## Abstrict There is described herein a method of obtaining a measure of flow
from an electromagnetic flow meter having at least one field generating
coil and potential sensing electrodes. A composite excitation waveform
is applied to the field-generating coil. The composite excitation
waveform comprises at least two frequency components and a plurality
of frequency components are simultaneously present for at least
a portion of the waveform. The potential sensed by the potential
sensing electrodes is sampled to produce a sampled output and the
sampled output is correlated with a composite waveform based on
the composite excitation waveform. The result of the correlating
is used to derive a measure of flow.
## Claims What is claimed is:
1. A method of obtaining a measure of flow from an electromagnetic
flow meter having at least one field generating coil and potential
sensing electrodes, the method comprising: applying, to the at least
one field generating coil, a composite excitation waveform comprising
at least two frequency components, wherein a plurality of frequency
components are simultaneously present for at least a portion of
the waveform; sampling a potential sensed by the potential sensing
electrodes to produce a sampled output; correlating the sampled
output with a composite waveform based on the composite excitation
waveform; and deriving a measure of flow based on the result of
the correlating.
2. A method according to claim 1 wherein the waveform comprises
a set of components, each having a pre-specified amplitude and frequency.
3. A method according to claim 1 wherein the composite waveform
is non-sinusoidal.
4. A method according to claim 1 wherein the composite waveform
is digitally synthesised.
5. A method according to claim 1 wherein correlation is performed
over a window which does not contain an integer number of periods
of all frequency components.
6. A method according to claim 5 wherein the window is shorter
than the period of the lowest frequency component.
7. A method according to claim 1 wherein correlating comprises
determining a measure of closeness of fit.
8. A method according to claim 1 further comprising deriving a
factor from said correlating, and deriving a measure of flow from
the factor.
9. A method according to claim 1 further comprising determining
a measure of a trend within a correlation window.
10. A method according to claim 9 further comprising subtracting
the measure of trend from the output sample.
11. A method according to claim 9 further comprising determining
a measure of measurement accuracy from the measure of trend.
12. A method according to claim 10 wherein the processing signal
is the output of an electromagnetic flow meter.
13. A method according to claim 10 wherein the physical quantity
is flow.
14. A method according to claim 10 further comprising determining
a measure of a linear trend.
## Description BACKGROUND OF THE INVENTION
The present invention relates to electromagnetic flow meters. However,
aspects of signal processing techniques disclosed herein may be
more broadly applied. The operating principles of Electromagnetic
Flow Meters are well known, discussed for example in GB-A-2380798.
Where the sensing electrodes are in contact with the fluid, due
to electrochemical or other effects, a DC potential is usually present
across the electrodes even when there is no coil excitation, i.e.
no field. That component of the signal is independent of the flow.
This inhibits the ability to determine the flow in a static fashion.
To overcome this some form of dynamic excitation to the coils is
typically provided in order to generate a dynamic component at the
electrodes that can be differentiated from the background DC (or
slowly varying) bias signal. This dynamic signal is normally pulsed
DC or an AC signal.
The bias signal will not generally be static. It may drift randomly
with time, flow and temperature. In some applications, particularly
for non-homogeneous fluids with inclusions, the signal may contain
large amplitude decaying exponential components, for example in
paper pulp and slurry applications, as discrete charged particles
occasionally touch the electrodes, changing the voltage instantaneously
and this voltage then discharges exponentially to the (drifting)
baseline.
The energy can be spread across a wide frequency range but most
applications have significant low frequency noise and this is often
more problematic due to the nature of the signal processing.
Therefore, one way to obtain flow measurements which are less susceptible
to such effects might be to use a high frequency excitation, above
most of the noise effects. It is found that a frequency above about
1 kHz would for most practical purposes exclude most bias effects.
However, using such a frequency would introduce its own problems.
Most notably, the magnetic circuit of the flow meter is less stable
at the higher frequencies. One factor that contributes to this is
that the losses in the magnetic circuit, which become more significant
at higher frequencies, are quite temperature dependent. Thus, for
a given excitation, the field strength generated may vary unpredictably.
This can be exacerbated because the excitation coils are usually
positioned outside a steel shell of the flow meter. A search coil
can be used to measure the actual field generated but this adds
significant cost and manufacturing difficulties.
