Enhancing manufacturing process by predicting component failures using machine learning

• Main Authors: Saadat, Raihanus, Syed Mohamad, Sharifah Mashita, Azmi, Athira, Keikhosrokiani, Pantea
• Format: Article
• Published: Springer 2022
• Online Access: http://psasir.upm.edu.my/id/eprint/101274/; https://link.springer.com/article/10.1007/s00521-022-07465-1

Manufacturers customize computer components upon receipt of sale orders. They perform burn-in tests on each unit of product before shipment to ensure a high standard of quality. Burn-in is normally associated with high production costs and slows down manufacturing operations. This study aims to enhance the manufacturing process by predicting test failure patterns using machine learning methods. By identifying the components that are likely to cause failures, manufacturers can accelerate the rectification process and improve delivery time which in turn leads to better customer service. This study hypothesized that the component of concern produces a higher test failure rate. To provide insight into the data and test the hypothesis, descriptive and predictive analytics are used at various stages. Predictive analytics was performed using machine learning via Naïve Bayes since it outperformed SVM and Random Forest classifier. For the descriptive analysis stage, a visual representation revealed many components (81) to be associated with a more than average test failure rate. Fisher’s exact test confirmed that 12 of them are statistically significant and worth studying their behaviour further. Moreover, an association rule mining exercise identified several combinations of modules that have a higher inclination with the test failure. For the predictive analytics stage, the Naïve Bayes classifier predicted test failure with 79% accuracy and 53% recall rate. Another Naïve Bayes classifier predicted error messages associated with a test failure with 68% recall rate over manually labelled error messages. However, a neural network-based automatic text classifier was developed and tested that yielded 66% accuracy. This analysis provides the foundation for a recommendation made that can reduce the burn test failure rate by 25% which is expected to increase further with the improved performance model upon training with a larger data set.