Quality Prediction Methods For Batch Processes

With the rapid development of information technology,the scale of the manufacturing process continues to expand,the complexity is increasing,and more and more data are being generated.Effective information is extracted from vast amounts of data to improve the safety and reliability of the production process and improve the quality of products that has become one of the research hotspots in the manufacturing industry.Quality forecasting based on the historical data produced in the production process,which using feature engineering techniques to build feature sets and establish efficient product batch quality prediction models,it is of great significance to control the production process and make corresponding decisions.Because the factors affecting the quality fluctuation have the characteristics of high dimension,non-linearity,and multi-period,the quality forecasting still faces many difficulties and challenges.It is a research hotspot for us to use feature engineering techniques to build feature and choose the appropriate regression algorithm to improve the accuracy of quality prediction for batch processes.The current study uses feature extraction methods to extract features from divided sub-periods,but this method cannot provide enough batch fluctuation information.In the feature selection,many researches have used the PCA method to select the optimal features.For nonlinear data,it is impossible to accurately select the variables that have a large impact on quality.For the high-dimensional,non-linear data in the production process,using the traditional linear regression method can not meet the industry's accurate prediction of quality.The research in this paper mainly design a batch of multi-stage feature engineering and quality prediction methods for the characteristics of unequal cycle data length,high dimension,multi-period,and nonlinearity in the intermittent production process.In order to solve the problem of unequal cycle data.First of all,looking for indicator variable replace time,the indicator variable value is reduced downwards,and the indicator variable corresponding feature values are sampled by means of exponentially moving weighted average method.In order to solve the problem of high dimension,Then dividing the indicator variable into multiple stages and use statistical methods to extract statistical features at different stages.Using XGBoost and Random Forest merging methods for feature selection;Finally,a multi-stage prediction model based on Stacking integration algorithm is established to predict product batch quality.This paper uses the manufacturing quality control data set provided by the Ariyoshi cloud intellectual task platform to analyze and visualize the historical data of the production process.According to the processing progress,the batch production process is divided into three stage,and the processing is completed.The multi-stage feature engineering and quality prediction methods designed in this paper are used for each stage that solved the problem of unequal batch data and high dimension,and achieved better prediction results for product batch quality.The experimental results show that the characteristic engineering and quality forecasting methods designed in this paper are correct and effective for the processing of complex data in batch industrial production processes and the prediction of product batch quality.