ML Stabilization Controller Using Differential Learning on Unstable Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning-based control systems for coherent beam combining in unstable systems require stable training conditions, are inefficient due to continuous dithering, and struggle with non-linear, time-variant systems, necessitating frequent retraining and additional perturbations.
Innovation Solution
A machine learning controller that learns differentially by mapping between differential observation and controller action spaces, allowing continuous learning and adaptation without the need for stable training conditions or retraining, using a neural network to predict actions based on differential measurements and target patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If machine learning is trained on absolute values of input and output, then the model can learn the system mapping, but training requires stable and reproducible states which are impractical due to system drift
Solution Approach 1:
Instead of training on absolute values of input and output, the patent inverts the approach by training on differential values (changes in input and output). This allows the model to learn system dynamics without requiring stable baseline states, making training practical in drifting systems while maintaining learning accuracy.
Solution Approach 2:
The patent transitions from static training (requiring stable states) to dynamic training by using differential measurements that capture system changes over time. This dynamic approach allows the model to adapt to drifting systems during training rather than requiring the system to be held static.
2Adaptability or versatility
If stochastic parallel gradient descent is used for model-free control, then the controller can search the parameter space, but continuous dithering causes additional perturbations and is very inefficient
Solution Approach 1:
The patent applies preliminary action by training the machine learning model offline before deployment. This preliminary training phase captures system dynamics without requiring continuous dithering during operation, eliminating the inefficiency of ongoing stochastic search while maintaining adaptability.
Solution Approach 2:
The patent replaces the mechanical dithering and searching process of stochastic parallel gradient descent with a learned mapping from the trained neural network. This substitution eliminates the need for continuous random perturbations while maintaining the ability to navigate parameter space efficiently.
3Reliability
If periodic re-training is performed to maintain accuracy, then the model adapts to system changes, but this requires stopping operation or additional computational resources
Solution Approach 1:
The patent enables continuous learning by training on differential data during normal system operation. This allows the model to adapt to system changes continuously without stopping operation, maintaining accuracy while eliminating retraining downtime through differential learning that works with drifting systems.
4Device complexity
If conventional model-based control is used, then mathematical representation optimizes controller design, but system identification becomes complex for non-linear, time-variant systems
Solution Approach 1:
The patent substitutes the complex system identification process with a data-driven machine learning approach. Instead of requiring mathematical representation and system identification for non-linear, time-variant systems, the neural network learns the mapping directly from differential input-output data, simplifying the control design process.
Data Source
AI summary
A machine learning (ML) controller and method for systems that can learn to stabilize them based on measurements of an unstable system. This allows for training on a system not yet controlled and for continuous learning as the stabilizer operates. The controller has improved performance on unstable systems compared to similar technologies, especially complex ones with many inputs and outputs. Furthermore, there is no need for modelling the physics, and the controller can adapt to un-analyzed or partially analyzed systems.