Soft Sensor Data Conditioning for Accurate Conformance Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data-driven solutions for monitoring and controlling complex industrial processes face challenges due to linear relationships between process variables and quality variables, which become inaccurate as processes become more complex, and are affected by anomalies in data, leading to delayed and less accurate predictions.
Innovation Solution
A system that conditions data by identifying and replacing missing and outlier variables, filtering redundant ones, and using machine learning models to capture both linear and non-linear relationships between process variables and quality variables, enabling the configuration of soft sensors for real-time conformance metric prediction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If linear relationships between process variables and quality variables are used for monitoring and control, then the solution is simple to implement, but the accuracy deteriorates as processes become more complex
Solution Approach 1:
The patent transforms the relationship model from linear to non-linear by applying mathematical transformations (logarithmic, exponential, polynomial) to the process variables. This allows the system to capture complex non-linear relationships while maintaining the structured approach of traditional linear models, thereby improving prediction accuracy without completely abandoning the simplicity of parametric modeling.
Solution Approach 2:
The patent introduces dynamic adaptation by continuously updating the non-linear model parameters based on incoming process data. The system adapts to changing process conditions by adjusting the non-linear relationships in real-time, allowing it to maintain high accuracy even as processes become more complex or operate in different regimes.
2Loss of information
If all available process variables are used for modeling, then more information is available for prediction, but redundant and correlated variables reduce model efficiency and accuracy
Solution Approach 1:
The patent extracts only the most relevant and non-correlated process variables from the full set of available variables. By identifying and removing redundant variables through correlation analysis, the system retains the essential information needed for accurate prediction while eliminating variables that would otherwise degrade model performance and increase computational burden.
Solution Approach 2:
The patent applies a selective approach where not all available variables are used, but rather a carefully chosen subset that provides the optimal balance between information content and model efficiency. This partial action principle allows the system to achieve high prediction accuracy with fewer variables, improving both modeling efficiency and real-time performance.
3Ease of manufacture
If data with anomalies (missing values and outliers) is used directly for modeling, then the process is simpler, but the prediction accuracy deteriorates
Solution Approach 1:
The patent performs preliminary data conditioning before modeling by detecting and correcting anomalies such as missing values and outliers. This pre-processing step ensures that the data fed into the non-linear model is clean and reliable, preventing anomalies from degrading prediction accuracy while maintaining a relatively simple overall process through automated detection and correction algorithms.
Data Source
AI summary
Approaches for conditioning one or more process variables for configuring soft sensors, are described. According to one example, a processor may receive unconditioned data comprising data points or process variables indicating one or more characteristics associated with a process. The unconditioned data may be supplemented with an auxiliary set of process variables on detecting one or more missing process variables within the unconditioned data. A modified unconditioned data may thus be obtained. Further, a conditioned set of process variables, empirically representing the one or more characteristics associated with the process, may be identified from within the modified unconditioned data. The conditioned set of process variables may be provided to a plurality of inferential modellers to configure one or more soft sensors. A soft sensor, from among the one or more soft sensors, may then be selected for predicting runtime conformance metric associated with an outcome of the process.