Remaining Useful Life Prediction Of Deep Neural Network Based On Genetic Algorithm Optimazation

In industrial production,Useful Life(RUL)prediction based on data-driven has strong practicability because of it does not require complex physical modeling and expert prior knowledge.Recently,deep neural networks are often used in RUL.However,the performance of deep neural networks is greatly affected by its structure and hyper-parameters.Due to a large of parameters,it is difficult to select appropriate parameters to obtain a high-performance deep neural network for different data.Evolutionary algorithms can automatically search the network structures with high performance for different data.Therefore,the thesis focuses on the design of evolutionary algorithm and the design of the connection of deep neural network to obtain an optimal neural network structure.The remaining life of the system is predicted using the optimal neural network structure.The research has important application value.The main research contents are as follows:(1)The selection strategy of genetic algorithm has a greater impact on finding a high-performance neural network.Therefore,a modified selection strategy of genetic algorithm is presented.Half of individuals are from better fitness individuals in population of father combined with child and the rest half of individuals are selected from their poor fitness individuals with a greater probability.And the structure and hyper-parameters of the neural network are used as variables.They are coded by binary and real number respectively.Further,dropout technology is used to optimize the network structure found.The modified genetic algorithm combines the BP algorithm to search optimal network structure for different data.Finally,the optimal deep neural network is used to predict RUL.Experimental results show that genetic algorithms with modified selection strategies can find higher-performance network structures.(2)For shortcoming of selection strategy and deleting connections used dropout technology in first research,a new modified selection strategy and the sparse strategy of neural network connections based on the weight are designed.At the same time,through analysis of the activation function in the first research,here,the activation function is optimized as a new variable in order to search higher performance structures.Finally,the results show that RUL predicted by the modified algorithm achieves better experimental results.