Hybrid Rolling Property Prediction for Sparse Sampling Conditions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for determining mechanical properties of rolling stock in rolling mills are limited by the inability to accurately account for all relevant process parameters, particularly in cooling sections, and have low extrapolation capabilities due to localized sampling and sparse data sets, leading to unreliable predictions outside the trained data range.
Innovation Solution
A hybrid model combining a physical production model and a statistical data model, using metallurgical model parameters and production data sets to simulate the production process and predict mechanical properties, while also incorporating sampling data to validate and refine the predictions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If statistical data models are trained using localized sampling data, then the model can predict mechanical properties without physical sampling, but the extrapolation capability is limited and predictions become unreliable outside the trained data range
Solution Approach 1:
The patent combines a physical production model that simulates the complete production process with a statistical data model that learns from sampling data. The physical model generates comprehensive process data including temperature profiles and cooling rates throughout the entire production line, while the statistical model captures correlations from actual measurements. This hybrid approach enables reliable predictions both within and outside the training data range by leveraging the physical model's ability to extrapolate based on process understanding.
Solution Approach 2:
The physical production model acts as an intermediary that generates synthetic process data and enhances the statistical model's input features. By simulating the complete production process physics, it provides the statistical model with comprehensive process parameters beyond what localized sampling can capture, thereby improving extrapolation capability while maintaining the efficiency of prediction-based methods.
2Measurement precision
If physical sampling is performed to measure mechanical properties, then accurate measurements are obtained, but the process becomes complex and time-consuming
Solution Approach 1:
The patent creates a virtual copy of the production process through the physical production model, which simulates temperature profiles, cooling rates, and metallurgical transformations. This digital twin approach allows mechanical properties to be predicted from simulated process data combined with statistical learning, eliminating the need for complex physical sampling while maintaining measurement accuracy through the hybrid model's comprehensive process understanding.
3Measurement precision
If comprehensive process parameters including cooling section data are included in the model, then prediction accuracy improves, but the data processing complexity increases
Solution Approach 1:
The patent divides the production process into distinct segments (heating, roughing, finishing, cooling) with the physical model simulating each section's specific physics. The statistical model then learns from these segmented process stages, allowing comprehensive process parameters to be processed systematically. This segmentation manages data complexity by organizing comprehensive process information into manageable stages while maintaining high prediction accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a method for determining the mechanical properties (Jf) of a first rolled product (2) using a hybrid model (10) comprising production data sets (Pi) of further rolled products (2'), a physical production model (12), and a statistical data model (14). From the production data set (Pf) of the first rolled product (2), a first mechanical, a further production, and a metallurgical data set (Cf, Ff, Kf) as well as a second mechanical data set (Sf) are determined. Furthermore, an averaged normalized distance value (df,m) to the production data sets (Pi) of the further rolled products (2') is determined, with which the mechanical properties of the rolled product (2) are determined as a weighted mean of the first and second mechanical data sets (Cf, Sf).In creating the hybrid model (10), further production data sets (Fi) of the other rolled goods (2') are determined using the physical production model (12) to train the statistical data model (14).