Hidden treasure
Drive data are a treasure trove of hidden information that can
help industries solve problems before they even happen
MICHAL ORKISZ, MACIEJ WNEK, PIEDER JOERG – As processes
become ever more complicated and margins thinner, minimizing downtime by ensuring that industrial machinery
operates correctly is as important as ever. Proper condition
monitoring of critical equipment can act as an early warning
system against impending problems. However, condition
monitoring is not used everywhere, often because of the
expense of installing proper sensors and cabling, especially
if the monitoring system needs to retrofi tted to existing
equipment. Another reason is that the task of selecting and
70 ABB review 1|10
interpreting the large quantities of data available in the most
effective way seems daunting as well as costly. ABB has
devised a way to easily access and process important data
without the burden of additional equipment, costs and
downtime. By extracting and processing data from existing
devices traditionally used in process industries, such as
drives, customers can prevent otherwise unforeseen problems from occurring and hence maximize the availability of
their machines.
ndustries are constantly under pressure to reduce costs while increasing
service and productivity. The most ef-
I
fective way of fulfi lling these aims is for
managers to know the state of their equipment – in particular the critical components – at all times and to use this information to quickly identify and rectify faults
before they spread to other parts of the
process
system helps predict the reliability of
equipment and the risk of failure. With so
much to gain, why is it that condition
monitoring is not used everywhere? One
reason is that existing equipment is often
already retrofi tted with a monitoring system and the installation of additional sensors and cabling could prove both complicated and expensive. Another reason
concerns the interpretation of results. In
many cases it may not be clear how to
use a set of data that gives information
about one aspect of a process to provide
information about another. For example,
determining the fractal dimension of a
certain phenomenon may be fairly straightforward but relating it to the condition of a
machine may not be so obvious.
Most processes use devices that are capable of collecting and producing relevant signals, which, if harvested and processed correctly, can also be used for
diagnostic purposes. Among others, one
such example is ABB’s family of ACS
variable-speed drives, which are often
[1]. A good condition monitoring
used to power critical equipment. The
drives are based on powerful controllers
that consume and provide tens, if not
hundreds, of signals with sub-millisecond
resolution.
To be useful for condition monitoring,
data needs to be obtained from the drive
inverter in one form or another. Internally
the signals – which include measured
and computed values such as speed,
frequency, torque, flux, current, power
and temperature, as well as parameters
such as configurable drive settings – are
stored in a regularly updated memory
table. Data can be retrieved from this table as OPC
ed into hardware data loggers.
Data loggers are programmable buffers
capable of storing values from several
selected variables concurrently with a
specified sampling rate, generally one
that is high enough to make the data
useful for spectral analysis. In normal operation, the newest data overwrites the
oldest until the loggers are triggered by
certain events, such as the occurrence
of a fault or an alarm, a selected variable
signal crossing a specified threshold or a
software command. As the buffers are
circular, some data
prior to and after
the trigger can be
retained. ABB’s
DriveMonitor
➔ 1 can read
tem
the contents of a
drive’s hardware
data logger. It consists of a hardware
module in the form
of an industrial PC and a software layer
that automatically collects and analyzes
drive signals and parameters
Data enhancement
Because the resolution has already been
determined and preprocessing has been
performed, drive signals are generally
available in a form not easily applicable
to diagnostic evaluation. It is therefore
necessary to employ a suite of “tricks” to
transform the data so that it becomes
useful for diagnostics.
True to their name, variable-speed drives
dynamically change the frequency of the
current supplied to the motor. The direct
torque control (DTC) method employed
in the drive produces a non-deterministic
1
values or they can be load-
Most processes use devices
TM
sys-
that are capable of collecting
and producing relevant signals which can be used for
diagnostic purposes.
[2].
1 ABB‘s DriveMonitor
switching pattern, so there is no such
thing as a constant switching frequency.
This makes the straightforward application of spectral analysis methods somewhat challenging. Because individual
spectra contain many hard-to-predict
components collected one after another,
the averaging of many spectra using
point-by-point averaging, for example, is
essential to obtain a “clean” spectrum.
In general, signals currently available
from the ACS drive are used primarily for
control purposes. Therefore some of the
preprocessing needed for condition
monitoring signals is missing. One such
process is anti-aliasing filtering. Data
points are sampled or computed at rates
up to 40 kHz, but can only be accessed
at lower rates (eg, by keeping every 40th
data point). In signal processing it is typical that frequencies above the so-called
Nyquist frequency – defined as half the
sampling rate – should be filtered out
prior to signal sampling. Skipping this
step means the peaks from the higher
frequencies will appear in the lower part
of the spectrum, making it very hard to
interpret. For example, signals containing frequencies of 400 Hz, 600 Hz,
Footnote
1 OPC stands for object linking and embedding
(OLE) for process control and represents an
industry standard that specifies the communication of real-time data between devices from
different manufacturers.
