A Comparative Study Of Deep Neural Network For Mobile Phone Sensor Activity Recognition

With the accelerated pace of modern life,due to academic pressure or work pressure young people often ignore their own health,and also because of the pressure of young people caused them have not a lot of time to accompany the elder at home.In the event of an accident,the elderly can't be timely relief,so it is necessary to monitor and identify people's daily activities to give feedback on their own health or to timely relief the accidental elderly.The use of sensors embedded in portable devices to collect information on human activities,through the analysis of information to identify human activities is one of the hot research.The deep neural network can directly analyze the original data without special treatment,and can carry out "self-learning",compared with the traditional classification method,the former has a unique advantage.This paper mainly introduces the different depth neural network structure,and then uses different networks to classify and identify the human activity signals collected by the sensors.This article has done the following work:Firstly,the convolutional neural network(CNN)is introduced with LeNet-5 network as an example,and the function of each part is analyzed.The training process of CNN is then introduced,and the activity identification process of WISDM human activity data set collected by the embedded mobile phone sensor is used.Note CNN is a more efficient and convenient algorithm than Logistic Regression,J48,and Multilayer Perceptron.Secondly,the recurrent neural network(RNN)is introduced.Since the RNN has the problem of gradient explosion / gradient disappearance,the long short-term memory network(LSTM)is introduced and its advantages in dealing with the time series of human activities are introduced in detail.We used LSTM to deal with UCI HAR mobile phone sensor data sets,and through multiple sets of contrast experiments under different parameters,we obtained a multi-layer LSTM network with good classification effect.Finally,the residual network(ResNet)is introduced.The structure of the basic unit "residual unit" and the signal processing of the multiple unit networks are analyzed,and the ResNet method is migrated from the processing image to the classification of WISDM human activity data sets,a number of comparative experiments were carried out.The correctness of activity identification was gradually improved with the deepening of the network,but the effect was slightly worse than that of CNN.The reason was discussed in this paper.