Adjustment mode decision based on support vector data description and evidence theory for assembly lines

Abstract

PurposeThree adjustment modes are alternatives for mixed-model assembly lines (MMALs) to improve their production plans according to constantly changing customer requirements. The purpose of this paper is to deal with the decision-making problem between these modes by proposing a novel multi-classification method. This method recommends appropriate adjustment modes for the assembly lines faced with different customer orders through machine learning from historical data.Design/methodology/approachThe decision-making method uses the classification model composed of an input layer, two intermediate layers and an output layer. The input layer describes the assembly line in a knowledge-intensive manner by presenting the impact degrees of production parameters on line performances. The first intermediate layer provides the support vector data description (SVDD) of each adjustment mode through historical data training. The second intermediate layer employs the Dempster–Shafer (D–S) theory to combine the posterior classification possibilities generated from different SVDDs. The output layer gives the adjustment mode with the maximum posterior possibility as the classification result according to Bayesian decision theory.FindingsThe proposed method achieves higher classification accuracies than the support vector machine methods and the traditional SVDD method in the numerical test consisting of data sets from the machine-learning repository and the case study of a diesel engine assembly line.Practical implicationsThis research recommends appropriate adjustment modes for MMALs in response to customer demand changes. According to the suggested adjustment mode, the managers can improve the line performance more effectively by using the well-designed optimization methods for a specific scope.Originality/valueThe adjustment mode decision belongs to the multi-classification problem featured with limited historical data. Although traditional SVDD methods can solve these problems by providing the posterior possibility of each classification result, they might have poor classification accuracies owing to the conflicts and uncertainties of these possibilities. This paper develops a novel classification model that integrates the SVDD method with the D–S theory. By handling the conflicts and uncertainties appropriately, this model achieves higher classification accuracies than traditional methods.