quindart
← All articles

Predictive Maintenance Software: What to Check Before Buying

Before buying predictive maintenance software, check the failure question, available data, warning quality and the team’s ability to act.

Two maintenance colleagues compare tablet and paper records beside a blue motor with an exterior sensor.

Start with a failure you could act on

Predictive maintenance software is often discussed as a way to discover equipment problems earlier. For a buyer, earlier matters only when the operation can do something useful with the warning. Name the asset, the failure mechanism of concern and the decision that additional notice would change. The response might be a planned inspection, a controlled maintenance window or further diagnosis by a qualified specialist. A dashboard alert without an owner does not create that response.

Consider a hypothetical plant with a compressor serving a production area. Maintenance suspects that a recurring temperature change sometimes precedes a problem, but operating load also varies. The initial task is to understand that observation and its context, not immediately purchase a failure-prediction model. Link maintenance records, production demands and operator observations. Decide which conditions warrant investigation under the site's existing procedures, without substituting an experimental model for safety protections or required maintenance.

Distinguish observation from a predictive claim

The U.S. Department of Energy's operations and maintenance guide describes condition-based approaches and performance trending. Its practical sequence includes identifying equipment, prioritizing it, examining useful measurements and assigning knowledgeable people to evaluate trends. This supports starting with an observation and maintenance process. It does not establish that a particular machine-learning product will predict failures in your plant or that a vendor's reported savings will transfer to your equipment. [1]

Ask a supplier exactly what its system produces. A threshold alert indicates that a measured value crossed a boundary. An anomaly score suggests behavior differs from a learned reference. A prediction of a particular failure within a stated period makes a more specific claim. Each needs different evidence and a different response. Clarify whether the proposed purchase is monitoring, diagnosis support or validated prediction, and ensure the proposal uses those terms consistently.

Check whether history can support the question

Inventory what exists before ordering additional sensors. Useful records may include asset identifiers, timestamps, operating hours, loads, product changes, inspections, fault descriptions and maintenance actions. Check whether the same asset has different names in different systems and whether clocks are aligned. A work order closed on Friday may describe a fault observed on Wednesday. Without that distinction, the apparent relationship between a measurement and a failure can be misleading.

Ask maintenance staff to review examples of normal operation, known faults and uncertain cases. Distinguish a confirmed failure from a precautionary replacement or a symptom that was never diagnosed. Missing failures and inconsistent labels limit what can be evaluated. There is no universal minimum number of records that makes a predictive model ready. The answer depends on the failure mechanism, variation, measurement quality and the claim being tested. Document where the evidence is too thin.

Preserve the operating context around measurements

A change in temperature or vibration can have different meanings at different loads, speeds or operating states. Record enough context to compare like with like, and involve people who understand the asset. In the hypothetical compressor case, a change associated with a new production pattern should not automatically be labeled deterioration. First ask whether the measurement method, environment or operating conditions changed and whether the observation is repeatable.

Define how readings will be collected, checked and retained, including units, sensor location, calibration responsibilities and gaps in collection. Assign ownership for reviewing suspect data. Protect connections to operational equipment according to the site's access and engineering requirements; a data project should not casually introduce write access to controls. Where instrumentation changes are needed, obtain the relevant technical and safety review before installation. More measurements are useful only when they improve an actionable understanding.

Evaluate warnings as an operational service

Run an initial evaluation without letting experimental predictions control equipment or replace established maintenance. Agree which assets and operating conditions are in scope, how warnings are reviewed and what counts as a useful advance notice. Evaluate against data not used to tune the model, with attention to time order and changing conditions. Ask the supplier to explain how it avoids using information that would not have been available when the warning was supposedly generated.

Track missed relevant events, false warnings, warning lead time and the work required to investigate each alert. A model that raises frequent alarms may consume the same scarce technicians needed for preventive work. A correct warning that arrives too late may not change the decision. Keep a record of what the team did and what was subsequently confirmed. Set stop or revision criteria before the trial, including cases where a simpler trend review proves adequate.

Buy the response process and continuing review

A procurement brief should state the failure question, available evidence, uncertainty, evaluation method and responsible people. Ask who maintains sensor quality, reviews model behavior, updates the system after equipment changes and supports users when an alert cannot be explained. Specify access to historical measurements and the ability to export them. Define escalation and service boundaries without assuming continuous expert coverage. The operating team needs a workable service, not merely permission to log into software.

Quindart describes an approach that connects business information with observations from the floor before selecting an intervention. For maintenance readiness, a useful proposed starting scope is one asset group, its maintenance history and the production decisions affected by downtime. The first deliverable could be a dependable event record and review workflow, with predictive modeling considered only if the evidence supports it. Bring an example of an unexplained stoppage and the records available today; that is a stronger starting point than a request to predict everything. [2]

Sources

  1. U.S. Department of Energy: Operations and Maintenance Best Practices Guide, Release 3.0
  2. Quindart: Factory software and integration