Dynamic rough set learning for reliable early warning in industrial time-series systems

Area of Science:

• Industrial Engineering
• Machine Learning
• Data Science

Background:

• Industrial early-warning systems need models to detect risks before failure and explain decision uncertainty.
• Current machine learning (ML) and deep learning (DL) models offer high predictive accuracy but limited insight into classification ambiguity.

Purpose of the Study:

• To propose a dynamic rough set learning framework for uncertainty-aware early warning in industrial time-series systems.
• To enhance the interpretability and uncertainty quantification of early-warning models.

Main Methods:

• Converting multivariate sensor data into sliding-window decision systems.
• Constructing time-dependent neighborhood rough approximations for evolving decision classes.
• Introducing a rough early-warning index and dynamic feature reduction for interpretability and efficiency.

Main Results:

• The dynamic rough set model achieved a test accuracy of 0.8647 and Macro-F1 score of 0.7895.
• While Random Forest and XGBoost had higher classification scores, the proposed method provided valuable uncertainty outputs.
• Dynamic rough feature reduction maintained high accuracy while significantly reducing the feature set from 118 to 40.

Conclusions:

• The proposed framework excels in early-warning scenarios prioritizing interpretability and uncertainty qualification alongside accuracy.
• The model's ability to identify ambiguous and transitional states is a key advantage over black-box classifiers.
• The framework demonstrates robustness to moderate noise, making it suitable for real-world industrial applications.