Predictive Maintenance Field Frameworks: The Evidence Chain
Twelve articles following one chain — asset, failure mode, signal, data, method, alert, governed work, assessed value — and the places a link quietly breaks without anyone noticing. Plus six free field cards. Each article ends in an exercise that costs an afternoon and no capital.
Predictive maintenance is usually described as a technology problem. It is better described as a chain of evidence, and a chain is only worth what its weakest link can carry.
The chain runs: asset → failure mode → signal → data → method → alert → governed work → assessed value. Every link is a claim someone has to be able to defend. The asset matters because of a consequence. The failure mode is specific enough to watch. The signal responds to that mode and not merely to something correlated with it. The data survives the join. The method earns belief in proportion to what it demands. The alert names an action. The work order gets raised and closed. And somebody, afterwards, checks whether any of it changed an outcome.
Break any one link and the rest still runs. That is the difficulty: a broken chain produces output. Dashboards populate, models train, alerts fire, and the programme quietly stops being evidence about machines and becomes evidence about itself.
This series is about where the links break, how to tell from inside the programme, and what it costs to repair one.
Each piece stands alone. Read in order, they follow the chain from the asset register to the value assessment. Every article ends with an exercise that costs an afternoon and no capital.
Where this comes from
The series draws on Predictive Maintenance: Practitioner Reference Frameworks and Planning Guide, a twelve-part reference running from maintenance outcomes and asset criticality, through failure modes, monitoring technologies and asset-class playbooks, into sensors, data foundations, signal processing and analytics, and out through alerts, work management, and pilot governance.
Each article cites the part it draws from. The book renders verdicts on monitoring evidence in three states — supported for this context, controlled trial, not demonstrated — and none of the three authorises anything on a plant. The articles keep that boundary. Where the underlying record supports a limit, they say so; where a number is an estimate from a named study of a named population, they say that too.
The series
- 01 The P–F Interval Is a Measurement Hypothesis
A P–F interval is not a property of the machine. It is a claim about one failure mode, one measurement contract and one response path, and it moves whenever any of them moves. What the evidence has to say before the number means anything.
- 02 The Trust Budget: Why Predictive Programmes Die at Month Three
Predictive programmes rarely die of bad algorithms. They die when the team's attention runs out — and attention is a budget every false alert spends. The six failure modes, the two invoices, and the ledger that shows which month the account went overdrawn.
- 03 MTBF: The Most Misread Number in Engineering
A drive datasheet quoting 1–2.5 million hours appears to promise 114 to 285 years of life. The reported fleet reality is an annualised failure rate of a percent or two. Both numbers are true, because MTBF is a population rate restated in hours — never a lifespan.
- 04 Predictive Maintenance Is a Data-Quality Problem First
The programme rarely fails at the model. It fails at the join, at the failure code nobody filled in, and at the reading stored without what the machine was doing at the time — none of which raises an error, because a broken data chain still produces output.
- 05 Bad Actors: The Vital Few Machines Eating the Maintenance Budget
Everything else in a predictive programme is an argument about the future. Bad actors are not — they have already confessed, in your own work-order history, with dates and costs attached. The only question is whether anyone has run the report, and read it properly.
- 06 From “Everything Is Important” to an Ordered List
Ask a plant which machines are critical and the honest answer is usually all of them. Criticality analysis is the method that turns that answer into a queue — by fixing the criteria, the anchors and the arithmetic before anybody starts scoring, and by naming what the ranking never gets to decide.
- 07 The Anatomy of an Alert Worth Acting On
An alert is a complete utterance or it is noise with a timestamp. The seven fields a message has to carry, the five-rung ladder that says what a finding has actually established, and the rung a monitoring programme is never allowed to climb.
- 08 Remaining Useful Life: A Window, Not a Date
A model returns a remaining-useful-life number. By the time it reaches the outage meeting it has become a date with a confidence percentage attached. Two transformations happened, and the data authorised neither of them.
- 09 Condition Evidence Doesn't Decide Spares
If the machine warns us weeks in advance, why is that seal sitting on a shelf earning nothing? It is the most seductive arithmetic in maintenance, and the bill arrives later and somewhere else. A condition record opens a spares review. It never decides the shelf.
- 10 "Read-Only" Is Not a Security Architecture
A path that cannot write reduces a defined class of write-path risk, and that reduction is real. It still establishes no secure architecture — because "no inbound session" is a property of a connection, not a property of a system.
- 11 Pre-Register the Pilot or Don't Run It
A verdict negotiated after the evidence arrives is not a verdict but a settlement. Write down the outcomes, the computation, the stop line, the questions the window cannot settle, and the date — before day zero, where nobody can edit them without a trace.
- 12 Six Predictive Maintenance Field Cards
Six one-page cards for the records a predictive programme keeps skipping: the detection evidence record, the criticality queue, the failure-mode sentence, the measurement contract, the alert anatomy and its ladder, and the pilot registration sheet. Questions and record fields only.
Predictive Maintenance — Practitioner Reference Frameworks and Planning Guide
The twelve-part reference this series draws on: foundations and the value case, asset criticality and strategy, failure modes and degradation, the monitoring technologies, asset-class playbooks, sensors and IIoT architecture, data foundations, signal processing, analytics and prediction models, alerts and diagnosis, work management and CMMS integration, and pilot execution through rollout and governance — 126 sections with 46 technical figures.