Professional Documents
Culture Documents
10.1 Introduction
The costs of lost production can be extremely high when machinery breakdowns lead to
unplanned stoppage of the plant. If you apply properly and manage it well, a condition
monitoring system is highly cost effective in minimizing maintenance downtime in two
ways:
- By providing advance warning, you get enough lead-time to prepare for action
- By ensuring that the plant and machinery will not deteriorate to a condition where
emergency action is required.
Condition monitoring is concerned with extracting information from machines and plants
to indicate their condition, and to enable them to be operated and maintained with safety
and economy. Formally condition monitoring is defined as the continuous or periodic
measurement and interpretation of data to indicate the condition of an item to determine
the need for maintenance.
In spite of the large number of techniques and the amount of instrumentation available,
there are only four basic methods of condition monitoring:
1. Visual monitoring : Machine components are usually inspected to determine their
condition
2. Performance monitoring: The condition is assessed by measurement of the
machine or the component to know how well it is performing its intended duty
3. Vibration Monitoring: The condition of the moving components in a machine is
assessed from the amount and nature of the vibration they generate
4. Wear debris monitoring
The condition of critical component surfaces, subject to loading and relative
movement is assessed from the wear debris, which they generate. They are usually oil
washed components, and the collection and analysis of debris is done via the lubricating
oil.
In the above methods, the existence of a problem is usually detected from the general
level of the measurement and its rate of change. The nature of the problem can generally
be determined from detailed analysis of the measurements obtained. Table10.1 presents
the details of the general methods of condition monitoring. The methods of monitoring
described in the table are in effect a mechanism of communication between a machine
and a monitoring engineer through the symptoms and their analysis using the human
senses (sight, smell, hearing, feel or general sense) or objective sensors in condition
monitoring and corresponding technological facilities to detect and diagnose the machine
ailments at the earliest with a very high success rate of detection.
1
Table 10.1 .Detection and definition of machine problems through condition monitoring
2
inside. The rate (frequency) and magnitude of the vibration is completely determined by
the excitation force, direction, and frequency and therefore vibration analysis can
determine the excitation forces at work in a machine. These forces are dependent upon
the machine condition, and knowledge of their characteristics and interactions allows one
to diagnose a machine condition accurately.
The vibration transducer is a device that produces an electric signal that is a replica or
analog of the vibratory motion it is subjected to. Different types of transducers respond to
different parameters of the vibration source such as proximity probe to displacement,
velocity probe to velocity and accelerometer to acceleration.
The proximity probe (Fig.10.1a), also called an “Eddy Current Probe” is a permanently
mounted unit, requires a signal conditioning amplifier to generate an output voltage
proportional to the distance between the transducer and the shaft.
The transducer measures relative displacement between the bearing and the journal and
does not measure total vibration level of the shaft or the housing. The displacement
transducer is very commonly installed in large machines with journal bearings where it is
used to detect bearing failure and to shut the machine down before catastrophic failure
occurs. These transducers are frequently used in pairs oriented in 90 degrees apart and
can be connected to the vertical and horizontal plates of an oscilloscope to display the
“orbit”, or path of the journal as it migrates around in the bearing. The frequency
response of the displacement transducer extends from DC (0 Hz) to about 1000 Hz.
The velocity probe (Fig.10.b) has a moving coil outside a stationary magnet, consists of a
coil of wire and a magnet so arranged that if the housing is moved, the magnet tends to
remain stationary due to its inertia.
3
Fig.10.b The Velocity Probe
The relative motion between the magnetic field and the coil induces a current that is
proportional to the velocity of motion. The unit thus produces a signal directly
proportional to vibration velocity. It is self-generating and needs no conditioning
electronics in order to operate, and it has a relatively low electrical output impedance
making it fairly insensitive to noise induction. It is relatively heavy and complex and thus
expensive, and it has poor frequency response, extending from about 10 Hz to 1000 Hz.
The accelerometer (Fig.10c) of the piezo-electric type can be considered the standard
vibration transducer for machine vibration measurement.
The seismic mass is clamped to the base by axial bolt bearing down on a circular spring.
The piezo-electric element (quartz being the most common) is squeezed between the
mass and the base. When a piezo-electric material experiences a force, it generates a
electric charge between its surfaces. The force on the crystal produces the output signal,
which is proportional to the acceleration of the transducer. Accelerometers are inherently
extremely linear in an amplitude sense; they have a large dynamic range. The frequency
range of the accelerometer is very wide, extending from very low frequencies in some
units to several tens of kHz. When mounting an accelerometer, it is important that the
vibration path from the source to the accelerometer is as short as possible, especially if
rolling element bearing vibration is being measured.
4
10.3.2 Displacement, Velocity and Acceleration: Equations of Motion
A vibration signal plotted as displacement versus frequency can be converted into a plot
of velocity versus frequency by a process of differentiation which involves multiplication
by frequency, and this means the vibration velocity at any frequency is proportional to
the displacement times the frequency. In order to obtain acceleration from velocity,
another differentiation is required, and this results in another multiplication by frequency.
It can be seen that the same vibration data plotted in displacement, velocity and
acceleration will have very different appearances. The displacement curve will greatly
emphasize the lowest frequencies, and the acceleration curve will greatly emphasize the
highest frequencies at the expense of the lowest ones. The three curves shown in Fig.10.2
display the same information, but the emphasis is changed.
5
The velocity curve is the most uniform in level over frequency. This is typical of most
rotating machinery and it is a good idea to select the units so the flattest curve is attained
providing the most visual information to the observer. Velocity is the most commonly
used parameter for machine diagnostic work and is the accepted standard.
For an object undergoing simple harmonic motion (SHM) the equation of motion:
Frequency Axis
When plotting vibration spectra from rotating machines we have several choices of units
for the frequency axis. Probably the most natural unit is the cycle per second (CPS) or
Hertz (Hz). Another unit in common use is related to RPM or cycles per minute (CPM).
