SOME INSTRUMENTATION CONSIDERATIONS IN ROLLING ELEMENT BEARING CONDITION ANALYSIS
Duncan L. Carter
Boulder Vibration Systems
This paper describes techniques commonly used to assess rolling element bearing condition using acceleration signals at frequencies over 1 Kilohertz. Signals sources which can be measured using these techniques include bearing damage, potential causes of bearing damage such as inadequate lubrication, loss of internal clearance, brinelling, and interference from non-bearing related sources. Detection of failed or failing bearings is one end point of the bearing test process. Delaying or preventing the untimely occurrence of those failures can be achieved using the types of test procedures reviewed in this paper, whether the tests are performed in engineering prototype test situations, during manufacturing and repair quality control tests, or as preventive maintenance tests performed in the field. High bearing emitted signals, especially high average level signals, do not necessarily indicate terminal bearing flaws. Bearing emitted signals caused by inadequate lubrication, loss of clearance, or other problems are frequently correctable before serious damage is done to the bearing or to the machine using the bearing.
The characteristics of the signals that rolling element bearings generate are, in may ways, very different from those measured in order related vibration analysis. Conventional vibration analysis assumes that signals are composed of individual or groups of sine waves with the peak value of each signal component equal to the square root of two times its rms average value. Rolling element bearing test and measurement procedures are usually better derived from a model that assumes discreet impulses or events, impacts analogous to individual hammer blows in a low frequency bump test. Martin Angelo (1) described the process as similar to the tapping of a brass bell with its resultant ringing response; this concept is very useful, especially if you understand that bearings make very poor quality, highly damped bells and that bells and bearings may be excited in other ways besides tapping. Using acceleration as a measurement parameter, most of the ringing response is contained within a range of frequencies from 2 Kilohertz to 15 Kilohertz; for the purposes of this paper, this frequency range will be referred to as the primary ringing mode band. One common misconception of the nature of these signals is that they result from harmonics of the flaw repetition frequencies. A single impact or event can produce the same damped ringing response. A harmonic of the repetition rate of a single event is a meaningless concept. The non-linear nature of the impact process effectively produces multiple harmonic series of the ringing frequencies. The exponential decay of each ring produces a continuous spread of modulating frequencies of the harmonic components of the ringing frequencies. The result is a continuous, complex, non-uniform distribution of frequencies, a spread spectrum using radio communications terminology. Figures 1 and 2 illustrate the ringing mechanism for a small bearing having a fractured outer race with the accelerometer mounted directly on the bearing housing. The data is not compensated for the rising response caused by the 25 KHz accelerometer first resonance.
The signal transmission characteristics of bearing materials, surrounding machine materials, and bearing lubricants profoundly affect the measurement process. At observed frequencies significantly lower than the primary ringing mode band, the bearing emitted signals are transmitted in a relatively unattenuated manner, exciting the machine structure globally in the manner of modal analysis bump tests. As observed frequency increases, signal attenuation with distance increases. At observed frequencies significantly higher than the primary ringing mode band, signal attenuation predominates. Resonant effects at higher frequencies, the result of signal transmission to and reflection from discontinuities in machine structures, become minimal and more localized. Within the primary ringing mode band, signal transmission and signal attenuation are both significant effects in the measurement process and in the machine damage creation process. Because of attenuation effects, bearing emitted signals are usually measured as close to the bearing as possible. Because of the same attenuation effects, damage produced by bearing problems usually occurs within, or close to, the bearing. As long as the integrity of the bearing lubricant film is maintained, the bearing is a very effective isolation device or attenuator for high frequency signals transmitted through the bearing from a source external to the bearing.
