Scholarlink

<p>Accurate traffic flow estimation is a major process in intelligent transportation systems (ITS), however it is generally limited by noisy and incomplete sensor data. This limitation is formulated as graph signal denoising and it aims to recover clean node signals from noisy observations defined on irregular graph structures. Conventional filtering and spectral models generally based on smoothness assumptions, and limits their efficiency in capturing nonlinear dependencies. To overcome these problems, this work presents a deep learning (DL) based denoising model Enhanced Variational Graph Autoencoder with Pied Kingfisher Optimizer (EVGAE-PKO). The method uses a dynamic improved graph attention (DIGAN) module for capturing high order structural dependencies and mitigating the over smoothing problem. Then, the graph convolutional network (GCN) branch extracts local neighborhood features that preserve important signal variations. The model is optimized through a variational learning objective combined with mean squared error (MSE) reconstruction loss; this enables the model to suppress noise and maintain major signal characteristics. Furthermore, a Pied Kingfisher Optimizer (PKO) is utilized for optimizing major hyperparameters that enhances model stability and reduces manual tuning costs. The proposed EVGAE-PKO model is evaluated by standard traffic flow denoising; on the METR-LA dataset, EVGAE-PKO achieves a MSE of 0.015 and SNR of 16.0 dB. Similarly, on the PEMS-BAY dataset, the model attains an MSE of 0.016 and SNR of 16.3 dB. These outcomes confirm that EVGAE-PKO achieves highly accurate traffic flow estimation and offers a reliable solution for real-world ITS applications. The source code used for training and evaluation is publicly available at: <a href="https://github.com/senthilswamy/Graph-Signal-Denoising">https://github.com/senthilswamy/Graph-Signal-Denoising</a>.</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

Enhanced Variational Graph Autoencoder with Attention and Bio-inspired Optimization for Graph Signal Denoising

  • 期刊论文

  • 发表日期：

    2026年7月22日

查看全文
  • Senthilkumar S,
  • S. Anitha

摘要

Accurate traffic flow estimation is a major process in intelligent transportation systems (ITS), however it is generally limited by noisy and incomplete sensor data. This limitation is formulated as graph signal denoising and it aims to recover clean node signals from noisy observations defined on irregular graph structures. Conventional filtering and spectral models generally based on smoothness assumptions, and limits their efficiency in capturing nonlinear dependencies. To overcome these problems, this work presents a deep learning (DL) based denoising model Enhanced Variational Graph Autoencoder with Pied Kingfisher Optimizer (EVGAE-PKO). The method uses a dynamic improved graph attention (DIGAN) module for capturing high order structural dependencies and mitigating the over smoothing problem. Then, the graph convolutional network (GCN) branch extracts local neighborhood features that preserve important signal variations. The model is optimized through a variational learning objective combined with mean squared error (MSE) reconstruction loss; this enables the model to suppress noise and maintain major signal characteristics. Furthermore, a Pied Kingfisher Optimizer (PKO) is utilized for optimizing major hyperparameters that enhances model stability and reduces manual tuning costs. The proposed EVGAE-PKO model is evaluated by standard traffic flow denoising; on the METR-LA dataset, EVGAE-PKO achieves a MSE of 0.015 and SNR of 16.0 dB. Similarly, on the PEMS-BAY dataset, the model attains an MSE of 0.016 and SNR of 16.3 dB. These outcomes confirm that EVGAE-PKO achieves highly accurate traffic flow estimation and offers a reliable solution for real-world ITS applications. The source code used for training and evaluation is publicly available at: https://github.com/senthilswamy/Graph-Signal-Denoising.