Rolling Mill Target Value Prediction Using RNN and Physical Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for determining production parameters in rolling mills fail to adequately consider the current state of the plant or its immediate production history, leading to inaccurate predictions due to changing environmental conditions and wear and tear, resulting in suboptimal product quality.

Innovation Solution

A method and device using a combination of machine-learned models, specifically a recurrent neural network (RNN) and a physical model, to determine target values for production parameters by incorporating material parameters and current plant conditions, allowing for near-optimal corrections based on long-term and short-term memory.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If lookup tables or physical models are used to determine production parameters, then target values can be determined in advance, but the current state of the plant and immediate production history are not adequately considered, leading to inaccuracies when plant conditions change

Engineering Contradiction:
Improveaccuracy of target value determinationVSAvoidadaptability to changing plant conditions
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The patent implements feedback by using a recurrent neural network that incorporates production history and current plant state information. The system continuously learns from past production data and adjusts target value predictions based on observed deviations, creating a closed-loop adaptive system that resolves the contradiction between using predetermined models and adapting to changing conditions.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent transforms static lookup tables and physical models into dynamic predictions by employing a recurrent neural network that evolves its predictions based on current plant state and production history. The system dynamically adjusts target values based on real-time conditions such as wear and tear, environmental changes, and immediate production outcomes, making the determination process adaptable rather than fixed.

Inventive Principle:
Principle #15Dynamics

2Productivity

If machine-learned models are used to predict production parameters, then average target values can be output, but individual corrections for current plant state and production history cannot be made

Engineering Contradiction:
Improveefficiency of target value determinationVSAvoidprecision of production parameter settings
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent introduces an intermediary correction mechanism between the base machine-learned model predictions and the final target values. The recurrent neural network acts as a mediator that takes average predictions from the base model and applies individual corrections based on current plant state and production history, thereby maintaining both efficiency and precision.

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent segments the target value determination into two distinct components: base target values from machine-learned models and individual corrections from a recurrent neural network. This segmentation allows the system to maintain the efficiency of automated modeling while adding precision through historical and contextual corrections, resolving the contradiction between speed and accuracy.

Inventive Principle:
Principle #1Segmentation

3Adaptability or versatility

If feedforward networks are used for each newly produced strip independently, then predictions are based on long-term inheritance, but short-term production history and current plant state are ignored

Engineering Contradiction:
Improveindependence of strip predictionsVSAvoidloss of production history information
Core Design Contradiction:
Adaptability or versatilityVSLoss of information

Solution Approach 1:

The patent ensures continuity of useful action by using a recurrent neural network that continuously processes production history and plant state information across multiple strips. Rather than treating each strip independently, the system maintains continuous learning and adaptation, preserving valuable production history information while still allowing for independent strip-specific predictions when appropriate.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP4645004A1Determination of a target value during the production of metal strips
Publication Date: 2025.11.05 PRIMETALS TECH GERMANY GMBH
  • EP4645004A1 patent drawingFigure 1~2
  • EP4645004A1 patent drawingFigure 3
  • EP4645004A1 patent drawing

AI summary

The present invention relates to a method (100) and a device (10) for determining a target value (S) for a selected production parameter (P) of a rolling mill (200) with at least one rolling stand group (230) for rolling a metal product (12) into a metal strip (14), as well as a rolling mill (200) with such a device (10), a computer program product, and a machine-learned model (40). A value (W) of a material parameter (M) for a predetermined reference material is provided (S1). A correction factor (c2) for the material parameter (M) is determined by a machine-learned model (40) based on a recurrent neural network (42) using previously known production parameters (P*) in the metal strip production process (S3).The target value (S) for the production parameter (P) of the rolling mill (200) can then be determined on the basis of a physical production parameter model (52) which incorporates the material parameter value (W) adjusted with the correction factor (c2) and target specifications (Z) for the rolling of the metal strip (14) (S4).