Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Self-degraded contrastive domain adaptation for industrial fault diagnosis with bi-imbalanced data

Gecheng Chen, Zeyu Yang, Chengwen Luo, Jianqiang Li

TL;DR

Industrial fault diagnosis faces simultaneous domain shift and bi-imbalance between normal and fault classes. Sd-CDA introduces a two-stage framework: (i) imbalance-aware contrastive learning with a pruned encoder (Ia-CLR) to elevate minority signals, and (ii) boundary-aware adversarial domain adaptation (Ba-ADA) using SupCon-DA and a pruned discriminator (PSupCon-DA) to push samples away from the domain boundary while maintaining discriminability. The method adopts a unified objective $\min_{\theta_g,\theta_c}\max_{\theta_d} L_g^p + \lambda_c L_c - \lambda_d L_d + \lambda_{bd} L_{bd}^p$, and provides practical guidelines for pruning proportions and learning rates. Empirical results on rolling bearing and industrial flow datasets show Sd-CDA consistently outperforms standard DA and imbalance-aware baselines, especially under inter-domain and intra-domain imbalance. These findings indicate Sd-CDA offers robust cross-domain fault diagnosis with reduced negative transfer and improved minority fault detection in real-world industrial settings.

Abstract

Modern industrial fault diagnosis tasks often face the combined challenge of distribution discrepancy and bi-imbalance. Existing domain adaptation approaches pay little attention to the prevailing bi-imbalance, leading to poor domain adaptation performance or even negative transfer. In this work, we propose a self-degraded contrastive domain adaptation (Sd-CDA) diagnosis framework to handle the domain discrepancy under the bi-imbalanced data. It first pre-trains the feature extractor via imbalance-aware contrastive learning based on model pruning to learn the feature representation efficiently in a self-supervised manner. Then it forces the samples away from the domain boundary based on supervised contrastive domain adversarial learning (SupCon-DA) and ensures the features generated by the feature extractor are discriminative enough. Furthermore, we propose the pruned contrastive domain adversarial learning (PSupCon-DA) to pay automatically re-weighted attention to the minorities to enhance the performance towards bi-imbalanced data. We show the superiority of the proposed method via two experiments.

Self-degraded contrastive domain adaptation for industrial fault diagnosis with bi-imbalanced data

TL;DR

Industrial fault diagnosis faces simultaneous domain shift and bi-imbalance between normal and fault classes. Sd-CDA introduces a two-stage framework: (i) imbalance-aware contrastive learning with a pruned encoder (Ia-CLR) to elevate minority signals, and (ii) boundary-aware adversarial domain adaptation (Ba-ADA) using SupCon-DA and a pruned discriminator (PSupCon-DA) to push samples away from the domain boundary while maintaining discriminability. The method adopts a unified objective , and provides practical guidelines for pruning proportions and learning rates. Empirical results on rolling bearing and industrial flow datasets show Sd-CDA consistently outperforms standard DA and imbalance-aware baselines, especially under inter-domain and intra-domain imbalance. These findings indicate Sd-CDA offers robust cross-domain fault diagnosis with reduced negative transfer and improved minority fault detection in real-world industrial settings.

Abstract

Modern industrial fault diagnosis tasks often face the combined challenge of distribution discrepancy and bi-imbalance. Existing domain adaptation approaches pay little attention to the prevailing bi-imbalance, leading to poor domain adaptation performance or even negative transfer. In this work, we propose a self-degraded contrastive domain adaptation (Sd-CDA) diagnosis framework to handle the domain discrepancy under the bi-imbalanced data. It first pre-trains the feature extractor via imbalance-aware contrastive learning based on model pruning to learn the feature representation efficiently in a self-supervised manner. Then it forces the samples away from the domain boundary based on supervised contrastive domain adversarial learning (SupCon-DA) and ensures the features generated by the feature extractor are discriminative enough. Furthermore, we propose the pruned contrastive domain adversarial learning (PSupCon-DA) to pay automatically re-weighted attention to the minorities to enhance the performance towards bi-imbalanced data. We show the superiority of the proposed method via two experiments.
Paper Structure (22 sections, 18 equations, 7 figures, 5 tables)

This paper contains 22 sections, 18 equations, 7 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Unsupervised domain adaptation
  4. Imbalanced domain adaptation
  5. Contrastive learning
  6. Methodologies
  7. Motivations and intuition
  8. Problem formulation and overall structure
  9. Imbalance-aware contrastive learning of representations
  10. Boundary-aware adversarial domain adaptation
  11. Training objectives and training procedures
  12. Implementation details
  13. Structure of neural networks
  14. Pruning methods
  15. Basic learning rate $\omega$ and loss weights $\lambda_c$, $\lambda_d$ and $\lambda_{bd}$
  16. ...and 7 more sections

Figures (7)

  • Figure 1: The intuition behind the proposed method. (a.1)-(a.2) illustrates the process of SimCLR, (b.1)-(b.2) shows the process of DANN and (c.1)-(c.3) introduces what the proposed Sd-CDA does.
  • Figure 2: The overall structure of Sd-CDA. It consists of two parts: imbalance-aware contrastive representation learning and boundary-aware adversarial domain adaptation.
  • Figure 3: Sample size for CWRU dataset in different settings
  • Figure 4: Class-wise diagnosis accuracy for CWRU dataset.
  • Figure 5: Sample ratio for Three-phase flow process in imbalanced settings
  • ...and 2 more figures