Neural MPC Controller Adaptation for Sudden Dynamics Changes
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing control systems for autonomous and user-controlled vehicles and aircraft struggle to rapidly adapt to sudden changes in system dynamics, particularly after damage, due to the need for extensive data and slow adaptation of online learning models.
Innovation Solution
A neural-Model Predictive Control (MPC) controller system that uses non-transitory electronic processors to generate control signals and predicted plant states, incorporating a forward module corrector as a neural network to rapidly adapt to changes in system dynamics with sparse data, allowing for quick recovery from damage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If on-line learning methods are used to adapt the controller, then the controller can gradually adapt with every data point, but it cannot adapt to sudden changes quickly enough
Solution Approach 1:
The system pre-generates multiple candidate controllers offline before damage occurs. When damage is detected, the system immediately selects from these pre-prepared candidates rather than learning online, enabling instant adaptation to sudden changes without the time penalty of gradual online learning.
Solution Approach 2:
The system transitions from static offline controller generation to dynamic online selection and hybridization. The controller pool is updated adaptively based on detected damage, and controllers are dynamically hybridized in real-time to optimize performance for the current damaged state, combining the speed of offline preparation with the flexibility of online adaptation.
2Measurement precision
If machine learning methods require a lot of data to relearn functional relationships, then accuracy can be maintained, but adaptation speed suffers due to data requirements
Solution Approach 1:
Multiple candidate controllers are pre-generated offline with high accuracy using extensive training data before deployment. When damage occurs, the system immediately selects from these pre-trained accurate controllers without needing to collect and process large amounts of new data, thus maintaining both accuracy and fast adaptation.
Solution Approach 2:
Instead of relearning from scratch with large datasets, the system creates and maintains multiple copies (candidate controllers) of control policies offline. These copies are then selected and combined based on the damaged state, avoiding the need to relearn functional relationships online and enabling fast adaptation while preserving accuracy.
3Ease of operation
If a forward model is used for model-predictive control, then optimal control commands can be computed, but adapting the model to sudden changes requires additional methods and time
Solution Approach 1:
Multiple candidate forward models are pre-generated offline representing different possible damaged states. When damage occurs, the system selects the appropriate pre-generated model that matches the damaged state, enabling immediate adaptation without needing to retrain or adjust the forward model online, thus maintaining both optimization capability and adaptability.
Solution Approach 2:
The system changes the active forward model parameters based on detected damage by selecting from a pool of pre-generated models with different parameters. This allows the model to adapt to sudden changes by switching between pre-configured parameter sets rather than continuously adjusting parameters online, maintaining both optimal control and rapid adaptation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A neural-Model Predictive Control (MPC) controller is described to control a dynamical system (i.e., "plant"). The MPC controller receives, in a base controller, a measurement of a current state of a plant and generates a control signal based on the measurement of the current state of the plant. A forward module receives the measurement of the current state of the plant and the control signal to generate a forward module prediction. A forward module corrector receives the measurement of the current state of the plant and the control signal from the base controller to generate an additive correction to the forward module prediction to generate a predicted plant state. Control sequences of length L of pairs of control signals and corresponding predicted plant states are generated until N s control sequences have been generated. A next plant control signal is generated based on the N s control sequences.