Forcing Frequencies
The value of vibration analysis of machinery is based on the fact that specific elements in
the rotating parts of any machine will produce forces in the machine that will cause
vibration at specific frequencies generating a vibration spectrum known as signature. One
of the most important of the forcing frequencies is the RPM of the shaft, and it arises
from the fact that any rotor will always have a certain amount of residual unbalance. This
imparts a radial centripetal force on the bearings, causing the structure to vibrate at the
1x, or fundamental, frequency. The so-called bearing tones which are characteristics of
every bearing geometry, are forces generated by defects in the races and rolling elements
of the bearing itself. Gear tooth-mesh frequencies come from the individual impacts of
the gear teeth against each other, and the tooth mesh frequency is equal to the number of
teeth on the gear times the gear RPM. Vane pass or blade pass frequencies are similar to
tooth mesh are equal to the number of vanes in an impeller or number of blades in a fan
times the RPM. Each forcing frequency will create a peak in the vibration spectrum, the
amplitude of the peak being dependent on the severity of the condition that causes it.
Thus the frequency indicates the type of problem and the amplitude indicates its severity.
Most machines have a relatively simple set of vibration forcing frequencies determined
by the geometry of the machine and speed. The existence of other frequencies than the
forcing frequencies, such as harmonics of 1x, in the vibration signature of the machine
indicates non-linear behaviour and the combined magnitude of these new frequencies is a
good indicator of the overall health of the machine. As a machine wears, its clearances
typically become greater and its vibration signature becomes more complex due to the
generation of harmonics and side bands. In trending the vibration level of a machine over
time, a rise in the level of a forcing frequency indicates a change in the mechanism
causing that particular forcing frequency, but does not necessarily indicate any damage to
the machine. For instance, an increase in 1x at a motor bearing indicates an increasing
6
unbalance condition, but if harmonics of 1x appear, this indicates damage, such as
increase in bearing clearance, looseness, or cracking of the structure. Therefore, a strong
1x vibration means the rotor should be balanced, but the appearance of harmonics of 1x
means the bearing and surrounding structure should also be inspected for damage.
Peak Amplitude (Pk) is the maximum excursion of the wave from the zero or equilibrium
point.
Peak-to-Peak Amplitude (Pk-Pk) is the distance from a negative peak to a positive peak.
Root Mean Square Amplitude (RMS) is the square root of the average of the squared
values of the waveform. In the case of the sine wave, the RMS value is 0.707 times the
peak value. The RMS value of a vibration signal is an important measure of its
amplitude.
Amplitude Probability Distribution
The vibration signature of a machine always has some random variation, i.e., its
instantaneous value is not predictable. Nevertheless, the probability of a given value
falling within a certain amplitude range is predictable in a statistical sense. The statistical
probability that the signal will be in any given division can be measured by noting the
time the signal spends in each division divided by the total time the signal is monitored.
The probability density is a measure of the distance away from the mean value the
amplitude will be, plotted against the amplitude. The most familiar probability density
curve is the famous normal, or Gaussian, distribution. The RMS value of a signal with
Gaussian distribution is equal to the Standard Deviation of the signal, σ (sigma). A
random vibration signal will produce a Gaussian distribution, and experience shows that
healthy machines also produce Gaussian distributions. As faults develop in machines, the
amplitude distribution curve changes shape; for instance, a small bearing fault will
introduce “spikes” in the vibration waveform, and this will increase the level of the
“tails” of the distribution curve.
Kurtosis
One mathematical representation of the deviation of an amplitude distribution from
Gaussian is the so-called “fourth moment”, or kurtosis. The Gaussian distribution has a
kurtosis of 3, and higher values of kurtosis indicate increased crest factor of the vibration
signal. Kurtosis is a valid measure of the degradation of a machine, but does not give any
indication of the diagnosis of the problem. It has been reported that kurtosis is especially
well suited to monitoring of reciprocating machines for fault detection. One possible
advantage of using kurtosis as a fault detection parameter is that it does not need to be
trended over time to be effective. A kurtosis of 3 is generally taken as indicating a
healthy machine, with higher values indicating progressive states of fault progression.
7
10.3.4 The Concept of Phase
Often when conducting routine vibration testing of machinery, the overall signal strength
(a broad-band reading) is measured as a first step. Actual machinery diagnostics however,
requires narrow-band signature analysis to identify specific peaks in vibration spectra. In
the same way that narrow-band signature analysis reveals another layer of information,
phase information can provide even more clues when diagnosing machinery and
structural problems.
Phase is a measure of relative time difference between two sine waves. Even though
phase is truly a time difference, it is almost always measured in terms of angle, either
degrees (°) or radians. This represents normalization of the time taken by one cycle of the
wave in question, without regard to its true time period. The phase difference between
two waveforms is often called a phase shift. A phase shift of 90° is a shift of ¼ of the
period of the wave. Phase shift may be considered positive or negative, i.e., one
waveform may be delayed relative to another one, or one waveform may be advanced
relative to another one. These conditions are called phase lag and phase lead respectively.
Phase can also be measured with reference to a particular time. An example of this is the
phase on an unbalance component in a rotor with reference to a fixed point on the rotor,
such as a key way. To measure this phase, a trigger pulse must be generated from a
certain reference point on the shaft. This trigger can be generated by a tachometer or
some type of optical or magnetic probe that senses a discontinuity on the rotor and is
sometimes called a “ tach ” pulse.
8
waveform of the vibration will not be a sinusoid, and may be very complex. In Fig. 10.3,
the high frequency and the low frequency vibration add together to make the complex
waveform and is measured and evaluated by Vibration Analyzers in the time domain or
frequency domain.
In simple cases like this it is relatively easy to find the frequencies by examination of the
waveform, but most vibration signals are far more complex than this and can be
extremely difficult to interpret than to look at the spectrum in the frequency domain
instead of the time domain as done in industrial practice.
Resonance
The “Bump Test”--The machine is impacted with a mass such as a wooden four by four
or the booted heel of the foot of a football player while recording vibration data. If a
resonance is present, the machine vibration will be at the natural frequency as it dies
away.
The “Run Up” or “Coast Down”--The machine is turned on, or turned off, while taking
vibration data. The time waveform will show maxima when the RPM matches natural
frequencies.
“Variable Speed Test”—With a machine whose speed can be varied over a wide range,
the speed can be varied while taking vibration and tachometer data. The data are
interpreted as in the run up test.
