woos 10 18436 Vian Anas Cater 2 oKF. Chapter 4: Signal Processing and Data Acquisition Introduction ‘This chapter covers signal processing and data acquisition. Understanding how the analogue signal from a transducer is transformed into a spectrum is an invaluable tool when assessing the vibration data. An understanding of the steps and the processes required to produce the spectrum gives the analyst a much better understanding of the data itself. Signal Processing and Data Acquisition Vibration Frequency Analysers Whenever a machine exhibits excessive vibration or a significant increase in vibration has been detected through periodic checks, the next step Is to perform @ complete analysis of the vibration to determine the cause. The key element for any analysis is determining the vibration frequencies that are present and understanding how these frequencies relate to the various machine components and the machine rpm. To do this a vibration frequency analyser is required. Vibration frequency analysers are available in a wide range of features and capabilities, however, all frequency analysers have certain common features and characteristics in terms of how the data {is processed and presented. It is therefore important for analyst to know and understand the basics, fof how their instruments work as well as their limitations. Many vibration problems have been incorrectly diagnosed or missed completely, simply because the vibration analyst didn’t understand the workings & limitations of the instruments that were being used. Historically, analogue or "swept-fiter” frequency analysers, such as those illustrated in figure 1, were first used from around the 1950's and continued in use until the 1980's. It was round about the late 1970's to early 1980's that advancing technology allowed the introduction of Digital or “FFT” frequency analysers. Both types of vibration frequency analysers are described in the sections to follow, as an understanding of analogue analysers is helpful in the later explanation of FFT analysers. Analogue (Swept-Filter) Frequency Analysers Coe : ron Oep bee 6 Figure 1. Typical Swept-Filter Analyser Chapter 4: Signal Processing and Data Acquisition Page 4-1 The canes of thitranng materi ae the cpt of SKE AB and ty at be earadace in ar Yo by ny mers een eons erin ranted1503836 vation Anais catego 2 olkKr. Analogue or swept-filter frequency analysers were the very earliest instruments available for performing vibration analysis. A swept-filter analyser works in much the same way as a radio. In almost any community, there are literally dozens of radio stations broadcasting programmes simultaneously. However, each station is broadcasting on its own assigned broadcast frequency. A radio has @ “tuner” that allows you to tune to a specific broadcast frequency so that you can listen to that particular station. The tuner is actually an electronic filter that accepts the broadcast frequency of the desired station while ignoring or rejecting all of the others. Or, if you are old enough, remember, the earliest televisions where the stationed had to be “tunedsin” by scrolling up and down through the available frequencies. ‘A swept-fter vibration frequency analyser works on exactly the same principle as a radio. Any machine is likely to generate (broadcast) many different frequencies of vibration simultaneously, depending on the rotating speeds, the various machine components, the operation of the machine ‘and the specific problem that is causing the vibration. The swept-filter vibration analyser includes a filter (tuner) that one can tune or “sweep” over a frequency range of interest to pick out of identify ‘each generated frequency of vibration. The only differences between a radio and 2 swept-filter vibration frequency analyser are: +The filter of a swept-fiter vibration analyser Is designed to respond to machinery vibration, frequencies and not radio frequencies + The swept-filter vibration analyser uses a transducer, such as an accelerometer attached to the machine to pick up the signal; @ radio uses an antenna to pick up the broadcast radio (RF) frequencies + The swept-fiter vibration analyser presents the measured vibration amplitude and frequency data on meters or as a hard copy printout of amplitude-versus-frequency (usually called vibration spectrum or signature); a radio presents data as sound through speakers A typical swept-fitter vibration analyser includes a filter dial that can be manually tuned (ike a radio) over @ wide range of vibration frequencies to identify the amplitudes and frequencies of machinery vibration. Selector switches are provided for selecting the desired amplitude parameter (displacement, velocity acceleration, or enveloped acceleration), full-scale amplitude range, and filter “bandwidth” characteristics. Most of these instruments include a display so that either the raw waveform or the amplitude and frequency information can be displayed. In most of these Instruments, a stroboscopic (strobe) light Is also included to allow phase analysis, dynamic balancing, frequency confirmation, and “slow motion” studies. ‘Some instruments of this type also incorporated a built-in printer for generating hard copy data. By simply pushing a button, the analyser filter automatically sweeps over the frequency range of Interest and the printer generates a spectrum. This enhancement to swept-filter analysers eliminates the tedious and time consuming task of manually tuningto each vibration frequency of Interest and manually logging the information on a data sheet, as was the practice with the first analysers of this type. Chapter 4: Signal Processing and Data Acquisition Page 4-2 ose ree 2003 ‘The contns this traning matralrthe epright SKF AB and may nat be reprtacedin ay frm ory ay ean [ven ea) nls permsion rae18018426 Viraton Anais Ctegoy 2 oKF: Characteristics of Tuneable Filters First, It Is Important to recognise that a tuneable or swept-filter does not reject all vibration frequencies except the one to which it is tuned. The filter is actually tuned to a narrow band of frequencies and accepts those frequencies within the band while increasingly rejecting those frequencies outside the band. Typical filter response curves are illustrated in figure 2. Lower Cutoft Frequency t g * oon 2 100% +4 Tee Frey te, ayy Fiter ndvicth tr ba Upper Cutoff % 30x Precuency 2 100% ese} J Frequency =e ‘<. PS Figure 2. A Comparison of Analysis Obtained with a Broad (10%) and Sharp (5%) Filter Note that the filter bandwidth is defined by the upper and lower cut-off frequencies. The cut-off frequencies of a filter are those frequencies above and below the tuned centre frequency of the filter where the response to a vibration signal is approximately 30% less than maximum (centre) response. In other words, at the cut-off frequency points, a signal would be reduced by 30% of its actual amplitude. Beyond the cut-off frequencies, vibration signals will be reduced by considerably Filters may be further subdivided into “constant bandwidth” or "percentage bandwidth” filters. With 2 constant bandwidth filter, the bandwidth between the upper and lower cut-off frequencies is a fixed frequency range such as 10Hz (60 cpm) regardless of the frequency to which the filter is, tuned, This range is spread over the frequency range, above and below the filter setting. For the 10Hz filter mentioned, this means that if the filter is tuned to SOHz, the actual cut-offs would be at 45Hz and SSHz. The actual width of a constant bandwidth may be fixed or adjustable, depending fon the instrument design. Compared to a constant bandwidth filter, the percentage bandwidth filter is expressed as the percentage between the fiter’s tuned centre frequency and the lower and upper cut-off frequencies. For example, the upper filter shown in Figure 3 has its lower cut-offs shown at + 5% Chapter 4: Signal Processing and Data Acquisition Page 4-3 esierea 2013 “Te contre of hi rang mata ate th opt of SE AB and ma ok reproduced a Yor ay ny mesns een sre) ures permasin gared,15018436 Viraton Amys agony 2 oKF. of the centre frequency. Therefore, If this filter was tuned to a vibration frequency of 1,000 CPM, the filter bandwidth would extend from 50 CPM below to 50 CPM above the centre frequency (5% (of 1,000 CPM = 50 CPM) or from 950 CPM to 1,050 CPM. This is a total frequency bandwidth of 100 CPM or 10% of the tuned centre frequency of 1,000 CPM. This would be known as a 10% bandwidth filter. If this same 10% filter is tuned to a frequency of 10,000 CPM, it would have a bandwidth from 9,500 CPM to 10,500 CPM or a bandwidth of 1,000 CPM (10% of 10,000 = 1,000). Ih other words, the actual frequency range of a percentage bandwidth filter depends on the centre frequency to which the filter is tuned. Understanding these characteristics are important in this respect. For example, assume that @ 10% bandwidth filter is tuned (centre frequency) to @ vibration at a frequency of SOHz (3000 cpm) with an amplitude of 10 mm/sec. In addition to this vibration, another vibration Is present at a frequency of 47.5Hz (2850cpm) and has an amplitude of 25 mm/sec. At this frequency, the filter attenuates (reduces) the signal level by 30%, or passes 70% of the signal. The net result will be two vibration signals coming through the filter at the same time; 10 mm/sec at a frequency of SOHz and 17.5 mm/sec (70% of 25 = 17.5) at the frequency of the stronger or predominant vibration, 47.5H2, Therefore, the amplitude of 17.5 mm/sec at the 47.5Hz will be the one that is Indicated on the instrument's display and the one that will trigger the strobe light and the 10 mm/sec at SOHz will not been seen, To get around this issue, most swept-filter analysers offered a choice of several different bandwidth filters and, like so many other choices that have to be made, there is trade-off in filter selection between the analysis time and frequency accuracy. A filter with a 10% bandwidth is generally considered a fairly wide or “broad” filter, and can be tuned fairly rapidly over a wide frequency range to quickly determine the predominant vibration frequencies present. Where more accuracy is needed in measuring vibration frequencies or for separating closely spaced vibration frequencies @ narrower or “sharp” filter would be selected. A 5% bandwidth (sharp) filter provides twice the frequency measurement accuracy and frequency separation compared to a broad 10% filter. However, the 5% sharp filter requires nearly twice the time to analyse @ given frequency range as the broad filter does. For most analysis applications, the sharp filter is recommended, since the