Análise de métodos de agrupamento de dados para detecção de anomalias na precificação e categorização de peças da indústria automotiva

Abstract

The manufacturing performance of a company to reduce costs is an extremely important action to ensure competitiveness and avoid wasting resources. The variables of processes and composition of products in the automotive industry generates a daily number of combinations of data configurations and scenarios that make their manual processing unfeasible. As a result, much of the knowledge generated ends up not being used directly in similar items, resulting in large differences in costs due to small structural and product design differences. The objective of this work is to compare the performance of clustering algorithms and provided pre-processing steps for the grouping of parts, considering physical manufacturing characteristics. Then, a costefficiency comparison of similar components is made, assisting in the decision making for the formation of strategies to reach the optimum point regarding the costs of these components. Grouping was performed using the following algorithms: K-Means, K-Medoids, Fuzzy C-Means - FCM, Hierarchical, Density Based Spatial Clustering of Applications with Noise - DBSCAN, Self Organizing Maps - SOM, Particle Swarm Optmization - PSO, Genetic Algorithm - GA and Differential Evolution - DE. As a comparison metric, the following indices were used: Sum of Squared Errors - SSE, Sum of Squares Within Clusters - SSW, Sum of Squares Between Clusters - SSB, Calinski-Harabasz - CH, WB and Silhouette index. The hierarchical algorithm was the one that obtained the best practical results, when checking the SI metric and in the general result scoring all different applied metrics.