Data Mining For PCB Template Scrap Rate Prediction And Material Feeding Optimization

Printed Circuit Board(PCB)is the base of electrical and electronic equipment.With the increase of customer's individualized demand,the order of PCB template is greatly increased.The corresponding production mode is transformed from the traditional mass production to the long tail production with the main characteristics of large-scale personalized production.More precise and reasonable material feeding for the orders automatically can reduce the comprehensive cost of workshop material,production,inventory,recycling and tardiness.The scrap rate and material feeding area of each individualized PCB template order are difficult to be accurately determined in advance of the production.Many factories undergo fluctuation in both surplus rate and supplemental feeding rate with high value due to empirical manual feeding in practice by heavily depending on their experience and knowledge.Individualized surplus template products can be placed in inventory or directly destroyed while frequent material feeding brings production varieties and the instability of scheduling and therefore increase the supplemental production cost and delivery tardiness compensation.Optimizing feeding strategy that can reduce the two objectives has scientific and engineering significance.In this paper,data mining(DM)approaches are presented to establish the scrap rate prediction model and optimize material feeding of PCB orders.Initially,current research status of data mining in manufacturing,especially for PCB fabrication,was introduced;and then the technical route was established based on the characteristics of PCB template.On this basis,the research was conducted from the following four aspects:1)The PCB scrap rate related parameters have been specified combining the experience of expert from factory.30,117 records accumulated between September 2013 to October 2016 have been exported from the enterprise resource plan system.The reduction of data has been conducted based on parameter derivation and transformation.Multiple Boxplot approach was developed for the outlier detection;on this basis,the basic statistics,correlation and significance analysis of the corresponding parameters were conducted,and the core variables of the prediction model of scrap rate were established.2)The parameters of scrap rate prediction model have been and selected by hypothesis testing and multiple linear regression(MLR).And then the samples were divided into training samples(20693)and test samples(8890)based on random sampling.On this basis,the MLR model for the prediction of scrap rate have been establishment;and the surplus rate and supplemental feeding rate were defined for the model evaluation and verification.3)The artificial neural network(ANN)for the prediction of scrap rate base on the parameters selected by MLR was developed;and the parameters of ANN were determined based on some initial experiments and analysis.The simulation feeding based on results of the two prediction models was conducted and the comparison analysis was carried out.4)The MLR prediction models was selected based the result of comparison and related feeding strategy was established.On this basis,the service oriented tool was developed and the implementation for verification was conducted.The final results indicate that the prediction model based material feeding is superior to the manual feeding in reducing surplus and supplement feeding.The systematic researches were conducted in this paper from the aspects of problem analysis,parameter specification/derivation/transformation,outliers detection,scrap rate related parameters selection,prediction model development and evaluation index determination as well as implementation and verification.The novel DM-based PCB template material feeding has been established and it can also facilitate other similar problems study.