Parallel stochastic configuration networks for large-scale data regression

Stochastic configuration network (SCN) is a new powerful approach for large-scale data processing which introduces a supervisory mechanism to configure the parameters of each hidden node stochastically. To enhance the accuracy and robustness of SCN, a parallel stochastic configuration network (PSCN) based on the beetle antennae search (BAS) and fuzzy evidence theory is presented. Firstly, we propose a fuzzy evidence theory for the data fusion of multiple neural networks; Secondly, to choose a suitable scale factor lambda of weights and biases, BAS, as a meta-heuristic algorithm which only uses one individual to search for optimal parameter, it is appropriate for parameter selection of SCN, termed as BAS-SCN. Thirdly, parallel training of multiple BAS-SCNs with different objective functions, then several preliminary results of BAS-SCNs are fused by the fuzzy evidence theory to obtain the final results of PSCN. Finally, a complicated real-world dataset (IEEE 2012 PHM) is used to verify the performance of the PSCN. Numerical experimental results show that BAS-SCN performs well in parameter optimization of SCN, PSCN not only has the advantages of BAS-SCN, but also has a higher accuracy and stronger robustness. PSCN has a better performance in the prediction of bearing remaining useful life. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

The study introduces a parallel Stochastic Configuration Network (PSCN) method based on fuzzy evidence theory and beetle antennae search, which is used for large-scale data processing to improve accuracy and robustness, especially excelling in predicting bearing remaining useful life.