1. Introduction
Rotating machinery is widely used in manufacturing, energy supply, rail transit, aerospace, and other industries and plays a fundamental role in industrial production [
1]. As an important part of rotating machinery, bearings are a kind of mechanical part that is used to reduce friction between other moving parts. Rotating machinery performance can be significantly impacted by bearing failures, particularly under severe working conditions [
2]. The incorrect handling of these failures can pose a serious threat to personal safety and have a major impact on social and economic development [
3]. Therefore, to maintain a safe operating environment and ensure high production efficiency, it is crucial to accurately identify the health status of the bearings.
The emergence of the industrial big data age has led to the collection and utilization of vast numbers of operational data in industrial production, which has accelerated the development of artificial intelligence techniques for fault diagnostics [
4]. It has been difficult to meet the requirements of rapid and accurate diagnosis with traditional signal-based and model-based fault diagnosis methods, such as wavelet transform, state estimation methods, and so on [
5,
6]. In order to automatically extract fault features and implement high-efficiency diagnosis, deep learning (DL)-based diagnosis methods have become research hotspots in recent years [
7,
8]. Compared with traditional methods based on signal analysis and machine learning, DL-based diagnosis methods can reduce human experience interference and have more advantages in the field of intelligent diagnosis [
9,
10]. In addition, deep learning has further evolved to encompass transfer learning, federated learning, meta-learning, and other advanced paradigms [
11]. These developments have significantly contributed to the enhancement of fault diagnosis methodologies, enabling the extraction of intricate fault features and the realization of high-efficiency diagnostic capabilities.
Currently, the following presumptions underlie effective applications of DL-based methods: (1) The fault samples are drawn from the same distribution of data; (2) A significant quantity of excellently labeled samples with faults is needed. Nevertheless, it is challenging to meet these presumptions in actual industrial scenarios due to the following issues: (1) Samples acquired under various conditions will have different distributions due to variations in working conditions, equipment wear, and ambient noise. In the end, diagnostic performance deteriorates when a model trained under one situation is applied to another; (2) Only a small number of fault data are available under particular operating conditions, and it is challenging to gather substantial historical fault data in advance [
12]. The use of intelligent defect diagnostic techniques in industry is restricted by these issues.
Transfer learning (TL) aims to address the shortcomings of deep learning in fault diagnosis and offers a feasible and promising resolution [
13]. TL allows models to apply knowledge learned from one task to another related task. As one of the important branches of TL, domain adaptation (DA) [
14] transfers knowledge by narrowing the gap between two domains and learning domain-invariant features. In the application of bearing fault diagnosis, Huang et al. [
15] developed a multi-source dense adaptation countermeasure network suitable for bearing fault diagnosis through integrated fusion. Jiao et al. [
16] constructed a shared feature generator network and double classifier network, added additional classifier differences to the adversarial domain adaption network, and achieved more effective cross-domain diagnosis in a domain distribution transfer scenario. Kuang et al. [
17] built an adversarial transfer learning network to diagnose data pertaining to class imbalance. Through adversarial training, they acquired information about class separation diagnosis for unbalanced data, and, through joint-distributed two-layer adversarial transfer learning, they acquired domain-invariant knowledge. While transfer-learning-based fault diagnosis techniques have shown promise, existing strategies primarily focus on minimizing distributional differences, overlooking variations in the underlying data structure of fault features. This oversight limits their effectiveness in capturing crucial structural information, thereby hindering performance in cross-domain fault diagnosis tasks.
To address these limitations, a multi-perception graph convolution transfer network (MPGCTN) is proposed. Unlike previous methods, the MPGCTN not only aims to minimize distributional differences but also aims to capture the intricate data structures inherent in fault features, which is effective for variable-working-condition bearing fault diagnosis across different domains. Central to the MPGCTN is the multi-perception graph convolutional network (MPGCN), which leverages multiple receptive fields to extract robust feature representations from complex data relationships. By integrating class labels, domain labels, and data structures into its architecture, the MPGCTN facilitates high-performance domain adaptation by effectively modeling the inherent characteristics of fault data. Specifically, the MPGCTN comprises three key modules: a graph generation module, a graph perception module, and a domain discrimination module. In the graph generation module, a one-dimensional CNN (1-D CNN) is used to capture features from the input, and then the instance graph is constructed through the graph generation layer. Then, in the graph perception module, the proposed MPGCN is utilized to identify the fault features that are hidden in the relationship data between nodes with varying scales. Finally, in the domain discrimination module, adversarial training of the classifier and domain discriminator is conducted to reduce structural differences between target domain and source domain samples for cross-domain diagnosis. Twelve transmission fault diagnosis tasks are constructed using two types of bearing datasets with four different working conditions to verify the effectiveness of this method. The following are the primary contributions of the suggested method:
(1) A multi-perception graph convolution transfer network (MPGCTN) is proposed which models class labels, domain labels, and data structure in deep networks and which achieves high-performance domain adaptation;
(2) A multi-perception graph convolutional network (MPGCN) is proposed to obtain a more robust feature representation by aggregating features learned from multiple receptive fields;
(3) The proposed method is evaluated on a cross-domain task of variable working conditions on two test platforms. Comparing the MPGCTN to other approaches, the results demonstrate its superior performance.
The rest of this paper is structured as follows:
Section 2 presents a concise overview of key concepts and theoretical foundations concerning graph convolution networks and domain adaptation. The method proposed in this study is thoroughly introduced in
Section 3. In
Section 4, a comparative analysis of the proposed model and a variety of other models is conducted using the CWRU and PU datasets. The paper’s conclusion and future directions are covered in
Section 5.
5. Conclusions
In this paper, a novel graph convolutional network is designed to extract multi-scale structural features, and an MPGCTN is proposed for bearing fault diagnosis. The three modules in the MPGCTN can realize the modeling of sample class labels, domain labels, and data structures so as to complete high-precision cross-domain diagnosis. The experimental results of two cases show that the classification accuracy of the MPGCTN is higher than that of other methods. However, some issues remain unresolved:
(1) This paper ignores the influence of the interpretability of graph neural networks. We cannot fully explain this model;
(2) The ability of the model to transfer diagnostic generalization between different devices is unexplored.
In order to enhance diagnostic performance, we will take these two issues into account in later research and create appropriate solutions.