The Fig. 10.4 shows an idealized response curve of a mechanical resonance of a very
simple system. The behaviour of a resonant system when subjected to an external force is
interesting. It depends strongly on the frequency of the excitation force. Examples of
resonance in machines are the so-called critical frequencies of rotating shafts. The phase
angle between the excitation source and the response of the structure is always 90° at
resonance. In the case of long rotors such as turbines, the natural frequencies are called
critical frequencies or critical speeds and care must be taken that machines are not
operated at speeds where 1x or 2x correspond to these critical frequencies
9
Fig.10.4 Amplitude and Phase in Resonance
Theoretically, a sine wave exists over infinite time and never changes. The mathematical
transform that converts the time domain waveform into frequency domain is called the
Fourier transform, and it compresses all the information in the sine wave over infinite
time into one point. The fact that the peak in the spectrum shown in Fig.10.5a has a finite
width is an artifact of the FFT analysis. A machine with unbalance has an excitation force
that is a sine wave at 1x, or once per revolution. If the machine were perfectly linear in
response, the resulting vibration would be a pure sine wave like the one shown in
Fig.10.5a. In many poorly balanced machines, the waveform does resemble a sine wave,
and there is a large vibration peak in the spectrum at 1x or one order.
10
Fig.10.5b shows a harmonic spectrum resulting from a periodic waveform,
(symmetrical), in this case a “clipped” sine wave both sides.
The spectrum contains equally spaced components, and the lowest of the components
above zero frequency is called the fundamental, and the others are called harmonics. It is
not uncommon in machine vibration signatures to see a waveform, which is clipped on
one side or something like the one shown here. What this usually means is there is
looseness in the machine, and something is restricting its motion in one or both
directions. This type of signal can occur in machine vibration if there is looseness in the
machine and motion is restricted in both directions. The spectrum seems to have
harmonics, but they are actually only the odd-numbered harmonics. All the even-
numbered harmonics are missing. Any periodic waveform that is symmetrical will have a
spectrum with only odd harmonics! The spectrum of a square wave would also look like
this. Sometimes the vibration spectrum of a machine will resemble this if there is extreme
looseness and the motion of the vibrating part is restricted at each extreme of
displacement. An unbalanced machine with a loose hold-down bolt is an example of this.
Fig. 10.5c shows a short impulse and its spectrum, which is continuous rather than
discrete. The energy in the spectrum is spread out continuously over a range of
frequencies rather than being concentrated only at specific frequencies. This is
characteristic of non-deterministic signals such as random noise and transients. A rotating
machine seldom produces a single impulse like this, but in the bump test, this type of
excitation is applied to the machine, its vibration response will not be a classic smooth
curve like this one, but it will be continuous with peaks corresponding to the natural
frequencies of the machine structure. This spectrum shows that the impulse is a good
input force to use in this type of test, for it contains energy over a continuous frequency
range.
11
Fig.10.5c The Spectrum of a Short Impulse
If the same impulse that produced the previous spectrum is repeated at a constant rate, the
resulting spectrum will have an overall envelope, as shown in Fig.10.5d with the same
shape as the spectrum of the single impulse, but it will consist of harmonics of the pulse
repetition frequency rather than a continuous spectrum. A rolling element bearing
produces this type of signal with a definite defect in one of the races. The impulses can be
very narrow, and they will always produce an extensive series of harmonics.
12
10.4.1 ISO Standards for Evaluation of Vibration Severity
Standards for evaluation of vibration severity are considered one of the most important
activities of ISO/TC 108. A wide variety of published standards describe acceptable
vibration limits, including the ISO/7919 series (5 parts) and ISO/10816 (6 parts). As
detailed in Table 10.2, ISO 7919-1 is the basic document describing the general
requirements for measurement and evaluation of machinery vibration using shaft
measurements. Similarly ISO-10816-1 is the basic document describing the general
requirements for evaluating machinery vibration using casing and /or foundation
measurements. Subsequent parts of each series of documents apply to different classes
and types of machinery, and include specific evaluation criteria used to assess vibration
severity. The evaluation procedures in the ISO/7919 and ISO/10816 series of standards
are limited to broadband measurements, without reference to frequency components or
phase. Measurement procedures are detailed in Part 1:general Guidelines of each series,
including shaft relative, shaft absolute, and pedestal vibration measurement.
13
Table 10.3. Typical Evaluation Criteria Zone Vibration Magnitude (Severity)
Vibration Magnitude is defined within this group of standards as the maximum value of
the broadband RMS velocity in the specified frequency range (typically from 10 to 1000
Hz), as evaluated on the structure at prescribed points. Note that other quantities such as
displacement or acceleration and peak values instead of RMS values are permitted, but
may not easily correlate to criteria based on RMS values. Evaluation zones are defined to
permit a qualitative assessment of the vibration, and to provide guidelines on possible
actions. Limited evaluation criteria are provided in an informative annex of Part 1 of each
standard as shown in Table 10.3.
14
Fig.10.6 Machinery Faults
10.5.1 Mass
Unbalance:
Force Unbalance Will be in-phase and steady. 1x always present and normally
dominates. Can be corrected by one balance weight in one plane
Overhung Rotor Causes high 1x vibration in both axial and radial. Often have both
Unbalance force and couple unbalance.
Eccentric Rotor The largest vibration occurs at 1x RPM of eccentric part. Comparative
Horizontal and vertical phase readings differ either by 0° or 180°.
Bent Rotor Causes high axial vibration with axial phase difference
tending toward 180°.The dominant vibration is normally at 1x if bent
near the shaft center , but at 2x if bent near coupling.
15
10.5.2
Misalignment:
Angular High axial vibration, 180° out-of-phase across the coupling, with both
Misalignment 1x and 2x. Sometimes 3x also.
Parallel Similar to angular, shows high radial, 180° out of phase across the
Misalignment coupling.2x larger than 1x.When severe higher harmonics present.
Misaligned Cocked bearing will generate considerable axial vibration. Will cause
Bearing Cocked twisting motion with180°phase shift top to bottom and /or side to side
On Shaft in the axial direction.
10.5.3
Mechanical
Looseness-3
Types
Type C Generated by improper fit between component parts which will cause
many harmonics. Often caused by a bearing liner loose in its cap,
excessive clearance or a loose impeller. Looseness will often cause sub
harmonic multiples.