Rolling element bearing test procedures currently in common use are, in essence, the same basic techniques that have been available for the last two decades or more. The simplest and probably earliest method of piezoelectric accelerometer based bearing tests used voltmeters to measure average signal levels and oscilloscopes for measurement for peak levels and for an assessment of the spectral content of the signals. Within limits imposed by interfering signals, a good assessment of bearing condition could made by considering the peak value of the time domain acceleration signal, a fair indicator of the potential of bearing flaws to produce damage, and the average value of the signal. One early example of this basic, broad band, equipment approach to bearing analysis in industrial applications was the IRD 330 with its optional accelerometer accessory, an instrument introduced in the late 1960s. On the IRD 330 which had an output jack for an external oscilloscope and/or headphones, average broad band acceleration was measured on a panel meter. The Metrix 5115 series analyzers, instruments of roughly the same vintage, had an accelerometer as their basic transducer and used an built-in oscilloscope as the display device. Simple, broad band peak and average reading acceleration input meter devices by manufacturers such as Bruel and Kjaer continue to be marketed and used very effectively. The peak and average values can be used to calculate the crest factor, the peak time domain signal divided by the root mean square (rms) average of the time domain signal. The crest factor is frequently used as a bearing diagnostic tool and is commonly expressed both as a numerical ratio and in decibels. High crest factor values usually indicate significant, discreet, bearing flaws. Medium levels crest factors usually indicate more generalized wear, lubrication problems, or loss of clearance in moderately worn bearings. Low crest factors usually indicate non-bearing signal sources, interference, although loss of clearance on new bearings can also produce relatively low crest factor acceleration signals.
In 1971, when I began working as a vibration analyst, almost all of the current bearing test techniques including the use of high pass filters below, in, or above the primary ringing response band, detection at accelerometer resonant frequencies, and envelope detection or demodulation of time domain bearing emitted signals were well established in literature. At that time, having the studied the IRD consultant's manual which was a very comprehensive compendium of published vibration literature and, having concluded from experience that the non-availability of an acceleration mode on my new IRD 350 was a severe handicap for bearing testing, I designed and installed a simple 1 Kilohertz high pass filtered accelerometer input device in that instrument. Adding the 1 Kilohertz high pass filter was a significant improvement over the more broad band method used with the 330 and its accelerometer accessory in that interference lower than 1 Kilohertz was suppressed while still essentially retaining the full magnitude of the bearing ringing response. Interference to bearing emitted signals in machines such as a. c. motors which also produce signals resulting from electrical excitation effects remained a problem. Continuous monitor devices reading average acceleration with optional high-pass filters were also available in that period. In the mid 70s, IRD introduced Spike Energy, a measurement system which used acceleration signals high pass filtered at 5500 Hertz, which effectively included the upper half of the primary ringing response band and higher frequencies. Since interference often decreases more rapidly with frequency than the bearing generated signals, the high pass filter method IRD and others have used can significantly improve the ratios of bearing emitted signals to interference; unfortunately, in many situations, the remaining interference levels can still mask low level bearing signals of interest. Since signal attenuation generally increases significantly with frequency, suppressing the lower frequency end of primary ringing response band increases the uncertainty of the actual overall bearing emitted signal levels. Transducer mounting induced errors, especially the negative effects of using magnets as mounts, also tend to increase with frequency.
During the preparation period for this paper, I measured bearing emitted signals from examples of differing bearing test problems on a client's machinery consisting of three systems, each having of a 1250 or a 1500 horsepower a. c. motor producing electrical excitation interfering signals, an Eaton Dynamatic clutch with no significant interfering signals, and a Joy Vane Axial Fan which had significant blade passage harmonics below 1 KHz. Each bearing test point used had the shortest physically practical distance between the transducer mount and the bearing. Signals at each point were recorded from permanently mounted transducers attached using either machined surfaces with studs or with machined adapters attached using commercial grade dental cement. Additional data from a magnetic mounted transducer was collected at one temporary point on the clutch output shaft which was used to test the pilot bearing, a relatively remote, inaccessible bearing connecting the clutch input and output shafts. Envelope detected acceleration high passed filtered at 2 KHz, and envelope detected acceleration high-pass filtered at 5.2 KHz, a compromise between the original implementations of spike energy and current spike energy and HFD implementations, were recorded simultaneously. On the fan bearings and on the three more accessible clutch bearings, all of which had short transmission paths between the bearing outer races and the accelerometers, the average level of the 5.2 KHz high passed filtered signal ranged from 40 to 70 per cent of the overall bearing emitted signals. At the clutch pilot bearing test point, the 5.2 KHz high pass filtered signals were roughly one-eighth of the overall bearing emitted signals measured at the same point. At the motor test points, the average level of the 5.2 KHz high pass filtered signal ranged from 30 to 40 per cent of the overall bearing emitted signals. However, using the 5.2 KHz high pass filter reduced the relative amount of 120 Hertz harmonic interference in each sample set by an average of roughly 40 per cent.
