Neural Control System Inversion With Regularized State Prediction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Model predictive control (MPC) methods in robotics and automation face challenges due to the complexity and noise in automatically generated system models by machine learning methods, leading to unsuitable control solutions that may not correspond to real-world dynamics.
Innovation Solution
A method involving a neural network with a first component trained to predict the target system's state and a second denoising neural network component to generate a regularizer, which is used to invert the first component and optimize control signals, ensuring the control system generates sensible outputs by directing the search towards typical operating conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If machine learning methods are used to automatically generate a model of the target system, then the model can be constructed without requiring detailed system knowledge and can be updated automatically, but the generated model becomes very complex and internally noisy, producing artefactual predictions that do not correspond to real-world dynamics
Solution Approach 1:
The patent segments the model construction process into two distinct components: (1) an automatic model generation component that creates the initial complex model from operational data, and (2) a human-in-the-loop verification component that reviews and validates the generated model. This segmentation allows the benefits of automatic generation while mitigating the harms of model complexity and noise through human expert review.
Solution Approach 2:
The patent introduces an intermediary human expert layer between the automatic model generation process and the final control system deployment. This intermediary verifies the generated models, filters out artefactual predictions, and ensures that only reliable models corresponding to real-world dynamics are used for control optimization.
2Extent of automation
If control actions are planned by optimizing the outcome predicted by the automatically generated model, then control signals can be generated automatically, but the optimization finds points in input space that are artefacts of the model rather than corresponding to dynamics of the real world system
Solution Approach 1:
The patent performs preliminary verification of the model before using it for control optimization. A human expert reviews and validates the automatically generated model in advance, ensuring it accurately represents real-world dynamics before the model is used for generating control actions. This preliminary action prevents artefactual optimizations.
Solution Approach 2:
The patent implements a feedback mechanism where the generated control actions are validated against real-world system behavior. If artefactual predictions are detected, the system returns to the model generation stage for correction, ensuring continuous improvement and reliability of control actions.
3Reliability
If a sufficiently accurate model of the target system is constructed for MPC applications, then reliable control signals can be generated, but constructing such a model is not feasible for large and/or complex systems and requires continuous adjusting when the system changes
Solution Approach 1:
The patent enables the system to automatically generate its own operational model from its own operational data without requiring external expert intervention for model construction. The system serves itself by learning from its operations, reducing the complexity burden while maintaining model accuracy through automatic adaptation.
Solution Approach 2:
The patent allows the model to automatically adapt to system changes by continuously learning from new operational data. When system parameters change (e.g., physical parts wearing down, operating conditions changing), the model automatically updates its parameters to reflect the new reality, eliminating the need for manual model re-construction.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
The invention relates to a method for generating a control system (120) for a target system (1 10), wherein: operational data (210) is received; a first neural model component (310) is trained with the received operational data for generating a prediction on a state of the target system (1 10) based on the received operational data; a second neural model component (320) is trained with the operational data for generating a regularizer for use in inverting the first neural model component; and the control system (120) is generated (330) by inverting the first neural model component by optimization and arranging to apply the regularizer generated with the second neural model component in the optimization. The invention relates also to a system and a computer program product.