TM
71Hidden treasure
1.4 kHz and 1.6 kHz that are sampled at
1 kHz all produce the same aliased spectrum with a peak at 400 Hz.
When it comes to monitoring drive-induced changes in the output frequency,
the high frequencies are important. Because they were not filtered out by the
anti-aliasing filter combined with the fact
that the drive’s output frequency is rarely
constant means they can be recovered.
This recovery process is illustrated in
The individual true spectrum containing
the original and aliased peaks, as computed from the measured data, is shown
➔ 2a. The x-axis is scaled so that the
in
output frequency is 1. This spectrum is
“unfolded” by appending copies of itself
(alternating between reversed and
straight) along multiples of the Nyquist
frequency. A number of unfolded spectra
for varying output frequencies are then
averaged so that previously aliased
peaks are returned to their original
➔ 2b.
place
Variable-speed drives are generally used
in applications where a process parameter needs to be controlled. The drive
changes the output frequency in response to an external request (eg, to
pump more water) or because of process
changes (eg, more load on a conveyor
belt increases the slip of an asynchronous motor) or perhaps because of a
combination of both. While traditional
spectral analysis methods assume constant frequency, frequency variations can
be handled using one of two approaches: selecting constant frequency moments or rescaling the time axis.
The first approach takes advantage of
the fact that data is available in large
quantities at any time. Most of it can actually be ignored in favor of keeping only
a few “good” data sets. The trick, however, is knowing what to keep and what
to throw away. A good criterion for selecting a suitable data set is that the output frequency should not change appreciably during the measurement, and only
a set of conditions that occur regularly in
the process should be considered for selection.
Sometimes the operating-point variations are so frequent that it is impossible
to find such a stretch of data for any
length of time. In such cases, the solu-
➔ 2.
tion is to convert
the data domain
from time to another quantity, such
as the electric field
2
angle.
To aid in
this transformation, various measurements can be
collected from the
drive inverter in
parallel with the
original signal. The
instantaneous value of the output
frequency
such measurement. This frequency is then integrated to yield
the angle of the
stator electric field,
which then replaces the original xvalue of each data
point. Further normalization can be
applied to the yvalues.
This transformation results in an x-axis
that is no longer equispaced and therefore the fast fourier transform (FFT) spectral approach cannot be used. Instead,
the Lomb periodogram method is employed [3]. This process, as applied to
one of the phase currents of a hoist machine, is illustrated in
nal with pronounced frequency and amplitude variability is shown in
RMS current value reported by the inverter is given in
stantaneous frequency is plotted in
The stator electric fi eld angle is shown
➔ 3d and its shape follows the trend
in
that the higher the frequency, the faster
the rate the angle increases. The regular
sinusoid shown by the solid mustard-colored waveform line in
the original current signal is normalized
(using point-by-point averaging) by the
RMS current value and its x-axis respaced
to refl ect the angle. This in turn leads to a
spectrum that is represented by a singlefrequency peak (solid line in
the raw data spectrum, shown by the
dotted line, is not represented by a singlefrequency peak.
Different transformations can be applied
depending on the information required.
3
is one
➔ 3b and the measured in-
2 An individual electric-torque spectrum
1.5
1.0
0.5
Torque (kNm)
0
0 5 10 15 20 25 30 35
2a With aliased peaks
1.5
1.0
0.5
Torque (kNm)
0
0 5 10 15 20 25 30 35
2b With an averaged “unfolded” spectrum
➔ 3. The original sig-
➔ 3a. The
➔ 3c.
➔ 3e results when
➔ 3f), while
Frequency (orders)
Frequency (orders)
The frequency
variations associated with variable-speed drives
can be handled
by either selecting constant frequency moments
or rescaling the
time axis.
72 ABB review 1|10
3 Normalization and transformation of variable frequency (and amplitude) current
2,000
1,000
¶
Current (A)
-1,000
-2,000
0 2 4 6 8 10 0 2 4 6 8 10 0 2 4 6 8 10
3a Original signal
15
10
5
Stator electric field angle (revs)
0
0 2 4 6 8 10
3d Integrated frequency (angle)
Time (s) Time (s) Time (s)
Time (s)
1,200
1,000
800
RMS current (A)
600
400
3b RMS current
2
1
0
-1
Normalized current
-2
0 2 4 6 8 10 12 14
3e Transformed signal
Angle (revs)
3.0
2.5
2.0
1.5
Frequency (Hz)
1.0
0.5
3c Instantaneous frequency
1.0
0.8
0.6
0.4
Normalized current
0.2
0
0 0.5 1.0 1.5 2.0 2.5 3.0
Angular frequency (orders)
3f Spectrum (raw signal is dotted;
transformed is solid)
For example, suppose engineers want to
know if certain motor defects such as imbalance, misalignment and bearing faults
are present. Rather than measuring the
instantaneous value of the output frequency, a motor speed signal may be acquired. After an analogous transformation, the x-axis represents the shaft angle,
which in turn facilities the search for motor defects related to the rotating speed.
