Scholarlink

This research focuses on anomaly detection in multivariate time series, specifically utilizing the Controlled Anomalies Time Series (CATS) dataset, which includes 5 million timestamps and 17 attributes related to telemetry and commands, featuring 200 known anomalies for testing. The study employs a comprehensive workflow that integrates both classical and advanced machine learning techniques, including Isolation Forest, Long Short-Term Memory (LSTM) networks, and Simple Recurrent Neural Networks (RNN). Data preprocessing involves standardization and marking known anomalies, followed by Exploratory Data Analysis (EDA) to visualize trends and correlations. To assess model performance, classifiers such as Random Forest and K-Nearest Neighbors (KNN) are utilized, focusing on metrics like precision, recall, and F1-score. Visualization techniques, including 3D scatter plots, illustrate the anomaly detection capabilities of model. Results indicate that both ensemble methods and deep learning architectures effectively identify anomalies, each exhibiting unique strengths. The research underscores the significance of feature selection and model tuning for optimal performance. Future work will aim to enhance detection accuracy by integrating additional features and refining tuning strategies.

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

Sequence Prediction and Anomaly Detection for Satellite Telemetry

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Gauri Ratnawat,
  • Kanak Sharma,
  • Vipin Jain

摘要

This research focuses on anomaly detection in multivariate time series, specifically utilizing the Controlled Anomalies Time Series (CATS) dataset, which includes 5 million timestamps and 17 attributes related to telemetry and commands, featuring 200 known anomalies for testing. The study employs a comprehensive workflow that integrates both classical and advanced machine learning techniques, including Isolation Forest, Long Short-Term Memory (LSTM) networks, and Simple Recurrent Neural Networks (RNN). Data preprocessing involves standardization and marking known anomalies, followed by Exploratory Data Analysis (EDA) to visualize trends and correlations. To assess model performance, classifiers such as Random Forest and K-Nearest Neighbors (KNN) are utilized, focusing on metrics like precision, recall, and F1-score. Visualization techniques, including 3D scatter plots, illustrate the anomaly detection capabilities of model. Results indicate that both ensemble methods and deep learning architectures effectively identify anomalies, each exhibiting unique strengths. The research underscores the significance of feature selection and model tuning for optimal performance. Future work will aim to enhance detection accuracy by integrating additional features and refining tuning strategies.