ESA Anomalies Dataset (ESA-AD)

ESA Anomalies Dataset（ESA-AD）是由欧洲空间局（ESA）创建的一个大规模、结构化的机器学习准备数据集，用于卫星遥测中的异常检测。该数据集包含来自三个不同ESA任务的真实生活遥测数据，其中两个任务被选为ESA-ADB的一部分。数据集由空间飞行器操作工程师（SOEs）和机器学习专家手动标注，并通过最先进的算法进行交叉验证。ESA-AD旨在解决卫星遥测中多变量时间序列异常检测的挑战，提供了一个包含超过7亿数据点的大规模数据集，总压缩数据量超过7GB。该数据集的应用领域包括卫星健康监测和自主卫星操作，旨在通过机器学习技术提高异常检测的准确性和效率。

ESA Anomalies Dataset (ESA-AD) is a large-scale, structured machine learning-ready dataset developed by the European Space Agency (ESA) for anomaly detection in satellite telemetry. This dataset comprises real-world telemetry data from three distinct ESA missions, two of which are included as part of ESA-ADB. It was manually annotated by Spacecraft Operations Engineers (SOEs) and machine learning specialists, and cross-validated using state-of-the-art algorithms. ESA-AD is designed to tackle the challenges of multivariate time series anomaly detection in satellite telemetry, offering a large-scale dataset containing over 700 million data points with a total compressed size exceeding 7 GB. Its application scenarios cover satellite health monitoring and autonomous satellite operations, with the goal of improving the accuracy and efficiency of anomaly detection through machine learning technologies.

搜集汇总

数据集介绍

构建方式

The ESA Anomalies Dataset (ESA-AD) was meticulously constructed through close collaboration between spacecraft operations engineers from the European Space Agency (ESA) and machine learning experts. The dataset comprises annotated real-life telemetry data from three distinct ESA missions, with two of these missions selected for benchmarking in the ESA Anomaly Detection Benchmark (ESA-ADB). The telemetry data was manually annotated by spacecraft operations engineers (SOEs) and machine learning experts, with cross-verification using state-of-the-art algorithms to ensure accuracy and reliability. The dataset is structured to be machine learning-ready, addressing the challenges of multivariate time series anomaly detection in satellite telemetry.

特点

The ESA Anomalies Dataset (ESA-AD) stands out due to its large-scale, high-dimensional, and diverse nature. It includes telemetry data from three ESA missions, covering a wide range of operational conditions and anomalies. The dataset is annotated with a variety of anomaly types, including univariate and multivariate anomalies, global and local anomalies, and point and subsequence anomalies. Additionally, the dataset incorporates rare nominal events, which are atypical but expected changes in telemetry that should not be alarmed. The inclusion of these events poses a unique challenge for anomaly detection algorithms, as they must distinguish between true anomalies and nominal events. The dataset also features irregular timestamps, varying sampling rates, and communication gaps, reflecting the raw telemetry data accessible to SOEs.

使用方法

The ESA Anomalies Dataset (ESA-AD) is designed to facilitate the development and validation of anomaly detection algorithms for satellite telemetry. Researchers and practitioners can use the dataset to train, validate, and test their algorithms in a controlled environment that closely mimics real-world operational scenarios. The dataset supports both supervised and unsupervised learning approaches, allowing for the evaluation of a wide range of machine learning techniques. The hierarchical evaluation pipeline, designed by machine learning experts, includes new metrics tailored for satellite telemetry, such as corrected event-wise F0.5-score, subsystem-aware F0.5-score, and anomaly detection timing quality curve (ADTQC). These metrics enable a comprehensive assessment of algorithm performance, focusing on practical aspects of mission control. The dataset is publicly available, ensuring reproducibility and encouraging further research in the field of satellite telemetry anomaly detection.

背景与挑战

背景概述

ESA Anomalies Dataset (ESA-AD) 是由欧洲空间局（ESA）与机器学习专家合作开发的一个用于卫星遥测异常检测的基准数据集。该数据集旨在解决卫星遥测中多元时间序列异常检测的挑战，特别是在缺乏可理解的基准数据集的情况下。ESA-AD 包含了来自三个不同 ESA 任务的真实遥测数据，其中两个任务的数据被用于基准测试。该数据集通过手动注释和使用最先进的算法进行交叉验证，确保了数据的质量和准确性。ESA-AD 的发布不仅为卫星遥测分析和一般时间序列异常检测领域提供了一个新的标准，还为研究人员和实践者提供了一个可直接应用于实际空间操作环境的验证平台。

当前挑战

ESA-AD 数据集面临的主要挑战包括：1) 卫星遥测数据的高维度、复杂性和多样性，例如不同采样频率、数据缺失、组件退化趋势以及操作模式的变化；2) 构建过程中遇到的挑战，如数据注释的复杂性和手动检测异常的高成本。此外，现有的时间序列异常检测数据集和基准存在诸多缺陷，无法对新兴的机器学习技术进行公正的评估。ESA-AD 通过引入新的评估管道和指标，旨在解决这些挑战，并为卫星遥测分析和一般时间序列异常检测领域建立新的标准。

常用场景

经典使用场景

ESA Anomalies Dataset (ESA-AD) 是一个用于卫星遥测数据异常检测的基准数据集，主要用于评估多元时间序列异常检测算法。该数据集包含了来自三个不同ESA任务的真实遥测数据，其中两个任务的数据用于基准测试。ESA-AD 提供了大规模、结构化的数据集，并由卫星操作工程师（SOEs）和机器学习专家手动标注，并通过最先进的算法进行交叉验证。

衍生相关工作

ESA-AD 的发布催生了许多相关的经典工作。例如，基于该数据集的基准测试结果为研究人员提供了改进现有算法的方向，推动了深度学习、变分自编码器（VAE）等技术在卫星遥测异常检测中的应用。此外，ESA-AD 还激发了对时间序列异常检测评估方法的进一步研究，提出了新的评估指标和层次化评估流程，以更好地模拟实际操作场景。

数据集最近研究

最新研究方向

近年来，ESA Anomalies Dataset (ESA-AD) 在卫星遥测异常检测领域引起了广泛关注。该数据集通过提供来自三个不同ESA任务的真实遥测数据，填补了多元时间序列异常检测领域中缺乏可理解基准的空白。最新的研究方向主要集中在开发和验证能够应对卫星遥测数据复杂性的算法，特别是针对高维度、多通道依赖、采样频率变化以及噪声和测量误差等问题。ESA-AD 的公开性和可复现性为研究人员提供了一个标准化的评估平台，推动了机器学习技术在卫星遥测分析中的应用。未来的研究将进一步探索如何利用该数据集改进现有算法，特别是在实时监测和自主卫星健康监测方面的应用。

相关研究论文

• European Space Agency Benchmark for Anomaly Detection in Satellite Telemetry

  欧洲空间局 · 2024年