Systems and records
Energy maintenance work order dataset: linking alarm, repair and outcome
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
Connect alarms, maintenance decisions and repair evidence by making the maintenance episode the unit, not the work order. Each episode joins one alarm occurrence and its acknowledgment history, the asset and reading revisions the operator saw, the decision and who made it, the isolation and maintenance-program rules in force that day, the repair evidence, and a dated outcome check showing whether the condition returned.
Record timeline
Energy maintenance work
Key takeaways
- A closed work order records that a job happened; it does not record the alarm that started it, the reason it was chosen, or whether the repair worked.
- Build one record per maintenance episode and hang the alarm event, the decision, the isolation reference, the work order and the outcome check off it.
- Store the reading the operator saw and any later revision separately, so a model learns from the information that was available at decision time.
- Record which rule version and which isolation procedure applied on the decision date, because they change what the permissible next step was.
- An episode passes only when its outcome comes from a later reading or alarm history, not from the work order status field.
- Strip detailed design, protection settings and network details before any linked history leaves the plant, and have a qualified reviewer confirm it.
Why can't a closed work order explain the decision behind it?
A work order tells you that someone was sent to an asset and closed a job. On its own it cannot tell a model which alarm prompted the visit, what the operator knew at the time, why a corrective job was chosen over waiting for the next outage, or whether the fault came back. Those facts live in other systems, and a dataset that leaves them behind teaches pattern matching on job descriptions rather than maintenance judgment.
The alarm side is already structured. In the OPC UA information model published by the OPC Foundation, every event carries an EventId, a SourceNode, a Time and ReceiveTime, a Message and a Severity, and alarm conditions add states such as active, acknowledged, confirmed, suppressed and shelved. The ISA-18.2 alarm management standard, adapted internationally as IEC 62682, treats alarms as a lifecycle with rationalization, monitoring and management of change, which means an alarm's meaning can change over time and the dataset has to record which definition was in force.
The maintenance side is usually a CMMS or EAM such as IBM Maximo, SAP Plant Maintenance or Infor EAM, and the readings sit in a historian. The gap is the join: almost no site stores an explicit link from the alarm occurrence to the work order it caused and on to a later check that the repair held.
What does one maintenance episode have to contain?
An episode is the stretch from the first alarm activation that started a maintenance response to the end of a fixed outcome window after the asset returned to service. It is the right unit because it holds a starting state, a decision, an action and a result, which is the shape of a task an AI system would be asked to perform or assess.
At a glance
What does one maintenance episode have to contain
ISO 14224, the reliability and maintenance data standard for the petroleum, petrochemical and natural gas industries, separates equipment data, failure data and maintenance data and gives normative failure modes. Even outside oil and gas, borrowing that split keeps an episode honest: the asset facts, the failure as observed and the work performed are recorded as different things rather than blended into one free-text field.
Set the outcome window per asset class before linking starts, not episode by episode. A cooling fan that fails again usually shows it within days, while a breaker mechanism or a pump seal may need weeks of operation before a repeat is meaningful. Writing the window and the expected reading range into the episode before the window opens stops anyone from choosing a window after the fact that makes the repair look successful.
- Starting state: the alarm occurrence, its acknowledgment history and the readings visible on the operator screen at that moment.
- Decision: what was chosen (monitor, adjust, corrective work order, planned outage), by whom in role terms, and the stated reason.
- Boundary: the isolation procedure, maintenance-program interval and any operating restriction that applied on the decision date.
- Action: the work order, its tasks, parts used, test results and the failure mode and cause codes entered at close.
- Outcome: the readings and alarm history for a fixed window after return to service, compared with an expected range written down before the window opened.
Which fields link the alarm, the work order and the outcome?
The table below is the linkage example for one episode. Field names are suggestions; the source and the validation rule are what make the record trustworthy. Every identifier should be a stable pseudonymous key whose mapping stays with the operator, so links survive later de-identification.
Readings need special care. Historians and meter systems revise values after calibration checks, gap filling or manual correction. Keep the value as first recorded and each later revision as a separate entry that points back to it, in the way the W3C PROV-O vocabulary describes one entity as a revision of an original. The decision is then judged against what the operator could actually see.
The OPC UA ClientUserId and Comment fields can hold a person's login and free text. Map the user to a role code and pass comments through the same review as work order notes.
