LIGO Document T2300283-v1

Implementing Nonlinear Control in a Classical Experiment to Reduce Measurement Noise

Document #:
LIGO-T2300283-v1
Document type:
T - Technical notes
Other Versions:
Abstract:
Precise temperature control in the presence of noisy environments and heat loss through complex channels, involving conduction, convection, and radiation, presents a significant challenge. Traditional control methods, such as PID control, struggle to maintain a desired set-point due to system non-linearity and large disturbances caused by day-to-day ambient temperature fluctuations. The optimal tuning parameters also vary with external factors, further affecting performance.
In this project, we propose an adaptive control approach for nonlinear systems using neural networks trained via reinforcement learning. By introducing nonlinearity into the controller, we aim to address the limitations of traditional methods. The neural network-based controller leverages the entire state space, overcoming the challenge of non-separable and non-linear actuation functions where system parameters lack linear relationships.
Our approach offers several advantages for precise temperature regulation. Through reinforcement learning, the neural network controller learns to effectively respond to varying ambient conditions, adapt control signals, and dynamically adjust to disturbances. This adaptability eliminates the need for fixed tuning parameters, ensuring robust performance across different operating scenarios.
Extensive simulations are conducted to evaluate the proposed approach in realistic scenarios with diverse environmental conditions. The results demonstrate superior performance compared to traditional PID control methods. The neural network-based adaptive control exhibits enhanced set-point tracking accuracy and reduced sensitivity to low frequency (~1/day) fluctuations.
The significance of this work lies in its potential to advance temperature control in various nonlinear systems. By combining neural networks and reinforcement learning, our approach offers a practical solution for achieving precise control in the presence of disturbances. This work opens doors for the application of adaptive control in a wide range of fields where accurate control is essential in the face of complex dynamics and external fluctuations.

Files in Document:
Other Files:

DCC Version 3.5.0, contact DCC Help