Name: ALEXANDRE ROSSETO LEMOS
Publication date: 01/10/2024
Examining board:
Name |
Role |
|---|---|
| CELSO JOSE MUNARO | Coorientador |
| FRANCISCO DE ASSIS BOLDT | Examinador Externo |
| KARIN SATIE KOMATI | Examinador Externo |
| THOMAS WALTER RAUBER | Presidente |
Summary: Failures in industrial processes can lead to equipment deterioration, pose risks to
worker safety, harm operational productivity, and reduce profitability, potentially
necessitating process shutdowns for maintenance. With the advancement of Artificial
Intelligence, the use of machine learning algorithms for fault detection tasks has
become increasingly prevalent. Industrial data are typically highly complex due
to the large number of features present in industrial processes, making feature
selection algorithms a crucial step. In this study, a feature selection methodology
was developed that operates in two stages: the first stage uses an algorithm for
the initial ranking of features, and the second stage employs Sequential Forward
Selection to obtain the optimized feature subset. The use of the SHAP algorithm for
the initial feature ranking was proposed, expanding its application beyond model
interpretability. The results obtained with SHAP were then compared to those
obtained using established algorithms in the literature for initial feature ranking:
mRMR and ReliefF. In 16 out of 21 faults analyzed, the use of SHAP yielded
superior results in terms of F1-score and FDR. The proposed method achieved an
average F1-score 3.34% and 5.96% higher and an average FDR 4.74% and 8.21%
higher than those obtained by mRMR and ReliefF, respectively. Furthermore, the
proposed methodology selected, on average, 34% fewer features than mRMR and
40% fewer features than ReliefF. The subsets obtained by the proposed method
were subsequently used in different classifiers, achieving, for the vast majority of
faults, similar results with an average F1-score above 85% and an average FDR
above 80%. The strong results of the final experiment reaffirm the quality of the
obtained feature subset and highlight the effective performance of using SHAP
as a feature importance ranking algorithm and the developed feature selection
methodology.
