A Comparison of Clustering Measures on Raw Signals of Welding Production Data

Production data from industries today have a heterogeneous structure, which makes it difficult to analyze and derive some viable inferences. Because of the varying pattern of data, whether labeled or unlabeled, numerical or categorical, any strict standard or optimization procedure using production data is a difficult task. Applying machine learning (ML) algorithms to analyze production data has therefore become an essential requirement for industries. In this study, production data obtained from welding seams is used. We analyze raw signals of electrical current and voltage in the form of arrays obtained from welding processes by applying clustering algorithms. Each process is represented by a group number in the procured data and the corresponding welds of a group are divided into optimal number of clusters on the basis of results given by metrics such as Silhouette Scores and Adjusted Rand's Index. As a validation of the metrics, we use Davies-Bouldin Index to compare and optimize our results. We conclude that a multi-clustering technique can be devised to profile clusters of welding data using only current and voltage signals.