Fuzzy jump wavelet neural network based on rule induction for dynamic nonlinear system identification with real data...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Fuzzy Wavelet Neural Networks (FWNNs) are effective for nonlinear system identification, handling data imprecision and modeling both global and local system properties.
• Existing FWNN models can be complex and may not always achieve optimal precision in practical applications.

Purpose of the Study:

• To introduce a novel Fuzzy Jump Wavelet Neural Network (FJWNN) model for identifying dynamic nonlinear-linear systems.
• To enhance the precision and efficiency of nonlinear system identification using an improved FWNN architecture.

Main Methods:

• The FJWNN model integrates Takagi-Sugeno-Kang type fuzzy rules with wavelet neurons and linear regressors.
• Orthogonal Least Squares (OLS) and Genetic Algorithms (GA) are employed for wavelet purification within sub-jump FWNNs.
• Fuzzy rule induction is utilized to optimize the model structure, reducing the number of rules, inputs, and parameters.

Main Results:

• The FJWNN model achieved high accuracy across various benchmarks, including piecewise function approximation (RRSE: 10e-5±6e-5) and nonlinear dynamic system modeling (RMSE: 2.6-4±2.6e-4).
• Significant performance was observed in Mackey-Glass time series prediction (RRSE: 1.59e-3±0.42e-3) and EMG signal modeling (VAF%: 98.24±0.71).
• The model demonstrated superior performance compared to state-of-the-art methods, validated by metrics like RMSE, RRSE, Rel ERR%, and VAF%.

Conclusions:

• The FJWNN model exhibits excellent accuracy and generalization capabilities while maintaining manageable network complexity.
• The integration of key wavelets with linear regressors and fuzzy rule induction contributes to the model's enhanced performance.
• The FJWNN presents a novel and effective tool for advanced nonlinear system identification in practical scenarios.