Scholarlink

<p>As data environments become more complex, defects in sensors present significant challenges for fault diagnosis, particularly under multivariate data missing (MDM). We propose a dual-dimensional adaptive sequential learning framework addressing two key aspects: (1) Data complexity quantification via generative adversarial imputation network (GAIN) and Gaussian mixture model (GMM) clustering, generating subsets stratified by incompleteness (measured by MDM complexity); (2) Multi-teacher knowledge distillation where a student model progressively learns from teachers, with supervision dynamically weighted by confidence scores. Inspired by human learning dynamics, this dual strategy integrates sample sequencing (curriculum difficulty) and supervision intensity (teacher dependency). Validation on satellite power systems confirms &gt;95 % diagnosis accuracy for mixed MDM data, outperforming nonsequential methods in accuracy and generalizability.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dual-dimensional adaptive sequential learning: Enhanced fault diagnosis under multivariate data missing

  • 期刊论文

  • 发表日期：

    2026年2月9日

查看全文
  • Laifa Tao,
  • Haifei Liu,
  • Shuting Zhou,
  • Xuanyuan Su,
  • Shangyu Li,
  • Cong Zhang,
  • Chen Lu,
  • Zhixuan Lian

摘要

As data environments become more complex, defects in sensors present significant challenges for fault diagnosis, particularly under multivariate data missing (MDM). We propose a dual-dimensional adaptive sequential learning framework addressing two key aspects: (1) Data complexity quantification via generative adversarial imputation network (GAIN) and Gaussian mixture model (GMM) clustering, generating subsets stratified by incompleteness (measured by MDM complexity); (2) Multi-teacher knowledge distillation where a student model progressively learns from teachers, with supervision dynamically weighted by confidence scores. Inspired by human learning dynamics, this dual strategy integrates sample sequencing (curriculum difficulty) and supervision intensity (teacher dependency). Validation on satellite power systems confirms >95 % diagnosis accuracy for mixed MDM data, outperforming nonsequential methods in accuracy and generalizability.