Our headline claim — Watsynq detects bearing defects 6 weeks before a conventional vibration analyst would — is the kind of statement that deserves scrutiny. Any measurement that lives in marketing copy needs a methodology behind it that holds up when examined carefully. This article explains exactly what the six-week figure means, how we measured it, and where it does and doesn't apply.
We're writing this partly because we've been asked, and partly because we think the field has a precision problem. "Early warning" is a phrase that appears in essentially every PdM vendor's materials without a consistent definition of what it means or how it's measured. We want to be specific.
Defining the measurement: what does "detection window" mean?
Detection window, as we use it, is the interval between two events:
- Tdetect: The point at which Watsynq's fusion model first generates a sustained anomaly signal on the affected asset — specifically, two consecutive RUL updates showing degradation trending outside the healthy operating band, with multi-sensor corroboration across at least two signal modalities.
- Tvibration-threshold: The point at which overall vibration velocity RMS would cross ISO 10816 Zone C/D threshold for that machine class, or — when we have it — the point at which a vibration analyst conducting a monthly route would have flagged the asset for follow-up action based on spectral analysis.
Detection window = Tvibration-threshold − Tdetect.
We are measuring the lead time between our detection and what a conventional monthly route-based program would detect. We are not measuring the lead time to failure, which is typically longer. A bearing that we first flag at Tdetect may still run for many additional weeks before functional failure — our claim is specifically about the comparison to the vibration-analyst detection point, not to the failure point itself.
This is an important distinction. "6 weeks before failure" and "6 weeks before a vibration analyst would catch it" are different claims. The second one is ours.
The dataset: 40+ confirmed failures
As of the time of writing, we have 47 confirmed bearing failure events across our deployment sites that we have analyzed for detection window calculation. "Confirmed" means: a bearing was removed and inspected, the inspection found evidence of the failure mode our model had identified (outer race spalling, inner race spalling, rolling element damage, cage failure, or lubrication-driven surface fatigue), and we have both the Watsynq sensor record and at minimum a monthly vibration route record for the affected asset covering the period prior to failure.
Of 47 events, 43 met our criteria for detection window calculation. Four were excluded: two because the vibration route record was incomplete for the relevant period, one because the failure was an infant mortality event following a recent bearing replacement (P-F interval too short to be meaningful), and one because a process upset caused an acute failure with no pre-failure degradation signature in any sensor modality.
For those curious about the breakdown: 31 of the 43 were outer race failures, consistent with the general prevalence of BPFO-driven spalling as the dominant mode in radially loaded rotating equipment. Eight were inner race failures. Four were rolling element failures. The detection window characteristics differ somewhat by failure mode, which we discuss below.
The distribution: median, not mean
The six-week figure is the median detection window across the 43 qualified events. The distribution is not symmetric.
The 10th percentile is approximately 3.5 weeks — meaning roughly 10% of detections gave less than 3.5 weeks of lead time over the vibration analyst reference point. The 90th percentile is approximately 10 weeks. The mean is slightly above the median at about 6.5 weeks, pulled upward by a handful of outer race cases where AE picked up the initial crack extremely early and vibration symptoms developed slowly.
We use the median because it's more resistant to the outliers in both directions: the unusually early AE detections that aren't representative of typical cases, and the few cases where a failure developed faster than the typical outer race progression curve.
When we say "6 weeks," we mean: in a typical (median) case, Watsynq will flag the degradation trend approximately 6 weeks before the vibration-based route program would flag it. Some cases will be shorter; some will be longer.
Why acoustic emission is the leading modality for outer race failures
The physics of why AE detection precedes vibration detection is worth explaining explicitly, because it's the core of why multi-sensor fusion outperforms vibration-only monitoring for early detection.
Outer race spalling begins as subsurface fatigue at the contact stress concentration zone beneath the raceway surface. At initiation, the crack is microscopic — a few hundred micrometers in depth, not visible to the eye and not generating measurable force variations. But the crack surfaces are in relative motion under the rolling contact stress cycle, and that motion generates ultrasonic stress wave emissions in the 100–500 kHz range. An acoustic emission sensor on or near the bearing housing picks up these bursts; an accelerometer does not, because the energy is attenuated rapidly through the bearing housing structure and appears far below the noise floor of conventional vibration instruments at the early crack stage.
