Feature Selection for Fault Detection in Industrial Processes Based on the SHAP Algorithm

We apply SHapley Additive exPlanations (SHAP) feature selection method to improve fault detection models in an industrial production context. An illustrative synthetic process and the benchmark Tennessee Eastman Chemical Plant Simulator (TES) were used for the experiments. The principal hypothesis of this work is that the proposed SHAP feature selection methodology produces a subset of features with better discriminatory information, or at least approximately the same, than the original feature set. For the synthetic data experiment, the subset obtained achieved approximately the same results while reducing the number of features. For the experiment using the Tennessee Eastman simulation, the methodology produced a reduced feature subset that was able to obtain, overall, better values of the
$F_{1}$
-score and the Fault Detection Rate (FDR, aka. recall) compared to the original feature set. The set obtained using the proposed SHAP feature selection resulted in an average
$F_{1}$
-score of 82.36% and an average FDR of 78.03% for all the faults. These results represent an increase of 2.77% and 1.84% of the
$F_{1}$
-score and FDR, respectively, when compared to the original feature set while using an average of six features, a reduction of the cardinality by approximately 97.12%.