Why Event-Triggered Waveform Capture Misses Degradation

Power quality meters in data centers have captured waveform data for decades. The standard architecture is event-triggered: the meter monitors a summary metric (RMS voltage, typically), and when it deviates beyond a configurable threshold, the meter stores a window of high-resolution waveform data around the trigger point. Pre-trigger buffers of 100–500 ms and post-trigger captures of 200 ms to 1 second are common.

This architecture is well-suited to its design purpose: documenting discrete power quality events for root cause analysis and compliance reporting. A voltage sag to 82% of nominal for 200 ms, a transient impulse of 1.5 kV lasting 2 ms, a momentary interruption: each produces a clean trigger, and the captured waveform provides the information needed to determine what happened and why.

The architecture has a structural assumption embedded in it: the interesting phenomena are discrete events that cross thresholds. For power quality, this assumption holds. For equipment health, it does not.

The event model vs. the degradation model

Power quality events and equipment degradation are different physical processes with different temporal characteristics.

Event-triggered capture is designed for the left column. It records when something unusual happens. The problem is that degradation does not produce anything unusual at any single point in time. A UPS rectifier whose 5th harmonic THD increased from 27.3% to 27.8% today is not unusual. The same rectifier whose 5th harmonic has increased from 26.1% to 29.4% over the past three weeks is degrading. The diagnostic information is in the trajectory, not in any individual measurement.

An event-triggered system cannot capture a trajectory. It captures instants. Between triggers, the waveform data is discarded. The slow, monotonic spectral drift that characterizes early-stage degradation never triggers a capture, because at no individual moment does it exceed any threshold.

By the time a degrading component produces an event that crosses a PQ threshold (a voltage sag caused by increased rectifier impedance, a current spike from a capacitor failing to filter a load transient), the degradation has already progressed from the early stage (where maintenance is straightforward) to the late stage (where failure is imminent or occurring).

What continuous capture means in practice

Continuous waveform capture is the opposite approach: capture every cycle, on every circuit, regardless of whether anything unusual is happening. Then compute spectral features from each cycle and store the resulting time series for trend analysis.

At 8 kHz sampling on a 60 Hz system, each electrical cycle is represented by 133 samples. A DFT of each cycle yields harmonic amplitudes at every integer multiple of 60 Hz through the 66th order (4,000 Hz Nyquist limit). Computing this 60 times per second produces a harmonic time series with one-cycle temporal resolution.

The raw waveform data is not stored indefinitely (at 8 kHz per channel, the data volume would be prohibitive). The per-cycle spectral features are stored: harmonic amplitudes H1 through H25, THD, power factor, crest factor, and several derived indices (H5/H7 ratio, triplen harmonic content, etc.). This compressed representation retains the diagnostic information while reducing storage requirements by roughly 100x compared to raw waveform retention.

The resulting time series, one spectral snapshot per cycle, 60 per second, accumulated over weeks, is what makes degradation detection possible. Each circuit builds a baseline distribution for every spectral feature. Degradation appears as a statistically significant departure from the baseline, detected through standard change-point and trend analysis methods.

A concrete comparison

Consider a 6-pulse UPS rectifier with an aging DC bus capacitor. The capacitor's ESR has increased from its nameplate 15 milliohms to 45 milliohms over 18 months of operation. The increased ESR extends the commutation overlap angle by approximately 3 degrees, which redistributes harmonic current from the 5th order to the 7th order.

Here is how each measurement architecture responds to this degradation:

Event-triggered PQ meter. The meter monitors output voltage THD against a threshold of, say, 5%. The rectifier's output voltage THD increases from 3.2% to 3.8% over three weeks. No threshold is crossed. No waveform capture is triggered. No data is stored. The degradation is invisible.

1 Hz BMS polling. The BMS polls output voltage, current, and power once per second. RMS values show no meaningful change (within 0.5% of baseline). The BMS reports normal operating parameters. The degradation is invisible for the same reason it was invisible at the PQ meter: the RMS values have not changed enough to notice, and the harmonic content that has changed is not measured.

Continuous 8 kHz capture. Every cycle's harmonic content is computed and stored. Over three weeks, the 5th harmonic amplitude decreases from 28.1% to 26.3% of fundamental, while the 7th increases from 11.4% to 14.1%. The H5/H7 ratio shifts from 2.46 to 1.87. This is a 24% change in a specific spectral ratio, well above the noise floor. The trend is detected, flagged, and correlated with the known signature for capacitor ESR increase.

The degradation is the same in all three cases. The difference is entirely in whether the measurement architecture captures the data that contains the diagnostic information.

The pre-event window

The practical value of continuous capture is the time it provides between detection and failure, time that can be used for planned maintenance rather than emergency response.

Return to the ESR degradation example above. The capacitor's ESR tripled over 18 months, from 15 to 45 milliohms. The resulting H5/H7 ratio shifted from 2.46 to 1.87, a 24% change in a specific spectral ratio. That shift was detectable within the first three weeks of accelerated degradation, well before the capacitor's condition affected rectifier output quality. During that window, no BMS alarm triggered and no PQ meter event fired, because the degradation was progressing below every threshold in the existing monitoring infrastructure. It was detectable only in the harmonic time-series data that continuous capture provides.

The advance warning window in cases like this is typically measured in weeks, not hours. That is the difference between scheduling maintenance during a planned window and scrambling at 2 AM.

This is not an argument against event-triggered monitoring. It is an argument for complementing it. Event-triggered capture tells you what happened during a power quality event. Continuous capture tells you what is happening to your equipment between events. The first is forensic. The second is prognostic. Both are necessary.

Limitations of continuous capture

Continuous capture solves the temporal coverage problem but introduces others.

Storage and compute cost. Even in compressed spectral form, storing per-cycle features for hundreds of circuits over months requires meaningful storage infrastructure. At roughly 200 bytes per second per circuit in compressed spectral form, a 500-circuit deployment generates approximately 3 TB per year. With aggressive compression and retention policies, this reduces to manageable levels, but it is not trivial.

Baseline quality. Trend detection accuracy depends on baseline quality. A circuit that was already degrading when monitoring began has a corrupted baseline. The degradation appears as normal operating behavior, and the system misses the existing condition. We address this with fleet-level comparison (a degrading unit's baseline looks different from the fleet median), but fleet comparison requires a population of similar equipment, which is not always available.

False positives from operational changes. Replacing a load on a circuit, changing a workload mix, or modifying the upstream power configuration all change the harmonic profile in ways that can mimic degradation. Our detection pipeline includes change-point classifiers that attempt to distinguish operational changes (abrupt, coincident with known maintenance windows) from degradation (gradual, monotonic), but the false positive rate for operational changes is approximately 5–8% of detected anomalies. Each anomaly requires human review to confirm or dismiss.

These are engineering challenges, not fundamental limitations. But they are worth acknowledging because they define the real-world reliability of the approach. Continuous capture gives you access to the diagnostic information. Extracting accurate, actionable conclusions from that information is a separate, harder problem that remains under active development.

Request a demo to compare continuous capture against your existing event-triggered monitoring.