Resources and content
Use cases September 10, 2026 18 min read

Data for predictive maintenance: what needs to exist before the AI

Before predicting a failure, you need to be able to explain the behavior of the asset that might have it.

Athenas — Data for predictive maintenance: what needs to exist before the AI
In this article

Predictive maintenance projects rarely fail for lack of algorithms. They fail because the available data base doesn't support the kind of question the model needs to answer. Before predicting a failure, you need to be able to describe the asset's normal behavior, what happened around it, and under what conditions each reading was produced.

The promise of predictive maintenance

Predictive maintenance is the practice of anticipating an asset's degradation or failure based on its observed behavior, instead of acting only by calendar (preventive) or after a breakdown (corrective). The goal is to identify a pattern change with enough lead time to plan the intervention.

The sources of information usually already exist on the plant floor, spread across different systems:

  • Vibration
  • Temperature
  • Pressure
  • Current
  • Speed
  • Energy consumption
  • Alarms
  • Operating states
  • Maintenance events
  • Failure history

The mistake of starting with the model

Teaching example

Millions of records and a poorly framed question

A hypothetical company has millions of sensor records accumulated in a historian. From that volume comes the idea: “let's use AI to predict failures.”

Before training any model, it's useful to check whether the base can answer the questions below:

Association
Which asset is the sensor associated with?
State
Was the asset running?
Condition
Under which operating condition?
Product
Which product was being processed?
Intervention
Was the equipment under maintenance?
Reliability
Is the value trustworthy?
History
When was the last maintenance?
Failure
What failure was recorded?
Precursor
Which variable changed beforehand?

Without these answers, the model can learn incomplete or mistaken patterns: treating a scheduled shutdown as an anomaly, learning the behavior of two different assets as if it were one, or associating degradation with a product change. Hypothetical scenario, for illustration purposes.

The data needs to know the asset

The same reading, with and without asset identity

Isolated data

Temperature = 96 °C

Contextualized data

Temperature
96 °C
Asset
Compressor 04
Plant
Plant A
Area
Compression Area
Line
Line 02
Product
Product X
Condition
Normal operation
Timestamp
14:32
Quality
valid

The second set is more useful for analysis because it allows legitimate comparison: with the same asset's history, with equivalent assets on another line, with behavior under the same product and under the same operating condition. Without a stable asset identity, time series from different pieces of equipment can get mixed and produce a pattern that doesn't exist on the plant floor.

Before applying models to the data, it's necessary to build a consistent view of the assets and the operation.

Maintenance history also has to enter the equation

Telemetry describes behavior; maintenance describes interventions and their consequences. A model that only sees sensors can't tell real degradation apart from an effect that follows a component swap.

Sources that need to be related

Behavior

  • Sensor telemetry
  • Operating states
  • Alarms

Intervention and outcome

  • Maintenance events
  • Maintenance orders
  • Recorded failures
  • Operating conditions
  • Asset context
An asset's chain of evidence
  1. 01SensorContinuous readings of the monitored variable.
  2. 02BehaviorPattern observed over time, with a baseline for normality.
  3. 03EventA deviation, alarm or state change that gets recorded.
  4. 04MaintenanceThe intervention performed, with scope, date and components.
  5. 05OutcomeSubsequent behavior: return to baseline, recurrence, or a new failure.
To do this, data coming from different OT and IT environments needs to be integrated into a coherent architecture.

The importance of data quality

Quality dimensions and the corresponding check
DimensionWhat to check
CompletenessAre there periods without data, collection gaps, or assets with no records?
ConsistencyDoes the same variable have different structures, names or scales across sources?
ValidityIs the value within a physically possible range for that asset?
TimelinessAre timestamps reliable, with a defined time zone and no silent delays?
UniquenessAre there duplicate records from resends or multiple collections from the same source?
AccuracyDoes the data properly represent the operational reality it's meant to measure?

Operational context prevents wrong interpretations

Consider a motor showing a temperature increase. That increase could mean:

  • Component degradation
  • Increased load
  • Product change
  • Speed change
  • Different ambient conditions
  • Off-standard operation
  • Running a different process

Without context, these situations can look identical in the time series: temperature rises in all of them. With context — product, load, setpoint, ambient condition, operating regime — it becomes possible to separate expected behavior from anomalous behavior. That separation is exactly what a detection model needs to learn.