actual time saved by using a broad filter is normally only @ matter of a few minutes. Also, when doing an in-depth analysis, the accuracy of analysis data generally takes precedence over analysis time. Figure 3 shows the comparative spectrums taken on a reciprocating air compressor operating at 41,800 RPM. The upper trace was collected using the broad (10%) and the lower trace was collected Using the sharp (5%) filter, Both plots cover a frequency range from 600 CPM up to 600,000 CPM. It Is clear to see that the sharp 5% filter provides greater frequency accuracy and separation of closely spaced frequencies than the broad 10% filter does. Chapter 4: Signal Processing and Data Acquisition Page 4-4 osarree 2009 ‘The contents fis tering mater re contig ol SKF AB and may net be rwpraacedin ay fr ry ay mean (ve xa} alee person atewoos 150 18636 Vain Aas ater 2 OKF-. DH-oED i, Figure 3. Example of print-out from Swept-Filter Analyser using Broad and Sharp Filters Digital (FFT) Frequency Analysers Analogue, swept-flter frequency analysers have been in use for many years and have been used to detect, identify, and solve many machinery problems. However, they have several disadvantages, ‘over modern digital (FFT) frequency analysers, including: + Analogue analysers are typically bulky and could weigh anywhere between 10kg & 20kg. By ‘comparison, most modern digital data instruments weigh less than 1 Ka. ‘+ Analogue analysers cannot store data or interface with software programs. Digital analysers ‘can store vibration data from literally hundreds of machines. ‘+ The features and capabilities of analogue instruments are generally governed during Initial, design and any upgrade normally required hardware changes. Digital analysers are essentially portable computers and enhancements can be made or features added by simply reprogramming the operating “firmware” or by installing new programmes. + Analogue analysers are somewhat limited in frequency accuracy and the ability to separate close vibration frequencies (because of the filter selections available), With the proper setup, digital analysers can measure vibration frequencies with accuracies up to 0.03Hz If required. ‘+ Analogue analysers are typically limited to a maximum frequency of 10 KHz (600,000 cpm), Whereas modern digital analysers can measure frequencies to 80 KHz. ‘+ Analogue analysers are slower than digital analysers. An analogue analyser may take from 30, ‘seconds to several minutes to print a frequency spectrum. Digital analysers can perform this, task in a fraction of the time, These are only a few of the advantages digital instruments (Including vibration data collectors) have over analogue instruments. However, it should be apparent that digital analysers are superior In many respects, and that is why, beginning in the early 1980's, they have essentially taken over, the field of vibration analysis. Chapter 4: Signal Processing and Data Acquisition Page 4-5 ‘The cores of hiring materi ee copy of SHE AB and may nt be sprosced nay Yor oy ny means een serene perma raredaos 150186 Viaton nabs Category 2 ukKF However, with thelr added power and features, digital analysers also require the analyst to make ‘more critical decisions in set up to ensure that all important data is presented and presented in a way that makes it possible to identify specific problems. What is FFT? When referring digital instruments for vibration data collection, this normally refers to FFT data collectors. The term “FFT” stands for “Fast Fourier Transform.” Nearly 200 years ago, French mathematician Baron Jean Baptiste Fourier (figure 4) established that any periodic function (which Includes machinery vibration signals) can be represented mathematically as a series of sines and, cosines. In other words, it is possible to take a vibration time waveform, whether simple or complex, and mathematically calculate the vibration frequencies present along with thelr amplitudes. Figure 4, Baron Jean Baptiste Fourier The process Is called the “Fourier Transform." Although the Fourier Transform can be done manually, the process is extremely time consuming. However, with the introduction of digital technology, the actual mathematical expression could be included onto computer chips, which means the transform can be carried out very fast. Hence the term Fast Fourier Transform (FFT). In order for an FFT analyser to produce a spectrum, there are several steps in processing the data, of which, the FFT is only one step. Figure 5 is a block diagram that shows the main components Incorporated in an FFT Analyser and therefore the steps that it goes through to finally produce the spectrum. Figure 5. Steps required to produce the spectrum Chapter 4; Signal Processing and Data Acquisition Page 4-6 osirres zm rearoced man orm rb ny ears een ere unless persion rarewzea 15018436 Viraton Anais Category 2 oKF. Analogue Signal ‘The FFT process begins with an analogue signal from a vibration transducer. This is also referred to {as the raw signal that is generated as a result of the vibration. Normally, the transducer will be a vibration accelerometer; however, signals from other types of transducers can be processed as well, such as velomitors, proximity probes, microphones, pressure transducers, current transformers, etc. Input ‘At the input to the instrument, the signal Is calibrated based on the sensitivity of the transducer. For example, a particular transducer may have a sensitivity of 100mv/g, which means for every g of acceleration, the transducer sends out 100mv. Therefore, when the data collector registers the amount of voltage, it must convert this to acceleration by applying the calibration based on that sensitivity. However, if @ transducer with a sensitivity of 10mv/g Is used when the instrument is calibrated for 100mv/g there will be errors in the calibration. For every g of acceleration, the transducer gives ‘out 10my, The instrument is expecting 100mv/g and therefore calculates that the vibration is only 0.19. The same principle applies if a low frequency accelerometer is used, for example with a ‘sensitivity of 500mv/g. Then, for every g of acceleration, the transducer gives out S0Omv and the instrument is expecting 100my, it calculates that there are 59's. ‘Also taking place at the input is the integration of the signal if this is required. Since an accelerometer Is often used for measuring the vibration, it may be necessary to convert the acceleration signal to velocity and this is achieved by “single integration” or, in some situations, it may be necessary to convert the signal into displacement and this is achieved by “double integration”. Anti-Alias Filter ‘One of the first decisions the operator must make in setting up a FFT analyser is to decide what range of vibration frequencies needs to be looked at or analysed. For example, assume that the decision was made to analyse a frequency range from 0 to 2,000 Hz (0 to 120,000 CPM), to include the maximum frequency of interest. In order for the signal to be measured, the instrument must ‘actually sample the incoming raw signal, If the sampling rate was set at 3000 Hz, then the difference between the sample rate and the Fmax is 1000 Hz. This frequency would then appear at 1000 Hz in the spectrum. This Is known as an Allas (false) frequency. In other words, Allasing is the result of sampling a signal too slowly. Although these frequencies do not actually occur within the selected frequency range (Fmax), they will appear as calculated vibration frequency ‘components in the FFT. ‘A practical example of this would be to point a strobe light at a rotating shaft and set the flash rate exactly at the rotating speed, when the shaft will appear to be stationary. If the strobe flash rate is not synchronised with the disc speed, the disc will appear to be rotating slowly. Similarly, if a sinusoidal waveform is sampled too slowly, It will appear to be a lower frequency. Chapter 4: Signal Processing and Data Acquisition Page 4-7 sie Fea 2003 “Te contents ths inng materi are the capri of SKF AB ard ay nt be eprodscnain frm or yay ears een ent) less permission rantedwi203 150 18426 Vraton Anais Category 2 oKF To avoid this problem, Anti-Alias filters are used. This effectively means applying a low-pass filter to the signal before it is sampled so that there will be no frequency components above half the sampling frequency. A sampling rate at exactly twice the maximum expected frequency Is known as the Nyquist frequency, after the Swedish American engineer. In practice, the sampling rate will bbe greater than twice (approximately 2.56 times) to allow for a low-pass filter without a sharp cut off. Modern digital analysers use anti-aliasing filters that remove all frequencies above 40% of the ‘sampling rate. It is important, however, that the analyst understands the characteristics of the particular analyser before beginning data acquisition. To illustrate the importance of the anti-aliasing filter, figure 6 shows two 2 vibration frequencies, 2 low frequency component at 1 x RPM, together with a high frequency component at 9 x RPM. Figure 6. Sampling 2 1x RPM and 9 x RPM In this example, it can be seen that if both signals were digitised at the same sample rate (red dots on blue waves), the signal occurring at 9 x RPM would appear to have the same frequency as the 1 x RPM component as illustrated by the blue dots. ‘The anti-aliasing filter takes care of both of the problems j.e. generating false frequencies and sampling too slowly. Types of Filters A low-pass filter passes only the low frequency components of the signal and blocks the high frequency components above the cut-off frequency. The cut-off frequency divides the pass band ‘and the stop band. Examples of low-pass filters Include Anti-aliasing filters and Filters for excluding higher frequencies, such as gear-meshing components (may be used for balancing) 4 high-pass filter passes only the high frequency components in the signal but stops the low frequency components. The cut-off frequency divides the pass band and the stop band. In other words, the frequency components lower than the cut-off frequency will be stopped by a high-pass filter. Example uses of high-pass filters are to exclude low-frequency transducer noise (sometimes Chapter 4: Signal Processing and Data Acquisition Page 4-8 osirrea2019 “we contr of hi trang materi are he opr af SKF AB ana may et be ‘predundin ay for or yay ears evn exc) ules pein granted15020436 Vian Arabi Cateary 2 ORF. due to thermal effects) & to exclude low frequency structural vibration influences that may contribute to high overall levels when