ABIGX: A Unified Framework for Explainable Fault Detection and Classification

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Fault detection and classification (FDC) models require robust explainability.
• Existing methods like contribution plots (CP) and reconstruction-based contribution (RBC) have limitations in general FDC applications.
• Gradient-based explanation methods struggle with fault class smearing, hindering accurate fault classification.

Purpose of the Study:

• To propose ABIGX (Adversarial fault reconstruction-Based Integrated Gradient eXplanation), a unified framework for explainable FDC.
• To extend the applicability of established fault diagnosis principles to general FDC models.
• To improve the precision and comprehensiveness of fault explanations.

Main Methods:

• Developed Adversarial Fault Reconstruction (AFR) by reframing fault reconstruction through adversarial attack perspectives.
• Introduced a novel fault index for both fault detection and classification tasks.
• Theoretically bridged ABIGX with CP and RBC, proving them as linear specifications of ABIGX in fault detection.

Main Results:

• Demonstrated that ABIGX effectively mitigates fault class smearing in classification tasks.
• Showcased ABIGX outperforming current gradient-based explanation methods.
• Validated the generality and accuracy of AFR through quantitative metrics and intuitive illustrations.

Conclusions:

• ABIGX provides more comprehensive and precise explanations for FDC models compared to existing methods.
• The proposed AFR method is general and accurate, enhancing explainability in FDC.
• ABIGX represents a significant advancement in explainable AI for fault detection and classification.