Articles
13 minutes
Copy Link
How Industrial Edge Data Feeds AI Root Cause Analysis in Legacy Factories
TL;DR
A legacy plant working from spreadsheets does not need a multi-year ERP replacement to begin AI-assisted root cause analysis. You can start with one signal and keep your existing systems.
ISA-95 provides a way to organize equipment and the information shared between plant and business systems. An approved edge connection supplies the motor-current readings.
Contextualization links those readings to a conveyor, work order, and production interval. Humble Ops sits downstream, using that context and operator knowledge to support a reviewed root cause decision.
In the illustrative walkthrough, a motor-current excursion occurs near package rejects. Investigators treat the association as a hypothesis, approve a corrective action, and compare later runs with a baseline. A change in results alone does not confirm the cause.
What ISA-95, industrial edge, and contextualization each actually do
ISA-95 organizes how manufacturing operations and enterprise systems describe equipment and exchange information. Its equipment hierarchy gives you a way to identify where production happened, while its enterprise and control interface helps define information such as orders and production activity. ISA-95 does not collect PLC readings or prescribe one way to connect your systems.
Industrial edge devices and OT/IT connections can provide access to plant readings. An edge device or gateway can read a signal through a plant-approved path and pass it to other applications, subject to the access and network boundaries your plant sets. Connectivity alone cannot tell an investigator which order was running when the reading occurred.
A motor-current reading becomes useful for RCA when it identifies the conveyor and the order running at the time. The plant or its data provider must join the signal to those records. An ontology can formalize the relationships across systems, but the plant still has to maintain the mappings. MES records may identify the production run, and ERP records may add business order details.
Humble Ops fits after that contextualized data becomes available. As a downstream decision-intelligence layer, it can bring factory records and frontline knowledge into an AI-assisted root cause investigation, where people review evidence and approve actions. Humble Ops does not replace the edge platform, connector, ontology work, MES, or ERP. The workflow described here does not assume a certified direct integration with any particular edge vendor.
How a conveyor motor-current signal supports root cause analysis
Local capture, timestamping, and data quality checks
Suppose a packaging conveyor begins drawing more current while the line records more rejected packages. A legacy PLC reports the motor reading, and a gateway collects it through an interface the plant has approved. The collector logs amperes and identifies whether the PLC, an intermediate interface, or the gateway supplied the time and quality status. If only the gateway supplies them, its record cannot establish the exact moment the motor current changed.
When the approved interface uses OPC UA, a DataValue can carry a value, quality status, source timestamp, and server timestamp. The source timestamp reflects when the data source stamped the value, while the server timestamp reflects when the server handled it. If the gateway stamps a PLC reading, you should label that time as gateway-generated rather than assume the PLC recorded the measurement time. Record which clocks supply each timestamp and whether they are synchronized before comparing motor readings with package reject events.
A missing sample could hide the start of the excursion, and a stale value could make it look as though the conveyor ran normally. Compare the expected sampling interval with the readings received, then mark gaps and uncertain values. Otherwise, even the order of the motor event and the rejects may be wrong.
Adding asset and order context
The hypothetical conveyor motor current reading needs an asset and production interval before you can compare it with package rejects. Map the PLC tag to the conveyor’s asset ID, then locate that conveyor within the equipment hierarchy described by ISA’s enterprise and control integration standard. Use the reading’s recorded time to find the work order running on that line, rather than the order displayed when someone reviews the data.
If the plant uses an MES, its execution records can identify the operation, active order, and product at that time. ERP records may add business order details, but an ERP order alone does not establish which conveyor handled a package. Record who maintains the tag-to-asset mapping and, where mappings are versioned, which version applied when the reading was captured. If a tag was remapped or a changeover fell near the reading, check the join before treating the current spike and reject events as part of the same production run.
Bringing in operator knowledge and shift context
An operator may remember increasing conveyor speed during the changeover. Compare the note's time with the work order and controller history before blaming motor wear. The recollection narrows the next check, but it cannot settle the cause on its own.
Keep the operator’s observation alongside the signal, reject events, and work order in the investigation record. The Resource Center guide to capturing frontline knowledge for decision intelligence provides further context. Record when the operator observed the change, which shift it affected, and whether the note was entered at the time or recalled later. You can then compare the account with equipment records and ask the next shift whether the same behavior recurred.
AI-assisted investigation and evidence-linked recommendations
Once the plant has matched the signal to the right order and collected the operator's account, Humble Ops can help sort the competing explanations. The plant must first make those records available through an approved path. A reviewer needs the source readings and notes behind each suggestion, not a confident-sounding verdict.
Gradually rising current across comparable runs could justify checking motor wear, while a recorded drive fault before the current change could justify inspecting the drive. Check whether the suspected tension sensor actually supplies the load reading before treating sensor drift as an explanation. If rejects appeared on only one order, compare its recipe and speed settings with unaffected orders. The investigation should also surface contrary evidence, such as normal current on earlier runs at the same speed.
The strongest recommendation may be a test, not a repair. Humble Ops could point maintenance to the drive fault log or a relevant sensor check, with the event times, order settings, and operator note attached. Maintenance and quality staff can then test what the records suggest before approving a change. The current spike and rejects occurring together do not settle which fault caused either symptom.
Human approval and corrective-action validation
Suppose the drive fault log warrants an inspection. The maintenance lead owns that check, while plant policy determines who approves any repair that changes line operation or product disposition. The inspection might support a drive-fault hypothesis, but the current reading and reject count cannot establish it by themselves.
Before work starts, capture who approved it and which competing explanations remain open. Maintenance logs the action time, changed settings or parts, affected work orders, and person who did the work in the plant's existing records. Without that timestamp, a later reviewer cannot reliably separate pre-repair readings from post-repair ones.