| Field | Source system | What it records | Validation rule |
|---|---|---|---|
| episode_id and asset_ref | Linkage table and asset register | One maintenance episode on one asset, with its class and location level | Both required; asset_ref must resolve in the register on the episode start date |
| alarm_ref | Alarm and event journal | Condition name, EventId, active, acknowledged and shelved timestamps, severity | Activation time must precede the decision time |
| alarm_definition_version | Alarm rationalization database | The setpoint, priority and response in force for that alarm on the day | Version effective date on or before activation |
| reading_snapshot | Historian | Tag values in a fixed window before the decision, as first recorded | Revisions stored separately with a revision_of pointer |
| decision | Shift log, control room log or work request | Chosen action, decision role and reason from a closed list plus short text | Required; reason code must be in the list version used that year |
| boundary_refs | Isolation records and maintenance program | Isolation procedure ID, program interval and any operating restriction | Required when the action involved hands-on work |
| work_order_ref | CMMS or EAM | Work order, tasks, parts, labor hours, test results, close codes | Created after the decision; closed after the last task |
| failure_coding | CMMS close-out | Failure mode, cause and method of detection | Codes drawn from one named code list and version |
| outcome_check | Historian and alarm journal | Readings and repeat alarms for a fixed window after return to service | Window and expected range set before the window opens |
| confidentiality_class | Linkage review | Whether the episode passed design, security and personal data review | Must be cleared before the episode enters a dataset |
Which rules and safety boundaries change the permissible next step?
A model that learns only from what technicians did will miss why some faster actions were not allowed. Record the boundary, not the safety procedure itself: a pointer to the procedure and the dated proof that it applied is enough to explain the decision, while the procedure text stays inside the operator.
Two examples show why dates matter. OSHA's lockout and tagout standard, 29 CFR 1910.147, requires periodic inspection of energy control procedures and a certification that identifies the equipment, the inspection date, the employees included and the inspector. NERC's protection system maintenance standard, PRC-005, accepts dated maintenance records, check-off lists, inspection records or work orders as evidence in the PRC-005-6 text cited here; NERC has since developed a PRC-005-7 revision, so confirm which version applies to your registration. An episode linked to the right procedure and interval version can show that a repair waited for a valid isolation or a program interval, not for lack of attention.
| Boundary | Record that proves it applied | How it changes the permissible next step | Linkage field |
|---|---|---|---|
| Hazardous energy isolation | Isolation procedure ID and its latest periodic inspection certification | Hands-on repair cannot start until isolation is applied and verified | boundary_refs.isolation |
| Protection system maintenance program | Program document version and component interval | Sets when testing is due and what evidence closes it | boundary_refs.program |
| Alarm shelving or suppression | Shelved or suppressed state with start and end time | A shelved alarm may not have been seen, so delay is not a decision | alarm_ref.states |
| Operating restriction | Derate, switching order or outage approval | Some repairs need an approved outage window first | boundary_refs.restriction |
| Alarm definition change | Rationalization record and effective date | An alarm reprioritized last month carries a different expected response | alarm_definition_version |
Worked example (illustrative): a transformer cooling alarm from activation to a checked repair
The episode below is illustrative. It does not describe a real site, asset or dataset; the times are offsets from the first alarm and the counts are invented to show structure.
A substation transformer raises a high top-oil temperature alarm on a hot afternoon. The operator acknowledges it within minutes, sees that one cooling fan group shows no current, and logs a decision to request corrective work rather than reduce load, because the temperature is rising slowly and is still below the trip level. The work order is created, the crew applies the isolation procedure for the fan control cabinet, replaces a failed contactor, tests the fan group and returns it to service. The alarm clears within the hour.
The outcome check is what turns this into training data. The expected range and a 14-day window were written down at close. Readings over the window stay within range and the alarm does not repeat, so the episode is marked verified. A later revision of one temperature reading, made after a sensor calibration check, is stored as a revision and does not change the decision snapshot.
| Offset | Event | Source record | Linked field |
|---|---|---|---|
| T+0 min | High top-oil temperature alarm becomes active | Alarm journal event | alarm_ref |
| T+4 min | Alarm acknowledged; fan group current shown at zero | Alarm journal and historian snapshot | alarm_ref and reading_snapshot |
| T+11 min | Decision logged: corrective work, no load reduction, reason code rising slowly below trip | Control room log | decision |
| T+2 h | Isolation applied under the cabinet procedure; contactor replaced | Isolation record and work order tasks | boundary_refs and work_order_ref |
| T+3 h | Fan group tested, work order closed with failure mode and cause codes | CMMS close-out | failure_coding |
| T+3 h to T+14 days | Readings within the expected range; no repeat alarm | Historian and alarm journal | outcome_check |
What does a completed work order with no proven outcome look like?
Now take a second illustrative episode on a similar asset. The work order is closed as completed with a cause code of no fault found. The same alarm returns six days later, but the new occurrence opens a fresh work order with no link to the first. A later reading revision overwrote the original value in place, so nobody can say what the operator saw.
Steps
What does a completed work order with no proven outcome look like
Every field is filled, yet the record teaches the wrong lesson: it labels a missed fault as a successful repair. The validator below is what catches it, and the right response is to merge the two occurrences into one episode with a failed first outcome, or to exclude it if the original reading cannot be recovered.
- Order check: alarm activation precedes the decision, which precedes work order creation, which precedes close.
- Outcome source check: outcome_check cites historian or alarm records, never the work order status alone.
- Repeat check: any recurrence of the same condition on the same asset inside the window marks the outcome failed and links the new occurrence.
- Revision check: every reading in the decision snapshot is an original value; revisions exist only as linked entries.
- Boundary check: hands-on actions carry an isolation reference whose certification date precedes the work.
