Predictive Model for Electrical Assembly Manufacturing Defect Dispositions
Open AccessDefects produced during the manufacturing of electrical assemblies require timeand resources to evaluate and determine next steps. After the identification of a defect, an engineer performs a review. Using their expertise and judgement, they decide if the assembly will be reworked, removed and replaced, retested, scrapped or determined to not be a defect. In the Defense Industry, there is a strong emphasis on continuous improvement to remain competitive. Industry 4.0 has increased the amount of manufacturing and quality data available for analysis and provides new opportunities to apply machine learning. This research is a practical application of machine learning methodologies to predict defect dispositions for a manufacturing center that produces electrical assemblies in the Defense Industry. The original data set included two sources from a quality database with 22,508 instances from January 2019 through June 2022. This data set was unbalanced with significantly more minor rework classifications than any other type. The imbalance was addressed through ensuring the test data set had the same proportion of classifications, choosing models that typically perform well with that condition, cross-validating the models, and measuring balanced accuracy in addition to overall accuracy. The four models trained, validated, and tested included: a simple baseline that chooses the majority class, Decision Tree, Random Forest, and AdaBoost. The baseline had an overall accuracy of 68% and a balanced accuracy of 17%. Decision Tree had an overall accuracy of 84% and a balanced accuracy of 61%. Random Forest had an overall accuracy of 89% and a balanced accuracy of 72%. AdaBoost had an overall accuracy of 80% and a balanced accuracy of 70%. Random Forest performed best in overall accuracy, balanced accuracy, precision, recall, and F-1 score. Recommendations for future research include collecting additional data to expand the data set, especially in the less common classes to improve class accuracies and balanced accuracy. Researchers may wish to focus on test failures and predicting which components will need to be reworked using the quality data from this research in combination with test data. Additionally, other features such as cost of assembly and yield could be explored to further improve accuracy performance.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.
| Thumbnail | Title | Date Uploaded | Visibility | Actions |
|---|---|---|---|---|
|
|
Rodriguez_gwu_0075A_16236.pdf | 2022-12-11 | Open Access |
|