Real-Time Intelligence with IBM Time Series Models on Confluent

基于 Confluent 的 IBM 时间序列模型实现实时智能

Foundation models transformed how enterprises unlock value from unstructured data. The bigger prize is streaming data, where the mission-critical decisions live: how much to order, which payment to stop, when the pump will fail, how hard to run the line, what happened the last time it looked like this. IBM and Confluent are now bringing that unlock stream-native, and the models are live in Early Access on Confluent Cloud, running where the data already moves, with Confluent Platform next.

基础模型改变了企业从非结构化数据中挖掘价值的方式。而更大的价值在于流数据，这里承载着关键任务决策：订购多少货物、拦截哪笔支付、泵何时会发生故障、生产线运行强度如何、以及上次出现类似情况时发生了什么。IBM 和 Confluent 现在正将这种能力以“流原生”（stream-native）的方式实现。这些模型目前已在 Confluent Cloud 上开启早期访问（Early Access），直接在数据流动的源头运行，Confluent Platform 的支持也将紧随其后。

Until now, those decisions have run on outdated economics: one bespoke model at a time and months of expert work on each. So teams model the few hundred series where the money is and cover the rest with safety margins, extra inventory, extra headroom, extra tolerance, acted on after the window has closed. That margin is the cost of a decision nobody could forecast, paid every cycle.

在此之前，这些决策依赖于过时的经济模式：一次只能处理一个定制模型，且每个模型都需要专家数月的开发工作。因此，团队只能对少数几个涉及核心利益的序列进行建模，其余部分则通过安全边际、额外库存、额外空间和额外容差来覆盖，而这些往往是在机会窗口关闭后才采取的行动。这种边际成本就是因无法预测而付出的代价，且每个周期都在重复支付。

A time series foundation model (TSFM) changes that. Trained once across vast, varied signals, it generalizes to a series it has never seen: give it a window of measurements and it tells you what comes next, how far behaviour sits from normal, which history looks like this one, and which settings best serve a target. Using one does not take an army of data scientists either: a demand planner, a fraud analyst or a process engineer can put these models to work on their own streams.

时间序列基础模型（TSFM）改变了这一现状。它通过在海量、多样的信号上进行一次性训练，能够泛化到从未见过的序列：只需提供一个测量窗口，它就能告诉你接下来会发生什么、行为偏离正常的程度、哪段历史记录与当前相似，以及哪些设置最能达成目标。使用它也不需要庞大的数据科学家团队：需求规划师、欺诈分析师或流程工程师都可以直接在自己的数据流上应用这些模型。

Around the models, IBM is building functions that shift the work left, so forecasting, anomaly detection, optimization and semantic intelligence arrive as capabilities you call rather than projects you build. Picture one tempering line in a chocolate factory, its temperature, speed and throughput sampled every few seconds and watched against fixed thresholds. Drop a foundation model into that stream and it forecasts the line’s output through the evening shift, so the planner sees a shortfall while there is still time to act.

围绕这些模型，IBM 正在构建“左移”（shift-left）功能，使预测、异常检测、优化和语义智能成为你可以直接调用的能力，而不是需要从头构建的项目。想象一下巧克力工厂的一条调温生产线，其温度、速度和吞吐量每隔几秒采样一次，并与固定阈值进行比对。将基础模型植入该数据流中，它就能预测晚班的生产输出，让规划人员在还有时间采取行动时就能发现产量不足。

It scores today’s run against how the line normally behaves on dark chocolate, so a slow drift surfaces before a bar blooms. It finds the closest match in plant history, so the engineer knows how the last runs like it turned out. It conditions on the settings the crew controls, and fine-tunes when the last points of accuracy are worth it. No data science team required, and the same model rolls to every line in every factory.

它会将当天的运行情况与生产黑巧克力时的正常表现进行比对，从而在产品出现瑕疵前发现细微的偏差。它能找到工厂历史记录中最接近的匹配项，让工程师了解上次类似运行的结果。它还能根据员工控制的设置进行条件分析，并在需要提升最后一点精度时进行微调。无需数据科学团队，同一个模型即可推广到每一家工厂的每一条生产线。

IBM ran these models before offering them, in its own products and operations first, then with design partners in cement, steel, pulp and paper, food and telecommunications. The numbers make the case: every point of accuracy is worth millions, productivity gains run 5 to 10×, and work that waited for specialists now sits with the domain experts who own the decision.

