Recovering traffic data from the corrupted noise: A doubly physics-regularized denoising diffusion model,Transportation Research Part C: Emerging Technologies

当前位置： X-MOL 学术 › Transp. Res. Part C Emerg. Technol. › 论文详情

Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)

Recovering traffic data from the corrupted noise: A doubly physics-regularized denoising diffusion model
Transportation Research Part C: Emerging Technologies ( IF 7.6 ) Pub Date : 2024-02-09 , DOI: 10.1016/j.trc.2024.104513
Zhenjie Zheng , Zhengli Wang , Zijian Hu , Zihan Wan , Wei Ma

Noise is inevitable in the collection of traffic data, which may cause accuracy and stability issues in smart mobility applications. In the literature, most of the existing studies on traffic data denoising assume that noises follow specific distributions (e.g., Gaussian) or structures (e.g., sparsity). However, various noises may coexist in the traffic data and their distributions or structures are usually unknown in the real world. This leads to the corrupted noise and brings a huge challenge to the traffic data recovery due to the irregular characteristics of such noise. In this research, we develop a doubly physics-regularized denoising diffusion model to address this issue. The proposed denoising diffusion model gradually adds noise to the data and then learns to recover the true value from noises. Each layer of the model can be perceived as a filter that removes noises with different scales, which helps to disentangle the coupling of noises. To make up for the insufficient information of noise distributions or structures, we incorporate prior traffic domain knowledge of the fundamental diagram into the denoising diffusion model using . Under suitable assumptions, we prove that our model can output an unbiased estimation of the true value. We validate our model using two real datasets collected in Japan and China. Results demonstrate that our model outperforms other denoising methods, such as filter-based methods, compressed sensing, robust principal component analysis, and machine learning methods.

更新日期：2024-02-09

点击分享查看原文

点击收藏

阅读更多本刊新发论文