10.5.4 Journal
Bearings:
Wear/Clearance Later stages of bearing wear are normally evidenced by the presence of
Problems whole series of running speed harmonics.
Oil Whirl/Whip Occurs at (0.42-0.48) x RPM and is often quite severe. Considered
Instability excessive when amplitude exceeds 50% of bearing clearances.
Changes in oil viscosity, lube pressure and external preloads can affect
oil whirl. Oil whip may occur if machine is operated at or above 2x
rotor critical frequency speed. More stable bearings are available for
high speeds to overcome this instability problem.
16
10.5.5
Rolling Element
Bearings
(4 Failure Phases):
Stage 1
Earliest indications of bearing problems appear in ultrasonic frequencies,
20kHz-60 kHz. These are frequencies evaluated by spike energy (gSE).
Stage 2 Slight bearing defects begin to ring bearing component natural frequencies (f
n) which occur in the 30k-120k CPM range. Sideband frequencies appear
above and below natural frequency peak at end of stage 2.
Towards the end, the amplitude of 1xRPM is even affected. It grows and
Stage 4 normally causes growth of many running speed harmonics. Natural frequency
peaks are replaced by random broadband high frequency noise floor.
Normal Gear
Spectrum Shows 1x and 2x RPM along with Gear Mesh Frequency (GMF). GMF
commonly will have running speed sidebands. Peaks are of low amplitude
and no natural frequencies are excited.
Gear Tooth
Wear A key indicator is excitation of the Gear natural frequency, along with
sidebands.
Tooth Load
GMF is often very sensitive to load. Analysis should be performed with the
system at maximum operating load.
Gear Eccentricity
and Backlash Fairly high amplitude sidebands around GMF often suggest gear eccentricity,
backlash or non-parallel shafts which allow the rotation of one gear to
modulate the running speed of the other.
Gear Misalignment
Excites second order or higher GMF harmonics and with sidebands at its
Cracked or Broken running speed.
Gear Tooth
Hunting Tooth Will generate a high amplitude 1x RPM which will excite the gear natural
Problems frequency (fn) and with sidebands at its running speed.
17
10.5.7 Belt Drive
Problems:
Worn, Loose or Belt frequencies are lower than the RPM of either the motor or the driven
Mismatched Belts machine. When they are worn, loose or mismatched, they normally cause 3 to
4 multiples of belt frequency. Often 2x belt frequency is the dominant peak.
Can cause high amplitudes if the belt natural frequency is near the motor or
driven machine RPM. Natural frequency can be altered by changing the belt
Belt Resonance tension or the belt length.
10.5.8 Electrical
Problems:
Stator
Eccentricity, Generate high vibration at 2x line frequency (2fL). Soft foot and
Shorted warped bases can produce an eccentric stator. Loose iron is due to
Laminations and stator support weakness or looseness. Shorted stator laminations cause
Loose Iron uneven, localized heating which grows with time.
Eccentric Air Gap Produces a variable air gap between rotor and stator which induces
pulsating vibration, between 2fL and closest running speed harmonic.
Rotor Problems Broken or cracked rotor bars or shorted rings, bad joints or shorted
rotor laminations will produce high 1x running speed vibration with
pole frequency sidebands (f P)
Phasing Problems Due to loose or broken connectors can cause vibration at 2x line
frequency (2fL) which will have sidebands around at 1/3rd fL.
Synchronous Loose stator coils will generate high vibration at Coil Pass Frequency
Motors (CPF) surrounded by 1x RPM sidebands.
DC Motor Broken field windings, bad SCRs and loose connections can be
Problems detected by higher than normal amplitudes as SCR firing frequency
(6fL) and harmonics. Loose or blown fuses and shorted control cards
can have peaks at 1x through 5x line frequency
18
10.5.9 Flow
Related
Problems:
Blade Pass and Blade Pass Frequency (BPF) is inherent in pumps, fans and
Vane Pass compressors. Large amplitude BPF and harmonics are generated if
non-uniform gap exists. In addition resonance can also create high
vibration.
10.5.11 Rotor Produces similar spectra to Mechanical Looseness when rotating parts
Rub contact stationary components. Rub may be partial or throughout the
whole revolution. Usually generates a series of frequencies, often
exciting one or more resonances. Subharmonics are also present.
10.6 Typical Spectra and phase relationships for some common faults
As summarized in the table 10.4 and also given in Fig. 10.6, different machinery faults
has different characteristics. Each fault has characteristic spectrum and has specific
relationship with the phase, the measurement of which is very important many a time to
distinguish similar type of problems.
The following subsections indicate the common faults spectra and their relationships with
phase.
19
10.6.1 Mass Unbalance
Force Unbalance
Typical Spectrum Phase Relationship
Couple Unbalance
Typical Spectrum Phase Relationship
Eccentric Rotor
20
Typical Spectrum Phase Relationship
Bent Shaft
Typical Spectrum Phase Relationship
10.6.2 Misalignment
Angular Misalignment
Typical Spectrum Phase Relationship
21
Parallel Misalignment
22
Type 'B'
Type 'C'
23
Oil Whirl Instability
Typical Spectrum Shaft Diagram
24
little before being stopped by the mounting bolts. When one analyzes the vibration
spectrum of a machine in the context of linearity and non-linearity, one may arrive at a
better understanding of why spectra look as they do and how the appearance of a
spectrum relates to machine health. We can make a very general statement that as
machines deteriorate and develop faults they become less linear in their responses. We
can also say that many machine faults create non-linearity. Therefore, also in general
terms, we can expect the spectra from a healthy machine to be relatively simple
compared to the spectra from a machine with faults, for example mechanical looseness.
Looseness, Foundation Cracks and broken Mounting Bolts can cause non-linearity in
machines that may result in harmonics in a spectrum. Sidebands in a spectrum are
another result of non-linearity, produced by amplitude modulation. Frequency
modulation is similar to amplitude modulation in that it also results in sidebands. In
amplitude modulation, the amplitude of the impacts go up and down repeatedly, in
frequency modulation, the rate of impacts gets faster and slower repetitively resulting in
the same pattern in the spectrum. Rolling element bearing wear, gear defects and motor-
bar defects will all produce sidebands. Rolling element bearings will also create non-
synchronous tones. There are new peaks that are not exact multiples (harmonics) of the
shaft rate.
