Scholarlink

Time-series analysis tasks often suffer from limited labeled data, motivating researchers to explore frequency-domain data augmentation methods to improve model generalization. However, existing methods generally apply uniform spectral perturbations, which may inadvertently distort discriminative frequency bands critical to downstream tasks, thus limiting their effectiveness. To address this issue, we propose Multi-Scale Dynamic Preservation Mapping (MSDPM), a frequency-aware augmentation framework designed to dynamically detect and preserve discriminative spectral regions. MSDPM leverages discrete wavelet transform to construct a multi-scale spectral representation and combines energy-entropy scoring with an intra-band self-attention mechanism to adaptively generate sample-specific preservation masks. These masks selectively perturb non-critical frequency components while preserving informative ones. Extensive experiments on four publicly available datasets (HAR, SleepEDF, OPPORTUNITY, WISDM) demonstrate that MSDPM consistently enhances model performance and robustness compared to the state-of-the-art SimPSI. Ablation studies further verify that both the multi-scale decomposition and the dynamic masking mechanisms are crucial for achieving superior results.

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

Multi-scale Dynamic Preservation Mapping: A Time Series Data Augmentation Approach

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuchao Lu,
  • Jinghua Yan,
  • Wei Zhang,
  • Lin Huang

摘要

Time-series analysis tasks often suffer from limited labeled data, motivating researchers to explore frequency-domain data augmentation methods to improve model generalization. However, existing methods generally apply uniform spectral perturbations, which may inadvertently distort discriminative frequency bands critical to downstream tasks, thus limiting their effectiveness. To address this issue, we propose Multi-Scale Dynamic Preservation Mapping (MSDPM), a frequency-aware augmentation framework designed to dynamically detect and preserve discriminative spectral regions. MSDPM leverages discrete wavelet transform to construct a multi-scale spectral representation and combines energy-entropy scoring with an intra-band self-attention mechanism to adaptively generate sample-specific preservation masks. These masks selectively perturb non-critical frequency components while preserving informative ones. Extensive experiments on four publicly available datasets (HAR, SleepEDF, OPPORTUNITY, WISDM) demonstrate that MSDPM consistently enhances model performance and robustness compared to the state-of-the-art SimPSI. Ablation studies further verify that both the multi-scale decomposition and the dynamic masking mechanisms are crucial for achieving superior results.