Root cause prediction for failures in semiconductor industry, a genetic algorithm-machine learning approach.

Area of Science:

• Engineering
• Computer Science
• Materials Science

Background:

• Failure analysis is crucial for electronic component quality and reliability.
• Predictive models for failure analysis require effective feature selection from textual data.
• Existing feature selection methods struggle with large datasets and textual data.

Purpose of the Study:

• To develop a predictive model for failure conclusions using discriminant features from failure descriptions.
• To address challenges in feature selection for textual datasets in failure analysis.
• To improve the accuracy and reliability of failure analysis in electronic component manufacturing.

Main Methods:

• Preprocessing textual data using Natural Language Processing and vectorization.
• Combining Genetic Algorithms (GA) with supervised learning classifiers (Decision Tree, Support Vector Machine).
• Utilizing an F1 score as a fitness function for handling unbalanced datasets, leading to GA-DT and GA-SVM algorithms.

Main Results:

• The proposed GA-DT method demonstrated superior performance in predicting failure conclusions compared to using all features or features selected by GA-SVM.
• Experiments on failure analysis datasets confirmed the effectiveness of the GA-DT approach.
• Quantitative metrics like BLEU score and cosine similarity validated the prediction accuracy.

Conclusions:

• The GA-DT model offers an effective solution for predicting failure conclusions by selecting discriminant features.
• This approach enhances the ability to identify component flaws and understand failure mechanisms.
• The findings contribute to improving product quality and reliability in electronic component manufacturing through advanced failure analysis.