Research On Sensor Activity Recognition Based On Improved Deep Recurrent Neural Network

With the rapid development of sensor technology,people perceive that the world has more than one type of data and information.The increase in the number of sensors is accompanied by an increase in the amount of data.The human activity recognition(HAR)based on sensor data also brings new challenges.It is necessary to extract effective information from big data.Deep learning method should be born with the wave of big data.It can automatically extract features from massive data,realize autonomous learning,and be more intelligent.It overcomes the disadvantages of traditional manual extraction features and has unparalleled advantages over traditional methods.First of all,this paper studies the deep learning method of deep convolution neural network,deep recurrent network model,convolution neural network and long short time memory network,which are three current mainstream sensor based activity recognition methods on different activity recognition data sets,and compare the recognition accuracy and network training speed of the recurrent network.Based on this,an improved recurrent network method is used to make up for deficiencies.Second,this article describes the gated recurrent network,bidirectional recurrent network,and attention mechanism.The bidirectional recurrent neural network guarantees the integrity of information at the time point.The attention mechanism ensures the enhancement of the expression of information features.In order to solve the problem of low accuracy of the recurrent neural network recognition,an adaptive attention mechanism model is combined with bidirectional recurrent neural networks are used to improve the accuracy of model recognition.In addition,using the abstract feature representation of the recurrent network in time depth and features extracted through the full convolutional network in both spatial and temporal domains,a dual-flow gated recurrent full convolutional neural network is constructed to help improve the performance of the model.Finally,due to the problem of long training time of recurrent neural network,a Simple Recurrent Unit(SRU)is used to solve the time dependence problem in the “door loop” of the recurrent network and accelerate the training of the network.In addition,the real environment sensor data used in this paper often have unbalanced and empty data.To solve this problem,a Focus loss function is introduced,this paper uses multi-class Focal loss to realize the solution of class imbalance in the field of multi-classification activity recognition under sensor data.Problems have also achieved good experimental results and improved model performance.