Dynamic Model Interpretation-Guided Online Active Learning Scheme for Real-Time Safety Assessment.

Area of Science:

• Industrial Safety
• Machine Learning
• Dynamic Systems

Background:

• Real-time safety assessment is vital for industrial processes, especially in non-stationary environments.
• Concept drift is a common challenge in dynamic systems, necessitating robust detection and adaptation methods.
• Existing safety assessment methods have limitations in handling complex concept drifts and incremental learning.

Purpose of the Study:

• To propose a novel dynamic model interpretation-guided online active learning scheme (DMI-LS).
• To address the limitations of existing methods in real-time safety assessment of dynamic systems with concept drift.
• To improve hazard prevention and reduce risks in industrial settings.

Main Methods:

• Implementation of a broad learning system for chunk data model updates.
• Development of a novel query strategy leveraging explainable artificial intelligence for ranking preference differences.
• Utilizing dynamic model interpretation to guide the online active learning process.

Main Results:

• The proposed DMI-LS demonstrates superior performance compared to advanced existing approaches.
• Experiments conducted on JiaoLong deep-sea manned submersible data validate the effectiveness of DMI-LS.
• The scheme shows significant improvements in handling concept drift and incremental learning scenarios.

Conclusions:

• DMI-LS offers an effective solution for real-time safety assessment in dynamic industrial environments.
• The integration of dynamic model interpretation and explainable AI enhances concept drift adaptation.
• The proposed method provides a robust framework for improving industrial safety and risk management.