Before the work begins, the plant should define a baseline for motor current and reject events using comparable runs, including line speed and package type. After the action, quality and production should observe a specified window of comparable production and check whether the symptom returns on later runs. They should also check whether another explanation, such as a relevant sensor fault, still fits the evidence. A lower reject rate after a repair supports the hypothesis, but confirmation requires a suitable test that connects the repaired fault to the original symptom and rules out plausible alternatives. If the evidence remains mixed, the record should retain the cause as unconfirmed.
Where data access and security boundaries belong in this workflow
A plant should approve how the conveyor motor current reading leaves the control network before an RCA tool receives it. The collection path might use an existing SCADA interface or a gateway, depending on the plant’s equipment and network design. Give the collector read access only to the required signals where the equipment supports it. Keep credentials controlled, log access, and separate network areas according to the plant’s security policy. NIST SP 800-82 Rev. 3 frames OT security around operational, reliability, and safety needs rather than one prescribed connection pattern.
The RCA layer needs approved access to the recorded signal and relevant order and operator context, not authority to change the conveyor. A recommendation should go to the responsible person for review and approval through the plant’s established process. Local PLC and safety logic retain control authority regardless of what the AI recommends.
Evaluation table: what to verify at each stage
Use this table to check whether your RCA pipeline can connect a production signal to a reviewed action and a measured outcome.
Dimension | What to verify | Failure or acceptance condition |
|---|---|---|
Connectivity | Identify the collection owner, approved read path, access permissions, and gaps in captured readings. | Accept when the plant can account for how readings arrive and when capture fails. |
Data quality | Check timestamp origin, clock synchronization, units, quality status, and sampling interval. | Reject or label stale and uncertain readings before ordering events. |
Context | Match the signal to the asset, active order, product, and production time using documented mapping versions. | Reject joins that cross an unverified shift or changeover boundary. |
Traceability | Link each proposed explanation to readings, operator observations, and counterevidence. | Accept when a reviewer can inspect the evidence behind the recommendation. |
Action ownership | Record who approved the action, who carried it out, and when production records changed. | Reject an unowned recommendation or an action without an approval record. |
Measured outcome | Set a comparable baseline, observation window, and recurrence check before the action. | Report the observed change separately from any causal conclusion; confirm a cause only when a suitable test rules out plausible alternatives. |
Implementation checklist for legacy plants
Choose one signal and one symptom. For the illustrative conveyor case, track motor current alongside package rejects on a single line.
Assign an owner for collection. Document the plant-approved read path, sampling interval, and how the collector handles missed or repeated readings.
Record the current reading’s units, quality status, and timestamp source. Check clock synchronization before comparing it with reject events.
Join the signal to the conveyor asset ID and active work order. Name the owner of each mapping and record its version.
Add operator observations, changeover notes, and evidence that might contradict the initial explanation.
Set read permissions, network boundaries, credential ownership, and access logging with the plant’s OT staff.
Run the RCA pilot in advisory mode. Assign an owner to review each proposal, obtain the approvals required by plant policy, and record who carried out the action and when.
Agree on a pre-action baseline, a comparable observation window after the action, and criteria for checking whether the symptom recurs.
Where Humble Ops fits and where it stops
Humble Ops sits downstream of factory data infrastructure. Its role in this workflow is to use available production context and frontline knowledge to support an investigation. Reviewers should be able to trace proposed actions to evidence and route decisions through the plant’s existing approval process. It does not replace an industrial edge platform, connector, ontology product, MES, or ERP.
For the upstream choices, see the industrial edge platform comparison for collection options and the edge and ontology architecture guide for operational context. Neither resource implies that Humble Ops has a certified direct integration with any specific edge vendor.
Book a Call with Humble
If you are working with legacy PLCs, SCADA, MES, and ERP systems, book a call with Humble to discuss one root cause investigation on your floor. Bring the signal you want to examine, the context you can already access, and the people who would approve a corrective action. The conversation can establish where Humble’s decision support might fit without assuming a particular integration path.
See If Humble Fits Your Floor
Take Humble’s fit test if you want to assess whether your plant is ready for a focused pilot. Use one recurring production problem to assess whether enough signal history, order context, and operator input are available for a focused investigation.
Subscribe for more manufacturing data and AI guidance
Get practical guidance on connecting factory data to operational decisions. Subscribe to the Humble newsletter for new manufacturing data and AI articles.
FAQ
Does ISA-95 mandate a specific integration pattern?
ISA-95 describes how enterprise and control functions relate and what information they exchange. It does not require a particular gateway, network path, or software vendor.
What is the difference between edge connectivity and contextualization?
Edge connectivity collects readings from plant equipment and makes them available through an approved data path. Contextualization links those readings to the relevant asset and production interval, plus an order or recipe when those records are available so you can investigate what was happening when a reading changed.
Can AI alone confirm root cause?
AI can compare evidence and propose explanations, but a time correlation cannot establish cause. Your operations and maintenance staff must check competing explanations, test the suspected fault, and assess results after an approved action.
Does Humble Ops replace MES, ERP, or edge platforms?
Humble Ops supports root cause decisions using contextualized factory data and frontline knowledge. Your MES and ERP retain their recordkeeping roles, while existing edge infrastructure handles equipment connectivity.
What access does an RCA layer need to plant systems?
An RCA layer needs approved access to relevant recorded signals, production records, and operator observations. It does not need authority to write to the PLC. Under plant-specific access controls consistent with NIST’s OT security guidance, you can keep its investigation access separate from PLC control and route corrective actions through human approval.