The `roll off` of the sensor head at these higher frequencies leads
to an uncertainty in the sensitivity of the sensor and to variations
in the phase between the field (and hence the electrode signal)
and the drive current. It is known to use a moderate frequency (e.g.
70 Hz) sinusoidal excitation and to demodulate the electrode signal
synchronously with the excitation signal. Quite apart from other
considerations, the phase angle at which the demodulation is (should
be) carried out is not constant and requires either manual or automatic
adjustment.
To summarise the problems a lower drive frequency can give much
better sensor stability but measurements are more easily corrupted
by the bias drift and the effect of inclusions in the fluid. A further
important problem is that a low excitation frequency limits the
rate at which new measurements can be updated--it gives a low flow
measurement bandwidth. A higher frequency assists in distinguishing
wanted signals from unwanted and also allows a more acceptable higher
flow measurement bandwidth but the sensor characteristics will not
be so precisely defined and are less stable. The frequency chosen
is therefore normally a compromise for a particular set of circumstances.
These problems have been previously addressed and it has been proposed
in the assignee's earlier filed UK patent application no. 0116168.6
to excite a meter with two frequencies simultaneously or quasi-simultaneously.
Each frequency component is separately extracted and a combined
measurement obtained in such a way as to enhance the better properties
of each measurement. A drawback is that measurement rate is still
limited by the lowest frequency, as explained in more detail below.
SUMMARY OF CERTAIN INVENTIVE EMBODIMENTS
According to a first aspect the invention provides a method of
obtaining a measure of flow from an electromagnetic flow meter comprising
applying a composite excitation waveform comprising a plurality
of frequency components; receiving an output from the meter; and
correlating sampled output data from the meter with a composite
correlation waveform based on the composite excitation waveform
to derive a measure of flow.
In this method, by taking the novel step of correlating with a
composite waveform, rather than the conventional method of correlating
each frequency component with a sine wave, a number of advantages
can be obtained. Although the composite waveform will effectively
contain multiple frequency components, it has been found that it
is not necessary to process each component separately and thus measurement
rate is not limited by the lowest frequency component. Furthermore,
additional information is obtained from the combination of the components
(the shape of the waveform) as well as the components themselves.
Preferably the meter is excited with a composite excitation waveform
obtained by combining at least two frequency components; in this
way the composite waveform can easily be selected to contain frequency
components which excite desired properties, for example a relatively
high and relatively low frequency. The composite waveform will preferably
not be completely sinusoidal. The composite waveform preferably
is substantially devoid of sharp turning points (these are associated
with numerous high frequencies which may behave unpredictably).
The composite waveform is preferably digitally synthesised. The
frequency components are preferably combined at pre-determined,
preferably similar, for example within a range of minimum to maximum
of 1 to 5 or less, preferably 1 to 3 or less, amplitudes. The amplitudes
(and phases) may be chosen to enhance the prominence of turning
points in the waveform.
The composite waveform preferably has a plurality of turning points
within a predetermined measurement window period; it is found that
multiple turning points or "features" may assist in obtaining
accuracy.
More than two components may be used and provision of three or
four components may enhance accuracy, by giving more features to
the waveform. However, it is preferable that the number of components
is limited, for example to less than about 10 preferably less than
5 components and/or that the highest frequency component with significant
amplitude has a frequency no greater than about 1 kHz. Diminishing
returns are achieved with greater than about 5 components, but in
some cases a few more components may be used.
Whilst an arbitrary waveform can theoretically be approximated
by an infinite number of sine and cosine waveforms, such approximations
tend to require a large number of harmonics to converge accurately
and this implies a large number of high frequencies which would
behave unpredictably in a real meter. By limiting the number of
components, most preferably by synthesising the composite waveform
from a discrete set of components, each having a chosen amplitude,
the effects of the individual components and the overall response
becomes more predictable.
Although the higher frequency may be an integral multiple of the
lower frequency, it is not necessary for the frequencies to be harmonically
related. Even when two frequencies are integral multiples of each
other, it is not necessary to include intervening harmonics. Thus
the composite waveform may comprise at least two frequencies which
are not harmonically related to each other. Alternatively, the composite
waveform may comprise two frequencies of which a higher frequency
is a harmonic of a lower frequency but in which at least one intervening
harmonic of the lower frequency is substantially absent. Preferably,
when the higher frequency is an odd harmonic, at least one intervening
odd harmonic may be substantially absent and/or when the higher
frequency is an even harmonic, at least one intervening even harmonic
may be substantially absent.
