Adaptive open set domain generalization network: Learning to diagnose unknown faults under unknown working conditions

https://doi.org/10.1016/j.ress.2022.108672Get rights and content

Highlights

  • An end-to-end network is proposed for novel fault detection.

  • Generalized and discriminated representations are obtained by metric learning.

  • The decision boundary is adaptive with individual class representation spaces.

Abstract

Recently, domain generalization techniques have been introduced to enhance the generalization capacity of fault diagnostic models under unknown working conditions. Most existing studies assume consistent machine health states between the training and testing data. However, fault modes in the testing phase are unpredictable, and unknown fault modes usually occur, hindering the wide applications of domain generalization-based fault diagnosis methods in industries. To address such problems, this paper proposes an adaptive open set domain generalization network to diagnose unknown faults under unknown working conditions. A local class cluster module is implemented to explore domain-invariant representation space and obtain discriminative representation structures by minimizing triplet loss. An outlier detection module learns optimal decision boundaries for individual class representation spaces to classify known fault modes and recognize unknown fault modes. Extensive experimental results on two test rigs demonstrated the effectiveness and superiority of the proposed method.

Introduction

With the development of Industry 4.0 and smart manufacturing, rotating machines are moving towards informatization and intelligentization [1,2]. Prognostics and health management (PHM) for machinery are essential strategies to improve the efficiency of condition-based maintenance [3]. As a crucial part of PHM, fault diagnosis is urgently required to ensure manufacturing system reliability and safety [4,5]. Owing to the significant rise of industrial Big Data analytics and the rapid advancement of computing power, data-driven methods have become an active research field in modern fault diagnosis [6]. Especially, deep learning-based fault diagnosis (DLFD) methods are gaining tremendous attention by virtue of high precision and fast responses [7].

Most DLFD methods perform well under the common assumption that the training and testing data are subject to the same distribution [8]. Unfortunately, this strict assumption barely holds in real-world applications because of distribution shifts caused by variations in operation conditions, environmental noise interference, and equipment degradation [9]. Distribution discrepancy between the training and testing data (which are denoted as source and target domains, respectively) generally results in remarkable deteriorations for deep diagnostic models regarding their effectiveness [10]. To address this issue, domain adaptation techniques have been incorporated to improve the robustness and adaptability of DLFD models, which bridge divergences of the source and target domains by aligning their distributions in high-level subspaces [11], [12], [13].

Although domain adaptation-based fault diagnosis (DAFD) methods have achieved notable performance, these methods still face obstacles in handling practical cross-domain fault diagnosis tasks [14]. Unlabeled target data that are essential for facilitating diagnostic knowledge transfer from source domains to target domains are prohibitively difficult to be collected in advance. Because concerned diagnostic tasks are generally under novel working conditions and target fault data are inaccessible during model training.

To get rid of the dependence on target domain data and generalize diagnostic models to unknown working conditions, domain generalization techniques have become an emerging research trend in intelligent condition monitoring and fault diagnosis for mechanical equipment. Domain generalization-based fault diagnosis (DGFD) methods generalize diagnostic knowledge from multiple source domains and are directly applied to unknown target tasks, as shown in Fig. 1(a).

The important assumption of most DGFD methods is the consistent label space across different domains, which requires the same machinery health states existing in both the source and target data. However, this assumption is still difficult to be satisfied in real-world scenarios. Because new fault modes may occur in the testing phase while collecting comprehensive datasets that contain all potential fault modes in the training phase is time-consuming and expensive, incomplete knowledge of the machinery faults in the training phase is common, and label shifts may appear between the training and testing data.

Totally, the above two issues (distribution and label shifts) jointly pose a challenging but valuable problem called open set domain generalization-based fault diagnosis (OSDGFD), where the target domain data are sampled from unknown working conditions, and unknown fault modes appear in the testing phase. However, most existing DGFD methods cannot detect unknown fault modes and will misclassify new fault modes, as shown in Fig. 1(b). In OSDGFD scenarios, diagnostic models are expected to correctly classify target samples if they are shared in the source label space and precisely recognize target outliers as unknown fault modes if they are not observed in source domains, as shown in Fig. 1(c). To the best of our knowledge, there is no research reported so far on the open set domain generalization scenario in the intelligent fault diagnosis.