10.8.1 Introduction:
It has been seen that new model-based fault diagnosis techniques are being developed
rapidly in order to meet the demand for increasingly intelligent condition monitoring
systems for the maintenance of modern industrial process. Recently, Markert et al.
(1997,2000) proposed a model-based method, which allows the online identification of
malfunctions in rotor systems. They presented the models in which equivalent loads due
to the faults such as rubbing and unbalance as virtual forces and moments acting on the
linear undamaged system model to generate a dynamic behaviour are identical to the
measured one of the damaged system. The identification is then performed by least
squares fitting in the time domain. Edwards et al. (2000) employed a model-based
identification in the frequency domain to identify an unbalance on a test-rig. A more
comprehensive approach to identify different types of faults has been reported by
Bachschmid and Pennacchi (2000).
For diagnosing the state of a machine usually signal based monitoring systems which are
explained in the previous chapter 9, as well as in previous sections of this chapter, are
used as good tools, although they do not fully utilize the information contained within the
vibration data. These approaches to machinery diagnostics are generic rather than
machine specific and the interpretation of the data is based on qualitative rather than
quantitative information. Contrary to signal based monitoring systems, model based
diagnostics systems developed in recent year utilize all information contained in the
continuously recorded vibration signals. These methods work either in the time or in the
frequency domain depending on the malfunction type and the operating state for which
the vibration data are available. Also it can be used together with or alternatively to
25
conventional signal based monitoring systems. In this section the model based method is
applied to identify the faults such as unbalance, crack in the rotor.
Mathematical Description
The occurrence of a fault changes the dynamic behaviour of the system, the extent of the
change depends on the vector β, which describes the fault parameters like type,
magnitude, location etc. of the fault. The fault-induced change in the vibrational
behaviour is represented by the additional loads acting on the undamaged system.
The system matrices remain unchanged and the rotor model stays linear. Only the
equivalent loads induce the change in the dynamic behaviour of the undamaged linear
rotor model. To identify the fault parameters, the difference of the theoretical fault model
and the measured equivalent loads will be minimised by a least squares fitting algorithm.
26
For calculating the fault induced residual vibrations, measured vibration data for both the
undamaged and damaged rotor system have to be available for the same operating and
measurement conditions. For example, differences in the rotor speeds, phase and the
sampling times have to be taken into account. Because directly matching data are usually
not available, some signal processing has to be done to achieve the same conditions.
Different rotor speeds, for instance, are compensated by adjusting the time scale of the
recorded normal vibrations to the time scale of the currently measured vibrations.
Differences in the sampling frequencies can be eliminated by interpolating the recorded
normal vibrations to the sampling times of the current measurement. The phase shifts are
avoided by recording a trigger signal [7,8].
For the model-based fault identification method, each fault has to be represented by a
mathematical model describing the relation between the fault parameters β and the
~
equivalent force ∆ F (t ) . Hence, ∆F(β,t) is a mathematical expression for the time
history of the forces acting on the individual DOF of the model. The fault vector β
contains the information about the type, position and magnitude of the fault. Some
common and important faults are given below briefly. A detail explanation on these
faults has been given in previous sections.
Unbalance
A single unbalance un with the phase angle δn acting on the rotor at position n is replaced
by the equivalent forces
at node n in horizontal and vertical directions. The equivalent forces on all other nodes
are zero. Hence, the fault parameters of the single unbalance are given by,
βunb = [n, un , δn ] T .
Crack
The fault vector β in this case contains the crack depth and location. The basic idea to
model a transverse crack in a shaft is to consider the change of the cracked element’s
stiffness. The changed stiffness of the cracked element is multiplied with the
displacement vector r(t) which yield in the equivalent force,
The transverse surface crack on the shaft element introduces considerable local flexibility
due to strain energy concentration in the vicinity of the crack tip under load. During the
shaft's rotation, the crack opens and closes, (breathing of crack) depending on the rotor
27
deflection [22]. The additional strain energy due to the crack results in a local flexibility
matrix C c , which will be C op and C HC for a fully open crack and half-open, half-closed
conditions, respectively. The total flexibility matrix for the cracked section is given as
[C] = [C0] + [Cc]. Where [C0] is the flexibility coefficients matrix for an element without
crack. The stiffness matrix of the cracked element can be written as [Kc] = [T][C]-1[T]T,
where T is the transformation matrix . The details of these have been given in the
previous section.
Coupling Misalignment
The coupling misalignment can be represented by equivalent loads, which are the forces
and moments acting on the two neighbouring nodes of the misaligned coupling. Similar
to unbalance the equivalent loads are rotating with rotor shaft. These loads are estimated
based on a system of linear equations as explained previously, depends on type and
amount of misalignment. The fault vector for the coupling element N, and deflections ∆X
and ∆Y and angular misalignment θ is given by,
Modal Expansion
For calculating the equivalent loads from the mathematical model of the rotor system
by equation (10.4), measured residual vibrations must be available at all degrees of
freedom (DOF) of the model. Since the vibrations are measured only at a few DOF in
practice, the vibrations at non-measured DOF must be estimated using the measured
vibrations. Therefore, the residual vibrations need to be reconstructed via modal
expansion from the directly measured vibrations such as ∆ ~ rM (t ) , at the measuring
positions. This technique is based on the approximation of the residual vibration by a
linear combination of only a few eigenvectors. Simultaneously, a set of equivalent loads
representing the malfunctions is calculated from the measured vibration signals using the
mathematical model of the undamaged rotor system.
By comparing the equivalent loads reconstructed from the current measurements to the
pre-calculated equivalent loads resulting from fault models, the type, amount and location
of the current fault can be estimated. The identification method is based on least squares
28
fitting algorithms in the time domain. The quality of the fit is used to find the probability
that the identified fault is present.
