Industrial Time-Series Forecasting

Overview

SMS Group needed to evaluate optimization opportunities in cold rolling mill factories using high-frequency raw sensor signals — but 10 minutes of signals could reach 10GB, making storage and analysis impractical at scale. Took the project from proof of concept to a production MVP over two years. The core of the system is an ETL pipeline: specialized Python workflows automatically detect steady production phases in raw sensor streams, transform 10GB/10-minute windows into compact high-value summaries, and load them into SQL for downstream modeling — dramatically cutting storage requirements while preserving critical signal information. Worked closely with domain experts through iterative sessions to align processing with industry standards. Formulated the optimization problem mathematically, designing a custom loss function and testing a wide range of ML approaches. The resulting forecasting and anomaly detection models outperformed standard methods by 30%. The MVP integrates end-to-end with factory systems via gRPC, stores results in SQL databases, and provides real-time dashboards for engineers and managers. It automatically cleans data, segments production phases, and retrains models continuously — supporting predictive maintenance and process optimization.

Results

• Took the project from proof of concept to production over two years
• Custom forecasting and anomaly detection models outperformed standard methods by 30%
• Increased factory efficiency by 2%
• Reduced storage requirements while preserving signal information from 10GB/10-minute windows
• Currently in active use at SMS Group's cold rolling mills
• End-to-end gRPC integration with factory systems and SQL-backed real-time dashboards

Figures