Wearable Sensor Activity Recognition Based On Deep Forest Research

With the rapid development of sensor technology,people also have great demand for wearable devices with sensors as components.The processing of sensor data collected by volunteers wearing wearable devices has become the main research field of human activity recognition.The traditional machine learning method needs to extract the features manually,and the feature relationship cannot be handled well.The neural network is limited to the non-differentiable,the calculation is complex,and it is impossible to make a reasonable theoretical explanation.this paper mainly studies the activity recognition model with both advantage of the tree and the deep neural network.Firstly,An activity recognition algorithm based on deep neural decision tree is constructed.Aiming at the uninterpretability of deep neural network and the indivisibility of decision tree model,this paper studies the deep neural decision tree model which can use random gradient descent training instead of greedy splitting.The experimental results show that the activity recognition model finally converges and the convergence speed is fast;the accuracy of activity recognition reaches 91.5%,which is better than the traditional decision tree method and the neural network of the same structure;the number of feature cut points is increased,and the recognition accuracy is improved by 3.56%.Secondly,An activity recognition classification model based on multi-grained cascade forest is designed.For traditional machine learning,the feature relationship cannot be handled well.The neural network model has excellent representation learning ability and difficult to adjust the parameters.This paper studies the multi-grained cascade forest model.The experimental results show that the multi-grained scanning layer increases the types of time series features,which is beneficial to improve the performance of the cascaded forest model;the cascade layer enhances the hierarchical representation of features.The recognition accuracy on WISDM reaches 98.25%.Changing the type of cascaded forest-based classifiers,the recognition accuracy on UCI-HAR increased by2.21%.Finally,The activity recognition algorithm based on multi-layer gradient boosting decision tree is studied.Aiming at the problem that neural network is limited to the use ofback propagation algorithm training and gradient boosting decision tree hierarchical distributed representation ability of the micro-module,this paper studies the multi-layer gradient boosting decision tree that can be jointly trained using the target propagation variant.Experiments show that the feature representation increases with the depth of the model,and the lower training loss does not over-fitting.The recognition accuracy on UCI-HAR reaches 92.87%,and reaches 96.20% on the natural data DAPHNet.Practical application significance.