An advantageous feature is that correlation may be performed over
a window which does not contain an integer number of periods of
all (or any) frequency components. The window may be shorter than
the period of the lowest frequency component; this enables measurement
to be obtained with a higher bandwidth than the lowest frequency.
Correlating preferably comprises determining a measure of closeness
of fit. A factor may be derived from said correlating and a measure
of flow may be derived from the factor. A measure of a trend within
a correlation window may be determined and a measure of measurement
accuracy may be determined from the measure of trend.
With traditional methods of measurement, it is difficult to reduce
sensitivity to trends in the baseline reading, e.g. due to zero
offsets. Pursuant to the invention, it has been appreciated that
in fact, particularly with the techniques disclosed herein (but
other techniques which give multiple measurement may be used), trends
(particularly baseline trends) in the meter data can usefully be
measured and can be used to determine or improve meter accuracy.
According to a further aspect, the invention provides a method
of processing the output of an electromagnetic flow meter comprising
correlating sampled output data with a signal to determine a measure
of flow and a measure of a trend within the data over a measurement
period.
The trend may provide data about the meter or may simply account
for interference during the measurement period.
A measure of a linear trend may be determined. Additionally or
alternatively a measure of an exponential trend (e.g. a decaying
exponential) may be determined. Advantageously an exponential or
linear trend may be approximated by a polynomial. In a preferred
arrangement a low order polynomial trend, preferably second or third
order, is fitted to the data--this can effectively remove linear
or exponential trends without interfering substantially with fitting
of the data. The sampled output data may be fitted with reference
data comprising an expected signal multiplied by an unknown factor
and a trend of unknown gradient (or polynomial coefficients) to
determine values for the factor and for the gradient (or polynomial)
which gives the best fit according to a predetermined fitting method,
preferably a least squares fitting method. Using higher order polynomials
to model trends trends leads to diminishing returns. Preferably
a measure of flow is determined from the value of the factor. Preferably
a measure of confidence or accuracy is determined from the gradient.
The predetermined fitting method may comprise determining a sum
of squares of differences between the reference data and sampled
output data. The reference data may include a constant unknown offset.
A value for the offset may be determined to improve the fit. However,
advantageously, the offset value may be determined but the measure
of accuracy or confidence is based primarily on the value of the
gradient. Calibration parameters for the meter may be stored, the
method further comprising adjusting at least one calibration parameter
for the meter based on a plurality of measurements of trend or accuracy
or confidence.
According to a further aspect, the invention may provide a method
of operating a flow meter comprising storing calibration parameters
for the meter and adjusting at least one stored parameter as successive
flow measurements are derived based on a plurality of measures of
accuracy of the flow measurements.
Preferably at least one stored calibration parameter is adjusted
to reduce a measure of trend or to improve a measure of accuracy
or confidence based on successive measurements.
Calibration parameters may include at least one of phase and amplitude
response. Preferably calibration parameters are stored for multiple
frequency components. At least one calibration parameter for a first
frequency component may be adjusted in response to measurements
whereas at least one corresponding calibration parameter for a second
frequency component may retained at an initial (e.g. factory calibration)
setting. The first frequency component may have a higher frequency
than the second frequency component. Preferably calibration parameters
for the lowest frequency component are all kept constant.
Weighted filtering may be performed on measurements of flow based
on the measure of trend or accuracy or confidence.
BRIEF DESCRIPTION OF THE DRAWINGS
An embodiment of the invention will now be described, by way of
example only, with reference to the accompanying drawings in which:
FIG. 1 shows an example time window with two components for the
case of two frequency components and a good fit;
FIG. 2 shows another example of a good fit with four frequency
components;
FIG. 3 shows an example similar to FIG. 2 but exhibiting a poor
fit;
FIG. 4 illustrates removal of a trend from data;
FIG. 5 shows the expected "pure" signal from the electrodes
of an electromagnetic flow meter of an embodiment;
FIG. 6 is a schematic diagram of apparatus according to an embodiment;
FIG. 7 illustrates error processing using an SSE approach;
FIG. 8 illustrates an enhanced method for determining errors;
FIG. 9 shows a flow signal as determined by an LSE fit algorithm;
FIG. 10 shows flow after filtering;
FIG. 11 shows a flow profile;
FIG. 12 shows the added noise zoomed in so the total window is
about 0.2 seconds long;
FIG. 13 shows derived flow values;
FIG. 14 shows the error estimates for each of several windows;
FIG. 15 shows noise added (over a 6 second window);
DETAILED DESCRIPTION OF CERTAIN INVENTIVE EMBODIMENTS
The method underlying the invention and a background method will
first be explained.
