AssetOpsBench
收藏资源简介:
AssetOpsBench是一个为工业资产操作和维护中AI代理进行基准测试的框架和环境,旨在开发、协调和评估针对工业4.0应用的特定领域的代理。该数据集包含140多个由人类编写的自然语言查询,涵盖传感器查询映射、异常解释和故障诊断等任务,以及一个基于CouchDB的物联网遥测系统和多源数据集的模拟工业环境,用于端到端的多代理工作流程基准测试。数据集的应用领域包括异常检测、维护调度、故障诊断等,旨在解决工业应用中的数据模式多样性、业务对象复杂性和任务协作等挑战。
AssetOpsBench is a framework and environment for benchmarking AI agents in industrial asset operation and maintenance, aiming to develop, coordinate and evaluate domain-specific agents for Industry 4.0 applications. This dataset contains over 140 human-written natural language queries covering tasks such as sensor query mapping, anomaly explanation, fault diagnosis and more, as well as a simulated industrial environment based on a CouchDB-powered IoT telemetry system and multi-source datasets, which is used for end-to-end multi-agent workflow benchmarking. Its application fields include anomaly detection, maintenance scheduling, fault diagnosis and other related scenarios, and it is designed to address key challenges in industrial applications such as diverse data patterns, complex business objects and task collaboration.
AssetOpsBench 数据集概述
数据集简介
- 名称: AssetOpsBench
- 领域: 工业资产运维任务自动化
- 特点: 提供统一框架和环境,用于开发、编排和评估领域特定AI代理
数据集内容
- 场景数量: 140+
- 场景类型:
- 单领域任务 (如: "列出MAIN站点Chiller 6的所有传感器")
- 端到端多步骤任务 (如: "基于MAIN站点数据预测2020-04-27周Chiller 9 Condenser Water Flow的预测值")
- 场景领域:
- IoT数据检索 (IoT)
- 故障模式和传感器关系发现 (FSMR)
- 时间序列异常检测 (TSFM)
- 工单生成 (WO)
AI代理与多代理框架
- 领域特定代理:
- IoT代理: 提供
get_sites,get_history,get_assets,get_sensors等工具 - FMSR代理: 提供
get_sensors,get_failure_modes,get_failure_sensor_mapping工具 - TSFM代理: 提供
forecasting,timeseries_anomaly_detection等工具 - WO代理: 提供
generate_word_order工具
- IoT代理: 提供
- 多代理编排框架:
- MetaAgent: 基于reAct的单代理工具代理
- AgentHive: 计划执行顺序工作流
评估与结果
- 评估模型: 7种大型语言模型
- 评估方法: 使用LLM judge (Llama-4-Maverick-17B)在6维标准上评估
- 结果展示: 包含MetaAgent的评估结果
数据获取与使用
- 场景数据: https://github.com/IBM/AssetOpsBench/tree/main/scenarios
- 测试数据集: src/assetopsbench/sample_data/chiller6_june2020_sensordata_couchdb.json
- 数据库要求: 需安装CouchDB (https://couchdb.apache.org/#download)
运行说明
- MetaAgent运行脚本: https://github.com/IBM/AssetOpsBench/tree/main/src/meta_agent/tests
- AgentHive运行脚本: https://github.com/IBM/AssetOpsBench/tree/main/src/agent_hive/tests

- 1AssetOpsBench: Benchmarking AI Agents for Task Automation in Industrial Asset Operations and MaintenanceIBM Research - Yorktown & Ireland · 2025年



