Scholarlink

<p>Circular arrays of flow and flame units are prevalent in turbines, aero-engines, and many other thermal-fluid devices, where effectively identifying collective dynamical modes of these systems is crucial. This study proposes a time-series learning framework for identifying dynamic patterns of circular laminar flame arrays using a Chrono-attention variational autoencoder (VAE) and a Wasserstein distance classifier (WDC). The Chrono-attention VAE-WDC framework first maps the high-dimensional spatial–temporal data extracted from validated numerical simulations into a two-dimensional latent representation through the bidirectional long short-term memory variational autoencoder (2Bi-LSTM-VAE) dimension reduction model, then utilizes the Gaussian kernel density estimation to transfer time-evolving phase points into a phase contour, and then calculates two-dimensional Wasserstein distance between phase contours of various combustion states, and finally performs the Wasserstein-space-based classification for the dynamical mode recognition. By comparing with other models, such as VAE and PCA, the 2Bi-LSTM-VAE model can produce a non-overlapping distribution of phase points, in favor of an effective mode classification and recognition in the two-dimensional latent space. This proposed time-series learning framework provides a new and artificial intelligence (AI)-aided tool to analyze complex dynamical behaviors of thermal-fluid configurations with rotational symmetry in propulsion applications.</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

Time-series deep learning for dynamical mode identification in complex flow systems: application to combustor-inspired circular flame arrays

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年2月24日

查看全文
  • Weiming Xu,
  • Tao Yang,
  • Peng Zhang

摘要

Circular arrays of flow and flame units are prevalent in turbines, aero-engines, and many other thermal-fluid devices, where effectively identifying collective dynamical modes of these systems is crucial. This study proposes a time-series learning framework for identifying dynamic patterns of circular laminar flame arrays using a Chrono-attention variational autoencoder (VAE) and a Wasserstein distance classifier (WDC). The Chrono-attention VAE-WDC framework first maps the high-dimensional spatial–temporal data extracted from validated numerical simulations into a two-dimensional latent representation through the bidirectional long short-term memory variational autoencoder (2Bi-LSTM-VAE) dimension reduction model, then utilizes the Gaussian kernel density estimation to transfer time-evolving phase points into a phase contour, and then calculates two-dimensional Wasserstein distance between phase contours of various combustion states, and finally performs the Wasserstein-space-based classification for the dynamical mode recognition. By comparing with other models, such as VAE and PCA, the 2Bi-LSTM-VAE model can produce a non-overlapping distribution of phase points, in favor of an effective mode classification and recognition in the two-dimensional latent space. This proposed time-series learning framework provides a new and artificial intelligence (AI)-aided tool to analyze complex dynamical behaviors of thermal-fluid configurations with rotational symmetry in propulsion applications.