Neural Network Controller Switching for Nonlinear System Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing control systems for noise and vibration, particularly those using linear models and filters, face performance limitations when dealing with nonlinearities in systems, leading to suboptimal control performance and increased computational costs with larger neural networks.
Innovation Solution
The method involves training multiple small neural networks to achieve good performance over specific ranges of nonlinear behavior and switching between them based on system metrics to achieve near-optimal performance at low real-time computational cost, using a processor and memory system to execute this method.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a single large neural network is used to model nonlinear system behavior, then control performance improves, but real-time computational cost increases
Solution Approach 1:
The patent divides a single large neural network into multiple smaller neural networks, each specialized for specific ranges of system operating conditions. This segmentation allows each small network to be computationally efficient while the collection covers the full range of nonlinear behaviors, resolving the contradiction between control performance and computational efficiency.
Solution Approach 2:
The system dynamically selects which small neural network to use based on current system metrics and operating conditions. This dynamic adaptation allows the controller to achieve near-optimal performance for current conditions without the computational burden of a single large network, as only one small network is active at any given time.
2Adaptability or versatility
If multiple small neural networks are trained for different nonlinear behavior ranges, then control performance across varying conditions improves, but system complexity increases
Solution Approach 1:
The patent segments the control task across multiple small neural networks, each handling specific ranges of system metrics. While this creates multiple network components, the segmentation allows each network to be simpler and more specialized, with a selection mechanism that manages complexity by activating only one network at a time based on current operating conditions.
Solution Approach 2:
The patent introduces a selection mechanism that acts as an intermediary between the multiple small neural networks and the system being controlled. This intermediary evaluates system metrics and selects the appropriate network, managing the complexity of having multiple networks while enabling versatile performance across different nonlinear behavior ranges.
3Productivity
If linear control filters and models are used, then computational simplicity is maintained, but control performance degrades in the presence of nonlinearities
Solution Approach 1:
The patent replaces traditional linear control filters and models with neural network-based controllers that can capture nonlinear system behavior. By using small neural networks selected based on system metrics, the solution maintains computational efficiency comparable to linear methods while significantly improving performance in nonlinear regimes through the nonlinear modeling capability of neural networks.
Data Source
Figure 1~2
Figure 3
Figure 4
AI summary
A computer-implemented method of controlling a system using an artificial neural network comprises obtaining (1002) data defining a plurality of trained artificial neural networks models. Each of the plurality of models is trained to output a system control signal based on at least one input signal. The method obtains (1008) at least one system metric relating to operation of the system, and selects (1010) one of the plurality of models using the at least one system metric. The selected model is used (1006) as a controller for the system.