Figures 3 and 4 plot relative average signal levels normalized to 2,000 Hertz for a small bearing with a minimal distance between the bearing and the accelerometer ( Figure 3) and a larger bearing with a long path between the bearing and the accelerometer (Figure 4), illustrating the additional, variable, attenuation created by applying high pass filters at frequencies above the low end of the primary ringing mode band.
During the mid 1970's, higher frequency time domain measurement techniques such as Shock Pulse and Acoustic Emission also became available. Shock Pulse and similar systems used the signals at the first resonances, typically 25 KHz to 35 KHz of normal accelerometers. One example of Acoustic Emission was the Metrix model 5366 portable analyzer, which used a special accelerometer having a first resonance at 100 KHz. Both methods used the resonant characteristics of the transducers as part of their filter circuits. Currently, Shock Pulse uses peak and average signal values as its output signal parameters although earlier Shock Pulse also included signal level related histogram data (2). The Metrix 5366 measured rms levels at 100 KHz and a flaw specific impulse parameter called Spikes Above Threshold which was the portion of the 100 KHz band signals greater than 3.5 times the rms average of the 100 KHz band signals and also included normal low frequency acceleration and velocity. The Metriz 5366 also included an envelope detected output from the acoustic emission signals suitable for driving normal spectrum analyzers. Both types of measurements were extremely sensitive to signal path attenuation and could still be adversely affected by interference in some situations such as paper machine steam leaks and reciprocating machine valve leaks.
In the early to mid 1970s, portable spectrum analyzers in industrial use were mostly manually tuned analog devices of very low resolution. A decade later, portable Fast Fourrier Transform (fft) analyzers having frequency resolutions from 200 to 800 lines were widely available from manufacturers of general purpose and vibration specific instrumentation. Most conventional fft spectrum analyzers provide a limited bearing analysis capability although that capability can be improved if the spectrum analyzer is driven from an external envelope detector. RMS average acceleration levels can be calculated from the square root of the sum of the squares of the acceleration levels at each spectral line in the ringing mode frequency range. If a sampling rate corresponding to 10 KHz full scale on the frequency axis is used, the bearing emitted rms acceleration could be calculated from the upper 70 to 85 percent of the individual spectral values. Peak values of acceleration signals can generally only be measured from the time domain signal peaks, although the minimal storage and display capacity of most spectrum analyzers limit the measurement to very small time blocks, usually making visual assessment of the signal impractical. One error that some practitioners make is to attempt to derive a peak value of the acceleration signal from the average value of the broad band signal, a technique that is valid if, and only if, the numerical relationship between peak value and the average value, the crest factor, is already known. The hallmark of the bearing flaw signal is its impulse character, its high peak to average ratio. Another error assumes that the value of the acceleration, either peak or average, is the displayed value of a peak on the spectrum curve within the band of interest. Another frequently confusing problem comes from the fact that the individual values of a distributed spectrum and the appearance of the spectrum plot are dependent on the details of the measurement process. For a constant sampling rate, increasing the number of samples in each sample block containing a spread spectrum decreases the magnitude of the individual spectral lines. If a single frequency signal, a sine wave, is also present in the frequency band being measured, that signal remains constant with changed resolution.
Recently, several manufacturers of vibration test equipment have begun to include envelope detectors as integral or accessory parts of their test systems. Two decades ago, while envelope detectors were available on a very small number of pieces of commercially available equipment, most industrial envelope detection was done by eye with an oscilloscope and data was recorded with oscilloscope cameras. Fortunately, technology has advanced! Micro-computer based test and measurement systems with onboard data storage, signal processing, and display capacity have become commonplace. Envelope detection in computer based systems serves several purposes. Other than peak and average level signal quantities, very little useful information about bearing condition can be inferred from the raw bearing emitted acceleration signal or its spectrum without demodulating the signal. Demodulation can effectively compress the signal data, allowing the capacity of the test system to be used more effectively. Using demodulation allows the simple application of signal processing techniques such as the fast fourrier transform to enhance the sensitivity of the detection process by providing the ability to distinguish impacts occurring at repetition frequencies calculated from the dimensions of the bearings being tested. If the envelope detection systems are designed and constructed in a manner that preserves the essential time domain characteristics of the bearing emitted signal, peak and average level data can be measured as readily or more readily than similar data from the raw signal. In addition, many analysis problems such as rubs can be solved by quick visual inspection of the time domain data.
