Skip to main navigation Skip to search Skip to main content

Meta-class-incremental transfer learning for lifelong cross-domain diagnosis of rotary machinery under varying working conditions

  • Beijing Institute of Technology
  • Chongqing University
  • Tsinghua University

Research output: Contribution to journal › Article › peer-review

Abstract

Class-incremental learning approaches strengthen lifelong fault diagnosis performance in industrial machinery systems under continuously emerging fault modes throughout the system lifecycle. However, several obstacles including cross-domain diagnosis difficulties, catastrophic forgetting, and constrained model generalization continue to impede the practical deployment of class-incremental learning. To overcome these obstacles, this work studies a lifelong cross-domain intelligent diagnosis method utilizing a meta-class-incremental transfer learning (MCITL) model. Firstly, a continuous wavelet transform (CWT)-tailored feature extractor is constructed by integrating a residual network with a convolutional block attention feature fusion module, where channel attention recalibrates multi-sensor CWT channel importance and spatial attention highlights fault-discriminative time-frequency regions. Then, an enhanced knowledge distillation (EKD) method adapted from relational distillation principles is integrated with a memory replay strategy to capture both inter-class and intra-class correlation relationships, thereby effectively mitigating catastrophic forgetting. Subsequently, a meta-class-incremental parameter learning strategy inspired by gradient-based meta-learning is devised to promote adaptive cross-domain representation learning and continual knowledge evolution across expanding class distributions, substantially improving model generalization in incremental transfer diagnostic settings. Finally, the proposed MCITL model for lifelong class-incremental transfer diagnosis was validated on two rotary machinery datasets involving speed and load variations. Extensive experimental results have shown the advantageous average accuracy of 98% and 90.54%, the lowest average forgetting rate of 1.11% and 4.63% for our presented MCITL method on two machinery datasets, respectively, outperforming several representative class-incremental learning approaches under various class-incremental transfer scenarios.

Original languageEnglish
Article number116270
JournalEngineering Applications of Artificial Intelligence
Volume184
DOIs
Publication statusPublished - 15 Nov 2026

Keywords

  • Class-incremental learning
  • Industrial machinery systems
  • Knowledge distillation
  • Lifelong intelligent diagnosis
  • Meta-learning
  • Transfer learning

Fingerprint

Dive into the research topics of 'Meta-class-incremental transfer learning for lifelong cross-domain diagnosis of rotary machinery under varying working conditions'. Together they form a unique fingerprint.

Cite this