As noted above, single frequency measurement in an electromagnetic
flow meter is a compromise. To assist in understanding the invention,
an earlier technique will first be explained in more detail which
applies two frequencies simultaneously and then resolves the two
components of the electrode signal separately. This gives some of
the benefits of each frequency. Such a system is disclosed in the
assignees earlier filed UK patent application no. 0116168.6 hereby
incorporated by reference.
As a specific example, consider a system using 10 Hz and 70 Hz
excitation simultaneously. At 10 Hz, the sensor characteristics
can be assumed to be stable but the results will be influenced by
noise. At 70 Hz, the signal is easier to recover but since the sensor
characteristics are less well defined, the flow signal is correspondingly
less well defined.
A basic system could use the zero measurement from the 10 Hz `channel`
as the reference zero. The 70 Hz channel could then be used to obtain
a measurement with the benefits of the higher frequency excitation
and the zero would be `anchored` to the 10 Hz value.
Doing the 10 Hz correlation in the presence of the 70 Hz component
is not problematic; the higher frequency is an integral multiple
of the lower frequency (this is a constraint of the method) and
an integer number of cycles of an interfering frequency in the correlation
window does not cause a problem. However, performing the 70 Hz correlation
is problematic. It is necessary to arrange that the correlation
window to contain an integer number of both the 70 Hz and the 10
Hz signals in order to get an accurate result. This means that the
window length for the 70 Hz signal should be at least 0.1 seconds
and hence measurement bandwidth is limited to that of the lower
frequency, 10 Hz. Thus, although some noise may have been reduced
and stability gained, the measurement bandwidth is still limited
by the lower frequency.
The assignee's earlier filed application in fact discloses a more
sophisticated technique with complementary high and low pass filters
but again the measurement bandwidth is limited by the lower frequency.
In conventional multi-frequency analysis, the fact that superposition
applies (or is assumed to apply) is a useful basis for simplifying
calculations; it allows the components to be resolved independently
even if applied simultaneously as if the others were not there (although
this is subject to some limitations).
In the present invention, a composite waveform is applied, and
this can be considered to be applying multiple excitation frequencies
simultaneously (the coil excitation is the sum of a number of discrete
sine waves). An important difference between embodiments of the
invention and prior multi-frequency techniques is that the electrode
signal is analysed for all components simultaneously by correlating
the actual electrode signal with the expected electrode signal,
rather than by detecting each frequency component separately.
To assist in explaining the invention an example window in time
is shown in FIG. 1.
The "curve 1" line shows the expected electrode signal,
the "curve 2 " noisy data is the actual electrode signal.
The magnitude of the correlation is then the magnitude of the (curve
1) curve that gives the best fit to the (curve 2) data. In the figure,
the fit is already optimal.
Note that in this example, there are two frequency components and
both have an integer number of cycles in the window. In this case,
resolving the two frequencies separately would give similar results
to treating the composite waveform as one signal. In this case,
the inventive method may offer an advantage in the processing but
it is noted that the measurement should be similar to that obtained
by treating the components separately.
FIG. 2 shows another example where the inventive method gives a
significant advantage. In this example, there are four frequency
components, two of them containing less than one complete cycle
in the window.
Provided the expected signal is known then one of the great benefits
of this technique is that it does not depend on any or all of the
components having an integer number of cycles within the window.
Conceptually, one may look for the `pattern` (curve 1) in the data
(curve 2). In principle, the more complex curve 1 the more distinct
a pattern being looking for.
Relative to the window above, the four frequency components have
frequencies 1 2 0.5 and 0.25. The four frequencies are 40 Hz (1
cycle in the window) 80 Hz, 20 Hz and 10 Hz. In this example, the
frequencies are harmonically related. In practice, using a digital
signal processor (DSP), it may often be convenient to produce frequencies
which are integer sub-multiples of a common frequency and this will
often mean that the higher frequencies are harmonics of lower frequencies
but it is not necessary for the components to be related.