Envelope detection can be accomplished using either digital or analog methods. Digital methods require that data be sampled first at rates suitable for the raw signal. For a 10 KHz bandwidth, contemporary analyzers typically sample raw acceleration at rates from 25 KHz to 40 KHz and output that signal to a buffer or a file. Next, small blocks of successive data samples, blocks of 16, for example can then be processed to find the sample having the maximum absolute acceleration value and then output that value to a second file or buffer for further use. With hardware envelope detectors, data can be sampled at a slower rate, typically 2 KHz to 4 KHz. Software envelope detection affords greater flexibility at the price of requiring increased sampling rates, increased temporary data storage, and processing. Hardware envelope detection requires significantly lesser digital hardware at the price of increased analog hardware and comparative test system inflexibility.
Cautions apply to the use of envelope detectors, especially as implemented by some manufacturers who capacity couple the output of the envelope detectors to the input of the next stage of the test system. Envelope detection should return the magnitude of the envelope, the edge or outline, of the signal wave form. Capacity coupling offsets the zero value of the signal so that equal areas of the signal wave form are above and below the zero signal axis. If capacity coupling is used, accurate peak and average signal information is lost. In addition, capacity coupling effectively adds a low frequency transient signal to the data that can impair time domain analysis of the signal, limiting the ability to diagnose low level phenomena. Static or very slow time varying level information useful to the diagnosis of problems such as rubs, seal problems, and loss of clearance is lost by capacity coupling. Some manufacturers systems focus only on the spectrum analysis aspects of envelope detection, usually assuming that the only bearing emitted signals of interest are those characteristic of discrete flaws or spalls. Other bearing emitted signals reflecting the loads placed on the bearings can provide valuable information on the causes driving the wear and failure process. Spectrum analysis routines such as the fft return average, and only average, spectral information about the bearing emitted signals. The ability to distinguish between intermittently occurring large signals and frequently or continuously occurring small signals is missing in spectrum analysis, but can be provided with time domain analysis. Flaws from surfaces continuously in the load zone of the bearing are weighed more heavily than flaws present intermittently in the bearing load zone. Likewise, because of attenuation effects, flaw impacts closer to the accelerometer are weighed more than more distant flaws within the same bearing.
Without interference to contend with, bearing analysis is a relatively simple process. Higher frequency and band limited filter schemes typically reduce, but not eliminate interference at the price of creating significant signal path related uncertainty about the overall signals levels and at the additional price of discouraging standardization of testing and condition assessment methods. Spectral analysis of envelope detected signals is a significant, though incomplete, improvement in the process. Current technology can provide a simpler and more effective technique. Detect the signal in the entire primary ringing mode band, minimizing signal attenuation effects. If interference is present, assuming that the interference can be characterized differently from the bearing signals, use digital signal processing techniques to remove the interference. Then, analyze the resultant signal in both the time domain and the frequency domain as though the interference were not present. The final data sets in this paper, taken from an ac motor having roughly 25,000 hours on the bearings and processed using an algorithm especially designed for envelope detected bearing signals, show the stages of the bearing signal extraction process. With the use of these filter techniques, the motor excitation signals which otherwise mask the low level, bearing condition related signals are eliminated. Variations of the same techniques may also be applied to the raw on the raw acceleration signals prior to de-modulation.
Figures 5 and 6 are time domain and frequency domain data recorded from the motor. Test equipment used was a Metrix 5276 (Wilcoxon 760 series) accelerometer, a Metrabyte DASG-1 analog to digital converter, and an IBM compatible 486-33 personal computer. The data is not compensated for the 35 KHz. accelerometer first resonance. Figures 7 and 8 are envelope detected time and frequency domain data from the same source. The signals were conditioned using a Boulder Vibration Systems two channel signal conditioner and band pass filtered from 2,000 to 15,000 Hertz before detection. Figures 9 and 10 are derived from the data from figure 7 and are the result of removing the motor excitation signals, primarily the 120 Hertz modulated slot pass frequencies, from that data by using a digital signal processing post detection filter routine (3).
1. Angelo, Martin, "Vibration Monitoring of Machines," Bruel &Kjaer Technical Review No. 1-1987, pp. 1-36.
2. John Frarey, telephone conversation.
3. Duncan L. Carter, Patent Pending ( U. S. Patent # issued Dec. 26, 1995).
Return to Inteltek home page Return to bibliography page