collecting routine machinery vibration data A band-pass filter is used to isolate distinct frequency bands so that components both above and below the high and low cut-off limits are excluded. Example uses of band-pass filters include acceleration enveloping for rolling element bearings & correlation of vibration measurements with noise measurements ‘The use of a tracking filter results in transfer function estimates that do not suffer from the FFT induced errors such as resolution and leakage. Tracking is the process of sampling the data based on a periodic signal, such as from a laser tachometer, This allows the vibration signals to be analysed on run-up or coast down modes, as well as allowing the user to look at any harmonic or sub-harmonic of the rotating speed Analogue to Digital (A/D) Converter Depending on the frequency range (Fmax) to be analysed, a certain period of the vibration time waveform is collected. This sample of waveform is called a “dataset.” In general, analysing a low frequency range (low Fmax) requires a longer period dataset than a high Fmax does. In other words, analysing a low Fmax takes more time than analysing a high Fmax. After the unwanted high frequencies have been filtered out by the anti-aliasing filter, the remaining analogue signal is converted from its analogue (continuous) format to a digital format necessary for mathematical (FFT) processing. This is the function of the A/D (analogue-to-digital) converter. Essentially, the A/D converter changes the analogue (continuous) signal into a series of data points or dots that, hopefully, will accurately define all of the significant features of the true analogue signal. Figure 7 shows the continuous analogue signal and the digitised signal. Figure 7. Continuous analogue and digitised signals ‘The number of data points into which the dataset can be divided can be selected by the operator; however, the maximum number is governed by the A/D converter used within the instrument. A/D converters are classified by their number of “bits,” such as 8-bit, 12-bit, 16-bit, etc. The higher the number of bits that are used in the A/D converter then the more data points that are available and ‘with more data points available, then the better the amplitude resolution or the better the ability to display both large and small amplitudes at the same time, Chapter 4: Signal Processing and Data Acquisition Page 4-9 ‘The content of thi traning materi ee the copyright of SKF AB and ay ot be ereduedin ay fom a yay mean even nec) less prison rane.waa 15018436 Vraon Anais Category 2 ORF. In other words, the greater the number of bits in the A/D converter, the better the instrument's dynamic range. Dynamic range Is defined in 1S013372:2004 as: the ratio of the largest ‘magnitude to the smallest magnitude that a transducer or analyser can detect. ‘The majority of spectra that are collected for routine condition monitoring are taken using 1,024 data points. This gives a reasonable compromise between data collection time and accuracy. ‘Another factor is the amount of instrument memory and computer storage available. It takes four times the memory and computer storage capacity to store data comprised of 4,096 data points as it does to store data comprised of 1,024 data points. Most modern FFT data collectors/analysers have very large internal memory and some also have external memory cards that can be changed over, so the limits of memory are not so relevant now. However, if the software package is on a network and the database Is being accessed by several different people at the same time, then having lots of very large, high resolution data can cause the database to slow down very quickly. In most situations, the vibration analyst usually isn't concerned with the number of data points used to define a vibration signal, the accuracy of the data is far more important. The accuracy of the data is determined by the frequency range and the number of lines of resolution which are both determined by the person who creates the measurements in the database. This will be discussed in ‘more detail in the section "Defining Spectral Parameters” Window Functions The next step in the FFT process is called “windowing,” Referring back to figure 7, itis clear that a digitised signal is an approximation of the continuous analogue signal. The difference Is clear when the amplitude at 0 seconds is compared to the amplitude at 0.75 seconds. This difference means that the FFT circuit will not be able to process that data correctly and random frequencies can appear on the spectrum, 21/T a) Waveform that is periodic in the time sample 1 Leakage in the spectrum b) Waveform that is nan periodic in the time sample Figure 8, Periodic & Non-Periodic signals ‘Chapter 4: Signal Processing and Data Acquisition Page 4-10 o sere 203 ceases yr oy yeas en eva) nes pen Bae20a 15018436 Viraton ais Category 2 oKF-. ‘This problem is taken care of by an electronic circuit within the instruments, called a window. There ‘are many different types of windows used in digital analysers and data collectors to handle analysis requirements and these include the Hanning, Flat Top, Rectangular or Uniform, Force, Exponential, Hamming, Kaiser Bessel and Cos4. The most commonly used of these are the Henning, Rectangular and the Flat Top. The Hanning Window ‘The Hanning window, named after its inventor Von Hann, has the shape of one cycle of a cosine ‘wave squared so It is always positive. The sampled signal values are multiplied by the Hanning function and the result is shown in the figure. Note that the ends of the time record are forced to zero regardless of what the input signal is doing. While the Hanning window does a good job of forcing the ends to zero, it also adds distortion to the wave form being analysed in the form of amplitude modulation, i.e., the variation in amplitude of the signal over the time record. Amplitude modulation in a waveform results in sidebands in its spectrum, and in the case of the Hanning window, these sidebands, or side lobes as they are called, effectively reduce the frequency resolution of the analyzer by 50%. It is as if the analyzer frequency “lines” are made wider. In figure 9, the curve is the actual filter shape that the Hanning weighting produces. Each line of the FFT analyzer has the shape of this curve (only one is ‘shown in the figure). Time Record Time Record Figure 9. The Hanning Window Figure 10 shows the effect that this has on the wave form and the process that takes place the application of the Hanning window, or Hanning Weighting. Chapter 4: Signal Processing and Data Acquisition Page 4-12 94 Fea2003 “The content his ain materi are he apy FSH AB ad may rt be reproscesin ay Yrm ob ay mean ern ett) less permission ranted.wi203 SO 18436 Vibration Analysis Category 2 No Window Function ‘Sample will be forced to ‘Applied start and end at zero Window Funetion Applied ‘The Sample Signal Appears To Be Continuous Figure 10. The application of the Hanning Window If @ signal component is at the exact frequency of an FFT line, it will be read at its correct ‘amplitude, but if it is at a frequency that Is one half of delta F (one half the distance between lines), it will be read at an amplitude that is too low by 1.5 dB because of attenuation of the signal. Or, in other words, some of the amplitude Is cut off Figure 11 illustrates this, \ [es Figure 11. The effect of the Hanning Window ‘The illustration shows this effect and also explains why side lobes are sometimes created by the Hanning window. As was discussed in the Vibration Analysis category 1 course, sidebands are generated by amplitude modulation. Side lobes are a similar phenomenon in that they will appear on either side of the main peak in the spectrum, Referring back to figure 11, because of the overlap of the windows, it should be clear that if a frequency Is close to the upper or lower limits of the window, some of the amplitude will be seen by the window in the adjacent spectral line The hhighest-level side lobes are about 32 dB down from the main frequency peak in the spectral line. ‘The measured amplitude of the Hanning weighted signal is also incorrect, because the weighting process removes essentially half of the signal level. However, this can be easily corrected simply by multiplying the spectral levels by two, and the FFT analyzer does this job. This process assumes ‘the amplitude of the signal is constant over the sampling interval. If it is not, as is the case with transient signal, the amplitude calculation will be in error, as shown in figure 12. Chapter 4: Signal Processing and Data Acquisition Page 4-12 sxe Fes2013 The contnt of th training materi ete capri of SKF AB and may ot be reproduced ay frm by my ear een eta) eles rion ratedvzos 15018436 viraton Aas Categary 2 oKF-. The tranciont is distorted by the window function. The Hanning Window Figure 12. Amplitude Error in Transient Signal The Hanning window should always be used with continuous signals, but should never be used with transients, The reason is that the window shape will distort the shape of the transient, and the frequency and phase content of a transient is intimately connected with its shape. The measured level will also be greatly distorted. Even if the transient were in the centre of the Hanning window, the measured level would be twice as great as the actual level because of the amplitude correction the analyzer applies when using the Hanning weighting, ‘A Hanning weighted signal actually is only half there, the other half of it having been removed by the windowing, This is not @ problem with a perfectly smooth and continuous signal like sinusoidal waveform, but most signals that analysts come across, such as machine vibration signatures, are not perfectly smooth. If a small change occurs in the signal near the beginning or end of the time record, it will either be analyzed at 2 much lower level than its true level or it may bbe missed altogether. For this reason, it is a good idea to employ overlap processing. This will be discussed later when defining spectral parameters will be covered. ‘The Flat Top Window Referring back to figure 11, where it can be seen that the Hanning function gives the filter a narrow rounded top and relatively wider base. This was @ desirable characteristic when the analyst is required to identify frequencies in the spectrum. However, if the main aim of the measurement is to measure the amplitude of the signal then the Hanning window is not the best choice because it ‘can attenuate the signal by up 1.5 db, or around 16%. If amplitude accuracy is more important, then the