×

 

Using Vibration Analysis to Test for Bearing Wear

Azima DLI

This article discusses an example of an outer race-bearing defect on a 1,250-ton (4,400-kilowatt) air-conditioning chiller. The defect was detected using off-the-shelf portable vibration analysis hardware and software.

The machine is a single-stage centrifugal compressor with a rolling element thrust bearing on the motor free end and a sleeve bearing on the motor drive end. This article discusses methods used to diagnose and trend the bearing defect using vibration spectra as an indicator of the bearing condition.

The diagnostic methods covered are universally applicable to other types of machines (pumps, fans, gears) with rolling contact bearings.

Vibration Analysis Tools

Tools used to measure vibration have improved significantly in the past 25 years. The sensor of choice for most vibration data collection on industrial machinery is an accelerometer. As the name implies, the output is proportional to acceleration; however, it is normally integrated to display in units of velocity. All data shown in this case history was collected using a triaxial accelerometer cluster shown in Figure 2.

Figure 1. Chiller showing the vibration test location on the motor bearing housings.

Figure 2. Sketch of a triaxial accelerometer cluster (left) that is stud-mounted to the bronze disk mounting pad (bottom right). The cluster contains three accelerometers, each mutually perpendicular to the others. Vibration data collection is activated using a bar code that automatically sets up the data collector, collects and then stores the data.

The notched mounting pad is adhered rigidly to the motor bearing housing, and the triaxial cluster is stud-mounted, allowing all three axes to be collected simultaneously.

The bar code and mounting configuration allow accurate and repeatable data to be collected every time, regardless of the skill of the person doing the data collection. As long as the test conditions are relatively consistent, the test data also should be consistent if the mechanical condition remains unchanged.

The signal processing techniques used by a vibration analyst vary based on the level of detail desired. In routine predictive maintenance programs, two types of data typically are collected.

Overall level: The overall level is a measure of the total vibration amplitude over a wide range of frequencies. An overall vibration measurement, also called a broadband level, is a single value that is relatively easy and cheap to collect, process, analyze and trend.

Narrowband: Using a Fast Fourier Transform (FFT) algorithm, the vibration signal is resolved into a number of discrete frequencies and displayed as a two-dimensional spectral plot of amplitude vs. frequency. All of the triaxial spectra shown in this case history have a frequency scale with 800 lines of resolution. In other words, each spectrum is composed of 800 individual peaks defining an amplitude at a specific frequency.

Vibration Acceptance Criteria

What is an acceptable level of vibration for the 1,250-ton (4400-kW) chiller? The chiller manufacturer uses a pass/fail broadband specification of 0.25 inches/second (peak). Vibration tolerances or specifications typically are expressed as either absolute or relative criteria.

The IRD General Machinery Vibration Severity Chart, summarized in Table 1, is the most common example of absolute criteria and is a good guide for engineers who do not have any historical data on a machine. Each increment of machinery condition is represented by change of a factor of two in the vibration velocity.

A third column provides an easy conversion between inches/second (peak) and the commonly used logarithmic unit of velocity decibels (VdB). Note that 6 dB also represents a change by a factor of two, so each machinery condition band is represented by a 6 dB differential. 

 

Machinery
Condition

Velocity
(in./speak)

Velocity
(VdB)

Very Rough

V > 0.628

VdB > 121

Rough

0.314 < V < 0.628

115 < VdB < 121

Slightly Rough

0.157 < V < 0.314

109 < VdB < 115

Fair

0.0785 < V < 0.157

103 < VdB < 109

Good

0.0392 < V < 0.0785

97 < VdB < 103

Very Good

0.0196 < V < 0.0392

91 < VdB < 97

Smooth

0.0098 < V < 0.0196

85 < VdB < 91

Very Smooth

0.0049 < V < 0.0098

79 < VdB < 85

Extremely Smooth

V < 0.0049

VdB < 79

Table 1. The IRD General Machinery Vibration Severity Chart is an improvement on a pass/fail criteria such as that used by the chiller manufacturer because it provides a graduated scale of machine condition. The tolerances in the chart are for frequencies between 1.6 and 1667 hertz and apply to most rotating machinery.

While absolute criteria are fine for acceptance testing, more sensitive criteria are needed for making accurate diagnoses as part of a predictive maintenance program. A relative criterion is best if your goal is to diagnose machine health accurately.

Relative criteria are defined by averaging vibration measurements taken on several identical machines, all tested under similar operating conditions. For this case history, the acceptance criteria were constructed by averaging spectral measurements from 12 identical chillers.

The bronze disk mounting pads were mounted in the same location on each machine and the data were collected with the chiller operating at about 80 percent rated load.

The test spectra were developed into average baseline spectral data after manual review of the data, so each test showed the machines to be in reasonably good condition. Due to variations of test data among machines in the sample, standard deviations (sigma) were calculated.