As explained before, the residual vibrations ∆ ~ rM (t ) are available only for a few DOF
of the model. The number M of the measurement locations is much smaller than the
number N, the DOF of the model. The data of the non-measurable DOF have to be
estimated from the measured signals. The measurable part ∆ ~ rM (t ) of the residual vector
~
is related to the full residual vector ∆ r (t ) by the measurement matrix C,
rM (t ) = C ∆ ~
∆~ r (t ) (10.8)
)
The full residual vector can be approximated by a set of mode shapes rk which are put
together in the reduced modal matrix
) ) ) )
φ = [ r1 , r2 , ........ rk ]. (10.9)
Logically, the number K of mode shapes used may not exceed the number M of
independently measured vibration signals contained in ~
rM (t ) , K ≤ M. The vector of
modal co-ordinates ∆q(t) is estimated by combining the measurement equation (5) with
modal representation
)
∆~r (t ) = φ ∆q(t) (10.10)
of the full residual vector and minimising the equation error by the lest squares method.
Eventually, the full residual vector at all DOF is estimated by
) ) ) )
∆~
r (t ) = { φ [ (C φ ) T (Cφ ) ] −1 [Cφ ]T } ∆ ~
rM (t ) = Q ∆ ~
rM (t ) (10.11)
This method is the reduced version of Minimum Dynamic Residual Expansion (MDRE)
technique. This method avoids the computational cost by sacrificing the accuracy little
bit. The MDRE technique is explained as follows (see for more details in [29]). The
Minimum Dynamic Residual Expansion (MDRE) technique estimates responses at all
DOF of the FE model by formulating a minimisation problem combining modelling and
measurement errors. Measurement errors are taken into account using a quadratic norm,
29
10.8.5 Fault Identification process
Only simple matrix multiplications and additions have to be carried out, for estimating
the equivalent loads from the measured vibration signals. Thus, it is very suitable for on-
line identification of any fault.
For identifying the fault parameters, the equivalents loads from measured vibrations and
those from the mathematical model should be compared. If there were no noise and no
~
errors due to modal expansion, the equivalent loads, ∆ F (t ) from measured data would
match exactly to loads ∆F(t) of a certain mathematical fault model. Since the measured
signals and their processing are always associated with some noise and inevitable errors,
the best fit between the two loads patterns is sought by adjusting the fault parameter fault
model β. The least squares algorithm is used in time domain to achieve the best curve
fitting. The objective function to be minimised for the measured equivalent forces
~
∆ F (t ) and the theoretical ones (or of a certain mathematical model) ∆Fi (βi , t), is given
as
~
∫ ∑ ∆F ( β , t ) − ∆F (t ) dt = Min
2
i i (10.14)
i
The algorithm iterates for the values of fault parameters βi for all suspected faults taken
into account. A small amount of fault is assumed to start with in the element no.1 and the
algorithm is iterated in the program. Then the process is repeated considering the fault in
the next element and so on for all the elements for the same amount of fault. The process
is repeated for different amounts by a small increment and preceded to do the iteration,
till objective function is satisfied. Thus the fault size and location can be identified. If
more faults are there, fault parameters βi for all suspected faults shall be taken into
account. This eventually leads to identification of the fault type, its position and the
extent. The least square technique can be used in the frequency domain as well. Matlab
version 6.0 has been used to solve the algorithm. More details of the algorithm can be
seen in references [4,5].
The quality of fit achieved can be used to estimate the probability of the different
identified faults. Two probability measures based on correlation functions have been
developed and successfully tested in [8]. The first probability measure p1, called
coherence, is the normalised correlation of the identified equivalent forces ∆Fi (βi , t) of a
~
particular fault with the measured equivalent forces ∆ F (t ) for lag τ = 0 ,
30
p1 = φ∆Fi , ∆F~ (0) / φ∆F , ∆F i (0) φ∆F~ (t ), ∆F~ (0)
i
(10.15)
~
Due to the normalisation by the auto-correlation functions of ∆Fi (βi , t) and ∆ F (t ) , the
coherence takes values between –1 ≤ p1≤ 1, where p1 = 1 means that ∆Fi (βi , t) matches
~
∆ F (t ) perfectly. The other probability measure p2 , called the intensity, measures the
contribution of the particular fault to the measured total equivalent forces ∆Fident = Σ ∆Fi
φ∆F , ∆F (0)
p2 = i i
. (10.16)
φ∆F
ident , ∆Fident
( 0)
The intensity measures takes values in the interval 0 ≤ p2 ≤ 1, where p2 = 1 signifies that
the identified fault is the only one present in the rotor system. Both measures p1 and p2
should be used simultaneously to evaluate the probability of a fault. Suitable threshold
values are p1 ≥ 10 % and p2 ≥ 20 % indicating that specific fault is present in the rotor
system [8]. The whole process of identification is shown in the Fig. 10.12. The measured
vibration signals ~ rM (t ) are the input and the fault parameters βi for each fault are the
output.
Actual measurement ~
rM (t )
Normal vibrations
Residual vibrations ∆ ~
rM (t )
ro (t ) (undamaged)
~ (t ) ,∆ r~& ( t ) , ∆ ~
∫ ..dt , d (..) / dt ⇒ ∆ &r& M M
rM (t )
~ System matrices
Equivalent forces ∆ F (t )
Mo, Do , Ko
31
Identified Faults βi
Probability Measures p1,2
Example 10.1: Identification of cracks: Let us consider a rotor system with two
flexible bearings and two rigid disks ( see table .. for data) as shown in Fig. 10.13. The
analysis has been carried out using FEM for flexural vibrations. Consider one crack only
at the mid of the rotor and in the center of the element 7 (see Fig. 10.13).
For typical result, the same rotor system (Fig. 10.13) passing the critical speed has been
considered in this analysis. The first two critical speeds of the rotor are 100 and 168 rad/s.
The acceleration “a” is chosen as 50 rad/s2. A crack at the mid of the rotor (element 7) is
considered for the study. Houbolt time marching technique is used to model the system
33
in time domain and the time response has been modelled until the system passes the
critical speed. Using model based method algorithm with the residual vibrations
estimated the equivalent loads as shown in Fig. 10.15a. From this figure, it is clear that
the equivalent forces are observed only at node 7 and 8, which are the nodes of the
cracked element (no.7). Hence, this method identified the crack at the exact location.