- Code check: failure mode and cause codes come from one named list and version for the whole dataset.
- Clearance check: confidentiality_class is set and approved before the episode is exported.
What must be stripped before linked maintenance history leaves the plant?
Linking makes records more useful and also more revealing, so the review happens after linkage, not before. FERC describes Critical Energy/Electric Infrastructure Information as specific engineering, vulnerability or detailed design information about critical infrastructure that could be useful to a person planning an attack. One-line diagrams, protection settings, network addresses and detailed equipment locations can fall into that territory, and a qualified reviewer should decide what an episode keeps.
In practice most episodes survive with coarse asset classes, voltage or capacity bands instead of exact ratings, pseudonymous site and asset keys, role codes in place of names, and free text that has been reviewed line by line. Utilities subject to NERC CIP requirements should also run the export past whoever owns their information protection program. Ownership and permission are separate questions: data the operator holds may still be subject to contracts with equipment vendors, contractors or co-owners, and counsel reviews those before any licensing discussion.
Which episodes should be prepared first, and which left out?
Start with asset classes where alarms, work orders and historian tags already share a stable asset key, such as transformers, breakers, pumps or compressors with a long history in one CMMS. Episodes there can be linked by rule, and a small double-reviewed sample shows how often the automatic join is wrong.
Make the join rules explicit and keep them with the dataset. A typical rule links an alarm occurrence to a work order when both name the same asset key, the work request cites the alarm or its tag, and the request was raised within a set time after activation. Where the request does not cite the alarm, a time-and-asset match alone is a candidate link that a reviewer confirms, and the episode records whether its link was cited, rule-matched or reviewer-confirmed. That one field lets a buyer filter to the most reliable episodes and lets the operator see which control rooms already log their reasoning well.
Leave out episodes whose original readings were overwritten, whose decision cannot be found in any log, or whose outcome window was cut short by a later unrelated outage. Mark them excluded with a reason code rather than deleting them, so the share of episodes that could not be linked is itself a reported property of the dataset. A buyer will trust a smaller set of verified episodes more than a large set where the outcome is the status field.
Report the counts that matter alongside the episodes: how many alarm occurrences were considered, how many became linked episodes, how many outcomes were verified, failed or left open, and how many were excluded and why. Those numbers describe the dataset more honestly than the row count of the work order table.
Bringing linked maintenance history to a licensing review
Energy operators considering a license can bring linked episodes to SourceX, whose role spans rights review, contracting, data preparation, delivery and payment. For an energy operator, the linkage table above is a practical way to show what the history contains without sending any of it.
Historian and CMMS data stay where they are until the operator has approved scope and terms and signed. Large historian extracts can stay in the operator's own storage or travel on encrypted drives rather than being hosted. Eligibility is settled first, by a short assessment that needs nothing from the historian.
Frequently asked questions
What should the work order, sensor and outcome linkage example include?
At minimum an episode ID, the asset reference, the alarm occurrence with its state times, the alarm definition version, a snapshot of readings before the decision, the decision and reason, isolation and program references, the work order and failure codes, and an outcome check from later readings. Each field needs a named source system and a validation rule.
How do you prove what the operator decided and whether the repair held?
The decision is proved by a dated entry in a control room log, shift log or work request that names the chosen action and reason. The outcome is proved by historian readings and the alarm journal over a window fixed in advance, not by the work order being marked complete.
Do readings that were corrected later still belong in the dataset?
Yes, but as revisions linked to the original value, not as replacements. The decision should be evaluated against the reading the operator saw, and the revision explains why the later record differs.
Do we need ISO 14224 failure codes to build episodes?
No. Any failure mode and cause list works if it is named, versioned and applied consistently across the dataset. ISO 14224 is a useful reference for separating equipment, failure and maintenance data, especially for operators in oil and gas.
Can alarm and work order data from a utility be shared at all?
Often yes, after review. Detailed design, protection settings, network details and anything a reviewer classes as critical infrastructure information should be removed or generalized, and contracts with vendors and co-owners need to be checked by counsel first.
Sources
- The OPC UA information model gives events an EventId, SourceNode, Time, ReceiveTime, Message and Severity, and gives alarm conditions active, acknowledged, confirmed, suppressed and shelving states. Source
- ISA-18.2 organizes alarm management as a lifecycle including rationalization, monitoring and management of change, and was adapted as IEC 62682. Source
- ISO 14224 separates equipment, failure and maintenance data for reliability and maintenance data collection and provides normative failure modes. Source
- NERC PRC-005-6 accepts dated maintenance records, check-off lists, inspection records or work orders as evidence of protection system maintenance. Source
- OSHA's lockout and tagout standard requires a periodic inspection certification naming the equipment, date, employees included and inspector. Source
- FERC defines Critical Energy/Electric Infrastructure Information as specific engineering, vulnerability or detailed design information about critical infrastructure. Source
- W3C PROV-O models a revision as a derivation in which the result is a revised version of an original entity. Source
See if your company qualifies
A short company assessment. No data uploads are needed.