SHAP-Based Framework for Temporal Detection of Sensor Drift in Gas Sensor Arrays

Abstrak

Sensor drift is a fundamental problem that undermines the long-term reliability of machine-learning models, as gradual changes in sensor response can silently degrade predictive performance. This challenge is particularly relevant in gas sensor arrays, where environmental variability and sensor aging strongly affect data quality. Ensuring effective monitoring of such drift is essential for industrial safety, environmental monitoring, and other applications that depend on stable sensing systems. To address this issue, we develop a SHAP (SHapley Additive Explanations)–based framework combined with an XGBoost classifier to detect drift at the feature level. The model is trained on structured sensor data, and SHAP values are computed batch-wise to capture temporal changes in feature relevance. Preprocessing enforces balanced class distributions and consistent concentration ranges, reducing potential confounding effects. Results from the Gas Sensor Array Drift dataset demonstrate that several sensor channels exhibit clear temporal variability in SHAP importance. Notably, these changes were observed even when accuracy and F1-score remained stable, highlighting that SHAP provides a complementary early-warning signal of drift. Case studies further confirm that the framework can capture both gradual and abrupt drift patterns. The research contribution of this study lies in providing an interpretable and time-resolved perspective on drift detection, enabling the localization of drift signals to specific sensors and features, independent of drift type or feature domain. This approach advances the state of the art by bridging explainable AI and drift monitoring. In conclusion, the framework supports practical deployment of adaptive sensing systems by making drift behaviour visible before conventional metrics degrade. Future work will focus on real-time scalability, broader datasets, and automated drift correction through recalibration or retraining.

Kata kunci

• SHapley Additive Explanations (SHAP)
• Sensor Drift
• Concept Drift
• Model Monitoring
• Time-Series Interpretability
• Feature Importance Analysis