Maintenance Log: Using Regression Analysis To Improve Data Trending Sensitivity

Fig. 1. One of the plant’s two condensate pumps

Trending data to determine the health of equipment can be challenging to say the least. No single technology or equipment performance parameter can tell you everything you need to know about the equipment’s condition. When you are involved in this type of activity, the biggest mistake you can make is to miss an adverse trend on critical equipment. To be effective (i.e. ensuring there are no unpredicted equipment failures), you must integrate predictive diagnostics with trending equipment parameters and time-based maintenance. This article describes how regression analysis, a statistical tool used to investigate the functional relationship between two or more variables, was used to monitor the condition of a condensate pump at a nuclear power plant.

The challenge(s)
Fig. 1 shows one of the plant’s two 50% capacity condensate units used to pump condensate from the main condenser to the feed water system. It’s an Ingersoll- Rand (Model 36APKD) three-stage, vertical centrifugal with a Siemens-Allis 6.9 kV – 2000 hp air cooled induction motor. Pump and motor are coupled with a flanged hub-style coupling.

The pumps and motors on these units are monitored by predictive technologies that include vibration analysis and oil analysis. Performance parameters, such as flow, pressure and bearing and stator temperatures, also are trended.Vibration data is collected monthly on the motor housing, as shown in a simplified diagram in Fig. 2.

The vibration data is normally trended versus time, comparing the data with established limits. During routine monitoring on April 21, 2005, the vibration was observed to increase 0.04 ips on the ‘B’ pump motor housing at the lower motor bearing. Fig. 3 is the trend chart for the overall vibration on the motor lower bearing. As indicated in the trend chart, the change in vibration was small and within the established Alert Limit of 0.3 ips. All other monitored parameters on the pump and motor were normal. Ten days later, though, on May 1, 2005, the motor shaft failed, causing a plant shutdown.

The motor on the ‘B’ pump was installed in October 2001- as part of a time based motor refurbishment strategy. The root cause analysis that was conducted on the failed shaft determined it sheared due to circumferential crack that developed near the top of the motor coupling hub. The crack had propagated from a subsurface defect (i.e. lack of weld fusion) in the shaft material. The shaft had an inadequate weld repair many years earlier. The postrepair examination of the shaft did not detect a sub-surface flaw in the weld.

During the course of the root cause investigation, the motor was replaced with a spare and the plant was returned to full power operation. The subsequent extent of condition evaluation identified the root cause also applied to the ‘A’ condensate pump motor that was in operation. The same weld repair was conducted on this motor shaft. The weld repair was performed by the same company and the same personnel. The problem, though, was how to monitor this pump motor for any indication of a developing shaft crack while ensuring safe and reliable plant operation until the next opportunity to replace the motor in April 2006. This was quite a challenge, especially in light of the fact that the routine vibration monitoring had identified, but not diagnosed, a shaft crack on the ‘B’ pump motor.

Those who have monitored equipment with vibration monitoring know that diagnosing a shaft crack on a vertical pump can be very difficult—particularly with the type
of limited vibration monitoring that was available on the motor in this case study (i.e. housing sensors). Fortunately, the Electric Power Research Institute (EPRI) is conducting tests on torsional vibration monitoring for vertical rotating equipment to improve the capability of shaft crack detection. As indicated in EPRI’s research, notable changes in vibration occur only after the crack has propagated ~50% through the shaft. At this point, the shaft stiffness decreases and the vibration changes in magnitude and phase at 1X and 2X operating speed. Periodic monitoring of the equipment on a monthly frequency, however, was not adequate to identify this failure mechanism. As indicated by the small change in vibration (0.04 ips) prior to the ‘B’ pump motor failure, the monitoring would need to be very sensitive to any changes.

A previous Maintenance Log article (“Get ‘Control’ of Your Data Trending,” pgs. 56-59, Maintenance Technology, November 2005), discussed using statistical control charts to trend equipment performance provides an especially sensitive method for identifying equipment degradation. This method also was employed in monitoring the ‘A’ pump.

As shown in Fig. 4, the vibration on the ‘A’ pump began to trend with a step change increase. The change in data was a concern because it was not initially known if this was a crack propagating in the shaft or some other unknown effect influencing the data. If it were a crack, prompt action would be crucial to prevent a catastrophic motor failure- and subsequent plant shutdown. If the data was being influenced by some other effect or was dependent on another parameter, it needed to be identified and accounted for so it would not hinder personnel in diagnosing an actual shaft crack. The trend also presented a complication in the use of control charts. Data dependence violates a fundamental rule in applying control charts that requires the data to be ‘in statistical control’ or independent of any other influence.

