Note: Descriptions are shown in the official language in which they were submitted.
214625~
W O 94/08226 PCT/EP93/0273~
- 1 -
AN APPARATUS FOR FUEL QUALITY MONITORING
The invention relates to an in-line fuel quality monitor to be
used to provide feed forward information on fuel quality for use in
the control (e.g. feed-forward control) of an engine management
system. Such an apparatus is advantageously applied as a small
light-weight instrument in cars in order to advise drivers or
engine of fuel quality.
Information obtained will be physical property data of
hydrocarbon products such as octane number, cetane number, vapour
pressure density and the like of the fuel, and for use in
lO dual-fuelling vehicles, the gasoline/alcohol ratio. As is known to
those skilled in the art, organic compounds have in the infra-red
spectral region (about l to about 300 ~m) a unique spectral
fingerprint.
The potential to obtain correlations between the physical and
15 chemical properties of materials, and their Near Infra Red (NIR)
spectra has already been disclosed. (Vide e.g. EP-A-0,304,232 and
EP-A-2,085,251).
An empirical model can be created by finding the spectral
trend in a large set of data known as a training set.
(N)IR spectroscopy is both rapid and reliable, and could
potentially be applied to make on-line real-time measurements. A
spectrometer can be used to obtain the spectra of a training set of
characterized unleaded gasolines. By the application of complex
multivariate statistical techniques such as Principal Component
25 Regression, Reduced Rank Regression and Partial Least Squares to
develop the model, the Research Octane Number (RON) of a given fuel
may be predicted. These techniques require all of the data points
provided by the spectrometer and predict well allowing for the
variability of the initial RON measurement. The use of (N)IR
30 technology coupled with an empirical model can be therefore used to
predict performance quality of a fuel. The application of these
21462~
techniques to an on-line real-time field instrument is, however, not
trivial. This i~ because the spectrometers use highly precise
optical moving parts and are extremely sensitive to dirty hostile
environments such as found in the petrochemical plant or a
distribution terminal. Instrument manufacturers are striving to
produce more robust spectrometers.
Despite improvements, the spectrometers which are very expensive
are non-ideal for on-line real-time monitoring due to their delicate
nature, labour costs and the harshness of the environment. A method
of simplifying the application of (N)IR techniques as well as the
statistical technique to analyse the data is necessary.
Now, a small, robust, cheap and reliable "non-moving parts"
instrument has been developed, that uses (near) infra-red techniques
(advantageously 0.78-30 ~m wavelength) advantageously coupled with a
neural network to measure physical property data of hydrocarbon
products such as (research) octane number, cetane number, density,
vapour pressure and the like or gasoline/alcohol ratio on-line and
in real time and that, in particular, easily can be applied in cars.
The invention therefore provides an apparatus for on-line
measuring physical property data of hydrocarbon products such as
octane number, cetane number, density, vapoùr pressure and the like
or gasoline/alcohol ratio, characterized by means for providing
(N)IR radiation in a predetermined spectral range; said means being
optically connected to the hydrocarbon product line in such a manner
that an optical path length is present in the hydrocarbon product
line;
means for detecting the light transmitted through the said
optical path;
means for providing the obtained signal to be input to a neural
network for spectral analysis and for correlating the spectral data
to the physical property data of hydrocarbon products such as octane
number, cetane number, density, vapour pressure and the like or
gasoline/alcohol ratio, wherein the number of layers of the neural
network is 2 to 5, and in case of three or more layers the number of
A~llEN~E~) SHEET
2 14 ~255
- 2a -
nodes of the input layer is from 3 to 10, the number of nodes of the
hidden layer(s) is from 1 to 10, and the number of nodes of the
output layer is from 1 to 3.
It is remarked that DE-A-3,716,793 discloses an apparatus for
on-line measuring physical property data of hydrocarbon products
such as gasoline/alcohol ratio.
However, the specific apparatus of the invention applying a
specific neural network has not been disclosed.
MCS7/T567lPCT
A~ENDED SHEEJ
~ 21962~
W O 94/08226 PCT/EP93/0273
- 3
As already indicated in the foregoing, the principle of the
invention is based upon the technique of (near) infra-red ((N)IR)
analysis, advantageously coupled with the technology of neural
networks. Generally, a neural network can be defined as a system,
wherein during a learning period a correlation between input- and
output variables is searched for. After sufficient examples have
been offered in this learning period the neural network is able to
produce the relevant output for an arbitrary input. Neural networks
have found applications e.g. for pattern recognition problems.
