Name: GABRIEL DOS SANTOS SERENO
Publication date: 11/07/2024
Examining board:
Name |
Role |
|---|---|
| CELSO JOSE MUNARO | Presidente |
| FRANCISCO DE ASSIS BOLDT | Examinador Externo |
| LUIZ AFFONSO HENDERSON GUEDES DE OLIVEIRA | Examinador Externo |
| THOMAS WALTER RAUBER | Coorientador |
Summary: In this study, specialized Random Forest classifiers are trained using the SHAP feature selection method to identify faults in industrial processes. The classifiers receive input data from the decomposition of the multiclass problem using the One-vs-All technique. This methodology was applied to two benchmarks: the Tennessee Eastman Process and the Continuous Stirred Tank Reactor. The SHAP method achieved a performance of approximately 89% in the F1-score metric for the Tennessee Eastman Process, selecting around four features on average per fault. For the Continuous Stirred Tank Reactor, the algorithm achieved 90% in the F1-score metric and selected about three features on average per fault. The study showed that the Recursive Feature Elimination (RFE) method obtained similar results compared to SHAP in both benchmarks. However, the RFE method tends to select more features to achieve the same performance. Finally, the study suggests that SHAP can reduce the dimensionality of the dataset while maintaining good performance in the F1-score metric.