From telemetry to the model

Full chain of a predictive application
  1. 01Sensors and PLCsSignal acquisition on the plant floor.
  2. 02SCADA / HistorianSupervision and recording of the time series.
  3. 03CollectionReliable transport, with buffering and resend after downtime.
  4. 04QualityDeclared rules, continuous verification and flagging of suspect readings.
  5. 05ContextualizationAssociation with asset, hierarchy, unit, condition, product and origin.
  6. 06HistoryA long, consistent series, with retention matching the asset's criticality.
  7. 07Feature engineeringBuilding the derived variables the model uses: windows, rates, deviations, aggregations.
  8. 08ModelTraining and validation with data representative of real conditions.
  9. 09InferenceRunning the model on new data, in the same format as training.
  10. 10Alert and decisionNotification with enough context for maintenance to decide on the action.

What if generative AI enters the operation?

Language models expand how people interact with industrial data, but they don't replace the base. The most direct applications are for querying and analysis support:

  • Natural-language queries on the state and history of the operation
  • Assisted investigation of a deviation, suggesting what to check
  • Explaining events based on the available records
  • Querying history without relying on manual exports
  • Analysis support, organizing scattered evidence
  • Generating operational summaries by period, line or asset

“Which pieces of equipment showed off-standard behavior in the last 24 hours?”

A reliable answer to that question doesn't depend on the language model, but on what sits underneath it: asset identification, available history, reliable timestamps, declared quality, operational context, rules that define what's standard, and access control over the right information. Without that, the answer can be fluent and still wrong.

A data architecture ready for AI

Three blocks of the architecture
  1. 01OperationSensors, machines and industrial systems that produce the data.
  2. 02Data layerCollection, contextualization, quality, organization and distribution under a common model.
  3. 03ApplicationsAnalytics, BI, maintenance, AI/ML and generative AI consuming the same base.

The central point is reuse: the same contextualized base meets different needs. The indicator that feeds an operations dashboard and the variable that feeds a detection model can come from the same origin, with the same meaning and the same traceability.

What changes when data is reusable

With and without a common base

Without a common base

  • Each project rebuilds integrations
  • Each project redoes cleaning
  • Each project redefines context
  • Each project identifies assets its own way
  • Each project assembles its own history

With a contextualized base

  • Projects reuse data
  • Projects reuse information models
  • Projects reuse context
  • Projects reuse the asset hierarchy
  • Projects reuse events and history

This isn't about promising an effort reduction in percentages. It's about where the effort is applied: building and maintaining a common base, or repeating the same preparation in every initiative.

The role of Athenas

Athenas Data Intelligence structures the layer that precedes and supports these applications: a base of contextualized industrial data, organized to be consumed by analytics, operational applications and artificial intelligence initiatives.

From industrial data to the intelligent application
  1. 01Industrial dataPlant floor sources and existing corporate systems.
  2. 02ContextualizationCatalog, unit normalization, ISA-95 hierarchy, quality and origin traceability.
  3. 03Reusable baseHistory and current state available to multiple consumers, with access governance.
  4. 04Athenas AIQueries over this contextualized base, crossing sources and assets, with traceable answer origin.
  5. 05Intelligent applicationsAnalytics, operational dashboards and modeling initiatives built on the same base.

Checklist: is your plant ready to apply AI to its assets?

Data prerequisites for AI applied to assets
  • Assets have consistent identification
  • Historical data is available
  • Maintenance events are recorded
  • Timestamps are reliable
  • Quality is monitored
  • Variables have context
  • There is an asset hierarchy
  • Different sources can be related
  • Failure history can be linked to the assets
  • Data can be consumed by different applications
  • There is access governance
  • There is observability over the data chain

The more items that are missing, the more the challenge lies in data engineering and the less it lies in the AI model.

Predictive maintenance starts long before the algorithm

The conversation about industrial AI usually starts with the model. In practice, the model depends on a much larger chain.

You need to know where the data came from, which asset it belongs to, under what condition it was produced, whether it's reliable, which events happened around it, and how this information relates to one another.

That's why a contextualized data architecture is not just infrastructure for analytics. It is part of the very capability to build reliable AI applications in industry.

Also read: Contextualized data in practice: turning numbers into useful information.
EAEquipe AthenasAthenas Data Intelligence
Share

See how Athenas organizes your operation's data

A roughly 40-minute demo showing source connectivity, the catalog, the ISA-95 UNS and real-time data distribution.