Interpretable PID parameter tuning for control engineering using general dynamic neural networks: An extensive...

Area of Science:

• Control Engineering
• Artificial Intelligence
• Machine Learning

Background:

• Modern automation relies on Proportional-Integral-Derivative (PID) controllers, which struggle with complex, non-linear systems.
• Existing neural network extensions to PID controllers lack stability guarantees and interpretability.
• General Dynamic Neural Networks (GDNN) offer a potential solution for advanced control.

Purpose of the Study:

• To evaluate the performance of GDNN-extended PID controllers on complex benchmark systems.
• To assess the interpretability and stability of these neural PID controllers.
• To demonstrate a scalable and interpretable approach for AI in control systems.

Main Methods:

• Implemented and tested General Dynamic Neural Networks (GDNN) to extend PID controllers.
• Evaluated performance on four benchmark control systems under various conditions (noise, disturbances).
• Applied bounded-input bounded-output stability analysis for interpretability assessment.

Main Results:

• GDNN-enhanced PID controllers outperformed standard PID in 15 out of 16 tasks.
• Neural PID controllers surpassed model-based control in 13 out of 16 tasks.
• Stability analysis provided understandable parameters for neural network controllers.

Conclusions:

• GDNN-enhanced PID controllers offer superior performance and scalability for complex control.
• The developed interpretability methods facilitate the adoption of AI in real-world control.
• This work advances the development of interpretable and safe artificial intelligence for automation.