An Interpretability Method for Broken Wire Detection

Area of Science:

• Industrial Safety
• Artificial Intelligence
• Non-Destructive Testing

Background:

• Wire rope integrity is crucial for industrial safety and equipment operation.
• Automated broken wire detection using deep learning, specifically YOLOv8, shows promise.
• The 'black box' nature of deep learning models presents a trust challenge in critical applications.

Purpose of the Study:

• To address the trust and interpretability challenges of deep learning models in wire rope broken wire detection.
• To propose and evaluate a novel perturbation-based interpretability method, ESTC.
• To validate the reliability of YOLOv8 for detecting wire breaks by comparing its decision-making process with expert knowledge.

Main Methods:

• Development of ESTC (Eliminating Splicing and Truncation Compensation), a perturbation-based interpretability technique.
• Comparison of ESTC with existing model-agnostic interpretability methods (LIME, RISE, D-RISE).
• Application of these methods to a YOLOv8 object detection model trained on electromagnetic signal images of wire rope.

Main Results:

• ESTC demonstrated objective superiority over LIME, RISE, and D-RISE in interpretability analysis.
• The interpretability analysis confirmed that the YOLOv8 model's predictions align with prior knowledge from manual rope inspection.
• The proposed ESTC method enhances the credibility of using object detection for broken wire detection.

Conclusions:

• The ESTC method provides a reliable way to interpret deep learning models used for wire rope defect detection.
• This interpretability boosts confidence in the practical application of AI for ensuring industrial safety.
• The study highlights the importance of interpretability in critical infrastructure monitoring.