Identification of Neurofuzzy models using GTLS parameter estimation

Area of Science:

• Engineering
• Computer Science
• Artificial Intelligence

Background:

• Neurofuzzy systems are powerful tools for nonlinear system identification.
• Accurate parameter estimation in these systems is challenging due to noise in measured data.
• Determining the region of validity for local models is crucial for effective neurofuzzy network design.

Purpose of the Study:

• To address the challenges of parameter estimation and model validity in neurofuzzy systems with noisy data.
• To propose a novel methodology for nonlinear system identification using generalized total least squares (GTLS).
• To enhance the performance of neurofuzzy networks through improved parameter estimation and partitioning.

Main Methods:

• Utilizing generalized total least squares (GTLS) methodologies for parameter estimation in the presence of noisy input channels.
• Employing an expectation-maximization algorithm for proper partitioning of local model validity regions.
• Leveraging residuals from GTLS parameter estimation within the expectation-maximization framework.

Main Results:

• Consistent parameter estimates are achieved even when input channels are subject to noise.
• The proposed expectation-maximization algorithm effectively determines the region of validity for local models.
• The performance of the nonlinear model is enhanced through a combination of weighted GTLS parameter estimation and residual-based partitioning.

Conclusions:

• The presented GTLS-based approach offers a robust solution for nonlinear system identification in neurofuzzy networks.
• The method demonstrates significant benefits in handling noisy data and defining model validity regions.
• Successful application in illustrative examples and an automotive context validates the algorithm's practical utility.