| The intelligent factory is an important part of intelligent manufacturing,and the prediction and evaluation of production system performance is the basis of intelligent factory production control and decision optimization.Intelligent factories realize the functions of Internet of things,information sharing,service collaboration and so on through technologies such as sensors and Internet of things.By collecting massive and multi-source manufacturing data,the evolution trend of the performance indicators of the manufacturing system and the influence rules between the data are mined,so as to realize the prediction of workshop performance,which is conducive to guiding the operation optimization and independent regulation of the workshop.Under the background of intelligent factory,this paper studies the performance prediction of production system based on data-driven,and the main research work is as follows:(1)Performance prediction problem analysis of production system based on data-driven.This paper briefly summarizes the characteristics of production system under intelligent factory,points out that obtaining effective information from massive manufacturing data and constructing reasonable prediction model are the key of research,and introduces the datadriven performance prediction architecture of production system.(2)Key information acquisition method in performance prediction for production system.In response to the characteristics of manufacturing data in intelligent factory,such as high dimension,nonlinearity and high redundancy,which affect the prediction accuracy and operation speed.It is necessary to select features as the key information for the performance prediction of production system.Thus,a two-stage filtered feature selection method jointly constructed by mutual information and cross mutation operator improved particle swarm optimization(GPSO)is proposed.In this method,Laplacian score and Pearson correlation are combined as the fitness function,and GPSO algorithm is used to search the optimal solution to avoid premature PSO algorithm.So as to the data highly related to performance indicators,removing redundant features and reconstruction features,and provide effective data support for later performance prediction,and provide effective data support for later performance prediction.In addition,PCA contribution rate is used to identify important factors of abnormal state.(3)Performance prediction method considering timing and strong noise.Aiming at the classical problem of low prediction accuracy caused by serious timing and strong noise,based on the key feature subset selected by feature selection,a long short term memory(LSTM)network prediction model based on improved depth self encoder is proposed to realize accurate prediction.Among them,noise reduction self encoder and sparse self encoder are used to construct depth self encoder to enhance feature learning ability and noise reduction ability;PSO algorithm is used to optimize LSTM network parameters and learning rate to improve the prediction effect.Taking the processing cycle of production performance index as the prediction object,an example shows that the prediction accuracy of the model is higher than that of other traditional models.(4)Design and implementation of prediction prototype system for production system performance.Thus,the application system of performance prediction for production system is provided for workshop managers of intelligent factory,and the function of manufacturing execution system of intelligent factory can be improved. |
