the proposed algorithm to three different types of datasets,and has achieved good application and research results.First,in order to exploit the advantages of convolutional networks,recurrent networks and support vector machines for different types of sensor data,this paper considered a model that combines the above three algorithms based on the idea of ensemble learning.The validity of the fusion model was verified on the UCI_HAR dataset without significant changes in test time.Secondly,aiming at the shortcomings of poor feature extraction capability and unreasonable use of features in RNN,this paper designs three attention layers for classification tasks.After carefully studying the effects of these three mechanisms on the accuracy of the model with different datasets,it is proved that the attention layer proposed in this paper can improve the ability of feature representation and the performance of RNN.Finally,the use of natural-type dataset labels has always been one of the difficulties in the field of activity recognition.This paper proposes to use the entire set of labelss to convert classification task to segmentation task.Drawing lessons from image segmentation and machine translation models,this paper designs two attention models based on segmentation tasks.It introduces RNN into activity recognition segmentation tasks,and achieves good results on two public datasets.