Flexible structure multiple modeling using irregular self-organizing maps neural network.

Area of Science:

• Computational intelligence
• Adaptive control systems
• Machine learning for system identification

Background:

• Traditional system identification methods often struggle with complex, nonlinear, or time-varying systems.
• The authors previously introduced the Multiple Model Self-Organizing Map (MMSOM) for system identification.
• Existing methods may lack flexibility in capturing the full range of plant model spaces, especially concave ones.

Purpose of the Study:

• To enhance the MMSOM identification method by incorporating an Irregular Self-Organizing Map (ISOM).
• To develop a more flexible and adaptive multiple modeling approach for system identification.
• To improve the ability to model concave linear model spaces of a plant.

Main Methods:

• The Multiple Modeling by Irregular Self-Organizing Map (MMISOM) utilizes an Irregular SOM (ISOM) as its core component.
• Instantaneous model parameters are adaptively computed and used as inputs to the neural network.
• The neural network learns local models, and the closest model to the plant output is selected at each instant.

Main Results:

• The MMISOM demonstrates increased flexibility in covering concave linear model spaces.
• The method allows for the addition of new models dynamically if the initial number is insufficient.
• The MMISOM effectively estimates parameters of local models for adaptive plant modeling.

Conclusions:

• The proposed MMISOM method provides a more adaptable and robust approach to system identification compared to previous methods.
• MMISOM's architecture, using ISOM, enhances its capability to model complex system dynamics, particularly concave spaces.
• This advancement contributes to more accurate and flexible real-time system modeling in adaptive control applications.