| Discrete manufacturing is an important form of enterprise production,which has the characteristics of flexible production paths and complex equipment relationships.In order to ensure the efficient progress of the discrete manufacturing process,further integrate production resources,and improve the transparency and intelligence of the production process,it is necessary to collect and store the production process data efficiently and flexibly.The collected production data contains a wealth of production information.In different data application scenarios,it is necessary to mine and analyze the production data according to specific application requirements.Aiming at the problem of product quality prediction,this paper uses feature engineering to filter data,and constructs new features according to data characteristics,forms a data set which directly related to product quality,and modeling using ensemble learning methods,realizes effective prediction of product quality.The specific work is as follows:1)For production data collection,the framework of cyber-physical system for discrete manufacturing is firstly summarized.The production data collection and equipment control methods are researched from the aspect of data collection and control of PLC,robots,CNC machine tools,and the data collection of sensors.Discussed the production data management scheme with Postgre SQL database as the core,and designed the data table according to the production data type.A prototype system for production data acquisition and management was developed to realize the data acquisition and management of Siemens PLC,Huazhong CNC machine tools,Huashu robots and PCB acceleration sensors.2)Aiming at the problem of low information density of production data,feature engineering is performed on the data to obtain effective features for product quality prediction.Numerical features are screened according to the gain of numerical features in the construction of the XGBoost model,correlation analysis,and the distribution of failure rates.According to the time characteristics,the production start time and the number of unqualified products are fitted through the cubic spline,then the fitted curve is derived to obtain the production line performance characteristics,and the production time,batch,production path and other characteristics are constructed.For categorical features,the repetitive features are removed based on the Hash calculation,the XGBoost algorithm is used to select the integer-encoded features,and the WOE method is used to encode the final retained features.The quality prediction effect of the data set obtained by using feature engineering is compared with that of the original data set,which illustrates the necessity of feature engineering.3)For product quality prediction,use ensemble learning methods for modeling.By analyzing the characteristics of discrete manufacturing data,four ensemble learning algorithms,Random Forest,XGBoost,Light GBM and Cat Boost,are selected.The Bayesian optimization algorithm based on TPE agent is used to select the key hyperparameters of the models,and compared with the random search method,the results show that the former has higher search efficiency and can obtain the better hyperparameters.The optimized hyperparameters were used to establish the selected ensemble learning models,and the product quality prediction performance of each model was compared.4)In order to further improve the accuracy of product quality prediction,the model ensemble method based on neural network and Focal Loss is studied,the back propagation mechanism is used to automatically adjust the ensemble weight of models,and the overall model’s nonlinear expression ability is improved by using hidden layers.Compared with the traditional mean ensemble and weighted ensemble methods,the proposed method has a better integration effect.Finally,the proposed ensemble model is compared with a single ensemble learning model to verify the effectiveness of the proposed method. |
PDF Full Text Request