To bridge this gap and further develop data-driven methods for better industrial applications, an Adaptive Open Set Domain Generalization Network (AOSDGN) is proposed in this study. The proposed AOSDGN consists of local class cluster-based representation learning and class-wise decision boundary-based outlier detection. In the local class cluster module, triplet loss minimization is introduced to boost intra-class compactness and inter-class separability, which mitigates the distribution discrepancy between multiple source domains and learns discriminate representations. In the outlier detection module, a mechanism of class-wise decision boundary is proposed, where decision boundary is adaptive with individual class representation spaces to contain known-class samples and reject unknown-class samples. The two modules benefit each other in the training process.

Two main contributions of this study are summarized as follows: (1) A significant scenario called open set domain generalization-based fault diagnosis is studied.. (2) A novel end-to-end network AOSDGN is proposed for intelligent fault diagnosis. The proposed method can precisely classify known machine heath states and effectively detect unknown fault modes under unknown working conditions.

The remainder of this paper is organized as follows. Section 2 reviews some related work. Section 3 describes the proposed method in detail. Section 4 presents the experimental results and analyses. Finally, Section 5 discusses main conclusions and prospects for future work.

Section snippets

Related work

This section briefly discusses the related fault diagnosis research literature, including open set-based, domain adaptation-based, open set domain adaptation-based, and domain generalization-based fault diagnosis methods. Comparisons of the proposed setting with the fault diagnosis settings of previous research efforts are listed in Table 1.

Scheirer et al. [19] first defined the open set recognition problem where new classes unseen in the training phase appear in the testing phase. Classifiers

Proposed method

Details of the proposed method are introduced in this section, and the framework is shown in Fig. 2. Both distribution shift and label shift appear in the OSDGFD problem. To address these two shifts, local class cluster-based representation learning and class-wise decision boundary-based outlier detection are designed, which explore domain-invariant representations and detect unknown fault modes, respectively.

Experimental study

In this section, comprehensive experiments are implemented on two datasets to demonstrate the efficiency and practical value of the proposed method. We compare different methods, analyze experimental results, and intuitively interpret the improvements of the proposed method.

Conclusion

In this paper, a valuable and realistic industrial application scenario called open set domain generalization-based fault diagnosis (OSDGFD) is concerned and addressed. Given the difficulties of OSDGFD, a novel intelligent diagnostic network called adaptive open set domain generalization network (AOSDGN) is presented. The local class cluster module based on triple loss can guide the feature generator to extract domain-invariant representations and push the known-class samples to the edge of

CRediT authorship contribution statement

Chao Zhao: Conceptualization, Methodology, Software, Writing – original draft. Weiming Shen: Supervision, Writing – review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

This research was support by Fundamental Research Funds for the Central Universities (Grant No.2021GCRC058).

References (40)

  • T. Zhou et al.

    Towards trustworthy machine fault diagnosis: a probabilistic Bayesian deep learning framework

    Reliab Eng Syst Saf

    (2022)
  • W. Song et al.

    Unsupervised fault diagnosis method based on iterative multi-manifold spectral clustering

    IET Collab Intell Manuf

    (2019)
  • W. Song et al.

    An early fault detection method of rotating machines based on unsupervised sequence segmentation convolutional neural network

    IEEE Trans Instrum Meas

    (2021)
  • Z. Chen et al.

    Domain adversarial transfer network for cross-domain fault diagnosis of rotary machinery

    IEEE Trans Instrum Meas

    (2020)
  • C. Zhao et al.

    A balanced and weighted alignment network for partial transfer fault diagnosis

    ISA Trans

    (2022)
  • G. Liu et al.

    Knowledge transfer in fault diagnosis of rotary machines

    IET Collab Intell Manuf

    (2022)
  • Z. Lei et al.

    An intelligent fault diagnosis method based on domain adaptation and its application for bearings under polytropic working conditions

    IEEE Trans Instrum Meas

    (2021)
  • M. Ragab et al.

    Adversarial multiple-target domain adaptation for fault classification

    IEEE Trans Instrum Meas

    (2020)
  • C. Chen et al.

    Domain adaptation based transfer learning for gear fault diagnosis under varying working conditions

    IEEE Trans Instrum Meas

    (2020)
  • C. Zhao et al.

    Adversarial mutual information-guided single domain generalization network for intelligent fault diagnosis

    IEEE Trans Ind Inf

    (2022)
  • Cited by (27)

    View all citing articles on Scopus
    View full text