NONLINEAR SYSTEM IDENTIFICATION BASED ON FUZZY RADIAL BASIS NEURAL NETWORK WITH MULTI-CONNECTED WEIGHT CONNECTIONS

This paper builds on our earlier radial basis function network with multiple connections (RBFMC) by placing it within a fuzzy inference framework for nonlinear system identification. The idea is inspired by the diversity of neurotransmitters found in biological neurons: instead of a single hidden-to-output weight, RBFMC gives each hidden unit a multi-dimensional connection whose components act as independent filters. Once fuzzy logic is added, each hidden neuron becomes a fuzzy rule, and its antecedent is built from several Gaussian membership functions, one per connection. The resulting Fuzzy RBFMC produces an interpretable, multi-filter description of local regions of the input space without sacrificing universal approximation capability, and it is trained through gradient descent with momentum, . regularisation, and early stopping. After describing the architecture, the cooperative and competitive interplay among multi-connection rules, and the training procedure, we evaluate the model on four benchmark nonlinear systems - Narendra-Li, the Box-Jenkins gas furnace, a continuous stirred tank reactor, and a high-dimensional NARX plant - comparing it against standard RBF, fuzzy RBF, RBFMC, and ANFIS. Across all four benchmarks, Fuzzy RBFMC achieves the best accuracy, the fastest convergence, and the most compact rule base, and it produces interpretable fuzzy rules that expose local system dynamics, a conclusion supported by ablation, noise-robustness, and statistical significance analyses.

Hali tarjima qilinmagan