Eventos Anais de eventos
COBEM 2021
26th International Congress of Mechanical Engineering
BAYESIAN NETWORKS AS PRODUCT CONFIGURATION SYSTEMS (PCSs) TO SELECT ELECTRIC MOTORS COMPONENTS
Submission Author:
Bruno Ziegler Haselein , SC , Brazil
Co-Authors:
Bruno Ziegler Haselein, Gabriel da Silva Sezerino, Jonny Silva
Presenter: Bruno Ziegler Haselein
doi://10.26678/ABCM.COBEM2021.COB2021-0926
Abstract
Electric motors are used in virtually any industrial application that requires conversion of electrical energy into mechanical energy due to their high efficiency. However, this product is highly customizable and has to comply with specific configurations according to the customer’s needs. Together, these factors result in high demand for industries that design and manufacture the product. The configure-to-order approach allows the user to define the product configuration at the time of the order and the supplier to develop a product that meets the customer's needs. In this context of Design for Mass Customization (DfMC), the use of Product Configuration Systems (PCSs) is a decisive factor in obtaining the benefits of the mass customization approach due to improvements in knowledge management and control of product variants. The objective of this study is to evaluate the use of Bayesian Networks (BNs) as a PCS inference engine to select features of electric motors in a multinational-level company. These networks are considered one of the most effective theoretical models in the fields of knowledge representation and reasoning with uncertainty. The approach combines theory of probabilities and theory of graphs, and has a rigorous mathematical consistency that allows making inferences according to observations and based on the solid rules of probability calculation. Initially, two BNs structures are tested: the naïve Bayes, which uses a fixed structure and therefore dispenses any search engine, and the Tree Augmented Naïve Bayes (TAN), which limits its search space to a set of structures that corresponds to the Bayes naïve plus edges that form a tree whose nodes are the explanatory attributes. Next, the PC algorithm is used to define a general BN structure. This study uses recent data to learn the BNs structures and Conditional Probability Tables (CPTs), and the model evaluation method is cross-validation. The results indicate that it is possible to accurately predict in the proposed network structures the value of a characteristic, e.g., bearing size, given observations about other values. Nevertheless, the accuracy to identify some features, e.g., flange type, is highly sensitive to the network structure. The proposed approach allows the company to identify potential product variants and, consequently, to reduce costs and to better meet customer needs. This work is a preliminary study of a dynamic PCS, in which the characteristic values are automatically suggested to the user based on the probabilities calculated in BNs.
Keywords
product configuration systems, Bayesian networks, electric motors, design for mass customization
