Deep regularized variational autoencoder for intelligent fault diagnosis of rotor-bearing system within entire life-cycle process

The performance of complex rotor-bearing system usually decreases with the development of the running time, which indicates that the rotor-bearing system usually goes through several stages (e.g., normal, slight fault, middle fault, and severe fault) in performance degradation process. Namely, the fault categories and severities of rotor-bearing system are difficult to identify in entire life-cycle, which indicates that traditional methods are insufficient in solving such problems. Hence, this paper proposes a novel deep learning model named deep regularized variational autoencoder (DRVAE) for intelligent fault diagnosis of rotor-bearing system. Within the new model, the regular terms are respectively appended to the loss function of variational autoencoder (VAE) through several regularized techniques (e.g., Laplacian, L12 norm and homotopy regularization), which can solve the overfitting problem of the original VAE and enhance feature learning capability of network model. The weighted operation of deep features learned from the regularized VAE is conducted to capture more discriminative fault information, and the hyper-parameters of DRVAE are determined adaptively by bird swarm algorithm (BSA), thus enable the DEVAE method to identify automatically fault categories and severities of rotorbearing system in entire life-cycle. The effectiveness of the proposed method is validated applying two cases of entire life-cycle test of rotor-bearing system. Experiments show that the proposed method achieves a satisfactory identification result for fault categories and seventies of rotor-bearing system. More importantly, the proposed method can greatly improve identification accuracy and feature learning performance compared with the original VAE and most representative several deep learning models. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

The performance of complex rotor-bearing system deteriorates with time, making it difficult to identify fault categories and severities throughout the entire life-cycle. The proposed DRVAE method utilizes deep learning to effectively improve identification accuracy and feature learning performance.