Diagnostic opportunities
Converted drive data can be analyzed
using two general methodologies that reveal different and important diagnostic
information. These methodologies are:
– Point-to-point variability within one
signal
– Signal-to-signal correlations
Point-to-point variability can be analyzed
via spectral analysis in which periodic
components are represented as peaks in
the spectrum while various system defects or conditions can manifest themselves as spectral features with different
frequencies. Signal-to-signal correlations, on the other hand, give information
about the operating point and any associated anomalies.
Other methods use acquired knowledge
about the normal behavior of a machine
or process, and any observed deviations
are immediately indicated. Irrespective of
which method is
used, their under-
4 A fragment of the torque-signal spectrum from a rolling mill.
On the horizontal axis, one equals the output frequency.
lying purpose is
more or less the
same – to produce
key performance
indicators (KPIs)
that give adequate
information about,
for example, the
health of a machine, process ro-
0.4
0.3
0.2
Torque (kNm)
0.1
0
0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6
F
Rot
bustness or supply
quality. The conclusions can also
be helpful in uncovering the root cause
of a problem once it has been identified.
Spectral analysis
Drives equipped with an active rectifi er
unit can use the spectra of supply voltages and currents to yield valuable information about the quality of the power
supply. Phase currents and voltages that
are measured concurrently enable engineers to check for possible unbalances,
phase shifts, harmonic distortions, etc.
Similarly, looking at the harmonic content
of the output current is a means of verifying the quality of the motor’s power supply. The drive provides information relevant to the motor (such as frequency,
torque, power, RMS current and fl ux) and
to the inverter operation (such as internal
DC voltage levels, speed error and switch-
“X”
2·F
Rot
2·“X”
Frequency (orders)
ing frequency). In fact the spectral analysis of data supplied by a drive is capable
of revealing more than is uncovered by
the “classical” analysis of electrical or vibration signals.
An example of an averaged torque spectrum from a rolling mill is shown in
➔ 4.
The horizontal axis is scaled so that the
output frequency equals 1. There are two
peaks related to the rotating frequency,
. In addition, a family of peaks exists
F
Rot
at an interharmonic frequency of “X” =
0.7742 (37.86 Hz) and 2“X” (1.5484), and
Footnotes
2 These domains are equivalent when the
frequency is constant.
3 The frequency the drive establishes on the
output current. The drive controls this frequency
so it knows its exact value.
73Hidden treasure
this likely corresponds to a resonance
frequency in the driven equipment. This
is an interesting piece of diagnostic information since such resonances accelerate equipment wear, which in turn could
negatively impact certain process quality
issues, such as the uniformity of rolled
metal thickness.
Transient phenomena
Spectral analysis also helps to reveal the
presence of transient phenomena in drive
data. As well as stationary oscillatory
components in the signals, other more
temporary events may also be present
that are indicators of potential problems.
For example, the raw torque signal from a
rolling mill, measured over the course of
4s is shown in
➔ 5a. Some form of ring-
ing, which lasts roughly half a second, is
evident after approximately 3s. The spectrum of this ringing fragment is given
in
➔ 5b where a 10 Hz frequency compo-
nent and its harmonics are obvious. The
source of this oscillation is unknown but
the spectrum has highlighted a potential
problem that needs to be investigated.
5 Transient phenomena in a torque signal.
35
Torque (kNm)
Time (days)
2.0
1.5
1.0
0.5
0
0 10 20 30 40 50 60 70
Frequency (Hz)
30
25
Torque (kNm)
20
0 1 2 3 4
5a The raw waveform with ringing 5b Spectrum of the ringing fragment
6 Time evolution of the torque/speed ratio (τ/n2) for a fan
Time (s)
0.87
0.85
0.83
0.81
Arb. units
0.79
0.77
0.75
0 1 2 3 4 5 6 7
The spectral analysis of data from
a drive is capable
of revealing more
than is uncovered
by the “classical”
analysis of electrical or vibration
signals.
While it is impractical to continuously collect high-frequency data, the periodic collection and examination of such signals
signifi cantly improves the chance of detecting unwanted temporary occurrences.
Operating-point tracking
Concurrently tracking operating-point
quantities (such as current, torque,
speed, power and frequency) in drive
data is an example of the signal-to-signal
correlation methodology mentioned previously. Analyzing the relationships between certain quantities can shed light
on both the operation of the machine
and the state of the process. The rela-
tionship between torque and speed,
governed by the fan laws, is a good example of a process-dependent relationship.
