Physics-Based Equipment Health Monitoring: Catching Degradation Before It Looks Like a Failure

Most equipment monitoring programs in Oil & Gas are built around thresholds: vibration above X mm/s, discharge temperature above Y°F, bearing temperature trending past Z. These thresholds work, but only after a problem has already progressed far enough to move a raw sensor reading outside its normal range. By the time a threshold trips, the failure mode is often already mechanically established.

The earlier signal exists—it's just not visible in the raw measurement. It's visible in the relationship between measurements, and finding it requires a physics model of how the equipment is supposed to behave.

Why Threshold-Based Monitoring Misses Early Degradation

A centrifugal compressor doesn't fail in a single dimension. Discharge pressure, discharge temperature, suction pressure, flow, and shaft speed are all coupled through the machine's thermodynamic and mechanical behavior. A gradual impeller fouling issue might leave every individual measurement within its normal operating band for months, while the relationship between flow and head—the actual performance curve—has already shifted.

Threshold monitoring, and even most statistical anomaly detection, treats each tag independently or looks for deviation from historical patterns in the raw signals. Neither approach is built to notice that the compressor is still hitting its discharge pressure setpoint by working harder than it should have to. The machine compensates, the alarms stay quiet, and the degradation compounds until something breaks in a way that can no longer be compensated for.

Physics-Based Residuals: A Different Signal

The alternative is to build a model of expected behavior from first principles, then monitor the residual—the gap between what the physics predicts and what the sensors report—rather than the raw signals themselves.

For a compressor, this means encoding the polytropic relationship between suction and discharge conditions, the manufacturer's performance curve relating head and flow to shaft speed, and the thermodynamic relationship linking compression ratio to discharge temperature. Given current suction pressure, suction temperature, flow, and speed, the physics model predicts what discharge pressure and discharge temperature should be. The residual is the difference between that prediction and the actual reading.

A healthy machine produces a residual that hovers near zero, bounded by sensor noise and normal manufacturing tolerance. A machine with fouling, wear, or a developing mechanical fault produces a residual that drifts—often weeks or months before any individual tag would have crossed a threshold. The residual isolates the part of the signal that thresholds and single-tag anomaly detection cannot see: the mismatch between the physics and the observation.

Building the Health Model

1. Performance baseline. Start from OEM performance curves and thermodynamic relationships (polytropic head, efficiency, compression ratio vs. temperature rise). Where OEM data is incomplete, calibrate the baseline against a period of known-healthy field operation.

2. Residual generation. Compute the physics-predicted value for each key output variable at every scan cycle, and track the residual against the measured value in real time.

3. Fault signature association. Different degradation mechanisms produce different residual patterns. Impeller fouling shows up primarily as an efficiency-side residual—more power draw for the same head. Bearing wear shows up in vibration spectral content correlated with shaft-speed harmonics, layered on top of the thermodynamic residual. A developing seal issue shows a distinct pressure-differential signature. Building a library of these signatures, grounded in the failure physics rather than in historical labeled examples, lets the system attribute a drifting residual to a probable cause rather than just flagging "anomaly."

4. Trend and threshold on the residual, not the raw signal. The residual itself becomes the monitored quantity. Because it has already removed the expected variation due to load, ambient conditions, and normal operating envelope changes, a much smaller drift is statistically significant—and much smaller drift is visible much earlier.

Why This Beats a Pure Data-Driven Approach

A purely data-driven anomaly detector needs examples of failure to learn what failure looks like—or it needs to assume that "different from history" means "bad," which produces false positives every time the machine is legitimately operated outside its usual envelope (a rate change, a startup after turnaround, a different gas composition). Physical failure modes in rotating equipment are also relatively rare events; there is rarely enough labeled failure data at a single site to train a reliable classifier.

The physics-based residual approach doesn't need failure examples. It only needs a correct model of healthy behavior, which can be built from engineering first principles and OEM data before a single failure has ever occurred at that asset. It also doesn't confuse a legitimate operating envelope change with degradation, because the physics model already accounts for the load- and condition-dependence of expected performance.

From Detection to Prognosis

Once a residual is being tracked reliably, the same framework extends naturally to prognosis. A residual that is drifting linearly can be extrapolated to estimate when it will cross a threshold that historically correlates with intervention need—giving a remaining-useful-life estimate grounded in the physical degradation mechanism rather than a generic survival curve fit to a fleet average.

This matters operationally: a maintenance planner can schedule a compressor overhaul during an already-planned turnaround three months out, rather than reacting to an alarm that fires two weeks before a forced outage. The value isn't just avoiding failure—it's converting an unplanned event into a planned one.

What Changes for Reliability Engineers

The practical shift is one of trust as much as technology. A reliability engineer who sees a residual chart—predicted discharge temperature vs. actual, diverging over eight weeks—can go look at the fouling mechanism and the operating history and understand exactly why the model is flagging concern. That's a fundamentally different conversation than being told a black-box anomaly score crossed 0.87. The physics gives the alert a mechanism, and a mechanism is something an engineer can act on with confidence.

---

WellBeyond.ai builds physics-based health monitoring for rotating and reciprocating equipment across upstream and midstream operations. Talk to us about the assets you're trying to keep out of unplanned downtime.