As those skilled in the art will appreciate, neural networks
are built up of layers of processing elements (similar to the
brain's neurons) each of which is weighted and connected to
elements in other layers (similar to the brain's synapses). A
network learns patterns by adjusting weights between the elements
whilst it is being trained with accurate qualified data.
According to an advantageous learning algorithm, training
errors, the difference between the actual and predicted result are
propagated backwards through the network to the hidden layers which
receive no feedback from training patterns. The weights of the
interconnections are adjusted in small steps in the direction of
the error, to mini~i~e the errors, and the training data is run
through again. This happens many times till the error reaches an
acceptable level, which is usually the repeatability of the initial
measurement.
In the following, the invention will particularly be described
referring to the prediction of octane number of gasoline, but it
will be appreciated by those skilled in the art that the invention
is not restricted thereto and could also be used for prediction of
vapour pressure, density, cetane number and the like.
Data analysis on the set of spectra corresponding to the
gasolines of the training set is done in the following manner:
1. The mean spectrum of the set is generated and the differences
between each individual spectrum and the mean are calculated.
2. The mean spectrum will be in the order of 5000 data points and
35 so the problem of analysis of a set of 100 fuels is very difficult.
-
W 0 94/082~ 1 4 ~ 2 5 ~ PCT/EP93/02735
_ - 4 -
A technique is required to allow data reduction to a manageable
number of problem variables.
3. In the case of neural network technology the data reduction is
performed by physical reduction in the number of measured
wavelengths. The data reduction is in the following manner: A
multivariate statistical technique such as e.~Principal Component
Analysis is used on the training set of gasQ ~ , to generate a
'property spectrum' which represents the re'~ative importance of
each spectral data point to the correlation with octane number. The
spectral measurement is then simplified to discrete wavelengths,
typically numbering between 5 and 10. The absorbance values are
used as the input to the neural network.
Advantageously, the second overtone (harmonic) region of the
(N)IR spectrum is chosen. This region covers 900-1300 nm
(wavelength) and is chosen as it is in this region that the best
balance between available information from the measurement and
component instrumentation stability and sensitivity can be
achieved.
A number of discrete wavelengths is converted to absorption
data, which are used as the input to a neural network.
Advantageously, the number of selected wavelengths is 5 for
fuels that do not contain alcohols as oxygenates or do not include
cetane ignition improver additions and 6 if the fuels do contain
alcohol as oxygenates or do include cetane ignition improver
additions. Advantageously, for cetane number measurement a
wavelength of 6-7 ~m is chosen in addition to monitor the
concentration of cetane ignition improver additive.
One of the wavelengths is advantageously used as a
transmission reference to correct for any instrumental drifts.
The re~inin~ wavelengths, corrected by the reference, are
converted to absorption data. This may be done logarithmically, and
the data can be mathematically scaled within predetermined bounds
for each wavelength. That is, extreme values expected for either
fuels, or more likely, process streams are used to provide the
~ 214625~
W O 94/08226 PCT/EP93/0273
- 5
range of acceptable absorbances at each wavelength against which
the scaling can be done for the fuel to be tested.
The neural network is trained on the entire data set by
repeated presentation of input and known outputs i.e. the infra-red
data for a gasoline and its octane number, to learn the
relationship between the two and the performance of its predictions
against the actual octane number data as measured by standard
engine methods is monitored.
Once the neural network has "learned" the relationship, the
data set should be split into a further training set and a
validation set that will not be used in the "learning" phase.
The instrument of the invention advantageously collects (N)IR
absorbances at five discrete wavelengths, selected to yield
information from the C-H bond vibrations structure known to
influence the octane rating of a gasoline. The measured absorbances
are normalized to one of the wavelengths which is chosen to provide
a baseline and does not contain hydrocarbon information. This
allows for changing ambient conditions (temperature, (N)IR source,
electronic drift etc.) and the remaining four measurements are
applied to the neural network.
The invention will now be described in more detail by way of
example by reference to the accompanying drawings, in which:
fig. 1 represents schematically an engine based on-line octane
analyzer; and
fig. 2 represents schematically a neural network
advantageously applied in the apparatus of the invention.