The ability to fit without requiring a whole number of cycles of
all of the frequencies provides significant advantages.
It is advantageous from a `distinctiveness` point of view to have
several turning points in the waveform within the window, hence
the choice of frequencies and window length.
Ignoring zero offsets and noise for the moment then the expected
signal from the electrodes should be of the form shown in FIG. 5
and the amplitude should be directly proportional to the value of
flow. For example, if FIG. 5 represents the received signal at a
flow of 1 m/s then a flow of 0.5 m/s should give half the amplitude,
a flow of 0 m/s would give a flat line and a flow of -1 m/s would
give the same signal as FIG. 5 but inverted.
Note that this waveform shape is not intended to be fixed or indeed
optimal. As noted above, the example happens to use harmonically
related frequencies. This is not necessary and may not be the most
advantageous arrangement. In any event it should be recognised that
the pattern being matched will change from window to window.
In this case of no zero offsets, the waveform can be fitted using
simple and well known Least Squares; the amplitude of the red curve
is the one that minimises the sum of the square of the errors between
the two sets of the data.
If the expected signal is P.sub.i and the received signal is represented
by Y.sub.i then one should find the factor F such that:
is a minimum. In this case, the only degree of freedom is F and
one may simply choose F to minimise the SSE.
However, working with real world data, it has been recognised that
the signal being fitted is continually corrupted by a varying offset
and the offset may vary within the window being fitted.
Two extra degrees of freedom have been introduced, namely a slope
and an offset. Now the SSE equation that is to be minimised becomes:
##EQU1##
Referring to FIG. 4 curve 1 represents the incoming data with
a clear trend on it. Curve 2 represents the same data after fitting
using the above equation to remove the trend.
This idea will seem deceptively simple on first inspection. Indeed
the technique could be applied to single frequency excitation systems
to achieve a certain amount of robustness to the incoming signal
having a superimposed trend.
However, trying to apply the technique to a multi-frequency system
where the signals are analysed separately would be problematic.
It would only work provided the expected signal does not have a
trend within the window. This means that there could only be whole
number of cycles of any applied frequencies within the window.
This technique works so well because the composite signal is fit
directly and simultaneously. Where it demonstrates its power very
effectively is on the `tails` of exponentials. As mentioned earlier,
the electrode signal could have a trend that is nothing to do with
the flow. Most frequently this would be either a lowish frequency
drift or the tail of an exponential. In the above graph, the magenta
curve is the measured data. It has a negative going trend that would
lead to a poor fit with the red reference curve. However, by doing
the fit with the slope and offset degrees of freedom, a good fit
is obtained (the blue data has the trend removed).
It should be noted that this is very different to just removing
any trend before the fit because even the expected signal has an
inherent trend by virtue of not having a whole number of cycles
of the waveforms in the window and so simply attempting to remove
a trend would corrupt the data.
In the mathematics, the scale factor that gives the best fit (peak
to trough) is the flow signal and the slope and offset can be thought
of as the best trend that can be applied to the signal to make the
peaks and troughs coincide.
The scale factor is F, the slope is M and the offset is C. The
offset C is thrown away, it offers little informational value. However,
M gives surprisingly useful information. Although at first sight
M would be expected to be no longer needed once the fit is achieved,
M in fact gives a very useful indication of the amount of corruption
seen by the system during the measurement window.
The difference between a poor fit (above) and a good fit (the previous
two graphs) is very apparent by eye. However, mathematically, the
quality can be much harder to judge in a reliable but computationally
simple manner. This is because there is a moderate level of random
noise on the data anyway and even a perfect fit has a moderate rms
error. A further aspect of this invention provides a method to determine
the quality of fit. This is done in a way that can be seen mathematically
to enhance the distinction between FIGS. 2 and 3.
We now consider a practical system. It is a sampled data system.
The sensor is being excited with 10 20 40 and 80 Hz (these may
be conveniently generated e.g. as submultiples of a 20 kHz or 20.48
kHz DSP sampling clock frequency with 250 or 256 samples respectively
for the 80 Hz signal and correspondingly more for the lower frequencies).
The acquisition system is acquiring samples continuously but the
processing is done on a window of data at a time.
