Abstract:
Process industries today enjoy significant benefits from advances made in the field of
simulation as well as technologies that model normal plant operation. Plants however continue
to suffer during abnormal operation such as start-up, shutdown and equipment failures leading to
production losses or personal injury.
One of the key elements that determines the operational safety of any plant is the training
and technical knowledge of its personnel regarding plant behaviour. Simulators play a vital role
in training personnel, preparing them for normal plant operation, abnormal plant operation as
well as emergency and accident situations. This would not only enhance plant safety, but also
decrease total downtime.
In order to create such a simulator, mathematical models are required for each of the
numerous components within the plant. This dissertation focuses on the development of a
mathematical model for a control valve; a component commonly found in various industries to
control and manipulate processes. Different modelling methods are compared, taking into
account applicable modelling criteria such as training data, algorithm complexity, oscillations
near endpoints, degree of system integration and model limitations. Based on these criteria,
fuzzy logic with the nearest neighbourhood clustering algorithm is chosen as an appropriate
modelling technique especially due to its ability to deal with large quantities of data.
In order to meaningfully train the fuzzy logic system (FLS), a comprehensive set of
physical operational data is required, covering all the different operational characteristics. To
capture physical data, the development of a data acquisition (DAQ) system is introduced using
two common DAQ systems to create a hybrid solution. Transducer signals are converted tom
mA to V, using custom developed signal conversion hardware. This will allow data to be
sampled by a standard DAQ card and processed by accompanying software. Two post-
processing software applications are created. The first application solves the governing
equations (mass rate of flow, Reynolds number, expansibility factor and choke status) and the
second application is used to graphically display the acquired and calculated data. A set of
experiments are conducted, covering all relevant working areas, to capture the behaviour of the
control valve. This is achieved using five initial pressures ranging from 200 kPa to 400 kPa in
increments of 50 kPa. At each initial pressure a set of unit step responses with valve command
signals ranging from 0 % to 100 % in increments of 5 % is acquired. 24 data files at each initial
pressure set (200, 250, 300, 350 and 400 kPa) are acquired.
Before training the FLS, the optimal fuzzy logic parameters need to be determined e.g.
radius (r), sigma (σ)the number of time delays, the time delay increments and the impact of the
input signals. Determining these parameters is an iterative process. Only a single data set,
with initial pressure of 300 kPa, is used to derive the optimal fuzzy logic parameters. Four
performance criteria namely maximum error average (MEA), mean square error (MSE), root
mean square error (RMSE) and coefficient of variation of the error residuals (CVRE) are used
as benchmarks to obtain the optimal fuzzy parameters.
During both the search for the optimal fuzzy parameters and the training of the fuzzy
models using these optimal fuzzy parameters, 70 % of the data are used for training and
verification while the remaining 30 % of the data are used for validation.
Once the optimal fuzzy parameters are obtained using only the single data set, it is used
to derive a number of fuzzy control valve models based on all the available data sets. All
derived fuzzy models use the same parameters, except for a unique random file sequence
associated with each of the models. The only prerequisite for the fuzzy models is that the
generated file sequence be truly random. Irrespective of the random file sequence, fuzzy
models with the same parameters, produce models with more or less the same performance.
Therefore the performance criteria (MEA, MSE, RMSE and CVRE), for each data file, in the
respective initial pressure sets, remains more or less the same.
This method is found to be very useful in deriving a dynamic fuzzy logic control valve
model. Averaging the performance criteria of these five models, an overall modelling accuracy
of 90 % is achieved.
It is recommended that a flow meter be installed to measure the mass rate of flow through
the pipe network. This eliminates the need for an orifice, differential pressure transducer and
the use of the first principle governing equation for the mass rate of flow. If a flow meter cannot
be installed, a differential pressure transmitter with large range should be considered
accompanied by a single orifice.