Research On Incremental Learning Of Relevance Vector Kernel Machine Over WSN

Wireless Sensor Network(WSN)is a new information acquisition method and processing technology.Because of its convenient deployment,low power consumption and low cost,it has been applied in many fields and has given more application space.And application value offers possibilities.The biggest feature of WSN is application-oriented and data-centric.Therefore,the purpose of deploying WSN is not only to collect data and transmit it to observers,but also to complete specific tasks such as tracking,identification,and early warning.Classification and regression are the most basic and important tasks in the many tasks that WSN will accomplish.Therefore,the machine learning method for solving the classification and regression problems has been more and more widely used in the WSN.In WSN,the training samples are scattered on the sensor nodes,and the sample data has the characteristics of continuous generation and sequence arrival.It is difficult to use traditional machine learning methods to train the training samples in the WSN effectively.Therefore,how to train machine learning methods in WSNs with strictly limited node energy and communication capabilities becomes a key problem to be solved in the WSN data fusion technology.This dissertation aims at WSN,based on the characteristics of nuclear learning machine,using incremental learning theory and data dimensionality reduction theory to reduce the training cost of machine learning method model in WSN,balance the node energy consumption in the training process,and reduce the participation in model training.In order to improve the efficiency of model training,we studied the incremental learning methods of sparse nuclear learning machines,the dimensionality reduction methods of training samples,and the energy dynamic balance methods in the incremental learning process of sparse nuclear learning machines.Specifically,the problems of incremental learning from sparse nuclear learning machines,data dimensionality reduction based on principal component analysis of nuclear learning machines,and energy balance strategies of WSN are studied in depth.The main research contents and research results of this paper are as follows:(1)Aiming at the strict limitation of node energy and communication bandwidth in WSN and the characteristics of training sample data,based on the probability theory,the construction and solution of incremental learning problem for sparse nuclear learning machine are studied.An incremental based on correlation vector nuclear learning machine is proposed.Learning algorithm.The simulation experiment verifies that the incremental learning algorithm can obtain the prediction accuracy rate that is very consistent with the traditional nuclear learning machine method,and can significantly improve the sparsity rate of the model.(2)The training and transmission costs caused by the high dimension of the original training sample in the WSN are based on the incremental learning algorithm of the correlation vector kernel learning machine.A data dimension reduction method based on principal component analysis of a nuclear learning machine is studied.The simulation experiment verifies that without changing the prediction accuracy of the model,the computational load in the model training process and the demand for the memory space size can be significantly reduced.(3)For the problem of node energy consumption imbalance in model training in WSN,based on the incremental learning algorithm of the correlation vector kernel learning machine,the clustering strategy that dynamically balances node energy balance during incremental learning of nuclear learning machine is studied.Compared with the traditional clustering strategy,the non-average clustering strategy is optimized,so that the energy consumption of nodes in the WSN will be more evenly consumed with the implementation of the heterogeneous clustering strategy.Simulation experiments verify that under the heterogeneous clustering strategy,the energy consumption at the nodes in the WSN can be significantly reduced.