Research On Deep Feature Fusion Network For Wearable Sensor Activity Recognition

The rapid development of sensor technology and the advent of the era of big data,make intelligent wearable devices appear more and more in people's daily life.How to use the deep learning technology to accurately identify sensor data has become the most active in the field of activity recognition.One of the important research directions.Compared to traditional methods,deep learning technology has great advantages in recognition accuracy.But due to the large amount of parameters and high operating costs,restricts its use in industry.In view of the above problems,this paper will use the feature fusion idea to carry out the following research from three aspects: model accuracy,parameter scale and running speed.Firstly,in order to more efficiently extract the spatial and temporal domains of sensor data,this paper designs a feature fusion model that can be calculated in parallel.A two-branch network model is constructed based on convolutional neural network and cyclic neural network.At the same time,in order to improve the accuracy of the model,the compression constraint unit and the attention mechanism are introduced to optimize each branch.Experiments show that this parallel feature fusion network structure can well realize the feature extraction of sensor data,thus further improving the classification accuracy of the model.Secondly,for the problem of excessive parameters of convolutional networks and stacked long and short memory networks,the two parameters of hybrid packet convolution and nested long and short memory networks are used to compress network parameters,and inter-group information fusion technology and switchable normalization operations are used.Guaranteed network accuracy.Then use two compression strategies to build the network,and through experiments,the two compression strategies used in this paper can effectively simplify the network parameters under the premise of ensuring accuracy.Finally,for the problem of excessive time cost of the training of the loop gated network,the simple recurrent unit is used for improvement.Through the optimization of the number of gates,the speed of the network is improved,and the unit is used to improvethe bidirectional cyclic gated network.At the same time,the operating cost of the network is reduced,and the attention mechanism is introduced to optimize the performance of the entire network.And through the experiments on the Skoda and WISDM datasets,the rationality of the improved techniques used was proved.