Identifying cause-and-effect relationships of manufacturing errors using sequence-to-sequence learning

Area of Science:

• Manufacturing Process Analysis
• Artificial Intelligence in Industry
• Automotive Production Systems

Background:

• Automated car-body production relies on sequential station operations with strict cycle times.
• Errors in one station can cascade, causing downstream delays and impacting order completion.
• Current methods lack automated capabilities to differentiate source errors from their knock-on effects.

Purpose of the Study:

• To develop a novel system for automatically distinguishing between source and knock-on errors in car manufacturing.
• To establish a causal relationship between identified errors using real-time production data.
• To benchmark deep learning models for this specific industrial application.

Main Methods:

• Utilized a production data acquisition system for real-time manufacturing condition monitoring.
• Developed a vehicle manufacturing analysis system employing deep learning techniques.
• Benchmarked three sequence-to-sequence models (Transformer, LSTM, GRU) and introduced a novel composite time-weighted action metric.

Main Results:

• The analysis revealed that 71.68% of production sequences contained either a source or a knock-on error.
• The Transformer model outperformed LSTM and GRU in sequence-to-sequence tasks within this domain.
• Increased prediction range for future action durations further enhanced the Transformer's performance.

Conclusions:

• The proposed deep learning system effectively links source and knock-on errors in car production.
• The Transformer architecture is well-suited for analyzing complex error propagation in manufacturing.
• Findings highlight significant error prevalence and the potential for AI-driven process optimization.