groundup.ai
22/7/2026Maintenance teams often manage several units of the same equipment.
A production facility may have:
Because the machines have the same specifications, it may seem reasonable to apply the same alarm settings and AI model to all of them.
However, equipment rarely operates under perfectly identical real-world conditions.
NASA’s C-MAPSS prognostics dataset represents fleets of engines of the same type, but each engine begins with different degrees of initial wear and manufacturing variation. These differences are treated as part of normal operation rather than as faults. All equipment behaves in a similar way. Two machines can share the same design while developing different healthy operating signatures.
For an AI monitoring system, those differences matter.
A machine baseline is a reference model of how an asset behaves during healthy operation.
Traditional condition monitoring may establish a baseline using measurements such as:
An AI-based baseline can analyse additional relationships and patterns across time, including:
After learning the machine’s normal patterns, the system can look for deviations that may indicate changing equipment condition.
Siemens describes AI condition monitoring that can consume machinery information such as power consumption, rotating speed and operating mode in addition to other time-series data for anomaly detection. lt does not automatically prove that a component is faulty. It indicates that the machine is behaving differently from the expected pattern and may need further investigation.
A machine’s measured behaviour depends partly on how and where it is installed.
Differences may include:
For example, an accelerometer mounted on a clean, flat bearing housing may produce a different signal from one mounted on a painted or uneven surface.
Similarly, vibration from nearby machines may travel through a shared structure and influence the data.
The machines may be identical, but their signal paths are different.
When comparing similar assets, maintenance teams should verify that sensor installation and data-collection methods are consistent.
A difference in readings may come from the machine, the process, the sensor installation or a combination of all three.
Foundation stiffness can influence the vibration measured on a machine.
One pump may be mounted on a solid concrete base. Another identical pump may sit on a more flexible steel structure.
Other differences may include:
Even relatively small installation differences can change the way mechanical energy travels through the asset.
A machine-specific baseline captures the behaviour of the complete installed system — not merely the equipment model listed on the nameplate.
When a machine’s baseline changes after alignment, foundation repair or coupling replacement, the new pattern should be reviewed and, where appropriate, established as a new healthy reference.
Machine signals often change with operating conditions.
A fan operating at one steady speed may have a predictable vibration profile. The same fan operating under variable-speed control may show a wider range of healthy vibration patterns.
SKF’s vibration-monitoring guidance notes that equipment may operate under different conditions involving load, speed, processed materials and environment. Operating conditions that can affect machine data include:
Changes in speed can also shift the frequencies visible in vibration data. A system that ignores speed may confuse a normal operating transition with a developing mechanical issue.
AI models should compare the machine against relevant operating states.
A reading collected at full load should not always be compared directly with data collected while the machine is idling.
Two identical machines may have very different maintenance histories.
Machine A may have recently received:
Machine B may still contain older components that remain serviceable.
Both machines may be healthy, but their normal vibration, sound and temperature signatures may differ.
Maintenance events can also produce step changes in sensor trends. A sudden reduction in vibration after alignment may be a positive change rather than suspicious data.
Maintenance records should be connected with condition-monitoring data whenever possible.
When technicians replace or adjust a major component, the AI system should be informed so the post-maintenance behaviour can be validated and, when necessary, used to update the baseline.
The environment around an asset can influence both the machine and the sensor readings.
Relevant factors include:
An identical motor operating near a furnace may run hotter than one in a ventilated utility room.
Temperature can also affect lubricant viscosity, material expansion and machine clearances.
This does not mean temperature increases should be ignored. It means the system needs context to distinguish environmental effects from abnormal equipment behaviour.
Maintenance teams should consider ambient and process temperature when reviewing thermal trends.
A temperature alarm is more useful when the system can determine whether the rise came from the surrounding environment, increased machine load or internal deterioration.
Identical machines may support different parts of the production process.
For example:
These process differences affect the forces placed on the equipment.
A process change can therefore alter vibration, temperature, sound, pressure or power consumption even when the machine remains healthy.
Condition-monitoring data should be reviewed alongside relevant process data.
Without production context, a model may incorrectly identify normal process variation as a mechanical anomaly.
Two machines installed on the same day will not necessarily age at the same rate.
One may experience:
Gradual wear can change the machine’s signal without causing immediate functional failure.
A useful AI system needs to recognise both short-term anomalies and longer-term changes in the machine’s condition.
Baselines should not be treated as permanent and unchanging.
They require governance. Maintenance teams should review whether a gradual shift represents normal ageing, a new operating state or developing deterioration that requires intervention.
Universal alarm limits are valuable.
OEM guidance, engineering judgement, safety limits and standards can help teams determine when vibration, temperature or another measurement has reached an unacceptable level.
ISO 20816-3 provides general requirements for evaluating vibration for several categories of industrial machinery under normal operating conditions. General thresholds and individual AI baselines answer different questions.
Has this measurement exceeded a predefined limit?
Has this machine moved away from its own normal behaviour?
Both questions are useful.
The problem occurs when a maintenance programme uses only the first one.
Consider two identical pumps:
Asset | Normal vibration | General alarm |
Pump A | 2.0 mm/s | 5.0 mm/s |
Pump B | 3.5 mm/s | 5.0 mm/s |
Pump A’s vibration gradually increases from 2.0 mm/s to 4.2 mm/s.
It remains below the general alarm level. However, the measurement is now more than twice its previous healthy level.
Pump B remains stable at approximately 3.5 mm/s.
A threshold-only system may continue showing both machines as acceptable.
A machine-specific system may identify that Pump A has experienced a meaningful deviation and alert the maintenance team earlier.
