Deep Convolutional Stack Autoencoder of Process Adaptive VMD Data With Robust Multikernel RVFLN for Power Quality Events Recognition

Abstract

In this article, an improved particle swarm optimization (PSO)-based variational mode decomposition (VMD) is proposed to compute the most informative band-limited intrinsic mode function (BLIMF) of highly nonstationary single as well as combined power quality events (PQEs). A novel reduced deep convolutional neural network (RDCNN) embedded with stack autoencoder, that is, RDCSAE structure is introduced to extract the most discriminative unsupervised feature data by importing the selected BLIMF of parameter-adaptive VMD (PAVMD) algorithm. A supervised robust multikernel random vector functional link network (RMRVFLN) method is proposed to further train the unsupervised features combined with the deep convoluted Fourier privileged data for the recognition of complex PQEs accurately. Automatic computation of minimum overlapped descriptive features, unified complex feature learning framework, outstanding recognition accuracy, robust antinoise performance, and quick PQEs recognition time prove the superiority of the proposed PAVMD-RDCSAE-RMRVFLN method over RDCNN, PAVMD-RDCNN, PAVMD-RDCSAE, PAVMD-RDCNN-RMRVFLN, and PAVMD-RDCSAE-MRVFLN methods. Finally, the architecture of the proposed method is designed, implemented, and tested in a fast digital field-programmable gate array (FPGA) embedded processor to validate the feasibility, practicability, and performances for the online PQEs monitoring.