The velocity pressure difference at the
Δ
output
density
p is proportional to the gas
ρ
and the square of the output
velocity V:
Δ
p = ρ⋅V2/2
Power P is equal to the pressure difference times the volumetric flow rate Q:
Δ
p⋅Q
P =
but it can also be expressed as a prod-
τ
uct of torque
τ
⋅n
P =
and rotating speed n:
In normal operation under constant
geometry, both Q and V are proportional to n, thus:
τ
= C⋅ρ⋅n2
where the constant C depends on the
fan’s geometry.
τ
It follows that the ratio
/n2 refl ects the
density of the gas and the fan’s geometry,
which rarely changes.
In ➔ 6 this ratio for a drive-powered fan
over a period of several days is plotted. The
oscillations (with a period of one day) refl ect
the daily variations in temperature and thus
the density of the pumped air. High density
(cold temperature) occurs at night while
low density (warmer temperature) is evident
during the day. The drive data alone enables the evolution of process variables,
such as inlet temperature, to be tracked. In
addition, comparing this data with values
from the control system (temperatures in
this case) can lead to the detection of any
unexpected discrepancies.
Tracking the operating point is possible
without having to employ any additional
hardware – the data is already available
in the drive. The analyzed data can be
presented directly or further analyzed by
using the principal component analysis
(PCA) technique described below.
Cyclic process analysis
Some processes powered by a variablespeed drive are cyclic in nature. A rolling
mill application is one such example
where torque and current abruptly jump
or increase as a slab is loaded onto the
rolls and then suddenly decrease as the
74 ABB review 1|10
7 A typical rolling mill torque profile
40
30
20
Torque (kNm)
10
0
-1.0 -0.5 0 0.5
Time (s)
7a Examples of torque up and down profiles 7b The two clusters represent torque increases
50
40
30
20
10
0
-10
principal component
-20
nd
2
-30
-40
-40 -20 0 20 40 60
1st principal component
and decreases
ABB’s medium-voltage AC drive
ACS 1000
Drives are but one
example of useful
diagnostic data
providers. Other
examples include
motor control
centers, protection
relays and intelligent fuses.
slab leaves. These jumps can be analyzed to detect any process instabilities
or divergence from normal behavior that
may be an indication of equipment wear
or material variations.
In order to extract only the most essential information, high-resolution data
gathered around torque jumps is processed using the PCA methodology
[4].
This technique reduces multidimensional
data sets to lower dimensions for analysis. These lower dimensions condense
the set-to-set variability. Typical rolling
mill torque profiles are shown in
Each profile in
➔ 7a, corresponding to
➔ 7.
one jump, is reduced to a single point as
shown in
➔ 7b. Jumps – or points – that
tend to cluster within certain boundaries
generally indicate the process is operating normally while those outside could
signify a problem. The full data set can
be saved for further examination at a
later stage or, if the analysis takes place
in real-time, more data can be collected.
Healthy machines, healthy processes
In today’s competitive world, unplanned
downtime can be disastrous for a company. That is why industries are constantly striving to maximize the availability of their machines. To do this
effectively, some form of condition monitoring needs to be in place so that maintenance can be scheduled or actions
taken to avoid the consequences of failure before it occurs. Condition monitoring is increasing in importance as engineering processes become more
automated and manpower is reduced.
The benefits of condition monitoring
need not come at the expense of having
to install additional equipment. Often the
data provided by devices for one purpose in a process can be used to satisfy
another at no extra cost. As an important
part of an industrial process, ABB drives
have access to and generate large quantities of data, which, when properly processed, can be used for condition monitoring and diagnostics. Drives are but
one example of useful diagnostic data
providers. Other examples include motor
control centers, protection relays and intelligent fuses. As well as being data providers, these devices are capable of using their onboard computational power
for analyses.
Michal Orkisz
ABB Corporate Research
Krakow, Poland
michal.orkisz@pl.abb.com
Maciej Wnek
ABB Low Voltage Products
Turgi, Switzerland
maciej.wnek@ch.abb.com
Pieder Joerg
ABB Discrete Automation and Motion
Turgi, Switzerland
pieder.joerg@ch.abb.com
References
[1] Mitchell, J. S. (2002). Physical Asset
Management Handbook (185). Clarion
Technical Publishers, United States.
[2] Wnek, M., Nowak, J., Orkisz, M., Budyn, M.,
Legnani, S. (2006). Efficient use of process and
diagnostic data for the lifecycle management.
Proceedings of Euromaintenance and 3rd World
Congress on Maintenance (73– 78). Basel,
Switzerland.
[3] Press, W.H., Flannery, B.P., Teukolsky, S.A.,
Vetterling, W.T. (1986). Numerical Recipes:
The Art of Scientific Computing. Cambridge
University Press.
[4] Jolliffe, I.T. (2002). Principal Component
Analysis. Springer.
75Hidden treasure
Loading...