The window length is one cycle of 40 Hz (25 mS) but the window
is moved along half a cycle at a time so the windows overlap (it
is an advantageous but optional feature that windows overlap, although
it is not necessary to use exactly half a window increment). While
this does not directly increase the bandwidth, it does give a higher
data rate making subsequent filtering easier.
From the result of each window is a flow value F, a slope value
M and an error estimate for the fit of the waveform within that
window. Sometimes, the M and error estimates will suggest that the
value of F is accurate and can be trusted highly. Other times M
and/or the error estimate will suggest that the value of F could
well be corrupt because there was excessive trend in the data or
the fit just was not very good.
This confidence value can be used to `weight` the data for filtering
purposes. Again, this filtering has proven to be very effective
on real world data. Different filtering and weighting algorithms
may be used, the choice may depend on application.
A simple algorithm is to hold the last value if the confidence
is too low. A monitor can be provided to ensure that the algorithm
adapts if the overall reading confidence for all the windows is
low.
In a more advanced algorithm, trend analysis is used to give a
better guess of the value if the confidence in the current reading
is low. In this respect, the filter could behave like a Kalman filter.
Equally, the M value/error estimate can be used to give a SEVAL
like output e.g. clear, blurred, dazzled as proposed by M. P. Henry
et al.
The ability to fit the incoming signal and remove a trend is a
unique feature of this kind of system, which can be applied to other
measurements, but would not have an equivalent with a single frequency
excitation.
The system depends on knowing the amplitude and phase of the incoming
components. This information represents the frequency response of
the sensor. However, it is appreciated that this is not necessarily
constant. An algorithm will now be described that can automatically
adapt to changing sensor characteristics.
An outer loop is responsible for updating the matrix of coefficients
representing the frequency response. By suitable averaging of the
residual errors of each window, any error in one of the coefficients
becomes apparent.
The fit that gives the best possible fit while keeping the amplitude
ratio and phase of the components fixed is used. However, if the
fit could be improved by allowing the matrix coefficients to be
adjusted (preferably by a small amount below a threshold) and the
degree of adjustment is consistent over a number of windows then
the adjustment is made.
Preferably the lower frequency, here 10 Hz coefficients are kept
fixed (but in some cases may be subject to a very low limit of adjustment).
Limits can be placed on the amount by which the other components
can be adjusted. Limits may include a limit on the maximum total
deviation of a given coefficient from a starting value and/or a
maximum adjustment of a given coefficient in a given adjustment
interval and/or a maximum total or weighted combination of adjustment
of coefficients and/or a combination of the preceding limits.
A method of determining measurement error will now be explained,
referring back to FIG. 1. It is clear by eye that there would be
no better fit than the one shown. However, the SSE (Sum of the Squared
Errors) is still significant because of the (relatively) high frequency
noise on the data being fitted. Now referring to FIG. 3 by eye,
it is clear that the fit is not very good. However, the SSE is not
very different from that for FIG. 1. For illustration, the SSE for
FIG. 1 is 0.121 while for FIG. 3 the SSE is 0.269 a little over
a factor of 2. Yet the fit is very much worse. This illustrates
the normal difficulty in measuring accuracy of measurement.
Since it is desired to use the measurement error as an aid to weighted
filtering, it is preferred that there be more `contrast` (by which
is meant a mathematically identifiable distinction) between the
error values for FIG. 1 and FIG. 3. This can be achieved by substantially
ignoring the high frequency residuals and considering primarily
the low frequency residuals. Conceptually therefore the residuals
could be filtered with a low pass filter before performing the SSE
calculations. Alternatively, it would be possible to do an FFT of
the residuals and measure only components below a certain frequency.
Both of these methods may be effective, but performing an FFT at
least is computationally intensive.
A computationally simple but highly effective method for obtaining
much more contrast will be described. This is based on the recognition
that the residual errors in FIG. 3 that are desired to measure are
low frequency and that, as a result, the error spends a long time
on the same side of the curve (ie either positive or negative).
In contrast, the high frequency residuals average to zero over a
very short interval.
Various embodiments of the method divide the window into a plurality
of intervals, preferably at least about 4 preferably no more than
about 20 typically 5-10 here 8 intervals. For each of these intervals,
the residual is summed algebraically. Referring to FIG. 1 it is
clear that this algebraic sum would be close to zero over any of
the intervals whereas for FIG. 3 it would be far from zero (either
large and negative or large and positive) over most of the intervals.
