Uncertainty-Aware Cross-Dataset Reliability Screening of Machine Learning Models for Drift-Affected Gas Sensor Arrays Under Missing and Corrupted Channels

An uncertainty-aware reliability screening pipeline was developed for machine learning models applied to drift-affected gas sensor arrays under missing and corrupted sensor-channel conditions. The pipeline consists of dataset-specific preprocessing, model training, simulated sensor degradation, confidence-based abstention, coverage-risk evaluation, AUROC analysis, and SHAP-based sensor-feature interpretation. Two public datasets were used: the UCI Gas Sensor Array Drift Dataset and the Long-Term E-Nose Dataset. Logistic Regression, Random Forest, and Multilayer Perceptron were evaluated under five degradation scenarios: missing values, Gaussian noise, sensor-channel dropout, stuck-at-faults, and combined degradation. Results showed that clean classification performance did not guarantee robustness. On the UCI Drift Dataset, Logistic Regression achieved the highest clean balanced accuracy of 73.36%, but decreased to 25.57% under combined degradation. At the dataset level, mean balanced accuracy decreased from 68.96% to 33.05% for the UCI Drift Dataset and from 57.98% to 52.81% for the Long-Term E-Nose Dataset. At a confidence level threshold of 0.90, Logistic Regression on the UCI Drift Dataset still produced a high accepted-prediction risk of 72.39% under combined degradation, indicating overconfidence despite corrupted inputs. SHAP analysis showed that the dominant UCI sensor group shifted from Sensor_2 to Sensor_10, while Sensor_48 remained dominant in the Long-Term E-Nose Dataset. Compared with conventional gas sensor drift studies that mainly emphasized classification accuracy or drift compensation, this work contributes a cross-dataset reliability-oriented evaluation protocol that combines degradation testing, abstention-based screening, and explainable sensor-feature analysis for more practical assessment of electronic nose models.