Robust Pipelined Wavelet Recurrent Neural Network Based on Hybrid Cost Function Over Improved Equalization Whale Optimization

Abstract

The effects of non-Gaussian noise, convergence speed, and free parameter values ​​are the main causes of performance degradation of pipeline wavelet recurrent neural networks (PWRNN) in current practical applications. In this paper, a robust PWRNN based on hybrid cost function (HCF) over improved equalization whale optimization (IEWO) is proposed to overcome the above problems. In which, HCF combined the mean square error criterion and generalized maximum correntropy criterion to solve the problem of slow convergence and the effects of non-Gaussian noise. In addition, a novel IEWO algorithm is designed to automatically calculate the free coefficients, especially the shape factors of the kernels, to suit the experimental environment, thereby improving the detection performance. Finally, simulation results have confirmed the superior performance of the proposed algorithm compared to other time series anomaly detection models.

Keywords: Pipeline wavelet recurrent neural network, Generalized maximum correntropy, Time Series, Whale Optimization Algorithm, Anomaly detection