To get a final error figure for the window, the absolute value
of each of the 8 interval errors is summed. This is because the
summation within any interval could be positive or negative and
an algebraic sum would largely cancel out.
There is thus provided a method of determining error between a
signal and a corresponding reference signal, each signal comprising
a first plurality of sampled values in a sampling interval, the
method comprising:
defining a second plurality of sub intervals, each sub interval
containing a respective subset of the first plurality of sampled
values;
determining an algebraic sum for each sub interval of the difference
between signal sample values and reference signal sample values;
determining an absolute sum of the algebraic sums for each sub
interval as an error estimate for the sampling interval.
The calculation may include other filtering or weighting and the
error estimate may be further processed, but one point of interest
is that the step of algebraic summing comprises a calculation or
operation in which differences of opposite sign substantially cancel
and the step of absolute summing comprises an operation or calculation
in which the algebraic sums add constructively. In place of an absolute
sum, a sum of square might be used--this would increase the effect
of a single large error on the end result.
Although the units are different, the error using this technique
is 2.203 and 48.033 for FIGS. 1 and 3 respectively; a much higher
contrast (20:1 instead of 2:1).
Dividing the window into 8 intervals is very appropriate in this
example because of the frequencies present (ie ratios of 1 2 4
and 8). If different frequencies were chosen then these intervals
could be optimised accordingly.
FIG. 7 shows the SSE approach. The graph represents 6 seconds worth
of 25 ms windows. Although there is high contrast between the peaks
and the `baseline`, these peaks only represent the very worst readings.
FIG. 8 shows the enhanced method for determining errors. Here,
the baseline is much lower and much better contrast exists between
good and bad readings.
On these two graphs, every point represents a processed window.
FIG. 9 shows the flow signal as determined by the LSE fit algorithm.
The final stage of the processing now is to use the error measurement
from graph 8 to filter the data in graph 9. Since it is clear that
the high peaks in FIG. 8 represent points with high measurement
errors, one can `filter out` these points with a suitable algorithm.
FIG. 10 shows the flow after filtering.
FIG. 11 shows the flow profile used for the examples in this example.
FIG. 15 shows the noise added. Although the results are partially
simulated, the noise added to the data is from a real flow meter
in a real application.
FIG. 12 shows the added noise zoomed in (the total window is about
0.2 seconds long).
FIG. 13 shows the derived flow values for approximately the same
interval. The intervals in FIG. 12 show the window numbers and correspond
to the x axis values of FIG. 13.
FIG. 14 shows the error estimates for each of the windows, demonstrating
a sensitive indicator of when the flow estimate is not perfect.
Note that the flow reading is sometimes good even though the error
indicator is high. This is logical because the error indicator is
really indicating the uncertainty in the flow reading, at least
in a qualitative way.
Apparatus embodying the method will now be described with reference
to FIG. 6.
The Microprocessor 10 maintains a `time` variable for the purpose
of calculating drive signals and expected signals. In this embodiment,
the microprocessor calculates a signal that is the sum of four predetermined
amplitude (here equal amplitude), distinct frequency sinewaves.
This signal is applied to the drive coil in the sensor 3 via power
amplifier 2. The electrode signals are amplified 4 and converted
into discrete samples by ADC 5. Software modules running on the
microprocessor implements the method described.
The mathematical drive signal is transformed by the frequency response
coefficients in transformation module 9 in order to determine the
expected signal back from the sensor. In other words, knowing the
signal applied to the drive coil and knowing the transfer function
of the sensor, this module calculates the waveform expected back
from the sensor in the absence of perturbations.
Correlation module 6 performs the correlation or LSE fitting. The
results of this fitting are at least a flow signal and a `confidence`
value that can be used to assist in the weighted filter module 7.
As an enhancement, the results from the correlation module 6 can
also be used by an adaptive adjustment module 8. Such a module,
if present, monitors the quality of fit over a period of time and
decides whether the coefficients used by transformation module 9
need adjustment, subject to predetermined constraints.
Embodiments of the invention provide one or more of the following
novel features: Using multiple frequencies simultaneously, fitting
multiple frequencies simultaneously, weighted filtering and adjusting
coefficients dynamically.
Modifications of detail may be made and features disclosed herein
may be provided independently or in other combinations. |