Scholarlink

This work focuses on general ML/AI assisted analytic processes for monitoring, detection, and classification of anomaly signals from multi-modality sensor data. Specifically, I will show series of variational autoencoders (VAE) including unsupervised AI transformers and related workflow pipelines applied to maritime surveillance in the different time scales of hours, minutes, and seconds. I show these developments using the distributed acoustic sensing (DAS) data set. The DAS data set is from the Sandia National Laboratories. DAS is a special type of fiber optic seafloor communications cables to interrogate the submarine environment at Arctic Alaska. Acoustic heatmaps can be generated to detect waves, ships, marine mammals, and other events. There are 18000 channels and sampled at 1kHZ for the data in 2022. The data set is used to demonstrate the VAE methodology for detecting anomaly, event, and classify objects. The results can enable processing and data analytical capabilities critical to actionable intelligence for mission planning and emerging behavior detection.

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

Knowledge Graphs (KG) Assisted Variational Autoencoder (VAE) for Large-Scale Anomaly and Event Detection

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ying Zhao

摘要

This work focuses on general ML/AI assisted analytic processes for monitoring, detection, and classification of anomaly signals from multi-modality sensor data. Specifically, I will show series of variational autoencoders (VAE) including unsupervised AI transformers and related workflow pipelines applied to maritime surveillance in the different time scales of hours, minutes, and seconds. I show these developments using the distributed acoustic sensing (DAS) data set. The DAS data set is from the Sandia National Laboratories. DAS is a special type of fiber optic seafloor communications cables to interrogate the submarine environment at Arctic Alaska. Acoustic heatmaps can be generated to detect waves, ships, marine mammals, and other events. There are 18000 channels and sampled at 1kHZ for the data in 2022. The data set is used to demonstrate the VAE methodology for detecting anomaly, event, and classify objects. The results can enable processing and data analytical capabilities critical to actionable intelligence for mission planning and emerging behavior detection.