Robust Process Model Identification via Noise Injection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current process model creation techniques for advanced control routines, such as model predictive control, face challenges in generating accurate parametric models due to convergence issues and the inability to account for process complexities, leading to model mismatch and inefficiencies in controller performance.
Innovation Solution
Introducing noise into the process data collected during the model generation process, specifically using zero-mean, evenly distributed noise, to aid in the creation of robust parametric process models, which can improve model convergence and accuracy even in situations with limited data or process complexities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional model creation techniques are used without adding noise, then model convergence is difficult to achieve in the presence of process complexities, but adding noise increases the complexity of the data processing
Solution Approach 1:
The patent converts the harmful effect of noise (which typically degrades signal quality) into a beneficial effect by deliberately adding noise to process data. This added noise excites the process dynamics and enables better identification of process characteristics, improving model convergence and reliability in the presence of process complexities.
Solution Approach 2:
The patent changes the parameter of the process data by adding noise with specific characteristics (zero-mean, evenly distributed). This parameter change transforms the data properties to enable better model identification and convergence, allowing the system to handle process complexities that would otherwise prevent reliable model creation.
2Reliability
If noise is added to process data to improve model robustness, then model accuracy improves in complex scenarios, but the measurement precision of the original data is degraded
Solution Approach 1:
The patent applies partial action by adding noise at a controlled level rather than completely replacing or heavily corrupting the original data. The noise is added in measured amounts to provide just enough excitation to reveal process dynamics without overwhelming the actual process signal, thus maintaining a balance between robustness and precision.
3Manufacturing precision
If more process data is collected to improve model accuracy, then the model better accounts for process complexities, but the time and resources required for data collection increase
Solution Approach 1:
The patent uses noise addition as a form of dynamic excitation analogous to mechanical vibration. By adding noise to the process data, the system excitedly probes the process dynamics and extracts meaningful information more efficiently, achieving better model accuracy without requiring extended data collection periods.
Data Source
AI summary
A robust method of creating process models for use in controller generation, such as in MPC controller generation, adds noise to the process data collected and used in the model generation process. In particular, a robust method of creating a parametric process model first collects process outputs based on known test input signals or sequences, adds random noise to the collected process data and then uses a standard or known technique to determine a process model from the collected process data. Unlike existing techniques for noise removal that focus on clean up of non-random noise prior to generating a process model, the addition of random, zero-mean noise to the process data enables, in many cases, the generation of an acceptable parametric process model in situations where no process model parameter convergence was otherwise obtained. Additionally, process models created using this technique generally have wider confidence intervals, therefore providing a model that works adequately in many process situations without needing to manually or graphically change the model.