As the crack propagates toward the surface and spalling begins, material dislodges from the raceway. Now the rolling element passes over the spall lip at the ball-pass-outer-race frequency, generating a periodic impact. This impact is at frequencies that an accelerometer can capture — the characteristic BPFO signature appears in the velocity spectrum. This is when the vibration analyst on a monthly route would first notice the fault.
The interval between "AE detects subsurface crack activity" and "velocity spectrum shows BPFO signature" is the detection window we're measuring. In the outer race cases in our dataset, this interval ranged from approximately 2 weeks (fast-progressing failures, typically in high-load, high-speed applications) to 14 weeks (slow-progressing failures in lightly loaded, lower-speed applications). The median of approximately 6 weeks reflects the mix of applications in our current deployment base — primarily medium-to-heavy industrial rotating equipment in refining, aggregate, and cement-adjacent applications.
The lube oil corroboration piece
For the 43 qualified events, we also tracked when oil particle count first showed elevated ferrous content. In 38 of 43 cases, the AE detection preceded the first elevated oil particle reading. The median lag between AE detection and oil particle elevation was approximately 2.5 weeks.
This isn't surprising — it takes time for debris from an early-stage spall to enter the oil stream in sufficient quantity to register against the healthy baseline. But it confirms that oil analysis alone, even at monthly sampling intervals, would have detected these failures later than AE — typically 3–4 weeks after AE detection, which means approximately 2–3 weeks before the vibration route threshold crossing.
The combination of AE and oil — even monthly oil — provides stronger confirmation than either alone. When both AE and oil are trending in the same direction, the probability of a true-positive finding approaches confidence levels where we're comfortable recommending a pre-failure intervention work order. When only AE is elevated, we treat it as a watch-and-verify condition rather than an immediate action item unless the AE trend rate is steep.
The cases that didn't fit: what we learned from them
The four excluded events are worth discussing because they represent the limits of the claim.
The infant mortality case was a bearing replacement after which the new bearing failed within 8 days. Post-removal inspection showed a false brinelling pattern consistent with improper installation — likely an axial impact during pressing that caused raceway indentation. This failure mode has no meaningful P-F interval; the damage was instantaneous and progressive failure began immediately. No monitoring approach with any reasonable sampling interval would have detected this 6 weeks in advance, because the "6 weeks" simply didn't exist.
The process upset case was a pump that experienced a water hammer event — a hydraulic shock load far in excess of design operating conditions. The shock caused immediate bearing cage fracture. Again, no pre-failure degradation period, no detection window to speak of. AE spiked dramatically at the time of the event; there was nothing to detect in advance.
These cases are honest limitations of the approach. We're not claiming to predict installation errors or process safety events. We're claiming to detect progressive degradation earlier than conventional monitoring. For the failure modes that actually involve progressive degradation — which represents the large majority of bearing failures in normal heavy industrial operation — the six-week figure is representative.
How we'll continue to update this number
The dataset will grow. We currently have monitoring active on several hundred assets across our deployment sites, and we expect to add additional confirmed failure events to the dataset over the coming months. We intend to re-evaluate the median detection window figure annually and update it based on the current dataset.
We also expect the number to vary by industry segment and equipment type as we accumulate more data across different applications. The 6-week median reflects our current deployment mix. As we deploy into equipment classes with different failure progression rates — larger slow-speed equipment where P-F intervals tend to be longer, or high-speed precision equipment where they tend to be shorter — the distribution will shift.
One thing we will not do is retroactively narrow the dataset to make the number look better. The four excluded events are excluded because they failed inclusion criteria defined before we ran the analysis, not because they made the number worse. If future events show shorter detection windows that pull the median down, the reported figure will reflect that.
That's the only way this number has any meaning: if it's measured the same way every time, and if we report what it actually says rather than what we'd prefer it to say.