solution to this problem is to use a window function that flattens the top of the filter. This will mean that there will be less amplitude error as more of the signal will be included in the window function, This window function is call a Flat Top window and it used when measuring the amplitude is more important than identifying the frequency of the component. The Flat Top window is shown in shown in figure 13. Chapter 4: Signal Processing and Data Acquisition Page 4-13 esi rea2009 The conn ofthis tring materal ae the copyright of SKF AB and may not be vesroced nary frm ob ay mean een ora) less permission granted150286 Vian Anais Category 2 ORF. Figure 13. The approximate shape of the filters when the Flat Top window is used ‘The main benefit to be gained from using this window function is that the amplitude error does not exceed 0.1 dB (1%). The improvement in amplitude accuracy is achieved though a compromise with the frequency resolution because of the widening of the base of the filters, This means that the ability to identify a small component which is close to a large component is lost. This is illustrated in figure 14, which has a close up view showing the comparison between the Hanning window and the Flat Top window. Figure 14. Comparison of the Hanning window and the Flat Top Window In figure 15, the approximate shape of the Flat Top (blue trace) window is shown and the result of this window is that there is the smearing, or leakage, of the frequencies through the spectrum (as indicated in the orange trace). sessesuy Figure 15. Flat Top window and leakage across the frequency range Chapter 4: Signal Processing and Data Acquisition Page 4-14 © ser Fee 2013 ‘The content of hs tening materi athe coy ol SKF AB an may net be preducedin ay fro by ean eer extrac) less person ranted15018436 Viration Arabs Cleary 2 oF. Uniform or Rectangular Window ‘The Rectangular, or Uniform, window does not alter the ends of the time record in the same way that the Hanning window does. In fact, this is also referred to as 2 self windowing function because it weights the entire time sample evenly. In other words, it doesn't really do anything to the input signal at all, This makes the Rectangular window very good for measuring transient events and for collecting time waveform measurements. When measuring transient events, the analyser can be set up $0 that the whole event occurs completely within the time record and therefore there is no eed to alter the signal in any way. Figure 16 Is a representation of a Rectangular window. Flat, “Uniform,” or “Rectangular” Window Figure 16. Uniform Window Since the input signal is not altered in any way, the Rectangular window does not generate any leakage in the spectrum when the transient event is wholly contained in the time sample. However, ia Rectangular window is used on a non-periodic input signal, it will not eliminate leakage and the bases of the peaks in the spectrum will be very wide. Also, the amplitude variation in this case can be as much as 36%. In the case where the input signal is a transient, as shown in figure 17, it will by definition begin and end at zero level, and as long as it is entirely within the time record, no truncation will occur and the analysis will be correct because the FFT sees the entire signal. It is very important that the entire transient fit into the record, and the record length is dependent upon the frequency range of the analysis. Most FFT analysers allow the user to see the time record on the screen, so it can be assured that this condition is met. Flat, “Uniform,” or “Rectangular” Window Figure 17. Uniform Window and Transient Chapter 4: Signal Processing and Data Acquisition Page 4-15 The contents of th tring material ae the copyright SKF AB an may net produce nay form er by ay ear even ene} uless erision is ranted20s 15018496 Vrain Ans Catgony 2 okKF Strategies for Choosing Windows Each window has its own characteristics, and different windows are used for different applications. To choose @ spectral window, you must estimate the type of signal frequency content i.e. continuous, random or transient. If the signal contains strong interfering frequency components distant from the frequency of interest, choose a window with a high side lobe roll-off rate. If there are strong interfering signals near the frequency of interest, choose a window with a low maximum side lobe level. If the frequency of interest contains two or more signals very near to each other, spectral resolution is important. In this case, it is best to choose a window with a very narrow main lobe. IF the amplitude accuracy of a single frequency component is more important than the exact location of the component in a given frequency bin, choose a window with a wide main lobe, If the signal ‘spectrum is rather flat or broadband in frequency content, use the Uniform window (no window). In general, the Hanning window is satisfactory in 95% of cases as it has good frequency resolution and reduced spectral leakage. ‘The Flat Top window has good amplitude accuracy, but because it has a wide main lobe, it has poor frequency resolution and more spectral leakage. The Flat Top window has a lower maximum side lobe level than the Hanning window, but the Hanning window has a faster roll off-rate. If you do not know the nature of the signal but you want to apply @ window, start with the Hanning window. If you are analyzing transient signals such as impact and response signals, it is better not to use the spectral windows, because these windows attenuate important Information at the beginning of the sample block. Instead, use the Force and Exponential windows, A Force window is useful in analysing shock stimuli because it removes stray signals at the end of the signal. The Exponential window is useful for analysing transient response signals because it damps the end of the signal, ensuring that the signal fully decays by the end of the sample block, The Force and Exponential windows will be discussed in more detail when discussing analysing resonance & carrying out Bump tests. Selecting a window function is not @ simple task. In fact, there is no universal approach for doing so. Figure 18 gives some initial guidance on the selection of windows. Signal Content Window Sine wave or combination of sine waves Hanning Sine waves (amplitude accuracy is important) Flat Top Narrowband random signal (vibration data) Hanning Broadband random (white noise) Uniform Closely spaced sine waves Uniform, Hanning Excitation signals (Hammer blow) Force Response signals Exponential Unknown content Hanning Figure 18. Initial Window Choice Based on Signal Content Chapter 4: Signal Processing and Data Acquisition Page 4-16 ser fes2003 “Toe cents this trdhog mater athe conv of SHF AB and may nt be preuce in aryform oy ny mean (even eras) ls erie arevaca 15018436 iraton Anais Category 2 oKF FFT Calculation The next step in the FFT process is the actual FFT calculation. This Is nothing more than a computer chip that has been programmed to carry out the Fourier Transform function to determine the individual frequencies of vibration present and, of course, the amplitude of each frequency component. This has already been discussed in Chapter 2, when the importance of frequency was discussed. As mentioned then, the two most important pieces of information that the user needs to define are the frequency range and the number of lines. ‘The FFT will then transform the digitised, windowed waveform into the frequency domain, effectively separating all the component parts of the waveform. The user define frequency range is, then applied and this is then divided equally by the number of lines, again user defined. The amplitude of all of the vibration frequencies within each line are computed together and shown as a single value at the centre of that line. Selecting Fmax ‘The first and perhaps most important decision you must make in defining the set up for an FFT is the Fmax or maximum range of frequencies that will be analysed and displayed. The selected Fmax ‘must be high enough to include all significant, problem-related frequencies. However, care must be taken to maintain a suitable resolution or a loss of accuracy of the measured frequency ‘components can occur. Therefore, the Fmax selected should be no higher than needed to detect problem-related vibration frequencies for a particular piece of equipment. ‘The best method for selecting the Fmax is to familiarise oneself with the equipment and all its ‘components. After doing so, it will be possible to calculate the possible defect frequencies. Once these are established, the frequency range can be determined. In some situations it may be necessary to collect data and compute two different spectra at the same location in order to gather all relevant information, Figure 19 shows to spectra the one on the left has a very low resolution and high Fmax, while the one on the right has a lower Fmax and higher resolution. In figure 19 the two spectra were collected on a motor-fan assembly running at 1,490 RPM. Both spectra were collected using 800 lines of resolution. The spectrum on the left was collected with an Fmax of 0 to 5,000 Hz (5 KHz) and shows some vibrations on the far left side of the frequency scale, However the frequency information is not very clear. The second FFT was taken with a much lower 0-500 Hz Fmax and clearly shows that there are considerably more than three frequencies of vibration present and now the frequency inform: nis clear. Chapter 4: Signal Processing and Data Acquistion Page 4-17 sie Fea2002 ‘The content of thi ang materia are he copy SKE AB and may rt reprocicesin ay frm ob ery mean even ete) sess permission ares.wos 1501496 Vraton Anata 2 oKF. Figure 19. Comparison of FFT’s with High and Low Fmax Selecting the Number of Lines of Resolution ‘The next important decision that must be made when taking a FFT is selecting the number of lines of resolution. This is similar to selecting the “broad” or “sharp: filter in the analogue or swept-flter analysers. The decision will not only determine the accuracy of the frequencies presented in the spectrum, it will also determine the amount of time required to perform the analysis. This ‘combination of Fmax and the number of lines will also determine the amount of storage that will be required for the data in both the instrument and the database. Unlike swept-filter analysers that may offer a choice of only two or three filter bandwidths, most FFT frequency analysers and data collectors offer a much larger choice of lines of resolution. ‘Typically FFT analysers may offer 25, 50, 100, 200, 