Relationship Transfer Domain Generalization Network for Rotating Machinery Fault Diagnosis Under Different Working Conditions

Self Cite

Predictive maintenance (PdM) is currently the most cost-effective maintenance method for industrial equipment, offering improved safety and availability of mechanical assets. A crucial component of PdM is the remaining useful life (RUL) prediction for machines, which has garnered increasing attention. With the rapid advancements in industrial Internet of Things (IoT) and artificial intelligence (AI) technologies, RUL prediction methods, particularly those based on pattern recognition (PR) technology, have made significant progress. However, a comprehensive review that systematically analyzes and summarizes these state-of-the-art PR-based prognostic methods is currently lacking. To address this gap, this paper presents a comprehensive review of PR-based RUL prediction methods. Firstly, it summarizes commonly used evaluation indicators based on accuracy metrics, prediction confidence metrics, and prediction stability metrics. Secondly, it provides a comprehensive analysis of typical machine learning methods and deep learning networks employed in RUL prediction. Furthermore, it delves into cutting-edge techniques, including advanced network models and frontier learning theories in RUL prediction. Finally, the paper concludes by discussing the current main challenges and prospects in the field. The intended audience of this article includes practitioners and researchers involved in machinery PdM, aiming to provide them with essential foundational knowledge and a technical overview of the subject matter.

Section: Cutting-edge Methods In DLmentioning

confidence: 99%

Section: Tl Methods Considering Domain Shiftmentioning

confidence: 99%

A comprehensive survey of machine remaining useful life prediction approaches based on pattern recognition: taxonomy and challenges

Zhou,

Yang,

Qian

et al. 2024

Self Cite

“…a ij k j (10) where, the discrete point column {x n , n ⩾ 0} is introduced, y n is the approximation of y(x n ), h is the step size, λ(x n , y n ; h) is called the incremental function, and k i is the value of the slope of the tangent line at point x i . According to the Runge condition of the Runge-Kutta method, we often assume that c i = s j =1 a ij , i = 1, 2, .…”

Section: Pre-activation Runge-kutta Residual Blockmentioning

confidence: 99%

“…Numerous deep learning methods, such as convolutional neural networks (CNNs) [3,4], transformer [5,6], generative adversarial networks (GANs) [7,8] and domain adaptation [9][10][11], have their own unique advantages and characteristics. Among them, CNNs are designed to process structured data, and have been widely applied in various cross-domain applications of neural networks and fault diagnosis due to their powerful feature extraction capabilities.…”

Section: Introductionmentioning

confidence: 99%

A rolling bearing fault diagnosis method based on Markov transition field and multi-scale Runge-Kutta residual network

Ding,

Rui,

Lei

et al. 2023

In order to address the problem that one- dimensional convolutional neural networks is difficult to extract the local correlation information and mine multi-scale information of rolling bearing fault signals under variable working conditions, a novel fault diagnosis method for rolling bearings based on Markov transition field (MTF) and multi-scale Runge–Kutta residual attention network (MRKRA-Net) is proposed in this paper. Firstly, the original signal is encoded into a two-dimensional image using the MTF method. Then, a multi-scale network is constructed using pre-activation Runge–Kutta residual blocks to extract multi-level features. Secondly, a feature-guided attention mechanism is designed and embedded into the network model to enhance its generalization ability. Finally, the MRKRA-Net model is validated on two different bearing datasets, and the results show that compared with other popular intelligent fault diagnosis methods, MRKRA-Net has higher fault diagnosis accuracy and stronger robustness under both given and variable working conditions.

“…Han et al [30] proposed a domain adversarial-based transfer learning method to deal with sparse target domain data and better performance was achieved by a further comparison. Further, a novel relationship transfer framework was proposed in [31], and several domain discriminators are designed to enhance performance and several domain discriminators are designed to enhance performance when dealing with unseen domains. Experimental results on a wind turbine planetary gearbox dataset and bearing dataset have verified the effectiveness of the proposed RT framework.…”

Section: Introductionmentioning

confidence: 99%

A two-stage multi-scale domain adversarial transfer learning method and its application in fault diagnosis

Zhang,

Huang,

Wang

et al. 2023

Domain adaptation (DA) based intelligent fault diagnosis (IFD) methods have attracted great attention in recent years. The key motivation of DA methods is to extract the domain-invariant features. In most cases, the one-dimensional convolutional neural networks (CNN) are adopted as the feature extractor, in which the kernels are usually single and fixed. However, the monitoring data for IFD usually involves various scale information, the feature representation extracted by above models may be incomprehensive. Moreover, the target domain data is only used for narrowing distribution discrepancy in an unsupervised way, which may lead to the ignorance of class information of target domain. To address these issues, in this paper, a two-stage multi-scale domain adversarial fault diagnosis method is proposed. A multi-scale feature extractor with different kernel sizes is designed to acquire more discriminative domain-invariant features. Meanwhile, the pseudo label learning is adopted for providing transfer learning process with the pseudo labels of target domain, which are generated by a pre-train network in the first stage and then are optimised through iterations in the second stage. The maximum mean discrepancy (MMD) is also adopted to enhance the ability of model for marginal distribution alignment, which can make the model more robust. Thirty-eight transfer tasks from two different datasets were conducted to evaluate the effectiveness of the proposed method. The experimental results demonstrated that the proposed method achieved higher average diagnosis accuracy compared with several popular methods. The superiority of our proposed method was further explained by visualization of learned features.