A fault diagnosis method for rolling bearings based on RDDAN under multivariable working conditions

Abstract

Mechanical equipment in actual motion can produce noise interference with the vibration signal of rolling bearings, which have non-constant load and speed. These factors lead to variable and unstable vibration signals of rolling bearings, so it is very difficult to accurately diagnose the actual running rolling bearings. In this paper, a Residual Denoising Dynamic Adaptive Network (RDDAN) is proposed, which uses the signal knowledge under known working conditions to diagnose the rolling bearing faults under unknown working conditions. The method mainly consists of data pre-processing, feature extraction, and dynamic distribution adaptation. First, Gaussian noise is added in the data pre-processing stage to emulate the noise perturbation in the reality of rolling bearing operation. Secondly, a Deep Residual Shrinkage Network (DRSN) is used for noise reduction and feature extraction. Finally, the marginal probability distribution and conditional probability distribution under different working conditions are calculated depending on the characteristics. The network is disciplined using the relative weight of the marginal probability distribution and the conditional probability distribution. And the fault classification results are output after multiple iterations. The method was tested on the Case Western Reserve University bearing dataset and the Machine Fault Simulator Magnum bearing dataset respectively. By comparing other models, RDDAN improves the average accuracy by about 23%. The results show that RDDAN can effectively solve the problem of inconsistent data distribution in rolling bearings under operating conditions influenced by multiple variables such as noise, load, and speed.

Discriminative subspace embedded dynamic geometrical and statistical alignment based on pseudo-label correction for cross-domain bearing fault diagnosis

“…Additionally, manually labeling data from different working conditions is timeconsuming and costly. Secondly, the collected monitoring data exhibit significant distribution differences across different machines, loads, speeds, and noise interference [11,12]. Consequently, there is a significant gap between the probability distributions of the training and test sets, which may substantially degrade the generalization and robustness of DL-based diagnostic models [13,14].…”

“…Additionally, the deep correlation alignment (DCORAL) method [16] aims to align the covariance matrices of two domains by using the correlation alignment (CORAL) metric. Inspired by the concept of generative adversarial networks, Shi et al [12] proposed a domain adversarial training framework that learns domain-invariant features by simultaneously minimizing the label prediction loss and maximizing the domain classification loss. This approach has demonstrated promising results in various DA tasks.…”

Mechanical varying non-stationary signal separation method based on instantaneous frequency estimation

“…In general, the existing signal processing methods are majorly applicable to the steady speed working condition. 1 Under variable speed, the fault vibration signal will lose the periodic regularity, and present the frequency modulation, amplitude modulation property, leading to the invalidity of classical methods. 2 With the development of non-stationary signal processing technology, scholars have gradually carried out research on the mechanical fault vibration signal separation.…”

A lightweight GAN-based fault diagnosis method based on knowledge distillation and deep transfer learning

“…By detecting and identifying bearing faults, maintenance can be performed before severe damage occurs. This proactive approach can help save time and costs while preventing potential safety hazards [6].…”