MAS suite · Predict

Predict occupies rung five of six maintenance strategies

Four rungs have to hold before failure probability pays back: cheap failures left alone, calendar intervals reviewed against findings, usage read from the field, and condition signal bound to asset records. This page is the ladder, the readiness check and the gates.

Operational data streaming out of industrial plant, representing failure-probability analytics in IBM Maximo Predict

The ladder

Six maintenance strategies, and what each one demands

Most estates stand on several rungs at once, and should. The right strategy for a light bulb is not the right strategy for a safety system.

  1. Rung 6

    Risk and financially optimised

    Requires
    Criticality and consequence agreed beyond engineering, against a budget the plan is constrained by.
    Skip it and
    The optimiser issues a plan the capital process cannot execute, and the room decides by volume.
  2. Rung 5

    Predictive

    Predict sits here
    Requires
    Coded failure history, a real population, and a written rule for what a probability changes.
    Skip it and
    The model returns a number, the plan does not move, and the licence pays for a screen.
  3. Rung 4

    Condition based

    Requires
    Reliable operating signal bound to asset records, with a named triage owner per shift.
    Skip it and
    Thresholds widen until the alerts stop, which is how a monitoring programme quietly ends.
  4. Rung 3

    Usage based

    Requires
    Meter readings reach the asset record from the field rather than estimated at month end.
    Skip it and
    Intervals drift on estimates, and the hardest-worked unit runs the same clock as the standby.
  5. Rung 2

    Calendar based

    Requires
    The interval reflects duty and consequence, and somebody revisits it against findings.
    Skip it and
    Crew hours go into inspections that find nothing while failures happen outside the routine.
  6. Rung 1

    Run to failure

    Requires
    Failure is cheap, safe, and stops nothing downstream.
    Skip it and
    A cheap component takes a production line with it, and the review asks who was watching.

Run this before you buy anything

Five statements about one asset class

Answer for a single class rather than the estate, in order, and stop at the first one that is false.

  1. 1

    We can name what failed, not only what was replaced, on most work orders for this class.

  2. 2

    We run enough near-identical units for a pattern to be a pattern rather than an anecdote.

  3. 3

    Condition or meter data already reaches the asset record for this class.

  4. 4

    One named person is allowed to change a PM interval on evidence.

  5. 5

    An unplanned failure on this class costs more than the programme to predict it.

If you stopped early

Where you stopped is the engagement. Coding discipline, condition monitoring or a decision right comes before any model, and we scope it as that rather than hiding it inside a model build.

If all five held

The class qualifies. Bring five years of work orders and we read the failure coding rather than the business case.

Demonstrated, not described

Five tests your reliability team watches before the model is accepted

A model nobody witnessed being tested is a claim. These are run on your data, with your people in the room.

  1. A1

    It recognises failures your engineers already know about

    Passes when
    Historical failures on the class appear in the top decile of ranked output for the weeks before they happened.
    Witnessed by
    Reliability engineer and maintenance planner
  2. A2

    Every input is readable by the reliability function

    Passes when
    Each feature traces to a named Manage field, a meter or a Monitor tag, listed on one page.
    Witnessed by
    Your data lead and reliability engineer
  3. A3

    The decision rule actually moves a plan

    Passes when
    A probability above the agreed threshold changes a PM interval or inspection scope on a test record in Manage.
    Witnessed by
    The planner who owns the schedule
  4. A4

    Your team can retrain it without us

    Passes when
    A retrain on newly coded failures completes and is compared against the previous version.
    Witnessed by
    Your data or reliability team
  5. A5

    An engineer can override it and be recorded doing so

    Passes when
    An override is entered with its reason and lands against the asset record, not in a meeting note.
    Witnessed by
    Reliability lead and MaxIron delivery lead

The write boundary

Three gates between a probability and a changed maintenance plan

Gate three is written before anybody sees a probability, which is the difference between an input and an interesting number.

  1. Gate 1 Qualification

    Signed by Your reliability lead and MaxIron delivery lead

    Passes

    One asset class with coded history, a real population and a cost of failure that justifies the work.

    Held back

    The largest bespoke machine on site, and any class nominated because it is visible.

  2. Gate 2 Model acceptance

    Signed by Your reliability engineer

    Passes

    A model whose features the reliability function can read and whose output matches failures they recognise.

    Held back

    A held-out score nobody in the room can interpret or challenge.

  3. Gate 3 Plan change

    Signed by The planner who owns the schedule

    Passes

    A PM interval, inspection scope or priority changed in Manage, with the probability recorded as the reason.

    Held back

    Any automatic change to a live maintenance plan without a named approver.

The arithmetic

Predict changes which outage you choose. It does not reduce the work.

The strip, the parts lead time, the outage negotiation with production and the failure coding at close-out all cost what they cost. What stops is defending a PM interval with nothing but the interval's own history, and being handed a failure window by the plant.

Conditions and boundaries

Four conditions the estate has to supply

It will not learn from history nobody wrote down

Free-text long descriptions and repair-only records do not train a model. Where coding is thin, coding is the scope.

It will not work on a population of one

The largest bespoke machine on site is the most requested and the worst candidate. Monitor and inspection engineering serve it better.

We do not build proprietary failure models

The MAS Predict model lifecycle does the modelling. We add data engineering, integration and operations, and nothing shortens the data work.

It will not make the decision

A probability is an input. An engineer challenges it, a planner acts, Manage records what changed. Without that authority granted, the honest advice is to wait.

IBM Maximo Predict, frequently asked questions

What does IBM Maximo Predict do?
Predict applies analytics to coded failure history and condition signal to estimate the probability that an asset fails inside a given window. The output is an input to a maintenance decision that planners and reliability engineers already own.
What has to be true before Predict pays back?
Failure history rich enough to learn from, a population large enough to generalise across, a cost of unplanned failure that justifies the work, and a named person allowed to change a plan on the output. Why predictive maintenance programmes fail sets out what happens when one is missing.
Do we need Monitor in place first?
On most asset classes, yes. Condition signal beside historical failure coding is materially stronger than history alone, so we commonly sequence Monitor then Predict on the same class. The order is argued in sequencing Monitor and Predict after Manage.
What if our failure coding is poor?
Then coding is the first engagement, and it is a data and process job: agree the failure classes that matter for the class in scope, code the recent history against them, and change the close-out step so new work orders arrive coded.
Does MaxIron build proprietary failure models?
No. We use the model lifecycle that ships with MAS Predict and add data engineering, asset domain knowledge, integration into Manage and managed operations. There is no MaxIron accelerator that shortens the data work.
We are an upstream operator. Is there anything specific to us?
ISO 14224 failure coding, the rotating-equipment populations where Predict lands first and the integrity-management context are covered on Predict for upstream oil and gas.

Bring five years of work orders for one asset class.

We read the failure coding rather than the business case, name the rung that class stands on today, and say whether Predict is the next rung or whether condition-based work still has room in it.

Bring this to the first call

  • The asset class, its population, and how similar the units really are
  • A multi-year work order export with failure codes included
  • Whatever meter or condition data exists, and how it is captured
  • What an unplanned failure on that class costs in production or consequence
  • The name of the person allowed to change a PM interval