Benchmark for welding gun fault prediction with multivariate time series data

Area of Science:

• Industrial Engineering
• Manufacturing Technology
• Machine Learning

Background:

• Machinery failures in resistance spot welding (RSW) guns disrupt automotive production lines, causing significant downtime and reliability issues.
• Accurate fault prediction for RSW guns is essential for developing effective predictive maintenance strategies.
• The complex behavior and data variability of RSW guns present challenges for traditional fault prediction methods.

Purpose of the Study:

• To establish a benchmark dataset for RSW gun fault prediction.
• To propose machine learning (ML) benchmarks for developing advanced fault prediction approaches.
• To provide insights into time series forecasting for industrial machinery health monitoring.

Main Methods:

• Collected a comprehensive dataset from hundreds of RSW guns at BMW Brilliance Automotive Ltd.'s Body-Shop.
• Utilized historical data to capture patterns preceding welding errors.
• Applied state-of-the-art machine learning time series forecasting methods for benchmark analysis.

Main Results:

• Developed a novel benchmark dataset specifically for welding gun fault prediction.
• Demonstrated the effectiveness of ML time series forecasting for identifying pre-failure patterns.
• Established baseline performance metrics for future RSW gun fault prediction research.

Conclusions:

• The created benchmark dataset and proposed ML approaches facilitate the development of robust RSW gun fault prediction systems.
• This work enhances predictive maintenance capabilities in automotive manufacturing, minimizing unplanned downtime.
• It encourages ML researchers to contribute to solving critical industrial fault prediction challenges.