A latent representation dual manifold regularization broad learning system with incremental learning capability for fault diagnosis

Abstract

Fault diagnosis models based on deep learning must spend a lot of time adjusting the model structure and parameters for retraining upon the occurrence of a new fault. To address this problem, a latent representation dual manifold regularization broad learning system (LRDMR-BLS) with incremental learning capability is proposed for fault diagnosis. The model uses the link information between data to guide feature selection via latent representation learning. Meanwhile, two manifold regularization terms are added to the objective function of latent representation learning and the objective function of BLS to maintain the local manifold structure of data and feature spaces. Finally, the incremental learning capability of the proposed model enables the proposed model to be updated quickly when a new fault occurs. The superiority of the proposed model is demonstrated by two chemical processes.