Metal Sheet Property Prediction With Hybrid ML and Metallurgical Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional techniques for predicting material characteristic values in metal sheet manufacturing fail to account for disturbances such as air and water temperature, leading to variations in material properties and difficulties in adjusting subsequent processes to achieve desired values, especially when limited training data is available.
Innovation Solution
A material characteristic value prediction system that incorporates a machine learning model and a metallurgical model to predict material characteristics by considering line output factors, disturbance factors, and component values, allowing for accurate adjustment of production conditions in processes like hot-rolling, cold-rolling, annealing, and coating.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional statistical probability models are used to predict material characteristic values, then the prediction process is simple, but the prediction accuracy is insufficient because disturbance factors are not considered
Solution Approach 1:
The patent applies a composite modeling approach by integrating multiple types of models (machine learning model and metallurgical model) to create a hybrid prediction system. This composite structure combines the pattern recognition capabilities of machine learning with the physical chemistry principles of metallurgical models, enabling accurate prediction of material characteristics while considering both manufacturing parameters and disturbance factors.
Solution Approach 2:
The patent introduces disturbance factors as intermediary variables that mediate between manufacturing parameters and material characteristic values. These disturbance factors (air temperature, water temperature, etc.) serve as intermediate elements that capture environmental influences, allowing the model to accurately predict material properties by incorporating these mediating effects.
2Manufacturing precision
If highly precise target manufacturing condition values are set before manufacturing starts, then product quality can meet target values, but conditions of subsequent processes cannot be changed in the course of manufacturing
Solution Approach 1:
The patent implements a dynamic prediction system that continuously forecasts material characteristic values during the manufacturing process. This dynamic capability allows the system to adapt to changing conditions and provide real-time guidance for adjusting subsequent process parameters, transforming static pre-set targets into dynamic adaptive control.
Solution Approach 2:
The patent performs preliminary prediction of material characteristic values before subsequent processing steps. By predicting outcomes in advance, the system enables proactive adjustment of future process conditions rather than reactive corrections, allowing optimal settings to be determined ahead of time while maintaining adaptability.
3Measurement precision
If machine learning algorithms are applied to construct prediction models, then high-level prediction of defect occurrence probability is achieved, but a large number of training data are required which makes it difficult to apply when there is little or no past manufacturing track records
Solution Approach 1:
The patent merges machine learning models with metallurgical models to create a hybrid system. This combination allows the integration of physical chemistry principles that can provide prediction capabilities with limited data, while machine learning components enhance pattern recognition. The synergistic effect reduces dependency on large training datasets while maintaining high prediction accuracy.
Solution Approach 2:
The patent employs a pragmatic approach by using readily available manufacturing data even if the quantity is limited. Rather than requiring extensive historical track records, the system effectively utilizes available data in conjunction with metallurgical principles, making the prediction system applicable from the early stages of manufacturing without requiring long-term data accumulation.
4Ease of operation
If conventional techniques consider only directly adjustable manufacturing conditions, then the control process is straightforward, but disturbances such as air temperature or water temperature significantly affect material characteristic values
Solution Approach 1:
The patent introduces disturbance factors as intermediary variables that mediate between manufacturing parameters and material characteristic values. These disturbance factors (air temperature, water temperature, etc.) serve as intermediate elements that capture environmental influences, allowing the model to accurately predict material properties by incorporating these mediating effects without complicating the control process.
Solution Approach 2:
The patent creates a universal prediction model that handles both directly controllable manufacturing parameters and uncontrollable disturbance factors within a single integrated framework. This multi-functional model simultaneously processes various input types (manufacturing conditions, environmental factors) and provides comprehensive predictions, eliminating the need for separate control mechanisms for different factor types.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A material characteristic value prediction system that can predict material characteristic values with high accuracy is provided. Also provided is a method of manufacturing a metal sheet that can improve the product yield rate, by changing manufacturing conditions of subsequent processes. The material characteristic value prediction system (100) includes a material characteristic value predictor configured to acquire input data including line output factors in a metal sheet manufacturing line, disturbance factors, and component values of a metal sheet being manufactured, and predict material characteristic values of the manufactured metal sheet using a prediction model configured to take the input data as inputs, wherein the prediction model includes a machine learning model generated by machine learning and configured to take the input data as inputs and output production condition factors, and a metallurgical model configured to take the production condition factors as inputs and output the material characteristic values.