Electrical Signature Analysis Comes to the Data Center
Electrical signature analysis has protected industrial motors for decades. Verdigris applies the same physics to data center infrastructure at 8 kHz, continuously.
Electrical signature analysis (ESA) is a fault-detection technique that infers equipment health from the spectral content of current waveforms. Motor Current Signature Analysis (MCSA), the motor-specific case of ESA, has been detecting rotor bar cracks, bearing defects, and winding faults in industrial motors since the 1980s. The physics is straightforward: a fault changes the machine's impedance, which changes the current waveform, which produces spectral components that were not present when the machine was healthy. IEEE Std 1415-2006, EPRI TR-108773, and foundational work by Schoen et al. (IEEE Transactions on Industry Applications, 1995) have catalogued the fault signatures for most common motor types.
What is interesting about applying this technique to data center infrastructure is not the physics (which transfers directly) but the operational constraints that make the standard industrial approach unworkable, and what changes when you remove those constraints.
The industrial ESA workflow and its limitations
In a petrochemical plant or power generation facility, ESA works like this: a reliability engineer brings a portable analyzer to the motor, clamps a current transformer around one phase conductor, collects 30–120 seconds of steady-state waveform data at 10–50 kHz, and runs the data through spectral analysis software. The analysis identifies characteristic fault frequencies:
- Broken rotor bars: Sidebands at f_line * (1 +/- 2s), where s is the per-unit slip. For a 4-pole, 1,785 RPM motor on 60 Hz, s = 0.0083, yielding sidebands at 59.0 Hz and 61.0 Hz.
- Bearing defects: Frequencies calculated from bearing geometry (BPFO, BPFI, BSF, FTF). For a typical 6310 bearing at 1,785 RPM, the outer race defect frequency is approximately 88.5 Hz.
- Stator winding faults: Negative sequence current component, typically measured as I2/I1 ratio. Healthy motors run below 1–2%.
- Eccentricity: Characteristic frequencies at f_line * (1 +/- (1-s)/p), where p is the pole pair count. For our 4-pole, 1,785 RPM motor, the formula simplifies to f_line +/- f_rotation, evaluating to sidebands at 30.25 Hz and 89.75 Hz (where f_rotation = 1,785 / 60 = 29.75 Hz).
This approach has prevented catastrophic failures across heavy industry for 40 years. The limitation is not the physics. The limitation is that someone has to physically go to the motor, attach probes, wait for steady-state operation, and then analyze the data. Most industrial programs test critical motors quarterly. Between tests, degradation proceeds undetected.
For a petrochemical plant with 50 critical motors, quarterly testing is manageable. For a data center with 200+ UPS modules, 80+ CRAC/CRAH compressor motors, cooling tower fans, chiller compressors, and PDU transformers, the math does not work. The labor cost of portable testing at this density would exceed the cost of the equipment being tested.
What translates and what does not
The spectral analysis principles from industrial MCSA translate directly to data center equipment, but the equipment types and their fault signatures differ.
| Equipment | Relevant fault modes | Spectral indicator | Industrial MCSA coverage |
|---|---|---|---|
| UPS rectifier (thyristor, 6-pulse) | Capacitor ESR rise, gate firing drift | 5th/7th harmonic ratio change, commutation notch width | Not applicable (MCSA is motor-specific) |
| UPS rectifier (IGBT, PWM) | DC bus capacitor degradation, IGBT degradation | Switching frequency sidebands, DC ripple spectral content | Not applicable |
| CRAC/CRAH compressor | Bearing wear, refrigerant loss, scroll wear | Inrush envelope change, steady-state slip frequency shift | Applicable (these are motors) |
| Cooling tower fan | Bearing wear, belt tension, blade imbalance | Rotational frequency sidebands, 2x running speed | Applicable |
| Dry-type transformer | Winding insulation degradation, core saturation | High-order harmonic transfer function change, 3rd harmonic increase | Not directly (transformers are not motors) |
| Chiller compressor | Bearing wear, refrigerant charge, valve seat | Modulated current at compressor RPM, inrush profile | Applicable |
The key insight is that MCSA as practiced in industry covers about 40% of the equipment types in a data center. The technique needs to be generalized from "motor current signature analysis" to "electrical signature analysis" (ESA). In data center contexts, ESA refers to continuous spectral analysis of any load's current waveform, extending beyond rotating machines to rectifiers, transformers, and switching power supplies.
For rectifiers and transformers, the fault physics is different from motors but equally well-understood. Capacitor degradation in a rectifier changes commutation timing, which redistributes harmonic energy. Transformer winding degradation alters the frequency-dependent impedance, which changes how harmonic currents transfer from primary to secondary. Both produce measurable spectral changes.
The resolution threshold
Industrial MCSA equipment samples at 10–50 kHz. Most data center BMS systems sample at 1 Hz or slower, polling meters rather than capturing waveforms. The gap between these two numbers explains why data centers have not adopted ESA.
At 1 Hz, you get one power measurement per second. You can trend average power, voltage, and current. You cannot resolve any frequency content above 0.5 Hz (Nyquist). The entire harmonic spectrum, and all the diagnostic information it contains, is invisible.
At 8 kHz (8,000 samples per second), you resolve 133 samples per cycle on a 60 Hz system. Per Nyquist, this captures frequency content up to 4 kHz, which corresponds to the 66th harmonic. In practice, the diagnostically useful range for data center equipment extends through roughly the 25th harmonic (1,500 Hz). 8 kHz sampling provides comfortable margin above that.
