Kobe Steel Uses MATLAB Web App Server and MATLAB Production Server to Improve Mill Efficiency
An Approach Enabling Operations and Maintenance Engineers to Rapidly Deploy Highly Customized Models
“[We] have developed a MATLAB application that allows users to build condition-based models while reviewing the data. This enables operational and maintenance engineers—or domain experts—to create customized models for each piece of equipment in an agile manner, even without knowledge of data science.”
Key Outcomes
- MATLAB enabled the end-to-end development of a user-friendly application with agile development, allowing operations and maintenance engineers to create, prototype, and deploy anomaly models in their operational workflow
- The trial run model demonstrated an accuracy of 89% in detecting anomalies
- Utilizing the MONAD system, on-site personnel perform equipment inspections and preventive maintenance based on alert information
Overview of an anomaly detection system with preprocessing, model deployment, and anomaly reporting.
Kobe Steel is one of Japan’s leading diversified manufacturers, providing a wide range of technologies, products, and services centered on its three core businesses: materials, machinery, and electric power. The company is currently working on increasing mill efficiency and ensuring the timely delivery of steel products by identifying and addressing anomalies in its manufacturing process.
Since the factory contains many pieces of equipment, each with a wide variety of operating methods and patterns, the team identified a key challenge: It is difficult for the data science team at the research institute to develop a generic anomaly detection model that can be applied across all locations. The solution was to delegate model development to the operations and maintenance engineers who knew the intricacies and quirks of the equipment in their mill. The team developed a MATLAB® web app called MONAD (Mode Oriented Novel Anomaly Detector) to allow operations and maintenance engineers to build easily interpretable two-variable Local Outlier Factor (LOF) models. Early trials were very promising, demonstrating the ability to detect initial bearing lubrication failures with an accuracy of 89%. Democratizing model building with the operations and maintenance engineers also allowed Kobe Steel to scale up quickly—in the first month alone, 90 anomaly detection models were prototyped.
Process data from production equipment is collected and preprocessed on AWS® before it is evaluated by the anomaly detection model deployed in MATLAB Production Server™. The results are published in a monitoring web dashboard powered by MATLAB Web App Server™. If an anomaly is detected, an email is sent to the equipment manager, who inspects the equipment and performs repairs as necessary before a malfunction shuts down the line. The on-site model designer is also notified by email and may modify the model in response to the latest data or conditions to improve performance.
The MONAD system developed in MATLAB allows Kobe Steel operations and maintenance engineers to rapidly deploy highly customized models, enabling the early detection of potential anomalies and allowing for a prompt response.
