View in EDS

Debiased prototype network with relaxed contrastive distillation strategy for rotating machinery few-shot domain incremental fault diagnosis.

Saved in:
Bibliographic Details
Title: Debiased prototype network with relaxed contrastive distillation strategy for rotating machinery few-shot domain incremental fault diagnosis.
Authors: Zhang, Yu1 (AUTHOR), Han, Dongying1 (AUTHOR) hspace@ysu.edu.cn, Xu, Xuefang2 (AUTHOR), Shi, Peiming2 (AUTHOR)
Source: Knowledge-Based Systems. Jan2026, Vol. 333, pN.PAG-N.PAG. 1p.
Subjects: Fault diagnosis, Rotating machinery, Generalization, Machine learning, Distillation
Abstract: Incrementally assimilating new information from data streams with evolving distributions and scarce labeled samples in a manner more aligned with general intelligent learning and cognition presents challenges for the construction of diagnostic models. In this context, conventional diagnostic models inevitably suffer from insufficient cross-domain generalization, overfitting, and catastrophic forgetting, rendering them ineffective for meeting diagnostic requirements. In this study, we define this complex yet practical problem as a few-shot domain incremental fault diagnosis (FS-DIFD) task and develop a debiased prototype network with relaxed contrastive distillation strategy (RCD-DPN) to address it. The RCD-DPN aims to alleviate the performance limitations encountered by conventional diagnostic models confronted with FS-DIFD tasks within a unified framework. Specifically, architecturally grounded in a prototype network, the debiased prototype network (DPN) attempts to take dominant biased shortcut features as guidance to mine more generic representations with domain-invariant properties. This enables it to suppress overfitting under conditions of scarce labeled samples while laying the foundation for the model to achieve few-shot continuous cross-domain generalization. The relaxed contrastive distillation (RCD) strategy primarily acts on subsequent few-shot domain incremental learning and attempts to impose consistency constraints from the perspective of category semantics, feature representation, and classification decisions in that process. This allows the model to maintain favorable stability to resist catastrophic forgetting of previously adapted domains while exhibiting adequate plasticity to accommodate new distributions and facilitate continuous knowledge transfer. Extensive FS-DIFD tasks constructed on two rotating machinery datasets covering various operating conditions are used to comprehensively evaluate the performance of the proposed RCD-DPN. [ABSTRACT FROM AUTHOR]
Copyright of Knowledge-Based Systems is the property of Elsevier B.V. and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)
Database: Engineering Source
Description
Abstract:Incrementally assimilating new information from data streams with evolving distributions and scarce labeled samples in a manner more aligned with general intelligent learning and cognition presents challenges for the construction of diagnostic models. In this context, conventional diagnostic models inevitably suffer from insufficient cross-domain generalization, overfitting, and catastrophic forgetting, rendering them ineffective for meeting diagnostic requirements. In this study, we define this complex yet practical problem as a few-shot domain incremental fault diagnosis (FS-DIFD) task and develop a debiased prototype network with relaxed contrastive distillation strategy (RCD-DPN) to address it. The RCD-DPN aims to alleviate the performance limitations encountered by conventional diagnostic models confronted with FS-DIFD tasks within a unified framework. Specifically, architecturally grounded in a prototype network, the debiased prototype network (DPN) attempts to take dominant biased shortcut features as guidance to mine more generic representations with domain-invariant properties. This enables it to suppress overfitting under conditions of scarce labeled samples while laying the foundation for the model to achieve few-shot continuous cross-domain generalization. The relaxed contrastive distillation (RCD) strategy primarily acts on subsequent few-shot domain incremental learning and attempts to impose consistency constraints from the perspective of category semantics, feature representation, and classification decisions in that process. This allows the model to maintain favorable stability to resist catastrophic forgetting of previously adapted domains while exhibiting adequate plasticity to accommodate new distributions and facilitate continuous knowledge transfer. Extensive FS-DIFD tasks constructed on two rotating machinery datasets covering various operating conditions are used to comprehensively evaluate the performance of the proposed RCD-DPN. [ABSTRACT FROM AUTHOR]
ISSN:09507051
DOI:10.1016/j.knosys.2025.115117