Learning the baseline: what happens during the learning phase and what to watch out for

Before our system can assess a spindle, it has to learn the baseline, i.e. the normal state of this spindle: how strongly does exactly this spindle vibrate, at exactly this measuring point, at this speed, when everything is in order? Every new measurement is then compared against this learned state. The learning phase is therefore not a prelude but the foundation of every later alert. This article explains what "healthy condition" means here, why the baseline is divided so finely, what happens with a pre-damaged spindle, how to proceed after a bearing change and why the baseline does not simply drift along by itself.

Why a learned baseline instead of fixed limits

Fixed limits from standards such as ISO 10816 or ISO 20816 apply to whole classes of machines. They are useful for the overall vibration of a machine, but too coarse for early bearing damage on an individual spindle. Two identical machining centres vibrate differently because of installation, foundation, age and machining. A value that is normal on one machine may already be conspicuous on the other. The background is covered in the article Vibration analysis on machine tools and ISO limits.

In addition, there is hardly any damage data for spindle bearings from which generally valid thresholds could be derived. The question "Is that already too much?" is therefore best answered as: "More than is usual for this spindle?" That is exactly what the learned baseline checks.

What "healthy condition" means in the learning phase

The baseline is learned from measurements taken during normal operation. This rests on one clear assumption: the machine is in good order during the learning phase. The system cannot check this itself. It learns what it measures. That is why, before starting, we clarify with your maintenance team whether the spindle is running normally and when work was last done on it.

Not every measurement goes into the baseline. The edge box only evaluates a measurement window if the speed reported by the CNC control is the same before and after the window, to within ±2 %, and above 300 rpm. Run-up, braking and standstill are therefore excluded. From each valid window, the edge box calculates the features: RMS, peak value, crest factor, kurtosis, the vibration at once and twice the rotational frequency and the amplitudes in the envelope spectrum at the bearing frequencies. What these features tell you is explained in the article RMS, crest factor and kurtosis.

For each feature, the platform determines two quantities from at least 20 such measurements: the median as the typical value and a robust spread as a measure of how much the value normally fluctuates.

The first line is the classic robust spread. Individual outliers, such as an impact during a tool change, barely shift it. The second line protects against features that are skewed upwards, such as spectral amplitudes. The third line is a lower bound: if a spindle runs extremely smoothly during the learning phase, even a tiny change would otherwise count as a large deviation. Both factors make the spread comparable with the standard deviation for normally distributed data.

Separate baselines per machine, measuring point and speed range

A single baseline per machine would be too imprecise. Our system therefore learns separately according to three attributes:

• Machine: each machine has its own baseline. One spindle is never compared with another.
• Measuring point: a sensor on the front bearing sees a different signal from one on the rear bearing. Each measuring point therefore has its own baseline.
• Speed range: vibration depends strongly on speed. The platform assigns each measurement to a speed class in steps of 100 rpm and only compares within the same class.

Each row is a baseline of its own. A speed class that is run for the first time only after several weeks then starts its own learning phase. The other classes are already being assessed in the meantime. In the dashboard, the card for each measuring point and speed shows whether it is still learning and how many of the 20 measurements are already available.

How long the learning phase takes

There is no fixed duration in days, because the system learns by measurements, not by time. How quickly 20 valid measurements come together depends on:

• how often the spindle runs in the speed class concerned,
• how long it runs there at a constant speed, because only such windows count,
• how the edge box is configured for your machine, i.e. how often it records a measurement window,
• the shift pattern, i.e. whether the machine runs on one shift or around the clock.

A spindle that runs the same operation at the same speed all day learns its main speed quickly. A machine with many small batches and changing speeds needs longer for each individual class. While a class is learning, it does not trigger any alerts. This is deliberate: an assessment against a half-learned baseline would be unreliable.

What happens if the spindle is already damaged during learning

This is the most important limitation of the method, and we state it openly. If a bearing is already damaged during the learning phase, the system learns that condition as the baseline. The vibration already present at the damage frequency is then considered normal and does not trigger an alert on its own.

Even so, the system is not completely blind. Bearing damage usually continues to grow. If the vibration rises above the learned baseline, the system reports the increase as it would for any other spindle. In addition, the edge box searches the envelope spectrum for the bearing frequencies in every measurement window, independently of the baseline. These findings are stored with every measurement. On their own, however, they do not trigger an alert today.

Relearning after a bearing change or spindle replacement

After a bearing change, an overhaul or a spindle replacement, the spindle vibrates differently from before, usually more quietly. Assessed against the old baseline, the system would then see a lot of headroom, and new damage would be noticed too late. The baseline therefore has to be relearned after such interventions.

Relearning is triggered by your plant's administrator. Maintenance engineers and users with read-only rights cannot do this. The following happens:

1. All baselines of this machine, i.e. every measuring point and every speed class, are reset to "learning".
2. The reported condition of the machine is reset to Normal.
3. Learning uses only measurements from this point onwards. Older measurements are not included.
4. The action is recorded in the audit log with the person and the time.

Previous measurements and alerts remain stored and can be retrieved via the data export. Only trigger relearning once the spindle is back in normal operation after the intervention, i.e. after running in and not during test runs. Relearning is also necessary if we extend the analysis on the edge box and the features change as a result. A baseline built from old features no longer fits the new measurements.

Learning along without learning the damage

A spindle changes even without damage: new operations, different speeds, a different shift pattern. A baseline that never takes this into account becomes unsuitable over time. The obvious solution would be to keep updating the baseline from the most recent measurements. For bearing monitoring, that is precisely what is dangerous. Bearing damage develops slowly. A baseline that blindly follows the latest measurements drifts along with the damage. The deviation stays small even though the spindle keeps getting worse.

Our system therefore uses guarded learning. Three rules apply:

1. Only healthy measurements count. Only measurements that were themselves assessed as "Normal" feed into the baseline. Whatever is conspicuous never becomes the new normal.
2. New operating states are added. If the spindle runs in a speed range the system does not yet know, it learns that range in addition. Over time, more measurements can be assessed and damage is noticed earlier.
3. The original baseline is kept as a reference. Every measurement is also compared with the baseline that was learned first. If a feature creeps upwards over a long period, the system reports this trend separately, even if it is unremarkable from one measurement to the next.

Today, the learned baseline in our system stays fixed and changes only through a deliberate relearning that a person triggers and that is logged. New speed ranges are already learned separately. Guarded learning according to the three rules above is the next step we are building. Every alert remains traceable: it states which baseline it was calculated against. How this fits with our stance on opaque methods is covered in the article fixed thresholds or a learned baseline.

The learning phase in the pilot

In the pilot, the learning phase starts after the edge box and sensor have been installed. During this time you can already see the trend of the features and the learning progress for each measuring point and speed in the dashboard, but no alerts yet for the classes that are still learning. After that, actual monitoring against the learned baseline begins. The overall process is described on the pages Process and requirements and Condition monitoring pilot project. What an alert looks like afterwards is shown in the article Alerts with an explanation.

Frequently asked questions