PREDICTIVE MAINTENANCE,
METAL PRODUCTION
BUSINESS QUESTION
A world leader in the metal industry wants to predict drifts in production lines. In particular, specific situations are hardly detected on time with standard SPC monitoring and can lead to plummeting production yield.
DATA
- 3 production lines
- 3 years of data, incl. voltage and intensity per second…
- Loads of additional metrics (T°, R, mass, chemical composition, …) although at a lower frequency
→ Billions of data points ~40’s GB
APPROACH
- Significant preprocessing work, due to poor data quality (sensor and database issues)
- Spectral transformation of data through a FFT applied to each anode life cycle (circa 25 days)
- HMM & ensemble methods where tested
- Several workshops organized, with R&D experts and business leaders, to share intermediate results and work on the business case
RESULTS
- Machine learning and features engineering proved efficient to anticipate drifts with 80% precision, 10 days before what standard SPC technics could achieved
- A strong business case, worth 10’s millions CF per year
- Big Data value was again validated on an industrial performance use-case.