Further, the symptoms of the present fault are found using the CWT of the time response
of equivalent force. From the Fig. 10.15b (the CWT of the estimated equivalent force) the
1/2 and 1/3 criticals (at 1 & 0.65s with a= 50 rad/s2), which are the characteristics of
crack can be observed clearly. Thus a model-based method together with the wavelet can
be used effectively to identify and monitor the crack in a rotor passing the critical speed.
If the experimental results are available one can easily apply this hybrid approach to
identify the crack. For the better estimation of crack depth, one can refine the modal
expansion, considering some weitage functions or some suitable curve fitting functions
etc., together with the mode shapes.
(s)
Fig. 10.15 Equivalent Forces (Crack) , a = 50 rad/s2 (a) Time (b) Wavelet
Normally, more than one fault can occur in a rotor. In the present study the identification
process is also simulated when unbalance and crack are acting simultaneously on the
rotor [45]. The Fig. 10.16 shows clearly the unbalance in disks 1 &2 at 4th and 9th nodes
along with crack in the 7th element in the same rotor (see Fig. 10.13 for rotor model).
Further, the symptoms of the present faults are found from the Fig. 10.16b using the FFTs
of time responses measured at the disk1 and at crack node. The sinusoidal signal of time
response of equivalent forces at disk1 as well as the FFT of which, clearly indicate the
unbalance (1X peak corresponds to 73.7 Hz, the running speed). It can also be observed
the 2X, 3X harmonic components clearly from the FFT of the estimated equivalent force
measured at crack node. These harmonics are the characteristics of crack. Thus the
combined approach could identify more than one fault also.
34
Estimated equivalent loads
Measurement at Disk1
A rotor system as shown in Fig. 10.17, has been considered for the identification of
coupling misalignment. The coupling (element no.7) has been modelled and the
misaligned forces and moments have been estimated and are applied at the coupling
nodes (7&8) using the modelling techniques given in the chapter 9. The excitation has
been assumed in two harmonics. The simulation is carried out using the model-based
method with the reduced basis dynamic expansion (RBDE) technique for D.O.F less than
56.
35
Fig. 10.17 Typical rotor-coupling-bearing system considered
The results are shown in the Table 10.6. The vibrations are considered at all the 56 DOF
(treated as a reference case), in the beginning. Misalignment values of 900 µm are
considered. The identification algorithm was searched for misalignment [46].
The vibration data at all the DOF of the rotor are estimated using the expansion
techniques. The results show that with decrease in measured vibration data (less DOF)
the error in estimating the misalignment values has increased. This is expected as the full
vibration data of unmeasured locations using the mode shape of undamaged rotor system
cannot be accurately estimated from fewer data points. However, the location has been
identified successfully. The results of estimated equivalent loads are shown in Fig. 10.18
and the corresponding FFT is shown in the Fig. 10.19. The coupling misalignment has
been identified successfully, since the equivalent loads have been observed
domineeringly in the Fig. 10.18 at the 7 & 8 nodes, which are the nodes of the coupling
element. The FFT indicates signal nature (1X and 2X components of running speed, 60
Hz) of coupling misalignment fault. Thus a model-based method together with the FFT
can be used effectively to identify and monitor the coupling misalignment in a rotor
system.
Table 10.6 Coupling misalignment identification (element 7)
36
Fig. 10.18. Estimated Equivalent Fig. 10.19 FFT of Estimated
Forces at Coupling nodes Equivalent force
The combined technique is suggested in this paper also to identify any fault in rotor
systems. This will be done in two parts. In the Ist part Digital Signal Processing (DSP)
techniques will be used. The results of the Ist part will be used to perturb the numerical
modal of the 2nd part (A Model-based) where a numerical relation between the fault
damage and specific fault parameter will be established. The detection is performed by
relating algorithmically a model of the system under investigation to the measurements.
A suitable analysis of the differences between model and measurements yields the
determination of the location and size of the fault.
The difficulties of the model-based procedures to distinguish between specific faults can
be overcome by the combined approach starting with better signal processing techniques.
However, in the following section on numerical results, in the I stage the faults are
identified using model-based technique and in the II stage the nature or type of fault is
ascertained by using FFT or wavelet transform.
37
crack1 2
Figure 10.21 Estimated Equivalent loads, for Cracks of depth 4mm in Elements 4 & 7
(a) FFT (b) Equivalent loads
38
10.10 Condition Monitoring: Neural Networks, GA and others
As a result of the varying challenges offered by different structures and systems,
significant research effort has been applied to condition monitoring with the emergence
of broad range of techniques, algorithms and methods. Rytter (1993) classified the
various methods based on the level of identification:
The emergence of shape memory alloys in the context of smart structures, Park et al.
(2001), suggests a fifth level in condition monitoring which would include so called “self
healing structures” (Carden and Fanning, 2004).
Artificial Neural network (ANN’s) are learning algorithms, in the form of computer
programs or hardware, constructed by combining processing elements. Architecture and
a method of training characterized ANN’s (Fig. 10.22). Network architecture refers to
the way processing elements are connected and the direction of the signal exchanged.
A processing elements or unit is a node where input signal converge and are transformed
to outputs via transfer or activation functions. The values of outputs are generally
multiplied by weight before reaching another node.
39
A helpful concept here is the error surface. Each of the N weights and thresholds
of the network (i.e the free parameter of the model) is taken to be a dimension in
space. The N+1 dimension is the network error. For any possible configuration of
weights the error can be plotted in the N+1 dimension, forming an error surface.
The objective of network training is to find the lowest point in this many-
dimensional surface. The location of the minimum on the error surface represents
the culmination of training with the network said to have converged. At this point
the network with its stored value of the weights and biased is said to have
“learnt”. There are many training algorithms for feedforward networks but the most
popular ones are those based on gradient descent methods.
"Backdrop" is short for "backpropagation of error". The term back propagation causes
much confusion. Strictly speaking, backpropagation refers to the method for computing
the error gradient for a feedforward network, a straightforward but elegant application of
the chain rule of elementary calculus. By extension, backpropagation or backprop
refers to a training method that uses backpropagation to compute the gradient. By further
extension, a backprop network is a feedforward network trained by backpropagation.
