Intelligent fault diagnosis of machine under noisy environment using ensemble orthogonal contractive auto-encoder

Abstract

Despite significant advances in auto-encoder (AE) based intelligent fault diagnosis recently, their assumption of high-quality sensor data is unrealistic due to noisy environment. Aiming at such problem, this work proposes a novel ensemble orthogonal contractive auto-encoder (EOCA) method to improve diagnosis accuracy. By designing a new cost function including contractive term and orthogonal constrain, OCA can learn robust and non-redundant features from noisy input. To make fault patterns separable, fisher discriminant analysis is cascaded with OCA for mapping these features into a new coordinate space with maximum class separability. Considering the limitation of a single model, sixteen different OCAs with complementary activation functions as well as a new ensemble strategy based on expert model are developed to produce a stable diagnosis result. Additionally, to prove the effectiveness, error bound of EOCA is analyzed and theoretical basis is found for supporting OCA design. EOCA is verified on three cases including three-phase flow facility, gear box and self-priming centrifugal pump. Experiments are conducted by introducing various strength of white Gaussian noise. Results show that EOCA is better than single OCA, canonical AE, sparse AE, denoising AE and several advanced ensemble methods under strong noise.