Robust self-organizing fuzzy neural network with data immunity evaluation for industrial process modeling

Dynamic property of neural networks can enable the structure or parameters to adapt to the different input data, resulting in substantial improvements in terms of computational efficiency, model accuracy, and adaptability. However, due to the inevitable existence of outliers and noise in practical industrial applications, the input data suffers from high levels of uncertainty, which restricts the model promotion of neural networks. Hence, a robust self-organizing fuzzy neural network with data immunity evaluation (RSOFNN-DIE) is developed for industrial process modeling in this paper. First, a data immunity evaluation strategy is employed to judge the impact of input data on neurons to determine the structure of RSOFNN-DIE. Then, a suitable structure can be obtained to enhance the model robustness of RSOFNN-DIE. Second, a parameter learning algorithm with evaluation penalty mechanism is introduced to update the model parameters of RSOFNN-DIE to accommodate to the change of input data. Then, the proposed model can alleviate the sensitivity of outliers and noise to improve its anti-interference ability. Third, the convergence and robustness of RSOFNN-DIE are analyzed in detail. Then, the above theoretical validation can further guarantee the effective implementation of RSOFNN-DIE in environments with outliers and noise. Finally, RSOFNN-DIE is tested in several industrial processes to reveal its superior model performance in regard to accuracy and robustness.