Metallurgical Process Control With Explainable Prediction Reliability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Data analytics models used in metallurgical plants are often 'black box' models, making it difficult to understand why predictions are made and how reliable they are, which limits their applicability in critical control systems. Additionally, integrating these models into running plants requires adaptations to ensure reliability, comprehensibility, and accountability.
Innovation Solution
A method that verifies online input data against offline training data ranges, calculates reliability and transparency information, and uses predefined rules to ensure predictions are reliable and valid, incorporating techniques like concept drift detection and explainable AI methods such as counterfactuals and anchors to provide transparent and comprehensible results.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If data analytics models are used for prediction, then prediction capability is improved, but transparency and reliability are worsened due to black box character
Solution Approach 1:
The patent introduces transparency information as an intermediary element that mediates between the black box model and the user. This transparency information includes explanations of model behavior, input data quality assessments, and reliability indicators that bridge the gap between the opaque prediction mechanism and the need for trustworthy decision-making.
Solution Approach 2:
The system implements feedback mechanisms where transparency information about model predictions, input data quality, and model behavior is continuously provided back to users. This feedback loop allows operators to understand model decisions, assess reliability, and make informed decisions about whether to trust and act on predictions.
2Reliability
If online input data is verified against offline training data ranges, then reliability is improved, but processing time and complexity are worsened
Solution Approach 1:
The patent performs verification of online input data against offline training data ranges as a preliminary action before the actual prediction process. By checking data quality and range compatibility in advance, the system prevents unreliable predictions from being generated, saving time that would otherwise be spent on processing and validating questionable predictions later.
Solution Approach 2:
The verification process is segmented into distinct steps: checking if input data falls within training data ranges, assessing input data quality, and evaluating transparency information. This segmentation allows the system to quickly identify and handle only the necessary verification steps rather than performing complete validation on every prediction.
3Loss of information
If transparency information is calculated and verified, then model comprehensibility is improved, but computational complexity is worsened
Solution Approach 1:
The patent extracts only the essential transparency information needed for model comprehensibility rather than calculating all possible model attributes. This selective extraction focuses on key aspects such as input data quality, model behavior explanations, and reliability indicators, reducing computational complexity while maintaining comprehensibility.
Solution Approach 2:
The system applies different levels of transparency information calculation based on local needs and contexts. Not all predictions require the same depth of transparency analysis, so the system adjusts the computational effort according to the specific prediction context, process criticality, and available resources.
4Productivity
If data driven models are integrated into running plants, then productivity is improved, but adaptability to new conditions is worsened due to training data limitations
Solution Approach 1:
The patent implements dynamic monitoring of input data ranges and model performance during operation. When drift in data characteristics is detected, the system can trigger model retraining or adaptation processes, allowing the model to evolve and adapt to new operating conditions while maintaining productivity benefits.
Solution Approach 2:
The system continuously monitors whether online input data falls within the ranges observed during offline training. This feedback mechanism detects when the plant operates under new conditions not represented in training data, triggering alerts for model retraining or adaptation to maintain both productivity and adaptability.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
The present invention relates to a method and a computer system for controlling a process of a metallurgical plant, wherein at least one process variable of a process of the metallurgical plant is predicted by means of a data driven model (1) for the prediction of the at least one process variable. The object of the present invention is to provide a method, which explains the result of the data model (1) in a comprehensible, reliable, and transparent manner. The problem is solved by a first step where input data are verified if the online input data are in a range that was present in the offline process data used for training of the data driven model (1). Furthermore, in a second step, at least one reliability information of the predicted process variable is calculated and compared with a predefined reliability information or transparency information of the at least one predicted process variable is determined.