Tell Me, What Do You See?-Interpretable Classification of Wiring Harness Branches with Deep Neural Networks

Area of Science:

• Robotics and Automation
• Computer Vision
• Machine Learning

Background:

• Industrial automation requires precise manipulation of deformable linear objects, such as wiring harnesses.
• Current machine vision systems lack interpretability, hindering trust and adoption in complex assembly tasks.

Purpose of the Study:

• To develop interpretable neural network architectures for classifying wiring harness branches.
• To provide insights into the decision-making process of machine learning models in industrial settings.

Main Methods:

• Proposed and tested several novel neural network architectures on a custom dataset.
• Conducted experiments evaluating modality, data fusion, data augmentation, and pretraining strategies.
• Integrated saliency maps to visualize and explain model predictions.

Main Results:

• Achieved high performance in classifying wiring harness branches.
• Saliency maps successfully revealed the features influencing the network's decisions.
• Demonstrated the effectiveness of proposed methods in enhancing model interpretability.

Conclusions:

• The developed interpretable neural networks are suitable for robotizing industrial operations involving wiring harnesses.
• Saliency maps offer a valuable tool for understanding and trusting AI predictions in manufacturing.
• This work advances the field of explainable AI in industrial machine vision.