Referring to fig. 1, an optical means 1 has been shown.
Advantageously, this optical means 1 comprises a plurality of
light-emitting diodes (LED), a filter and a lens-holder. For
reasons of clarity, mechanical connections of the analyzer to the
engine or to the car have not been shown.
The means 1 is connected through any suitable optical
connecting means 2 (advantageously a multi-way fibre bundle) to an
in-line gasoline cell 3 fitted in any suitable manner in a
hydrocarbon product line (not shown).
21~625~ --
W O 94/08226 - PCT/EP93/02735
- 6
Further, a photodetector 4 is present and provides the
obtained signal to be input to the processing electronics and
neural network for spectral analysis. In fig. 1, there are shown
5 LED's; however, any suitable number can be applied. For reasons
of clarity the processing electronics and neural network for
spectral analysis are not shown. Advantageously the geometry of the
apparatus of the invention is such that it can be applied in cars
as an engine-based instrument. -
~
Advantageously, as shown in fig. 2 ~the network used has a
three-layer architecture which, for example, comprises four input
nodes, 2 hidden nodes in a layer between the input A and output B,
and one output node. This is called a (4, 2, l) network. The
spectral data are presented as inputs A to the input nodes, wherein
the product quality information B is the output.
As known to those skilled in the art the nodes possess certain
weights of interconnections, and may be biased.
The weights and biases of the network can be stored and used
to analyze input data comprising the measured infra-red absorbances
and correlate the pattern to the octane number of a gasoline. Thus,
for a prediction which utilizes the network algorithm to describe
octane number from (N)IR-data, important parameters, having been
trained and successfully tested against the validation set, are the
weights of interconnection between the nodes and the biases at the
hidden and output nodes.
These can be interrogated and then implemented in the network
algorithm for the octane number analysis of future fuel samples.
For multiple outputs, a neural network algorithm is
implemented for each output. The implementation is by software code
on a microprocessor chip, and is therefore flexible to any changes
in network parameters which can be easily re-programmed.
In addition to unleaded motor gasoline, the instrument can
produce results for leaded fuels, provided that the lead content is
known. A simple numerical correction can be added to the octane
number predicted.
21 ~ 62~s
W O 94/08226 P ~ /EP93/0~735
- 7
It will be appreciated by those skilled in the art that the
network architectures applied may vary in the precise number of
nodes that~are present in each layer, or even in the number of
actual layers. Advantageously, 2 to 5 layers are applied.
According to the invention advantageously the number of nodes
of the input layer ranges from 3-10, the number of nodes of the
hidden layer(s) ranges from 1-10, and the number of nodes of the
output layer ranges from 1-3. More in particular, (3, 5, 1), (6, 6,
3) and (6, 6, 6, 3) networks could be applied.
The operation of the apparatus of the invention is as follows:
Five light emitting diodes (LED's) provide the near infra-red
radiation e.g. in the spectral range of 1-2.0 microns. The light
from the LED's is collimated and passed through interference
filters (one for each LED) which transmit light at selected
wavelengths in the near-infra-red spectral region (e.g.
1-1.5 microns). Advantageously, for gasoline the five wavelengths
are 1106 nm, 1150 nm, 1170 nm, 1190 nm and 1219 nm, the
normalization wavelength being 1106 nm due to gasoline having
minir-l absorbance at this wavelength, thus giving a good baseline
measurement. It will be appreciated that for gasoline/alcohol other
wavelengths are needed: advantageously 1766 nm and 1730 nm. These
may be required in addition to the others. An optical fibre bundle
(five into one) collects the filtered light through the filters and
delivers the light, from the selected LED, to the hydrocarbon
product line.
The LED selection can be achieved by electronic pulses, to
allow rapid measurements (<1 second) achieved by pulsing the LED's
one by one. Advantageously, optical windows are placed in the
in-line cell of the fuel line, to allow a 10-30 mm, advantageously
20 mm optical path length. An indium gallium arsenide detector is
mounted to detect the light transmitted through the optical path,
and provide the obtained signal to be input to the processing
- electronics and neural network for spectral analysis.
Various modifications of the present invention will become
apparent to those skilled in the art from the foregoing
W O 94/08226 21 4 6 2 S 5 PCr/EP93/02735 -
description. Such modifications are intended to fall within the
scope of the appended claims.
