Scholarlink

<p>Existing studies on multi-source domain deep transfer learning (DTL) methods inadequately leverage the distribution distance between multi-source and target domains. This limitation leads to incomplete cross-domain alignment and negative transfer problems in practical conditions. Therefore, a staged deep multi-source domain adaptation (SDMDA) framework based on loss dynamic adjustment is proposed. Primarily, the dilated temporal convolutional network (DTCN) common feature extractor and joint classifiers are proposed to achieve domain-invariant deep feature extraction and classification, respectively. Moreover, a loss dynamic adjustment mechanism (LDAM) is proposed, which uses Wasserstein distance between the predicted labels of each domain combination to dynamically adjust the difference alignment loss weights for the whole source domains during batch-wise training iterations. On this basis, a negative transfer inhibition mechanism (NTIM) is proposed to strengthen the alignment of target domain feature distribution in the subsequent phase of the training process. The experimental results on three datasets demonstrate excellent performance of the SDMDA framework.</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

A staged deep multi-source domain adaptation framework based on loss dynamic adjustment for bearing fault diagnosis

  • 期刊论文

  • 发表日期：

    2025年8月28日

查看全文
  • Dongpo Wang,
  • Li Zhang,
  • Jianwei Yang,
  • Wenbo Yue

摘要

Existing studies on multi-source domain deep transfer learning (DTL) methods inadequately leverage the distribution distance between multi-source and target domains. This limitation leads to incomplete cross-domain alignment and negative transfer problems in practical conditions. Therefore, a staged deep multi-source domain adaptation (SDMDA) framework based on loss dynamic adjustment is proposed. Primarily, the dilated temporal convolutional network (DTCN) common feature extractor and joint classifiers are proposed to achieve domain-invariant deep feature extraction and classification, respectively. Moreover, a loss dynamic adjustment mechanism (LDAM) is proposed, which uses Wasserstein distance between the predicted labels of each domain combination to dynamically adjust the difference alignment loss weights for the whole source domains during batch-wise training iterations. On this basis, a negative transfer inhibition mechanism (NTIM) is proposed to strengthen the alignment of target domain feature distribution in the subsequent phase of the training process. The experimental results on three datasets demonstrate excellent performance of the SDMDA framework.