Online Inferential Modeling for Dynamic Plant KPI Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Inferential models used in process industries face limitations such as reliance on steady-state data, inability to generate dynamic predictions, high maintenance costs, and difficulty in automation, which hinders their effectiveness in equipment performance management and asset optimization.
Innovation Solution
The development and deployment of high-fidelity dynamic inferential models using computer systems that automate data screening, input selection, and model training, combined with online health monitoring and adaptation techniques, enabling accurate and predictive future product quality and KPI estimations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If first-principles equations are used to build inferential models, then reliability of simulation and prediction is improved, but development cost and maintenance difficulty increase
Solution Approach 1:
The patent introduces an automated model calibration system that acts as an intermediary between process data and inferential models. This system automatically retrieves process data, calibrates model parameters, and updates models without requiring manual intervention from process engineers, thereby reducing the complexity burden of first-principles models while maintaining their predictive reliability
Solution Approach 2:
The inferential models are equipped with automated self-calibration capabilities that allow them to automatically adjust their parameters based on incoming process data. The system performs self-diagnosis and self-updating, reducing the need for external expert intervention and making first-principles models more maintainable
2Reliability
If online model re-calibration is performed frequently to sustain performance under varying operating conditions, then model accuracy is improved, but cost and operational challenge increase
Solution Approach 1:
The patent implements continuous automated model calibration that operates in the background without interrupting process operations. The system continuously retrieves process data, calibrates model parameters, and updates models in an ongoing manner, eliminating the need for periodic shutdowns or manual re-calibration events, thereby maintaining accuracy without time loss
Solution Approach 2:
The system establishes a feedback loop where model performance is continuously monitored against actual process data, and calibration actions are automatically triggered when performance degradation is detected. This feedback mechanism ensures model accuracy is maintained only when necessary, reducing unnecessary calibration operations and associated costs
3Reliability
If intensive user inputs and expertise are required for model development and deployment, then model quality is improved, but ease of operation deteriorates
Solution Approach 1:
The patent implements automated workflows where the system itself performs data retrieval, model configuration, calibration, and deployment tasks that traditionally required expert intervention. The automated system selects appropriate models, configures parameters, and manages the entire development lifecycle, maintaining model quality while eliminating the need for intensive user inputs
Solution Approach 2:
The patent creates a universal automated platform that handles multiple aspects of inferential model development (data retrieval, model selection, calibration, validation, deployment) within a single integrated system. This multi-functional platform reduces the need for specialized expertise across multiple disciplines by consolidating tasks into an automated workflow that can be operated by general personnel
4Measurement precision
If raw measured plant operational data is used for model calibration, then model accuracy is improved, but data quality issues (spikes, off-sensors, shutdowns) worsen the calibration process
Solution Approach 1:
The patent implements preliminary data screening and quality assessment steps that automatically identify and flag problematic data segments (spikes, off-sensor readings, shutdown periods) before they are used for model calibration. The system prepares cleaned and validated datasets in advance, ensuring that only high-quality data reaches the calibration process, thereby maintaining accuracy while eliminating the harmful effects of poor data quality
Solution Approach 2:
The patent introduces an automated data quality assessment and preprocessing system that acts as an intermediary between raw plant data and the model calibration process. This intermediary layer filters, validates, and cleanses data automatically, removing spikes, handling missing sensor readings, and excluding shutdown periods, thereby protecting the calibration process from data quality issues while maintaining the use of real operational data
Data Source
Figure 1
Figure 2A
Figure 2B
AI summary
Embodiments are directed to systems that build and deploy inferential models for generating predictions of a plant process. The systems select input variables and an output variable for the plant process. The systems load continuous measurements for the selected input variables. For the selected output variable, the systems load measurements of type: continuous from the subject plant process, intermittent from an online analyzer, or intermittent from lab data. If continuous or analyzer measurements are loaded, the systems build a FIR model with a subspace ID technique using continuous output measurements. From intermittent analyzer measurements, the systems generate continuous output measurements using interpolation. If lab data is loaded, the systems build a hybrid FIR model with subspace ID and PLS techniques, using continuous measurements of a reference variable correlated to the selected output variable. The systems deploy the built model to generate continuous key performance indicators for predicting the plant process.