Access Type

Open Access Dissertation

Date of Award

January 2022

Degree Type

Dissertation

Degree Name

Ph.D.

Department

Industrial and Manufacturing Engineering

First Advisor

Kyoung-Yun Kim

Abstract

A significant information gap is prevalent between the design domain and the manufacturing domain. The designers lack manufacturing awareness as they have little knowledge regarding the manufacturability of their designs. The welding domain falls prey to this issue as designers lack manufacturing awareness and welding engineers lack weldability of the product assembly. Data-driven techniques have shown promising results in the analysis and understanding of complex welding processes. Data analytics play a significant role to turn data into valuable insights to assist in the weldability certification decision-making or weldability prediction for Resistance Spot Welding (RSW) as well. We have used machine learning algorithms for welding quality/weld diameter prediction as an alternative solution to the expensive time-consuming destructive welding certification testing. However, to successfully perform the associated data analytics, domain knowledge is essential to construct more ‘sense-making’ analytics models, as often the models cannot properly capture the nuances of the domain and do not properly indicate the relationship among the RSW concepts and parameters. Thus, machine learning models developed from rough experimental data often do not provide models meaningful and sensible to the domain expert. We have developed a data-driven semantic RSW weldability prediction pipeline to effectively capture the semantics/ domain knowledge and efficiently transfer the information captured from real-life data to the design modeling platform. To achieve this feat, an RSW welding ontology has been developed. The extracted machine-learning decision rules are converted into semantic rules for building the knowledge-based semantic prediction tool for predicting weld quality (weld diameter). The combined semantic platform containing formal knowledge and welding quality prediction will help the welding experts understand the weldability of the targeted product assemblies. To aid in this process, The target product assembly, along with the process parameters and the predicted response parameter is visualized using a modeling platform. In this manner, the domain experts can gain insight into the target welded product with the aid of the visualization platform and can decide if the final product is adequately welded as well as if the prediction performance is suitable based on the current process parameters. We have introduced a recursive approach to this pipeline, so that, consulting with the formal semantic platform and the visualization platform recursively, the domain experts share their opinions on how the prediction performance can be improved and if any change is needed in the data collection process. By making the suggested changes to the dataset (adding any significant process parameter or removing any non-significant process parameter), we have found out that the prediction performance can be improved from the experiment. It has enhanced the qualitative knowledge of the manufacturing processes and potentially increases the confidence of the experts in the data-driven models. The framework will help the design and welding engineers increase manufacturing awareness and help them in making design decisions in the early product life cycle phase.

Download

Share

COinS