The resolution question is not theoretical. Here is what you can and cannot see at different sampling rates:
| Sampling rate | Nyquist limit | Highest resolvable harmonic | Can detect bearing faults | Can detect capacitor degradation | Can detect winding faults |
|---|---|---|---|---|---|
| 1 Hz (BMS) | 0.5 Hz | None | No | No | No |
| 1 kHz (basic digital submeters) | 500 Hz | 8th | No | Partially (5th and 7th harmonics resolvable) | No |
| 8 kHz (waveform capture) | 4,000 Hz | 66th | Yes | Yes | Yes |
| 50 kHz (portable analyzer) | 25,000 Hz | 416th | Yes | Yes | Yes (plus partial discharge) |
8 kHz is not as high as a portable analyzer, and that matters for one specific diagnostic: partial discharge detection in transformer insulation, which produces spectral content in the 50–500 kHz range. At 8 kHz, we cannot see partial discharge directly. We can see its downstream effects on the harmonic transfer function, but with lower sensitivity and longer detection latency than a dedicated PD instrument. This is a real limitation.
For every other degradation mode in the table, 8 kHz is sufficient. And it has an advantage that no portable analyzer can match: it runs continuously.
What continuous monitoring changes
The difference between quarterly measurement and continuous measurement is not quantitative. It is qualitative. They answer different questions.
Quarterly ESA answers: "Is this motor healthy right now?" Continuous ESA answers: "How is this motor's health changing over time?" The second question is more useful because degradation is a process, not an event.
In field validation at a Fortune 50 telecom deployment, continuous harmonic trending identified units exhibiting gradual spectral changes consistent with capacitor degradation, with a 21-day advance warning window, while zero BMS alarms had triggered. The absolute THD values remained within manufacturer specifications. The trend was the diagnostic indicator, not the instantaneous value.
Continuous monitoring also enables fleet-level statistical analysis. When you have harmonic baselines for hundreds of identical rectifiers, you can compute fleet distributions for every spectral feature. A single rectifier whose 7th harmonic is drifting above the 95th percentile of the fleet distribution is statistically anomalous, even if its absolute value is unremarkable. This population-based approach catches degradation that unit-level threshold comparison misses.
Complementary detection, not replacement
ESA detects faults through their electrical manifestation. Vibration analysis detects faults through their mechanical manifestation. Thermal imaging detects faults through their thermal manifestation. These are different physical observables of the same underlying degradation process, and they become detectable at different points in the degradation timeline.
For bearing wear in a CRAC compressor motor, the typical detection sequence is:
- Electrical (ESA): Current sidebands at bearing defect frequencies appear. Detection possible at early-stage spalling. Lead time: 4–12 weeks before functional failure.
- Vibration: Acceleration peaks at bearing defect frequencies. Detection possible at moderate spalling. Lead time: 2–6 weeks.
- Thermal: Elevated bearing housing temperature. Detection possible at advanced degradation. Lead time: days to 2 weeks.
- BMS alarm: Current or temperature exceeds threshold. Lead time: hours to days, sometimes zero.
ESA catches the fault earliest because electrical impedance changes are the most sensitive indicator of mechanical condition change. But this sequence is idealized. Real degradation does not always follow the textbook timeline. We have seen bearing faults that produced clear vibration signatures before any electrical change was detectable. Relying on any single modality is a mistake.
The useful framing is not "ESA is better than vibration analysis." It is: "Each measurement modality provides an independent observation of equipment health. More independent observations reduce the probability of missing a developing fault." For data center operators, adding the electrical dimension to an existing thermal and threshold-based monitoring stack meaningfully extends the aggregate warning window.
Open questions
Three aspects of ESA applied to data center infrastructure remain active areas of investigation.
Rectifier topology evolution. Newer UPS designs use active IGBT front-ends with PWM switching rather than thyristor-based 6-pulse or 12-pulse rectifiers. The harmonic profiles of IGBT rectifiers are structurally different (lower low-order harmonics, higher switching-frequency content), and the degradation signatures are less well-characterized in the literature. We are building the baseline dataset for these newer topologies, but the sample sizes are small compared to the thyristor-based fleet.
Environmental confounds. Ambient temperature, input voltage quality, and loading level all affect harmonic content independently of equipment health. A 5% increase in input voltage THD from the utility can shift a rectifier's output harmonics in ways that resemble capacitor degradation. Proper baselining requires multivariate normalization, and we are not confident that our normalization captures every significant confound.
Diagnostic specificity for compressors. Motor bearing faults produce well-characterized spectral signatures with extensive literature validation. Refrigerant charge loss, scroll wear, and valve seat degradation produce subtler, less well-documented spectral changes. Our detection models for these non-bearing compressor faults have higher false positive rates (roughly 3x) than bearing fault detection, and we need more field-validated failure cases to improve them.
The physics of electrical signature analysis is well-established. Extending it from quarterly industrial motor testing to continuous, fleet-wide data center monitoring introduces engineering challenges that are partly solved and partly open. We publish these limitations because they matter for setting realistic expectations about what ESA can deliver today.
Key takeaways
Talk to an engineer about applying continuous electrical signature analysis to your infrastructure.