Standard backprop can be used for incremental (on-line) training (in which the weights
are updated after processing each case) but it does not converge to a stationary point of
the error surface. To obtain convergence, the learning rate must be slowly reduced. This
methodology is called "stochastic approximation. Training a network with
backpropagation involves three stages (Yeganarayana, 1999):
net=newff([-1 2; 0 5],[3,1],{'tansig','purelin'},'traingd');
This command creates the network object and also initializes the weights and biases of
the network; therefore the network is ready for training. There are times when you may
wish to re-initialize the weights, or to perform a custominitialization. Before training a
feedforward network, the weights and biases must be initialized. The initial weights and
biases are created with the command nit. This function takes a network object as input
and returns a network object with all weights and biases initialized. The specific
technique which is used to initialize a given network will depend on how the network
parameters net.initFcn and net layers{i}.initFcn are set.
An adaptive learning rate requires some changes in the training procedure used by
trained. First, the initial network output and error are calculated. At each epoch new
weights and biases are calculated using the current learning rate.
Neural networks have been applied successfully in many diverse applications including
vibration based damage identification (Feng and Bahng, 1999; Mangal et al., 1996;
Waszczyszyn and Ziemianski, 2001; Zubaybi et al., 2002). In general, neural networks
are particularly applicable to problems where a significant database of information is
available, but difficult to specify an explicit algorithm.
Ramu and Johnson (1995) and Pandey and Barai (1995) both applied back propagation
neural networks to identify damage. In both cases the network was found to be effective
except that the topology of the network was found to be critically important for
performance.
41
10.10.2 Genetic Algorithm Methods
Genetic algorithms are methods for optimisation of functions based on the random
variation and selection of a population of solutions. They are part of what may be
described as evolutionary algorithms, which have been developed since the 1950s
(Michalewicz and Fogel, 2000). Their advantage over traditional hill-climbing
optimisation algorithms is that they are capable of tackling multi-modal solution
topologies that are typical of damage identification problems. Many authors, for example
Chiang and Lai [26] and Moslem and Nafaspour (2002), describe a two-stage process
where the RFV is used to locate damage initially and then in a second stage a GA is used
to quantify the damage in the identified elements successfully. Ostachowicz et al. (2002)
identified the location and magnitude of an added concentrated mass on a simulated
rectangular plate by using the shifts in the first four natural frequencies. A genetic
algorithm was employed to overcome the problem of multiple peaks in the objective
function.
Farrar and Doebling (1999) suggested that the vibration based damage detection problem
is fundamentally one of statistical pattern recognition. In their opinion to advance the
state of the art in vibration based damage detection the developments of non-model based
pattern recognition methods are needed to supplement the existing model based
techniques.
This concept of non-model based identification has spurred interest in the use of Novelty
detection for condition monitoring. Novelty detection is concerned with the identification
of any deviations in measured data relative to data measured under normal operating
conditions. Features derived from measurements taken from a structure in its undamaged
state will have a distribution with an associated mean and variance. If the structure is
damaged, then there may be a change in the mean, the variance, or both. Statistical
process control provides a framework for monitoring the distribution of the features and
identifying new data that is inconsistent with the past -'outlier analysis'. If all other
variables can be eliminated then a change in the distribution characteristics of the features
will indicate damage. It is important to note that the detection of damage, rather than
location and quantification, is the objective of using statistical pattern recognition.
Worden et al. (2000), Fugate et al. (2000) and Fanning and Carden (2001) all considered
statistical process control approaches to damage detection. To create the features for
monitoring, the authors used autoregressive functions fitted to history response data of an
undamaged state. The mean and variance of the residuals of the autoregressive model
were used to form the statistical process control charts. The same autoregressive model
was then fitted to subsequent time history responses of damaged states. The resulting
residuals were plotted on the control charts and those points lying outside of control
limits were counted as outliers and used to indicate a change in the system. This approach
was found to be effective in identifying damage in each case and, in the case of Fanning
42
and Carden (2001), it was demonstrated clearly that this statistical pattern recognition
approach was clearly more effective than other single/few sensor algorithms.
10.11 Summary
The concept of linear and non-linear behaviour gives us another way to think about a
vibration spectrum and how its appearance relates to machine faults. Healthy machines
should respond more linearly than machines with faults, which is to say, as machines
develop faults they will likely respond less linearly. As they become less linear we begin
to see more and larger harmonics and / or sidebands in our spectra. Because we may not
know all of the details about the design of a machine or how its spectra will appear when
it is healthy, it is still best to trend information over time and also to compare the
standards. This is to say, look for more and larger harmonics and new peaks that are not
there before as an indication that the health of the machine is deteriorating. For predictive
maintenance applications where the goal is machinery health monitoring, it is sufficient
to realize that the problem is complex. One can routinely trend the vibration spectra until
it becomes severe. But for root cause analysis, one must exercise caution and perform a
detailed analysis. For example the changes that occur with shifts of speed and
misalignment clearly show the truth behind misalignment vibration spectra. There is no
strong evidence of a signature typical of misalignment and hence require complex
modeling of nonlinear machinery dynamics. This applies to many more complex faults or
a combination of faults or malfunctions in rotating systems. More research is expected in
this area in the near future.
The model based identification technique with modal expansion/RBDE has been
successfully demonstrated for different faults. The fault-induced change of the rotor
system is taken into account by equivalent loads in the mathematical model. The
equivalent loads are virtual forces and moments acting on the linear undamaged system
to generate a dynamic behaviour identical to the measured one of the damaged system.
The effectiveness of the identification process depends to good extent on the number of
measured locations (DOF). For the better estimation of faults, one need to refine the
modal expansion (shape expansion techniques in general) , considering some weitage
functions or some suitable curve fitting functions etc., together with the mode shapes.
More research scope is there in this area.
The advanced signal processing such as the continuous wavelet transform (CWT) has
been used to extract the characteristic features of faults such as crack and coupling
misalignment successfully from the time response. Thus the nature and symptoms of the
faults are ascertained by using FFT or wavelet. Results presented are only based on
simulated work; if experimental data are available one can use them easily to identify the
faults. For on-line rotor fault identification a combined (hybrid) model and signal-based
approach is suggested.
43