400, 800, 1,600, 3,200 lines, 6400 and 12,800 lines of resolution. Some of the latest data collectors, such as the SKF Microlog CMVA 80 (AX) can Row process 25,600 lines in the analyser mode. Most spectra taken for routine condition monitoring ‘and general machinery analysis are taken using either 400 or 800 lines of resolution. However, when detailed investigations are carried out, the number of lines is usually a minimum of 1,600 ‘and more likely to be either 3,200 or 6,400. When the choice of Fmax and the number of lines have been made, the selected frequency range is then divided by the selected number of lines. For example, assume that 400 lines of resolution were chosen for a spectrum with an Fmax of 2000 Hz (120,000 CPM). Therefore, looking at figure 20, 2000 Hz divided by 400 lines will mean that each line in the spectrum will cover a range of 5 Hz. These spectral lines have also been known as “Bins” or “Cells” in the past however these terms are less common now. (SEF Se, = Chapter 4: Signal Processing and Data Acquisition Page 4-18 sir ren 2002 “The consents fis valing materi are he apt 0 SKF AB ae may nt be reprocucedin ay frm oy ay mean even eect lees permis i raredzoe 150 18496 Vibration Anais Catsoy2 olerF. ‘The concept of “lines of resolution” can be compared to the swept-filter of an analogue frequency analyser. Where this instrument has one filter that can be moved or tuned over the frequency range, the FFT analyser uses many filters (lines or bins, also sometimes called bandwidth) stacked side-by-side to cover the selected frequency range. In the example, there would effectively be 400 Individual filters, each with a 5 Hz width to cover the selected 2,000 Hz frequency range. Each vibration within the 5 Hz is then processed together and the total is shown as a single value at the centre of the line. Therefore, the spectrum that is seen after the data has been processed is, essentially made of 400 individual values with a line joining them. This is an important fact because this determines the accuracy of the data that Is presented. This also explains why, when the cursor is placed on a peak in the spectrum, the frequency stated is actually the frequency at the centre of the line and not necessarily the frequency of the vibration. For example, assume that the machine being analysed is a 2-pole electric motor. Any two pole ‘motor will try to run at a speed that Is relatively close to the actual line frequency. Unless this is an AC synchronous motor, there will be some slippage present. Electrical line frequencies will be either 50 Hz or 6OHz, depending on which country or industry the analyst Is working. For 50 Hz supplies, typical running speeds are between 2,880 rpm and 2,950 rpm. For 60 Hz supplies, typical running speeds are between 3,520 rpm and 3,550 rpm. For this example, assume that the electrical supply (line frequency) Is 50 Hz and the actual speed of the motor is 2,940 rpm, which is 49 Hz Remember that mechanical problems such as misalignment, looseness, etc., will result in vibration ‘occurring at 1x and 2x RPM or 49 Hz and 98 Hz in this case. In addition to this, electric motors: such as this may also have electrical problems such as open or shorted windings, unequal air gap, ‘or broken rotor bars that cause vibration because of unbalanced magnetic forces between the motor rotor and field (stator). These and other electrical problems will result in vibration occurring at 1x and 2x the AC electrical line frequency (50 Hz and 100 Hz). In this case, since the motor Is 2 ‘two-pole motor, the rotating speed of the magnetic field in the stator will be at 1x line frequency or exactly 50 Hz. ‘Therefore, both the 1x rpm and the 1x line frequency (Fi), which are only separated by 1 Hz, will fall nto the same line of the FFT. Similarly, the 2x rpm and the 2x line frequency, which are also separated by only 2 H2, will also fall into the same line of the FFT. In this situation, it should be lear that with resolution of 5 Hz, it will not be possible to distinguish between mechanical or electrical problems, which can lead to a misdiagnosis and therefore an incorrect work order being put in place. ‘The reason for this error is the accuracy of the displayed data. The accuracy of the data is ‘effectively % (one half) of the resolution. The accuracy of the data Is calculated as follows. meimey) = mn _ = re In order to resolve this issue, the resolution will need to be changed. This can be changed by elther, reducing the frequency range and maintaining the same number of lines, or by increasing Chapter 4: Signal Processing and Data Acquisition Page 4-19 © ser FeB2003 ‘The content of is tang materi are the copyright of SKF AB and may nt be reproduced in ary fr by ny eas (ven eva unless ermssion rated,za 150 12426 Vraton Anais Category 2 okKF the number of lines and maintaining the same frequency range. In order to separate two closely spaced frequencies, there is a “Rule of Thumb" which says “select a resolution such that 3x Resolution is less than the difference between the frequencies” or, set a resolution that is less than 1/3 of the difference between the frequencies. In the case of the previous example, this will mean that the resolution would need to be less than 0.33 Hz. When trying to separate out electrical and mechanical frequencies in 2-pole electric motors, there is a commonly used spectral set up that has a frequency range of 200 Hz and 3,200 lines of resolution. This set up results in a resolution of 0.0625 Hz (or 3.75 cpm) and will completely separate the mechanical and electrical frequencies, as well as identifying any sidebands that are present. ‘This concept of accuracy of the data when placing a cursor onto @ peak in the spectrum also explains why, when using a harmonic cursor, the higher harmonics don’t line up with peaks on the spectrum. In the previous example, with a running speed of 49 Hz and a resolution of 5 Hz, the lower harmonic cursors will likely line with the peaks at 1x, 2x & 3x rpm as the difference will be minimal. However, when looking at the higher harmonics, around 10x rpm, the difference here is not 1 Hz, but will be 10 Hz and will mean that the harmonic cursors don't align with peaks, but look slightly higher above the frequency of the peaks. Averaging At the beginning of the FFT process, a sample or dataset of vibration time waveform data was collected that, hopefully, represents the true vibration behaviour of the machine being analysed. Unfortunately, while this sample of data was being collected, it is quite possible that some transient event may have occurred to introduce vibration characteristics that do not represent the normal vibration behaviour of the machine. Perhaps the machine was bumped while the data was being collected or maybe a machine nearby was started or shut down, ‘Therefore, because of the possibility of transient conditions, the normal practice is to perform the FFT process on more than one dataset and average the results, This will tend to minimise the influence of transient signals so that the real characteristics of machine vibration can be determined. This process is called spectral averaging. Some FFT analysers provide a choice of ‘averaging techniques, such as linear, RMS (sometimes called Power), exponential, peak hold, and synchronous time averaging, For most routine analysis work, simple linear averaging Is used. ks. at ccseenpiay Figure 20. Limiting the Effects of Transients by Increasing the Number of Averages. Chapter 4: Signal Processing and Data Acquisition Page 4-20 sxe 92019 ‘The content of hi trang material are the emit of SKF AB and may nt be "eproced nary frm ar by ay means (oven exacts less permission rated,wana 108436 Vatn Anais Cates 2 oKF. Figure 20 shows comparative FFT's taken on a motor direct coupled to a reciprocating compressor. One FFT was taken using a single dataset or “one average.” The second FFT was taken using 64 datasets and averaging the frequency and amplitude characteristics from these datasets. Considerable differences can be seen in the comparative data, not only in the amplitudes of certain frequencies, but the actual frequencies displayed. In the FFT from the single sample (top), a significant vibration component can be seen at a frequency of approximately 5x RPM, However, the frequency component is virtually non-existent in the FFT displayed after averaging 64 datasets of data, The component at Sx RPM that appeared in the single sample average was apparently 2 transient vibration and only there during the time the dataset was collected, Obviously, this vibration component is not representative of the machine's. normal vibration characteristics and could confuse the analysis process. RMS Averaging (Also known as Power Averaging) RMS, or Power, averaging can be thought of as averaging the amplitude of each component in the spectrum. A power spectrum is one that does not provide any phase information but will emphasise the difference in amplitudes of each frequency component. The amplitude of each frequency component in a power spectrum is directly proportional to the square of that component in an FFT spectrum. Therefore, for the purpose of emphasising the FFT spectral amplitude, instruments will first calculate the power spectrum and then take the square root of each component. There are two practical consequences when the spectrum is calculated in this manner: 1. The power spectrum will require less calculation time because the calculations for the square root are not performed. 