A working-condition-robust method for two-stage health measurement of rolling bearings based on energy entropy distribution and dynamic adversarial transfer network

Abstract

Abstract Accurate and robust health measurement for rolling bearings under variable working conditions has great significance in guaranteeing the safe and stable operation of rotating machinery. In this paper, a two-stage and working-condition-robust health measurement method is proposed, systematically blending energy entropy theory, a deep-learning approach and transfer-learning technology. In the first stage, a state boundary of energy entropy is systematically deduced based on an adaptive variational mode decomposition (VMD) improved fruit fly optimization algorithm (IFOA) and the principle of statistical analysis to detect abnormal states in bearings, where the IFOA is developed to search for the optimal parameters of the VMD with high efficiency. In the second stage, if a fault exists, a hybrid robust auto-encoder adopting a multi-layer and deep structure is constructed to strengthen the feature extraction capacity and automatically capture valuable and robust fault features from original samples. Considering the insufficiently labeled samples and significant data distribution discrepancy, a novel dynamic adversarial transfer network (DATN) is designed to extract the transferable and domain-invariant features between source and target datasets and achieve accurate fault identification. Specifically, a dynamic adversarial coefficient based on Wasserstein distance is provided in the DATN to quantitatively evaluate the relative importance of marginal and conditional distributions. Extensive experiments on two rolling bearing datasets validate the superior performance of the proposed method compared with other state-of-the-art identification models and transfer-learning approaches.