IBM 在推出这些模型之前，先在自己的产品和运营中进行了测试，随后与水泥、钢铁、纸浆造纸、食品和电信行业的合作伙伴共同设计。数据证明了一切：每一个精度点的提升都价值数百万美元，生产力提升了 5 到 10 倍，而原本需要等待专家处理的工作，现在直接由掌握决策权的领域专家完成。

Now that proof meets real-time context: IBM brings frontier models that understand how signals behave, 44M+ downloads behind them, and Confluent brings the live state of the business and reach to every system that acts. Together they run stream-native, hosted in Confluent Cloud and called from Flink. Access opens on Confluent Cloud on AWS. Confluent Platform follows, bringing the same models and capabilities to on-premises and hybrid environments.

现在，这种验证与实时上下文相结合：IBM 带来了能够理解信号行为的前沿模型（拥有超过 4400 万次下载量），而 Confluent 则带来了业务的实时状态，并将其触达至每一个执行系统。它们共同以“流原生”方式运行，托管在 Confluent Cloud 中，并通过 Flink 调用。目前已在 AWS 上的 Confluent Cloud 开放访问。Confluent Platform 也将紧随其后，将相同的模型和能力带到本地和混合云环境中。

Time series intelligence meets real-time context with zero configuration, built-in governance and efficiency. The months usually spent wiring a model into production are months you keep: Granite reads the signal, Confluent supplies the context, the governance and the delivery to everything downstream. A signal’s value decays with time: a pump caught drifting today is a work order, the same pump next week is an outage.

时间序列智能与实时上下文相结合，实现了零配置、内置治理和高效运行。通常用于将模型部署到生产环境的数月时间现在被节省了下来：Granite 读取信号，Confluent 提供上下文、治理以及向下游交付的能力。信号的价值随时间衰减：今天发现泵出现偏差是一个工单，而下周再发现同一个泵的问题，可能就是一次停机事故。

Forecasting and detection are stateful: the next value only means something against recent history, and an anomaly only exists against a running sense of normal. Flink manages that state, keyed per series and fault tolerant, so each model gets the history it needs without a separate data store or a database hit per call. This is where the value compounds.

预测和检测是有状态的：下一个数值只有在参考近期历史记录时才有意义，异常也只有在与持续运行的“正常状态”对比时才存在。Flink 管理着这种状态，按序列进行键控且具备容错能力，因此每个模型都能获得所需的历史数据，而无需单独的数据存储或每次调用都访问数据库。这就是价值倍增的地方。

Confluent’s data streaming platform puts business data in motion and makes it usable for ML. The platform continuously streams, connects, governs, and processes real-time data, capturing live business signals that IBM Granite Time Series models use for forecasting, anomaly detection, similarity search, classification, gap-filling and optimization.

Confluent 的数据流平台让业务数据“动”起来，并使其可用于机器学习。该平台持续流式传输、连接、治理和处理实时数据，捕获 IBM Granite 时间序列模型所需的实时业务信号，用于预测、异常检测、相似性搜索、分类、缺口填充和优化。

Confluent provides what you need to implement streaming use cases quickly, reliably, and securely, so you can focus on developing real-time ML applications rather than managing data infrastructure. Confluent Cloud, the cloud deployment of Confluent’s data streaming platform, provides native inference, which allows you to run IBM Granite Time Series models directly within Apache Flink® on Confluent, providing greater flexibility, security, and cost efficiency for real-time data processing while unifying data and ML workflows.

Confluent 提供了快速、可靠且安全地实现流式用例所需的一切，让你能够专注于开发实时机器学习应用，而不是管理数据基础设施。Confluent Cloud 作为 Confluent 数据流平台的云部署版本，提供了原生推理功能，允许你在 Confluent 上的 Apache Flink® 中直接运行 IBM Granite 时间序列模型，在统一数据和机器学习工作流的同时，为实时数据处理提供更高的灵活性、安全性和成本效益。

The benefits include: 优势包括：

• Real-time intelligence where the data lives: Run forecasting and anomaly detection directly on streaming data, at the very moment business conditions change, without extracting time-series data into a separate ML platform or data warehouse.

• 数据驻留地的实时智能： 直接在流数据上运行预测和异常检测，在业务条件变化的瞬间即可获得洞察，无需将时间序列数据提取到单独的机器学习平台或数据仓库中。

• Zero configuration: Confluent manages model serving, infrastructure, scaling, and runtime operations, so there is no provider credential to manage or glue between data pipelines and the model. Call IBM Granite Time Series models directly from Flink SQL for real-time anomaly detection and forecasting.

• 零配置： Confluent 管理模型服务、基础设施、扩展和运行时操作，无需管理提供商凭据，也无需在数据管道和模型之间进行繁琐的对接。直接通过 Flink SQL 调用 IBM Granite 时间序列模型，即可实现实时异常检测和预测。

• Fresh, enriched context: Confluent continuously captures and processes data into an up-to-date view of the current state of the business.

• 新鲜且丰富的上下文： Confluent 持续捕获并处理数据，形成对当前业务状态的最新视图。