Gas Turbine Dynamic Inversion Control With RL Model Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing gas turbine engine control systems struggle with non-uniform responses across engines and lack real-time adaptability, particularly during degradation, necessitating manual offsets for fan speed variations.
Innovation Solution
A system utilizing reinforcement learning and a linear state space model inversion, combined with a machine learning network, continuously updates engine control parameters in real-time to achieve uniform response and stability, eliminating the need for manual offsets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If traditional control systems are used for gas turbine engines, then the control structure is simple and easy to implement, but the system lacks real-time adaptability and produces non-uniform responses across engines
Solution Approach 1:
The control system transitions from static to dynamic by continuously updating the state space model parameters in real-time based on actual engine sensor data. The model parameters are adapted online to reflect current engine conditions, enabling the system to handle degradation and variations dynamically rather than relying on fixed pre-programmed parameters.
Solution Approach 2:
The system implements feedback by comparing actual sensor measurements with model predictions and using the differences to update the state space model parameters. This closed-loop approach allows the control system to learn from actual engine behavior and adjust accordingly, improving adaptability while maintaining a relatively simple control architecture.
2Reliability
If manual offsets are applied for fan speed variations, then uniform response can be achieved, but the operation becomes more complex and requires manual intervention
Solution Approach 1:
The control system performs self-adjustment by automatically updating its internal state space model parameters based on sensor feedback without requiring manual intervention. The system serves itself by learning from actual engine responses and adapting its control strategy, eliminating the need for operators to manually apply offsets while maintaining uniform response across engines.
Solution Approach 2:
The system achieves response uniformity by dynamically changing the parameters of the state space model based on actual engine behavior. Instead of applying manual offsets to control inputs, the system modifies the underlying model parameters to reflect actual engine characteristics, thereby achieving uniform responses through automatic parameter adaptation rather than manual adjustment.
3Reliability
If the control system does not adapt to engine degradation, then the control logic remains simple, but the engine performance and stability deteriorate over time
Solution Approach 1:
The system prepares for potential degradation by continuously maintaining an updated state space model that reflects current engine conditions. Rather than reacting to degradation after it occurs, the system proactively adapts its parameters in real-time, preventing performance deterioration before it impacts stability. This preliminary adaptation approach maintains simple control logic while improving reliability.
4Adaptability or versatility
If reinforcement learning is implemented for dynamic inversion control, then real-time adaptability and uniform response are achieved, but the computational requirements and processing time increase
Solution Approach 1:
The system applies reinforcement learning selectively to update only the critical parameters of the state space model that have the greatest impact on control performance. Rather than performing complete model re-identification or extensive computations, the system focuses computational resources on the most influential parameters, achieving real-time adaptability with reduced processing time and computational burden.
Data Source
AI summary
There are provided systems and methods for inversion control of turbine engines. For example, there is provided a processor-implemented method that includes a processor and a memory. The memory includes instructions which, when executed by the processor, cause the system at least to perform: simulating, by a simulated state space model, a desired dynamic response based on the sensor data and a control input; inverting the desired dynamic response as output by the state space model; determining an error between a perceived dynamic response and the inverted desired dynamic response; correcting the state space model for the determined error based on updating the one or more model parameters using a machine learning network; generating the desired dynamics based on the updated state space model; and controlling the engine based on the state space model.