Then, an average plus one sigma spectral mask was calculated. Statistically, about 85 percent of the vibration amplitudes should fall below the average plus one sigma mask. Unlike vendor defined pass/fail criteria or absolute criteria such as the IRD chart, the relative criteria allow the machines themselves to define what is acceptable.

Figure 3 is an example of one axis of the relative acceptance criteria used in this case history. Figure 4 shows the average and average plus sigma level for all three axes for both the low- and high-range data. If one or more peaks in an 800-line spectrum exceeds the average plus sigma amplitude criterion, that fact alone does not indicate a significant problem (see the section on diagnostic strategies).

    

Figure 3. The relative acceptance criteria was generated specifically for the 1,250-ton chiller by averaging spectra from 12 identical chillers. The average (bottom spectrum in blue) and average plus one standard deviation (top spectrum in red) criteria shown are for the radial axis. The IRD Severity Chart criteria are superimposed for comparison.

Figure 4. Average baseline data from two separate facilities: Plant A (top 6 spectra) and Plant B (bottom 6 spectra). Each plant has 18 identical model chillers. The running speed of the Plant B data reflects a different average running speed (2,982 rpm), which reflects in input line frequency of 50 Hz common to Europe. This average and average plus sigma data is from the free end motor bearing.

Anti-Friction Bearing Problems

Bearing problems are some of the most common types of faults diagnosed by vibration analysis programs. In the early stages, a defective bearing will produce vibration components with frequencies that are not multiples of shaft rotation rate.

The exact values of these non-synchronous frequencies are based on bearing geometry. They can be calculated if the race and roller dimensions are known. In practice, most facility engineers do not document the manufacturer and model number for bearings in machinery, and so must rely on other methods to determine bearing frequencies. A set of rules for determining approximate values of bearing tone frequencies follows:

Ball Pass Frequency Outer Race (BPFO)
= # of rollers x shaft speed x 0.4

Ball Pass Frequency Inner Race (BPFI)
= # of rollers x shaft speed x 0.6

Fundamental Train Frequency (FTF)
= speed x 0.4 

The chiller manufacturer stated that the ball bearing in this case is an SKF 7318. The bearing table confirms that 4.9xM and 7.1xM are the outer and inner ball pass frequencies, respectively. These frequencies are presented as orders, where the suffix “xM” represents “times motor shaft rotation rate”. Looking at the frequencies and the rules of thumb, it is evident that this bearing has 12 rolling elements.

Ball-bearing wear becomes increasingly evident as harmonics (integer multiples) of these ball-pass frequency peaks occur in the vibration spectra. Depending on the nature of the defect, there also may be strong 1xM sidebands around the bearing tones or their harmonics. Extreme bearing wear creates an abnormally high noise floor in the high-range spectra between about 70 and 100 times shaft rotation rate (70xM to 100xM).

Commonly Used Diagnostic Strategies

When evaluating bearing wear using vibration analysis, a consistent method of collecting and analyzing the data is necessary. Broadband vibration meters normally cannot detect bearing wear until the later stages. As a result, the maintenance technician has little warning to plan a repair, and risks catastrophic failure.

Diagnostic techniques that rely on high-frequency phenomena such as shock pulse or ultrasonic energy are very effective at early detection of bearing wear. However, they typically lack the ability to detect the many machinery faults that manifest themselves with vibration at lower frequencies.

Triaxial narrow band spectral analysis using relative acceptance criteria can accurately provide early warning of bearing defects. It also can be used to diagnose a wide range of other machinery problems such as imbalance, misalignment, impeller clearance problem, looseness and resonance. In this case, the third technique is used.

The narrow band data required for analysis includes two frequency ranges of high-resolution (800 lines) triaxial spectral data measured at a single point on a rigid part of the bearing housing. The low-range data is typically 0 to 10 orders of shaft speed where the fundamental bearing tones are easily seen.

The high-range data is typically 0 to 100 orders of shaft speed, where the bearing tone harmonics and high-frequency noise floor are seen. While this may seem like an overkill of data for a single location, the mounting method and the bar code allow for simple and efficient data collection. Furthermore, processing the data on a PC is fast and automatic.

The narrow band fault model for a bearing defect, as taught in most basic vibration analysis courses, includes the following four conditions:

  1. Harmonic series of peaks with a non-synchronous fundamental frequency.
  2. 1xM and/or fundamental train frequency (FTF) sidebands around any of the peaks in (1).
  3. Increased high-frequency noise floor magnitude.
  4. Harmonics series of the shaft rotation rate caused by excessive bearing looseness.

Vibration Analysis (Human vs. Computer)

The sheer quantity of narrow band data needed to accurately diagnose bearing defects requires that all the routine data reduction and logic functions are accomplished by a personal computer with an expert system. It is not time-economical for a human analyst to accomplish all the routine tasks that a computer can do in mere seconds.