2. The contrast between high and low amplitudes is emphasised because the quantities are squared. This technique is trying to determine the average amplitude of the signal and it is @ very valuable technique for determining the average power in any one of the filters in an instrument, Therefore, by collecting a higher number of averages, the estimated power level becomes more accurate, However, if there is a small signal that is to be measured in the presence of a lot of noise, the signal-to-noise ratio will not be improved when RMS averaging is used. This is because RMS averaging does not “filter out” the noise, it includes it in the calculation. Therefore, the accuracy of the total power (signal-plus-notse) that is measured will be improved with RMS averaging, but the ‘signal-to-noise ratio will not be improved. This type of averaging does not reduce unwanted background noise. If the Instrumentation supports this feature, it would be used when the amplitudes of the frequency components in the spectrum vary significantly. The result of an RMS average of successive spectra is an improved estimate of the average level of the amplitude of each frequency component because it causes varying amplitudes to converge to their average value. ‘Chapter 4: Signal Processing and Data Acquisition Page 4-21 o sie re82002 ‘The cntent of his waig maar are the opt of SKF AB and may not be reprodcedn any Frm or by my rman een exec) less persion grantedwos 15018696 Vrain Arabi Category2 oF. Linear Averaging Unlike RMS averaging, linear averaging will improve the signal-to-noise ratio. This is because in ‘many situations, the amplitudes of fundamental and harmonic frequencies are reasonably constant although they are often accompanied by random frequencies that have varying amplitudes, such as cavitation for example. Even if the amplitudes are low, the discrete frequency amplitudes usually remain reasonably constant, while the amplitudes of the random frequencies will vary between zero and a maximum value, Therefore, when large numbers of averages are collected, these random frequency amplitudes will tend to approach zero, while the discrete amplitudes remain reasonably constant, Linear averaging is where the value of each line of the spectrum is added to the value of the corresponding line for each sample that is collected. The total value is divided by the number of spectra collected and the result is the average value for that line. This is shown in Figure 4.17, where four (4) samples have been collected and the result seen in the database is the average of all four samples. Figure 24. Linear averaging of four spectra near averaging is @ useful method to use when trying to obtain repeatable data for trending the condition of equipment in a condition monitoring programme. There are usually four, eight, or 16 averages specified in the software programme. The disadvantage of using linear averaging is that because the vibrations measured at each location are a summation of the total vibration that is seen by the transducer, it does not separate the vibrations that originate from that point from those that are generated elsewhere and are only sensed at that point. ‘Synchronous Time Averaging ‘The difference between linear averaging and synchronous time averaging Is that a trigger source is required in order to perform a synchronous time average. There Is no need for a trigger source when using linear averaging. The trigger source (reflective tape on a rotating shaft for example) will provide a synchronising signal to a triggering device. The most commen types of triggering devices are photocells and laser tachometers. Synchronous time averaging should be performed with a high number of averages than for normal routine condition monitoring readings. This technique will normally only be applied when a detailed analysis of a piece of equipment is being Chapter 4: Signal Processing and Data Acquisition Page 4-22 oskrrep2019 “The contents f his raining mari at th opal SHF AB nd may nat be reproucein any Fr rb any ran (ven erat)woos 15018436 Viraton Anais Cteony 2 oer. carried out. In such situations, it is common to use between 64 and 128 averages. If the equipment is a quite stable environment, 32 or 64 averages may be sufficient to perform the analysis. However, where there Is a lot of background noise, or random vibrations from nearby equipment, at least 64 and preferably 128 averages are recommended. The result of this technique is a spectrum that will show a very much improved signal-to-noise ratio. ‘The synchronising signal will trigger the start of the time sample and, as long as the trigger source ‘and the triggering device are not moved, the sample will be collected at the exact same point in the cycle every time. Therefore, the periodic part of the input where the trigger is targeted will always remain the same in each time record taken. This means that any random, or non synchronous vibration, will have different amplitudes each time the sample is collected. When the actual average spectrum Is calculated, all the non-synchronous components will be averaged out and the components that are seen in the averaged spectrum will be those that are produced from that rotating component. ‘The non synchronous frequencies will not be displayed because they are not direct integer multiples of the 1x rpm synchronising signal. Therefore, bearing defect frequencies, electrical line frequencies and their multiples, as well as other harmonically distorted signals, such as 1x rpm and multiples from 2 nearby shaft, will be forced to zero. This will happen even if the difference in speed is very small if enough samples are collected This is a very useful method for analysing equipment that has many components that rotate ‘at almost the same speed, such as banks of fans, rolls on paper machines or gears and shafts in @ gearbox. After the results are transformed into the spectrum, the running speed and harmonics will be clearly seen and can be accurately measured because of the improved signal-to-noise ratio (reduced noise floor). Exponential Averaging ‘When there are transient events occurring and if it is available, exponential averaging is the best ‘option. In this process, the most recent amplitude values receive the most weight in the computation, while the oldest values have the least influence in the computation, Overlap Processing It is important to remember that the time sample Is not constant and is deliberately varied in order ‘to change the frequency span of the instrument. When a wide frequency span is required, the time record must be shorter so that the whole frequency span Is seen. Therefore, a certain amount of time will be required in order to capture the first time sample, no matter what instrument Is being used and irrespective of whether or not a 32-bit DSP chip is installed or not. The time that is required to collect this first time sample was given by Equation 13: Chapter 4: Signal Processing and Data Acquisition Page 4-23 esi Fe82013 “he cntet this traning materi are th copyright of SAF AB and ray nt be repratcedn an former by any means even extras) ules persons rantedAfter the time sample has been captured, a finite time will be required to calculate and display the spectrum and this is where the speed of the processor and the instrument will have an affect. This fact cannot be ignored and is especially important when low frequency measurements are to be collected. In such situations, the gathering of the time record may require much more time than the actual spectral calculation and during the actual collection of the data, the FFT processor sits waiting for the data to be delivered, ‘Assume that a measurement Is to be collected on a machine that runs at 480 rpm and that there ‘are several components that all run at approximately the same speed. In order to identify the running speeds of the components on this machine 2 spectrum with a relatively high resolution would be needed. Therefore, settings such as a frequency range of 20x running speed, 1600 lines ‘and 8 averages could be used and using Equation 4, the time to collect the data can be calculated. Equation 14 To collect this spectrum it would take 1 minute and 20 seconds assuming that there is no overlap Processing. The time it takes for the FFT circuitry to process the data and produce a spectrum may only be around 5 seconds (depending on the instrument used). Therefore, it would take 1 minute 20 seconds to collect the time sample and 5 seconds to process the data, giving a total of 1 minute 25 seconds before the data is seen on the screen of the instrument, However, the FFT processor has been sitting idle for the first 1 minute and 20 seconds of this. This can be seen in Figure 4.18 where the time sample and the FFT calculation are shown as blocks of time. [Total time for data collection is 2 Osecs_|10secs | 10secs |40secs | 0secs |10secs |0secs |a0secs FFT calculation takes 5 sec Total time for measurement is 1 min 25 sec in 20 sec Figure 22. The total time to collect a measurement Chapter 4: Signal Processing and Data Acquisition Page 4-24 © see Fee 2013 “he contents of his traning mati are the corti of SKF AS an ay net be presen ay frm aby ay means een eta) uss person ranted.aes 130 28436 Vraton Anais Cleon 2 moKF. This may not seem like @ big problem on its own, but remember that this is only one location on the machine, What about all the other locations that are to be collected on the machine? If this was. a motor driven pump for example, there could be horizontal, vertical and axial measurements on all four bearings, which would give a total of twelve measurements each taking 1 min 25 seconds and a total time of around 15 minutes to collect the data on this machine, Now, the rest of the equipment in the plant must be considered because four machines of the same type will mean that \t will take around one (1) hour to collect the data. For routine monthly (or weekly) data collection, this Is unacceptable. No organisation has the luxury of unlimited time or resources to be able to spend this amount of time collecting vibration data. Therefore, to speed up the data collection process, itis very common to “overlap” a new time sample with some of the previous time sample data. In doing so, the time it takes to collect all the samples is greatly reduced and the total time spent collecting the data is also reduced. This is shown in Figure 4.19, where the data is overlapped by 50% 20 ja0secs | 5 aosacs | 5 [sosece | js0secs | 5 ja0secs | 5 a0secs_ | 5 ae ansecs | [tates ase Total time for measurement is 50 seconds Figure 4.19 Data collection time with 50 % overlap In essence, a full length time record is first captured. Then portions of new time records are stored ‘and added to portions of the old data. With an overlap of 50%, half of the new time record is taken in and combined with half of the old time record, and since only half of the new time record is taken in, this reduces the data collection time to just over a half of what it would be if there was no ‘overlap processing. In reality, analysts should not use any more than 50% overlap for routine data collection because it will affect the quality of the data that is collected. Exercise 4.1 ‘A machine is to be monitored in a condition monitoring programme with the following spectral settings: Freq. Span 30,000 cpm # Lines 800 # Averages 8 Chapter 4: Signal Processing and Data Acquisition Page 4-25 546 Fea 2012 The contents of th tring materi ae the copyright f SKF AB anny et be reproduces in my frm aby years even eta) ulss permission rantedwos 15038496 Vain Andis Ctegary 2 olrF. Calculate the following data collection times: 2) No overlap processing b) 50% overlap processing ©) 75% overlap processing In order to do this, two time buffers are required in the analyzer. For 50% overlap, the sequence of events is as follows: when the first buffer is half full (i., it contains half the samples of a time record), the second buffer is connected to the data stream and also begins to collect samples. As soon as the first buffer is full, the FFT is calculated and the buffer begins to take data again. When the second buffer is filled, the FFT is again calculated on its contents and the result sent to the spectrum-averaging buffer. This process continues on until the desired number of averages is collected. However, using overlap processing will mean that there is a trade off between time and accuracy because an overlapped time record contains old data, the instrument adjustment is not completely correct. The results will indicate the direction and the amount of change, but the instrument will require a full time record after the change for the new spectrum to be displayed accurately. ‘Therefore, for measurements such as transient measurements which occur shorter than the time record, overlap processing is useless. When a transient event is longer than the time record, the real-time bandwidth of the analyser and spectrum recorder Is usually the limiting factor. If this is rot the case, overlap processing allows more spectra to be generated from the transient and this will improve the spectral data captured during the transient event, Overlap processing can greatly reduce the amount of time it tales to collect measurements and will eliminate most of the time that is wasted when the FFT circuitry is waiting on data to process. For ‘most measurements that occur during a vibration survey, up to 50% overlap processing can be used with @ good degree of confidence. And, this Is particularly useful for low-frequency ‘measurements that would otherwise require excessive data collection times, ‘The next question usually asked is: “How many datasets are required to a spectral average?” While hho absolute answer can be given to this question, some reasonable guidelines can be offered based on experience. First, the number of spectral averages is @ compromise between the accuracy and validity of analysis data and data collection time, Tt should be obvious that it will take more time to perform an 8 average FFT than a 4 average FFT. Fortunately, the number of averages does Chapter 4: Signal Processing and Data Acquisition Page 4-26 sir rea2002 “The conan fs traning materi are he opr af SKF AB nc may nol be reproduced nny frm by ny meas [oven evacuees antewi203- 15018436 Viraton Anais Category 2 oKF. not affect the amount of instrument and computer storage required. This is only governed by the number of lines of resolution and the frequency range selected for the FFT. ‘The following are some general guidelines that may be helpful in selecting the number of spectral averages for FFT's obtained for predictive maintenance and detailed machinery analysi 1. For general machines included in a predictive maintenance programme, such as fans, blowers, pumps, motors, etc., FFT's using four sample averages are normally adequate. Although there may be some affects from transient conditions, the ultimate goal of routine vibration checks in a predictive maintenance programme is to detect potential problems. Once the possibility of a developing problem has been detected, a thorough follow-up analysis should be carried out to verify and identify the problem. If periodic checks reveal a potential problem that is actually the result of a momentary or transient condition, a detailed analysis will indicate this. 2. When dealing with very high frequencies of vibration, such as gear-mesh frequencies and those from defective rolling element bearings, around 8 averages are usually recommended. Sources of high frequency vibration tend to be somewhat more erratic and variable than problems that cause lower frequencies such as unbalance and misalignment. Since higher Fmax FFT's take less time, there Is no significant increase in data collection time using a slightly higher number of spectral averages, 3. When performing a detailed analysis of a machine's vibration, analysis time is rarely a major concern, In other words, whether it takes five minutes or ten minutes to collect the detailed data needed to analyse the problem Is of little concern. Therefore, when performing a detailed analysis, a minimum of four to eight averages are recommended for general types of machines. When dealing with high frequencies of vibration, such as gear-mesh frequencies, 2 minimum of eight averages Is recommended. 4, Where the appropriate number of spectral averages is not known, a simple comparison can be performed to determine the number best sulted for data collection and analysis on a specific machine. Simply take and observe a FFT taken with four averages and compare it to a FFT taken with eight averages. If the four and eight average FFT's appear different, take a FFT with sixteen averages and compare it to the one taken with eight averages. If the four and eight sample averages are nearly the same In appearance, then a four sample average is probably sufficient. Display or Storage of Data ‘The final step in the FFT process is the graphic display of the machine's amplitude-versus- frequency vibration characteristics on the screen of the FFT analyser. ‘Although the FFT data can be studied and examined directly on the instrument screen in the field ‘or plant, the normal procedure is to store the data in the instrument along with FFT’s taken at other measurement points and machines and then download the data to the associated predictive maintenance software for more precise, detailed study. ‘This way, several FFT's can be displayed simultaneously and hard copies made for a more thorough comparison and analysis. Chapter 4: Signal Processing and Data Acquisition Page 4-27 sir Fe8 2003 “The contents fis raining material are the apt of SH AB ane may nt be reprocuendin ay frm ob oy meas evn ext] sess permission rantedw2ca 15018436 Vibration Anais Category 2 oKF: Figure 23. Typical Data Display from an FFT Data Collector. Signal Processing Methods for Rolling Element Bearings ‘Acceleration Enveloping HFD (High Frequency Detection) ‘Shock Pulse Spike Eneray Envelope Signal Processing When viewed in the frequency domain, repetitious gear mesh and bearing defects typically emit low energy vibration signals in the lower frequencies, especially in the early states of failure, However, because these defects emit impact signals that are highly pulse shaped and of very short duration, they generate harmonics out to very high frequencies. Generally, low frequency vibration from unbalance or misalignment masks the low amplitude bearing vibration signals and the bearing faults are not detected, For these reasons, signal processing techniques have been developed to filter out low frequency vibration energy and isolate and emphasize vibration signals occurring in specific high frequency ranges. Overall Vibration Signal (Including Bearing Fault Signal) A ei Bearing Fault Signal Figure 24. Bearing Fault Signal. Chapter 4: Signal Processing and Data Acquisition Page 4-28 see Fe8 2012 The contents hi teaing materi ae the copyright of SKF AB an may net be reproduced in ay frm orby ay meas even eat nes permision rataos 15018436 Viraton Anais Category 2 olKrF. ‘The technique of envelope spectrum analysis makes use of the high frequency resonance’s that are excited on the bearing components, etc., to act as both a mechanical amplifier and a carrier of the low Frequency fault condition information, thus giving an improved signal to noise ratio. If a surface irregularity exists on a rolling element or bearing race, then a burst of high frequency |bration is generated each time that defect contacts the race or element. This high frequency bration is a “ringing response" of the bearing housing to the impact of the rolling element or race with the surface irregularity. These bursts of vibration will repeat periodically at a rate determined by the bearing geometry and the machine running speed. Such repetition rates are known as bearing defect frequencies. There are four frequencies associated with a rolling element bearing: Inner race frequency (BPFI) Outer race frequency (BPFO) Ball spin frequency (BSF) Fundamental Train frequency (FTF also known as cage Freq.) The following table gives the whole range of calculations involved to determine these frequencies. B INNER RACE BPFI aC + cos °) X RPM OUTERRACE = BPFO= s - 3 cos ) X RPM Po 7 2 Bs! aa Ba BALL (OR ROLLER} ae ( 1 ( Lege cos 0) )x RPM. 1 Ba CAGE ce TC - Hos 0) X RPM =.35 -45 X RPM No =Number of Balls or Rollers Pa = Bearing Pitch Diameter Ba = Ball or Roller Diameter 8 — =Contact Angle (degrees) Figure 25. Table of Defect Frequencies Chanter 4; Signal Processing and Data Acquisition Page 4-29 esi rea2002 “he cntants f thc train mater are th cepyrht of SAF AB and ay et reproduced in ry frm oy ry mean een ore les ersson awi203 'SO 18436 Vibration Analysis Category 2 1 Typical Velocity Spectrum 2 - Acceleration Enveloping Process. Low frequency events ftered out, high frequency fault harmonics emphasised due to low signal to noise ratio, these high frequency harmorics are then summed and folded backto the defect’s fundamental frequency 10 kHe olerF. 