A Data-Driven Methodology to Assess Raw Materials Impact on Manufacturing Systems Breakdowns

Abstract

Data-driven Prognostics and Health Management (PHM) become a crucial layer in the realm of predictive maintenance (PM). However, many industries develop PM technologies based on the monitoring of machine data to anticipate failures without considering the injected raw material. In reality, non-compliant material characteristics can affect the manufacturing tools leading to machine breakdowns and poor quality product. To cope with this situation, this paper proposes a new methodology that helps operators predicting machine breakdowns. In detail, the methodology starts by implementing an Extract, Transform, Load (ETL) process which aims to create a new and reliable dataset from heterogeneous sources. Then, a feature selection method is used for dimensionality reduction and keep only useful information. After that, the selected features are injected to Machine Learning (ML) algorithms to predict system breakdown occurrences. Finally, the novelty in this study, an auto-labeling algorithm based on material data and machine breakdown predictions is proposed. This algorithm aims to enhance raw material stock management, scheduling their consumption accordingly and thus reducing machine breakdowns. The developed methodology is applied to a real dataset of a French company, SCODER, that shows and pointed out promising perspectives in PM.

Keywords

• prognostics and health management
• manufacturing
• raw material data
• extract-transform-load
• features selection
• machine learning
• auto-labeling