Predictive Maintenance for CRAC and CRAH Units [/prɪˈdɪktɪv ˈmeɪntənəns fər ˈkrak ənd ˈkrah ˈjunɪts/] n - Computer room air conditioning and computer room air handling units are the workhorses of precision cooling. They run without rest, oft in redundant pairs, and their failures announce themselves subtly before they turn acute. A vibrating bearing writes its spectral signature weeks before it seizes. A fouled filter raises a measurable pressure before the airflow collapses. A degrading compressor shows its distress in discharge temperature and current draw long before it trips.

Bearings, Read by Vibration

Rotating equipment vibrates at characteristic frequencies set by the geometry of bearings, gears, and imbalance. For a bearing of $N$ rolling elements, ball diameter $d$, pitch circle diameter $D$, and contact angle $\varphi$, the ball pass frequency on the outer race is:

$$f_{\text{BPFO}}=\frac{N}{2}{f}_r\left(1-\frac{d}{D}\cos \varphi \right)$$

where $f_r$ is the shaft rotation frequency. Early damage appears as a small rise in amplitude at this frequency and its harmonics. A healthy bearing shows a clean spectrum; a damaged one shows sidebands.

In practice the exact geometry is not always known. The commoner course is to track the overall RMS vibration level, or the kurtosis of the signal, over time. Kurtosis is sensitive to impulsive events (a cracked roller striking a race), which precede the steady degradation visible in RMS.

$$\text{Kurt}(x)=\frac{\mathbb{E}[(x-\mu {)}^4]}{{\sigma }^4}$$

A rising kurtosis trend in rolling-element bearings foretells failure. Control charts detect when it has moved beyond its baseline distribution. See Anomaly Detection for Infrastructure Sensors for the statistical machinery.

The Compressor’s Confession

A reciprocating or scroll compressor in a CRAC unit signals degradation through changes in:

• Suction and discharge pressure ratio (compression efficiency).
• Motor current draw at fixed load.
• Discharge temperature above the adiabatic ideal.

Each of these may be modeled as a time series. The natural baseline is a regression on ambient conditions and load:

$$y_t={\beta }_0+{\beta }_1{T}_{\text{amb}}+{\beta }_2{Q}_{\text{load}}+{ϵ}_t$$

The residual ${ϵ}_t$ captures departures from expected performance. A slowly rising trend in the residual marks degradation; a sudden jump may mark a refrigerant leak or a failed valve.

An EWMA control chart on the residuals is a simple, legible monitor. It detects a shift in the mean of the residual process:

$$Z_t=\lambda {ϵ}_t+(1-\lambda )Z_{t-1}$$

An alert fires when $Z_t$ passes a control limit calibrated to the desired false-positive rate. See EWMA and RiskMetrics for the method and Tuning alert thresholds for setting limits without drowning in nuisance alarms.

The Fouling of Filters

A clogged filter strangles the airflow, and the diagnostic signal is the differential pressure across the filter media. That pressure rises as the filter loads, at a rate set by the particulate concentration in the air and the volume of flow.

A simple model:

$$\Delta P(t)=\Delta P_0+k{\int }_0^t\dot{V}(\tau )C(\tau )d\tau$$

where $\dot{V}$ is volumetric airflow and $C$ the particulate concentration. In practice $C$ goes unmeasured, so the model reduces to a cumulative airflow integral. When the predicted pressure crosses the replacement threshold, the filter change is scheduled before the pressure can trouble the airflow.

Remaining Life and the Order of Work

All three degradation signals, vibration, compressor performance residuals, and filter differential pressure, may be joined into one estimate of remaining useful life. The standard approach weds a Weibull survival model of historical failures to a degradation model that advances the survival clock according to present condition.

$$S(t)=\exp \left(-{\left(\frac{t}{\lambda }\right)}^{\kappa }\right)$$

where $\lambda$ and $\kappa$ are scale and shape parameters estimated from failure history. The condition-based update uses the degradation signal to shift the effective age. This feeds directly into Remaining Useful Life Estimation.

Scheduling maintenance across a fleet of units becomes an optimization: sequence the planned outages to hold N+1 redundancy at all times while minimizing total travel and parts cost. The predictive model furnishes the input (expected time to failure for each unit), and the scheduler the output (the optimal maintenance date for each unit, given the predicted failure distribution and the redundancy constraint).