This does not confirm the cause. The change may relate to alignment, imbalance, looseness, bearing condition, process load or another factor.
But it gives technicians an opportunity to investigate before the measurement reaches a critical level.
Now consider a machine whose normal vibration is naturally close to a general alarm threshold because of its foundation, operating load or installation arrangement.
If the same alarm is applied without context, the machine may repeatedly generate warnings despite remaining stable.
This can lead to:
When teams receive too many low-value alerts, they may begin ignoring the system — including the alerts that genuinely matter.
The objective should not be to generate as many alerts as possible.
The objective is to generate actionable alerts supported by meaningful evidence.
No.
Machine-specific AI baselines should complement — not replace — OEM limits, machinery-protection systems, engineering standards and safety procedures.
A practical monitoring strategy can use several layers:
These help protect people and equipment from dangerous operating conditions and may initiate alarms or shutdowns.
These provide established reference levels for evaluating equipment condition.
These identify smaller or earlier deviations from the individual asset’s normal behaviour.
Maintenance engineers, reliability teams or specialists review the evidence and determine the probable cause and appropriate response.
AI is most valuable when it strengthens these layers rather than attempting to replace them.
A reliable baseline should represent the machine across its genuine healthy operating range.
The process should include:
A baseline should be reviewed after events such as:
A baseline should not automatically reset every time the machine behaves abnormally. Doing so could teach the AI that deterioration is normal.
The change should first be investigated and validated.
As industrial facilities deploy condition-monitoring systems across hundreds or even thousands of assets, manually creating and maintaining a baseline for every machine becomes increasingly difficult.
The future of industrial condition monitoring is automatic baselining.
An automatic baselining system continuously analyses healthy machine data and learns the normal operating behaviour of each individual asset. Instead of requiring engineers to manually define one fixed operating range, the system can identify patterns across different:
This allows each machine to develop its own operating profile while still being evaluated against OEM recommendations, engineering thresholds and machinery-protection limits.
Automatic baselining is particularly valuable for large industrial facilities. A plant may contain hundreds of motors, pumps, fans, compressors and gearboxes operating under different conditions. Creating and regularly updating each baseline manually would require significant engineering time.
With automatic baselining, the monitoring system can:
However, an automatic baseline should not mean an uncontrolled baseline.
The system should not automatically absorb every change into the machine’s normal behaviour. An unexplained increase in vibration, temperature or sound may represent developing deterioration. Before a significant change becomes part of a new baseline, it should be investigated, supported by operational context and, where necessary, validated by the maintenance team.
This creates a practical partnership between AI and industrial expertise.
The AI continuously learns, compares and identifies meaningful changes. Maintenance and reliability teams provide context, confirm maintenance activities and determine whether a new operating pattern is healthy, temporary or potentially harmful.
As industrial AI develops, baselines are likely to become more adaptive, contextual and increasingly automated. The goal is not simply to apply the same alarm settings to more machines. It is to give every asset an intelligent reference that reflects how it actually operates in the real world.
This is the future of industrial condition monitoring: scalable automatic baselining, supported by engineering guardrails and human judgement.
Before deploying AI condition monitoring, maintenance and reliability teams should ask:
These questions help determine whether the solution is providing genuine condition insight or merely applying generic limits through a new interface.
Yes. Differences in foundation stiffness, alignment, mounting, process load, speed, component history and environmental conditions can produce different healthy vibration levels.
The important question is not whether the measurements are identical, but whether each machine is stable within its expected operating behaviour.
An AI baseline is a learned representation of a machine’s healthy operating behaviour. It may include vibration, temperature, sound and process variables across different loads and operating states.
The system compares new data with the baseline to identify unusual changes.
Not necessarily.
An anomaly means that the machine’s behaviour differs from the expected pattern. Maintenance teams should review the evidence, operating context and physical condition before confirming a fault.
Fixed thresholds can identify measurements that exceed an established limit, but they may miss smaller changes that are significant for a particular machine.
They may also create false alarms when an asset’s stable normal behaviour is naturally close to the threshold.
Not necessarily an entirely separate AI architecture, but each asset should have an individual baseline or asset-specific context.
Shared models can provide useful knowledge across similar equipment, while individual baselines account for each machine’s installation and operating behaviour.
Yes, but baseline changes should be controlled.
A verified repair, permanent process adjustment or equipment modification may create a legitimate new normal. However, an unexplained increase in vibration or temperature should not automatically be absorbed into the baseline.
There is no universal duration.
The required period depends on the machine’s operating cycle, variability, available data and the number of operating states that need to be represented. A machine with a steady operating condition may require less data than one experiencing multiple products, speeds and loads.
The system should include operational context such as speed, load, pressure, flow, power or production mode.
Research and industrial condition-monitoring systems recognise varying operating conditions as an important challenge because operational changes can affect measured features and mask or imitate faults. Final Takeaway
Two machines can be identical in design but different in behaviour.
Their normal signatures are shaped by:
Universal thresholds remain useful as engineering guardrails. But comparing every machine against one fixed value can produce missed warnings, false positives and low-confidence alerts.
Machine-specific AI baselines add another layer of intelligence by asking:
Is this individual machine behaving differently from the way it normally behaves under comparable conditions?
For maintenance teams, that distinction can provide earlier, more relevant and more actionable insight.
Because in the real world, identical machines do not always have an identical normal.
Bergabunglah dengan Wall of ❤️ kami — bagikan dan tag kami!
Kami menyoroti para pemimpin yang mendorong inovasi dan kinerja, memberi Anda visibilitas sambil membentuk masa depan AI
Artikel Terkait