The first step is to order normalize the spectra with respect to the shaft rotating speed. The next step is to extract the amplitudes for the principal forcing frequencies and other peaks in the test data. The third step is to use the fault model described earlier to identify the pattern for a bearing defect. The last step is to determine the severity of the bearing defect, given that the pattern exists.

The expert system logic used to identify the fault and its severity was created as part of an empirical process whereby the automated diagnosis on a large population of machines was compared to the diagnosis that a vibration expert would make on those same machines.

The expert logic rules are, therefore, a reflection of exactly how a vibration expert would analyze the data. Because of this, an expert system can be no better than the vibration analyst who creates it. Because all the steps outlined in the previous paragraph are routine and repetitive, a modern computer can rapidly perform all four analysis steps in a few seconds per machine. The expert system then generates a text-based report for the chiller motor bearing wear as shown in Figure 5.

Figure 5. Computer-generated Diagnostic Results for Chiller 4

Manual Review of the Spectral Data

Figure 6 shows the motor free end data collected on March 25 on Chiller 4. The Avg + sigma mask (red), are superimposed for comparison. The x-axis is scaled in units of orders where the order “1” represents the rotational rate of the motor.

Note the prominent non-synchronous peaks in the axial low range data at 4.9xM and 7.1xM. These peaks correspond to the ball bearing pass frequencies for the motor bearing. In the high-range spectra, harmonic markers are placed on each peak of the harmonic series with 4.9xM spacing.

Note that the harmonic series has strong amplitudes in all three axes of the high range while only prominent in the axial direction in the low range data. The high-frequency noise floor is about the same or lower than the one sigma mask telling us that the bearing wear is not to a point of near failure; however, the strong harmonic series implies that a defect is present.

Figure 6. Triaxial vibration spectra for the motor free end bearing on Chiller 4. The three spectra on the left are low-range data and the three spectra on the right are high-range data. The marked peaks in the high-range data represent a 4.9xM spacing that is a BPFO harmonic series.

Prognostics

There is little doubt whether a bearing fault exists. The machine operators can identify this condition by sound and touch. A more important question is: How much remaining life does this bearing have? The clever vibration analyst will not make a conclusion until a trend has developed and several data sets are available. For each diagnosis, the expert system calculates a severity score based on three things.

  1. The number of peaks that support the diagnosis.
  2. The amount by which the criterion is exceeded for each peak in (1).
  3. The absolute amplitude of each peak in (1).

A trend of the severity score is a good indicator of machine health. For each diagnostic rule template, the severity score is mapped into a severity scale that includes slight, moderate, serious and extreme.

The algorithm used for determining severity was derived empirically for each diagnosis based on a large amount of test results that had been manually analyzed by vibration experts. If the indications (bearing tones, harmonics, sidebands and noise floor) of bearing wear increase over time, the expert system will recognize this fact and the severity assigned to the bearing wear diagnosis will increase. In the trend plot (Figure 7), the motor bearing wear fault severity has leveled off at a low serious.

  

Figure 7. Expert System Trend Plot for Chiller 4

The life expectancy of the bearing is mostly a function of the force on the bearing and the condition does not seem to be getting worse over the past year. Vibration analysis provides quantifiable evidence of the bearing condition and allows the owner to know the condition of his machinery. As the defect worsens, the bearing wear defect pattern will become more prominent. The expert system will respond by assigning a higher severity to the diagnosis.

The chiller owner had called in a service engineer from the chiller manufacturer to measure vibration. Using an overall vibration meter, the technician told the owner that no problem existed with the bearing. The chiller manufacturer specifies an overall vibration maximum limit of 0.25 inch/second (peak). In this case, the overall level is probably less than the limit. This highlights one of the problems with using broadband measurements for predictive maintenance.

The broadband measurements completely missed the indications of bearing wear because broadband measurement is sensitive primarily to the highest peak seen in the spectrum. It does not see the BPFO harmonic series.

Conclusion
To learn from this machine, and to obtain the maximum life from the bearing, the facility engineer must closely monitor this machine and compare its data to the average plus one standard deviation for this model of chiller. The expert system severity score is an excellent way to consistently trend the bearing health because it always applies the same logic and looks at a number of features in the data.

When the severity increases toward the extreme level and a bearing replacement is ordered, the bearing should be saved and cut open for inspection. Possibly then, this case history would be considered complete.

About Azima DLI
Azima DLI is a leader and premier provider of predictive machine condition monitoring and analysis services that align with customers’ high standards for reliability, availability and uptime. Azima DLI’s WATCHMAN Reliability Services utilize flexible deployment models, proven diagnostic software and unmatched analytical expertise to deliver sustainable, scalable and cost-effective condition-based maintenance programs. The company’s bundled solutions enable customers to choose comprehensive, proven programs that ensure asset availability and maximize productivity. Azima DLI is headquartered in Woburn, Mass., with offices across the U.S. and international representation in Asia-Pacific, Central America, Europe and South America. For more information, visit www.azimadli.com.

Subscribe to Machinery Lubrication