Research On The Predicting Method Of Complex Product Duration Based On Deep Learning

In the background of intelligent manufacturing leading the transformation and upgrading of traditional manufacturing,“predictive manufacturing” based on industrial big data technology is one of the key topics.And the forecast result of the complex product duration is an important indicator for the manufacturing enterprise to formulate a reasonable production plan,if accurate prediction of the duration of complex products can be achieved,it will not only improve the on-time delivery rate of the enterprise effectively,but also provide a technical reference for the manufacturing enterprise to move towards “the predictive manufacturing”.In the discrete manufacturing industry,the prediction of the duration of complex products mainly has the following problems: 1)due to the reliance on expert experience to predict the duration of complex products,and the lack of data support and scientific theory,the accuracy of prediction results is quite low;2)the prediction of duration is hard because of the traditional data analysis methods are failed to accurately confirm the influencing factors of complex product duration.Aiming at the problems above,this paper proposed a method for predicting the duration of complex products based on deep learning.The main research contents are as follows:(1)analysis and mining of key influence factors of complex product duration.Investigated complex product design and development,material distribution,production planning,production processing,assembly and other related links,collated the production data and experience of complex products,adopted statistical analysis methods to explore the data mechanism of complex products,and completed data preprocessing;Then proposed a feature extraction method based on machine learning method for influencing factors in the duration,this method employs RF-RFE,PCA,and k-means clustering algorithms to implement important feature selection,linear feature extraction,and non-linear feature extraction of influencing factors of complex product durations;At last,adopted a stack autoencoder to perform unsupervised feature extraction on influencing factors of complex product duration.These two methods realized the analysis and mining of the key influence factors of the complex product duration from multiple angles and dimensions.(2)constructed a complex product duration prediction model based on the deep neural network optimized differential evolution algorithm.According to the mining results of key factors which affecting complex product duration,a complex product duration prediction model based on deep neural networks was proposed,and a differential evolution algorithm was used to optimize the number of nodes in each hidden layer of the deep neural network to establish an adaptive complex product duration prediction model.Aiming at the difficulty in setting the learning rate of deep neural networks,an adaptive learning rate optimization method based on the loss function is proposed,which accelerates the convergence rate of the complex product duration prediction model.(3)constructed a complex product duration prediction model based on integrated deep neural network.Against the problem of weak stability of a single complex product duration prediction model,an integrated method based on k-means clustering and Stacking integration is proposed.Firstly,the constructed deep neural network duration prediction model is screened initially by k-means clustering.Then the Stacking integration strategy is used to fuse the deep neural network duration prediction model to improve the prediction accuracy and generalization ability of the complex product duration prediction model.Based on the research contents above,firstly,the data preprocessing was completed by combining the knowledge of empirical mechanisms of complex products;Secondly,the key influencing factors of the complex product duration was mined;Then,a prediction model of the complex product duration was established.Finally,the circuit breaker assembly process is taken as an example to verify the effectiveness of the method proposed in this paper.