The full power of TDengine, now free forever for up to 5,000 tags.

Explore More

How to Connect PI System, OPC UA, and MQTT Data for Industrial AI

Juno Qiu

November 4, 2026 /

Connecting PI System, OPC UA, and MQTT data for industrial AI starts with a simple architecture question: which source supplies each piece of evidence, and how will consumers understand it? PI System may hold established operational history, an OPC UA server may expose equipment information, and MQTT may carry telemetry published through a broker. They can coexist in one data architecture without performing identical roles.

The objective is not to connect every available signal. It is to give an analyst or AI workflow the right observations, history, and context for a specific industrial question while respecting plant network and operational boundaries.

Start with the question and source ownership

Suppose a reliability team wants to investigate recurring production-line interruptions. Identify the equipment, the decision to be supported, the required time period, and the people who own each data source. A useful source inventory records the measurement, asset, unit, source system, update behavior, and any known limitations.

Agree which system is authoritative for each signal. The same pressure value might appear in both PI history and a current OPC UA feed. Without a clear rule, downstream analysis may count the same observation twice or compare records with different timestamps as though they were separate measurements.

Understand the role of each source

PI System: established operational history

An existing PI System may contain historical process values and associated asset information. It can remain a source of record while selected data is made available to another analytics environment. Determine which tags and time ranges are needed, how the source exposes them, and whether the permitted access method supports both historical retrieval and ongoing updates.

History is more than a file export. Check what the source actually retains, including sampling or compression behavior, data quality, and later corrections. Do not assume that copying time-series values also copies every dashboard, calculation, event, or asset relationship.

OPC UA: access to industrial source information

An OPC UA server can expose measurements and metadata through an address space. A client may browse and read or subscribe to selected nodes, depending on server and client support. Record the server identity, node mapping, source timestamps, status, and available engineering metadata.

Not every server exposes the same information or supports the same services. The plant team should confirm what is available and what collection load is acceptable. OPC UA is a family of specifications, so a particular client/server implementation should not be treated as the only possible OPC UA architecture.

MQTT: distribute messages through a broker

MQTT publishers send messages to topics, and subscribers receive matching messages through a broker. The message payload and topic convention must be agreed by the applications involved. A topic name alone does not establish an asset model, unit, timestamp meaning, or data-quality policy.

Confirm publisher identity, authorization, message format, expected frequency, and the way consumers handle repeated or delayed messages. MQTT delivery settings apply to particular protocol exchanges; they do not by themselves prove that a historian or downstream AI dataset has complete history.

Design the paths and preserve meaning

A possible architecture collects selected PI history, reads approved OPC UA nodes, and subscribes to selected MQTT topics. A data platform or integration layer then maps the records to assets and makes the resulting data available to authorized analytics consumers. Depending on the deployment, these functions can be distributed across several components.

TDengine provides possible data ingestion paths for this architecture. Its taosX data connectors include PI System, OPC UA, and MQTT sources, while TDengine Cloud documents its own Data In options. Available connectors, source versions, historical data behavior, and deployment requirements need to be evaluated for the intended environment. For a product-specific coexistence example, see Integrate PI System Deployments with TDengine.

These connections can provide access to selected data, but they do not automatically establish a shared asset model, an OPC UA-to-MQTT bridge, or a complete AI workflow. Connecting PI data does not imply automatic PI Vision migration or complete PI feature parity.

At each boundary, document the original source, asset identifier, measurement name, units, timestamp meaning, and quality status. Map different names to a common asset only after the asset owner confirms that they refer to the same physical object. Industrial data contextualization explains why these relationships matter when signals come from different systems.

Keep raw source identity available alongside any standardized name. If a tag or MQTT payload changes, an engineer should be able to trace a downstream value back to the source and understand which mapping produced it. A common model needs an owner and a change process; creating a hierarchy once is not enough.

Plan for gaps, duplicates, and security

Design the integration around the actual network zones and approved paths. Limit access to necessary sources and consumers, authenticate connections, and agree who manages certificates or credentials. OT and security teams should review the design against the plant’s safety and availability requirements before any connection is made.

Then test interruptions. A collector may buffer data that it has already received, but buffering cannot recover measurements the source never captured. Determine whether the source supports historical retrieval, how a gap is detected, what happens when a connection returns, and how duplicates or late records are reconciled. No single protocol setting guarantees end-to-end completeness.

Example: one line, three data paths

A plant wants engineers to investigate a packaging line’s recurring stops. PI holds established process history; an OPC UA server exposes machine states and selected counters; an edge device publishes vibration measurements through MQTT. The integration team selects one machine and a small set of measurements for a pilot.

They map all three sources to the same equipment identity, record each timestamp and unit, and confirm whether the machine state already appears in PI. An engineer compares one known stop with source records and checks that the combined view preserves the sequence of events. If an MQTT gap occurs, the team records what was captured and whether recovery is actually possible.

The combined evidence may help an investigation, but it does not automatically identify a root cause or make an AI recommendation correct. An engineer still needs to assess the quality and operating context of the evidence.

Validate before scaling

Before expanding beyond the pilot, the team should be able to trace and check the sources expected to provide evidence relevant to a known production-line stop. The pilot should also establish whether source load, historical coverage, access permissions, and recovery behavior meet the site’s operational requirements. If engineers cannot determine which system supplied a measurement, whether a record was delayed, or where a gap occurred, adding more equipment will increase the integration scope without resolving the underlying data-quality problem. Assign owners for monitoring and mapping changes before scaling.

Related Articles