Investigation
An investigation into the data trend was initiated by site personnel. The motor component engineer and vibration analyst observed a perceived relationship with ambient temperature.

For example, when the area temperature conditions were warmer, higher vibration readings were observed. To validate this perception, a regression analysis was conducted on the pump historical vibration and the various temperature parameters that are related to pump motor operation. (Refer to the simplified model shown in Fig. 2). The temperatures considered included: service water temperature (used to cool the pump bearing lube oil); motor stator temperature; bearing temperature; ambient temperature.

As noted in the simplified model, the stator, service water and bearing temperatures all varied with ambient temperature. This indicated that the ambient temperature was the independent variable of interest.As the regression analysis was conducted, the best correlation was obtained with ambient temperature. Additionally, a 24-hr. average ambient temperature was used, since there was no practical means to measure the temperature right at the pump—and the structural influence from temperature would be a lagging effect.

Fig. 5 shows the regression of pump vibration to the 24-hr. average ambient temperature from May through November 2005. The May through July data is shown by the dark diamond symbols on the scatter chart. As the ambient temperature increases the vibration increases, indicating a direct relationship. The ‘Goodness of Fit’ statistic of 0.76 indicates an acceptable correlation between these parameters.With this relationship identified, the step increase in the vibration was validated to be a result of seasonal changes in ambient temperature and not a developing shaft crack. With this relationship identified, the regression line was used in the monitoring of the motor for a developing shaft failure.

As the pump motor was monitored through the summer of 2005, the data followed the same regression line until late July, when the data started to deviate from this known relationship. The late July and August data is shown by the purple square symbols in the scatter chart. Once again, an investigation was conducted into the data deviation. A walk-down of the equipment identified the upper bearing reservoir as being two inches below the fill mark on the local site glass. A small oil leak that had developed and gone unnoticed by Operations personnel had slowly decreased the oil level. It was later estimated that about 1/2 cup of oil a day had leaked from the 10-gal. reservoir. Since it was a small leak, the impact on bearing temperature only amounted to ~ 5 F over a two month period. This change went undetected with normal temperature trending (i.e. plotting temperatures versus time). Once oil was added to the reservoir, the vibration and temperature relationship returned to normal, as shown by the September data (triangle symbols) in Fig. 5.

As the pump motor was monitored through the fall of 2005, the vibration began to deviate again in October. The deviation from the known relationship is shown by the circle symbol in Fig. 5. The oil level was checked and found to be within an acceptable operating range.Further evaluation was conducted on the bearing temperature. The data of bearing temperature versus service water temperature was plotted on the XY scatter chart shown in Fig. 6.

The bearing temperatures were not following the known relationship with service water temperature. This can be observed through comparison of the 2004 and 2005 data in Fig. 6.There was a ~10 F lag in temperatures from 2004 to 2005. As the seasonal temperatures decreased, there was not the same corresponding decrease in bearing temperature. Further investigation determined that this ‘lagging effect’ was caused by degrading lube oil cooler performance.High magnesium levels in the lake water, used by the service water system, were plating out on the lube oil cooler tubes. The fouling reduced the lube oil cooler performance and resulted in the higher bearing temperatures,which indirectly affected the motor vibration.

The conclusion(s)
The use of control charting and regression analysis provided the sensitivity that was required for monitoring the motor condition of the nuclear power plant’s condensate pump. As noted, the control chart coupled with regression analysis identified equipment performance issues that would normally have gone undetected with normal trending data versus time. This approach, coupled with conservative decision- making, provided plant personnel and management a reasonable assurance that a developing shaft crack could be detected and acted on before catastrophic failure and a resulting plant trip. As a result, the plant was operated safely and reliably under the root cause extent of condition on the ‘A’ pump motor.

Epilogue
The ‘A’ motor was operated until April 2006 when the plant was shut down for a planned refueling outage. The subject motor was then replaced with a spare. MT

Daryl Gruver is a senior consultant with First Energy. Formerly a supervisor of Component Engineering at Progress Energy’s Shearon Harris facility, he has a B.S. in Nuclear Engin-eering from Penn State and an M.S. in Nuclear Engineering from the University of Cincinnati.He holds a Level II ASNT certification in Vibration Analysis and Thermography. Phone: 440-280-5934; e-mail: dgruver@firstenergycorp.com