Line-based process duration prediction using the vehicles features for the automotive industry

Abstract

Purpose: In this study, unlike the traditional approach, using regression, regression and classification trees, and fully connected artificial neural networks, which are among statistical and artificial intelligence techniques, the waiting durations of the vehicle in the buffer zones, process durations of the welding, assembly, dyehouse lines were predicted. Theory and Methods: The techniques used in this study are Multiple Linear Regression (MLR), Classification and Regression Trees (CART) and Artificial Neural Networks (ANN). First of all, data obtained from different data sources were combined, metadata was extracted, and divided to include line-based processing times and attributes. Using the metadata obtained, the datasets were digitized in both sequential numeric and binary form. Forward selection, backward elimination, stepwise selection and coefficient of determination selection methods are the feature selection methods in MLR technique that were used to find significant features and to reduce insignificant ones in datasets. Numeric, binary, reduced numeric and reduced binary datasets obtained for 7 process lines and 3 buffer zones were used to create process time prediction models with MLR, CART and ANN techniques. In the beginning, 45 models were created, the technique with the lowest error was determined, and 16 new models were created for the technique and re-run. Results: Among the 45 models which are the initial models, the models with the lowest error for the process duration estimation were determined and the results were found to be acceptable. In these models, it is observed that the CART technique applies the models with the lowest error. In the CART technique, 16 new models were created, and these models improved two process lines, one buffer zone. Conclusion: In this study, the time estimation of 7 different lines and 3 different buffer zones was carried out with 3 different techniques over 4 different data set structures by using 43 different features of vehicles produced. As a result of the predictions, it is seen that the tree structure gave the best result for 8 lines and the regression for 2 lines. It is observed that artificial intelligence approach gives acceptable results in buffer zones other than Welding Buffer Lower Floor, which is one of the buffer areas used only as waiting zones, where there is no production. The most important advantage of these techniques is that they can produce predictive results for the vehicle to be produced for the first time and for which there is no knowledge. In other words, when a new vehicle goes into mass production, the process duration prediction can be performed using only the attributes of 150 vehicles whose process time is unknown.