2 ~ Acceleration Envelaping Process, Band Pass Fitter is specified according to defect harmonic frequencies of interest 10 kHe 4 ~ Acceleration Enveloped Spectrum Defect fundamental frequency Resulting enveloped signal, measured in gE Fundammentat Figure 26. Enveloping Process Enveloped acceleration, or “demodulation,” is a signal processing technique that greatly enhances ‘an analyst's ability to determine the condition of rotating equipment. This signal processing method enables the detection of repetitive low energy impulse signals, such as bearing and gear mesh defects, much earlier than traditional analysis techniques. » Note: By detecting bearing problems early, maintenance personnel have the opportunity to apply proactive corrective measures that can extend the bearing's service life. In very simplified terms, the enveloped acceleration process first filters out the machine's low frequency vibration noise and looks for harmonics of repetitious signals occurring in a higher frequency band specified by the user, as seen in Figure 26 #2. These defect signal harmonics are present in the early stages of failure and typically occur at the very high frequencies, anywhere from 5 to 40 times the fundamental defect frequency (Figure 20 #3). ‘Once these harmonics are detected, an enveloper, in effect, sums the harmonic signals together and then folds the enhanced or summed signal back to the defect’s fundamental frequency range. ‘The result is a frequency spectrum with a much more obvious and enhanced display of the defect’s fundamental fault frequency (Figure 20 #4) >» Note: Because a defect's impulse signals are generated more forcefully in the bearing's load zone, enveloped acceleration measurements produce the best results when the data collection sensor is positioned in, or as close to, the load zone as possible. Chapter 4: Signal Processing and Data Acquisition Page 4-30 ‘© ser Fee 2033 “The contents of his rig materi ar he att of SKF AB an may nt be mre ay fro yan means een extra) lr perrion rantedwos 1501836 Viation Arb Cteory2 oKF e Figure 27. Bearing Loaded Zone Selecting the Enveloping Filter Bandpass Filter SHz-100Hz |, Fecelerometer| ‘50 H2—a KHE ‘500 H2—40 kHz ‘5 kHz — 40 KHZ Felt Rolls, Bearings Bearings _ Gears Frequeney-He Filter #1 Filter #2 Filter #3 Filter #4 Figure 28. The four Bearing Envelop Filters For SKF Bearing Envelope, there are four filters available Filter | Frequency Band ‘Speed Range Analysing Range 1 5-100 Hz 0-50 rpm 0-10 Hz 2 '50-1,000 Hz 25-500 rpm 0-100 Hz 3 '500-10,000 Hz 250-5,000 rpm 0-1,000 Hz 4 5,000-40,000 Hz 2,500-... rpm 0-10,000 Hz Figure 29. Examples of Enveloping Filters Chapter 4: Signal Processing and Data Acquisition Page 4-31 sir Fea2013 “Te contents fh ining mater are th capt SHE AB and mayne be reprocscaain my frm ory ay ear even ent) unless permission graredoa 150 18496 Vrain Anais Category 2 ORF. The key to detecting bearing faults is to enhance the low amplitude bearing defect harmonics, without including the high amplitude rotational vibration signals. To accomplish this, the user selects from a range of band pass filters, with choices dependent upon data collector and associated software, for the envelope measurement based on shaft speed and the frequency range of the bearing or gear mesh frequencies. ‘The purpose of the band pass filter is to reject the low frequency; high energy rotational signals generated by the machine's unbalance, misalignment, looseness, etc., and highlight the high frequency harmonic components of the bearing's defect frequencies (5 - 40 X defect frequency). This process results in a significant improvement of the signal to noise ratio, allowing the very small defect harmonic components to be detected, summed, and converted by the envelope process into the fundamental defect frequencies and their associated harmonics. The result is a spectrum that enhances any repetitive bearing defect signals and suppresses all low frequency and non-repetitive signals. It is important to remember that enveloped spectra are enhanced. The amplitudes are relative to the summation of the energy in the specified filter band and are dependent on variables such as sensor location and shaft speed. If these variables remain constant, the enveloped acceleration signals can be trended. As the defect worsens, the signal's strength increases. Over time, a trend plot can be used to show increasing amplitudes at the defect frequencies. Figure 30, Displaying bearing frequencies on an envelope spectrum Very early, microscopic bearing defects are not visible on normal acceleration or velocity vibration spectra due to the following reasons: + No vibration yet exists + If vibration does exist, it occurs in the bearing defect frequency range (which may not be shown by the FFT) ‘+ Its of such small amplitude that itis hidden by low frequency rotational vibrations Some guidelines for Envelope readings are included below, however please note that these ‘guidelines dependent on shaft diameter and rotational speed. Chapter 4 Signal Processing and Data Acquisition Page 4-32 0 sie Fee 20a “The contents of his ning material are he capaho HF AB and may not be reproduced in ay fro by means eer extrac) else permission granted‘Wwi203 15018426 Vraton Aniston 2 oKF ‘Shaft Diameter and Speed Figure 31. Guidelines for envelope alarms settings SEE™ Specti SEE is an acronym for Spectral Emitted Energy, or the eneray that is acoustically emitted due to very high frequency signals caused by faults in a bearing or gearbox. SEE technology provides very ‘early bearing and gear mesh fault detection by measuring these very high frequency acoustic emissions and also detects bearing lubrication problems stemming from contamination, inadequate lubrication, or the breakdown of the lubricant film by a local defect. These measurements produce ‘an overall numeric value and an FFT spectrum. | Emitted Energy ‘Operating in the very high frequency range, SEE measurements detect structural stress initiated by lubrication problems and bearing surface contact. Early indications of inadequate lubrication provide the analyst time to apply corrective lubrication actions, prolonging the bearing's service life, > Note: As with acceleration enveloping measurements, SEE measurements should be performed in the bearing's load zone, or as close to the load zone as possible. ‘The SEE sensor is not held directly against the machine's surface. Instead, it uses a film of grease or oil between the sensor and the surface to help conduct the acoustic signals from the machine to the sensor. Consistency in SEE data acquisition Is critical. Permanent mounted SEE sensors are preferred, as they assure consistent probe placement and pressure, and consistent fluid coupling. (oem Mesttog Samet How Figure 32. SEE Sensor Chapter 4: Signal Processing and Data Acquisition Page 4-33 “The catet Wis nng material are th apt SHE AB ane may net be reproduce ay fr by ay meas even ext) less permission raedcos 5 SSO 18426 Vraton nays Category 2 KF: SEE Numerical Analysis Depending on the monitoring instrument, SEE technology provides either numerical or spectrum output (or both). Because SEE signals indicate deterioration in the proximity of the monitoring sensor, a clear understanding of machinery components in the area is essential. For example, if a gearbox is adjacent to the bearing where the SEE measurement Is performed, acoustic emissions from the gearbox may add to the bearing's SEE measurement. In this case, a high SEE reading may be a normal reading for a good bearing condition (due to additional SEE’s transmitted from the gearbox). In general, a “higher than normal” SEE measurement indicates a problem, If no SEE signals are initially present, use the following as a guideline for severity of the SEE signal. + 0-3: Noiidentifiable problems + 3-20: Lubrication problem, contamination, bearing defect with light load, or a small bearing defect with normal load + 20-100: Bearing defect or contamination + 1004: Severe bearing problem Use these figures as guidelines only. Measurement trending experience determines valid figures for your SEE monitored machinery. SEE Spectrum Analysis Very early, microscopic bearing defects are not visible on normal acceleration or velocity vibration ‘spectra because the vibration doesn't yet exist, or if it does exist, it is in the bearing defect harmonic range which may be outside the frequency range of the spectrum, or the amplitudes are 0 small that is is masked by other defects (such as unbalance, misalignment etc). SEE measurements monitor the ultrasonic frequencies at which these acoustic emissions occur (250 - 500 kHz) and filter out all the low frequency rotational events. The very early bearing defect's repetitive acoustic signals are enhanced and shown as peaks at the bearing’s defect frequencies. With a bearing, if no peaks exist in the SEE spectra, no acoustic SEE signals are present. > Not There are non-bearing related instances where SEE spectra do no show obvious peaks, but where the overall level of the SEE spectrum may indicate a problem (for example, cavitation or friction due to fretting). SUNH6 REDIF Ang 024 Fry 9 1 I. 1 Ce OAS Se THR EE (6 OE 2 (TE dove ain rer Figure 33. Example of SEE Spectrum Chapter 4: Signal Processing and Data Acquisition Page 4-34 © sxe Feo2019 ‘The comnts ofthis wang material ar he cpt of SHF AB andy notte reprocces ny fr ob ay aan een ent] sole prs ranewas 1501836 Viraton Anais Cteory2 okKF HFD ~ High Frequency Detecti ‘hp 02712.Dae Tan: $QU19040 2 FM Se ee esas ronsnoea ansagss ens roniases ranges cans Figure 34. HFD Trend ‘The High Frequency Detection (HFD) processing method displays a numerical value for high frequency, low amplitude vibration generated by small flaws within a rolling element bearing. HFD provides early warning of bearing problems. HFD measures and displays a numerical value proportional to the overall signal level within a high frequency band pass. The detecting sensor’s resonant frequency is within the band pass and is used to amplify the low level signal generated by the impact of small flaws. The HFD sensor's resonance frequency is around 25 kHz and measures overall vibration in a frequency range between 5 kHz and 40 kHz, Readings are performed using an acceleration sensor and are registered in G's (RMS or 0-peak). HFD measurements are taken over a period of time with a portable data collector or with Permanent on-line data acquisition equipment. As the HFD program develops, collected HFD measurements determine normal and alarm values for the specific equipment being monitored. In addition, HFD values can be trended in a trend plot to more easily detect relative changes in amplitudes. Unlike enveloping, HFD signal processing does not Indicate where the fault is (e.g., inner race, outer race, cage, rollers). However, by providing a numerical value for overall energy within the high frequency band, HFD offers an easy to use method for trending and detection of bearing problems, > Note: HFD readings are also affected by seal rubs, metal-to-metal contact, and cavitation. ‘Spike Energy (gSE) ‘This process was developed in the 1970's by IRD Mechanalysis and it Is designed to mechanically and electrically respond only to high frequency problems that excite the resonant frequency of the mounted transducer. The acceleration signal is processed through a band pass filter that has a definable low frequency corner. The filtered data is then passed through a peak detection circuit that is set to look for a peak-peak repetition rate. The result is a numerical value which is the product of the number and amplitude of the impacts in a unit of time and is expressed as Spike Chapter 4: Signal Processing and Data Acquisition Page 4-35 © si Fe02019 “The contents his aig materi are the opiate SHE AB and may not be eroducedin ay former y any means (even ext unless permissions ranted