RI2AP: Robust and Interpretable 2D Anomaly Prediction in Assembly Pipelines.

Area of Science:

• Manufacturing
• Industrial Engineering
• Machine Learning

Background:

• Anomaly prediction is vital for manufacturing efficiency and cost reduction.
• Current machine learning methods struggle with rare events and complex dependencies, yielding suboptimal F1 scores.
• High-fidelity simulation data is scarce and expensive, limiting traditional ML model training.

Purpose of the Study:

• To introduce a novel method, Robust and Interpretable 2D Anomaly Prediction (RI2AP), for enhanced anomaly prediction.
• To address the dual challenges of predicting anomaly occurrence and understanding inter-dependencies.
• To provide interpretable insights into sensor data for domain experts.

Main Methods:

• Developed Robust and Interpretable 2D Anomaly Prediction (RI2AP).
• Utilized a causal-influence framework for model interpretability.
• Validated on rocket assembly simulations and real-world manufacturing data.

Main Results:

• RI2AP demonstrated up to a 30-point F1 measure improvement over existing ML methods.
• The method effectively predicts anomaly occurrences and their dependencies.
• The interpretation mechanism provided actionable insights to domain experts.

Conclusions:

• RI2AP significantly advances anomaly prediction in manufacturing assembly lines.
• The model's interpretability enhances trust and utility for industrial applications.
• RI2AP shows strong potential for real-world deployment in complex manufacturing settings.