Computer Aided Monitoring

A session on Introduction to Artificial Neural Network & its application

A case study on performance monitoring of a Gas Turbine used in Power plant
Presented by:- 1) Mr. Yogesh S. Suryavanshi
2) Mr. Dimpesh Silarpuriya
3) Mr. Ramashankar Gupta

Introduction to ANN( Artificial Neural Network)

ARTIFICIAL INLELLIGENCE:- Can Machines Think?
The answer to above question is yes. Machines can be made think like human brains.
Artificial intelligence is one of the branch of science which ascribes intelligence to a computer system.
The purpose is to artificially resemble a Brains capacity to draw conclusions, plan, solve problems, learn etc.

Among the several approaches within field of AI most common approaches are1) symbolic AI approach 2) connectionist AI approach
Artificial Neural Network (ANN) is one of the representative example of connectionist AI approach.
ANN is a non-linear statistical data modelling tool mimicking the neural structure of the human brain & It basically
learn from experience i.e. adaptive.
It is a computing system made up of a number of simple, highly interconnected processing elements, which process
information by their dynamic state response to external inputs.
ANN can be used to solve a variety of tasks including classification, regression, general estimation problems etc.
ANN consists of group of interconnected artificial neurons processing information in parallel.
Neural networks are typically organized in layers. Layers are
made up of a number of interconnected 'nodes' which contain an
'activation function'. Patterns are presented to the network via
the 'input layer', which communicates to one or more 'hidden
layers' where the actual processing is done via a system of
weighted 'connections'. The hidden layers then link to an 'output
layer' where the answer is output as shown in the graphic below

The performance of a network can be improved by rendering it aware of its output. Through a performance

feedback loop that includes a cost function. The feedback is used to adjust the network parameters through
systematic procedures called learning or training rules, in order to improve the system output with respect to the
desired goal. This learning procedure is illustrated in below fig.

The progress of learning or training the network is iterative process & involves modifying the strength of
connections between elements. The two main learning types are:1) Supervised learning:- in this type both inputs & desired outputs are known. The network can measure its
predictive performance for given inputs. (classification)
2) Unsupervised learning:- In this type the targets are unknown & the ANN has to find the underlying
relationships within the data set by itself & build cluster of data. (Clustering)

 Types of Neural Networks:1) Feed-forward networks:- It is the simplest network where information travel only in single direction i.e. linearly from
input to output. this type is used for static mapping.

2) Feed-back Networks:- in this type a later processing stage can propagate signals to an earlier stage. This type is used for
dynamic or time dependent mapping

Artificial Neuron:- It is simple in construction. It is the building block in an ANN. Below fig. shows a schematic of an
artificial neuron

Mathematical representation

ya & yb = incoming signals from other nodes or neurons (joined by propogation function normally standard summation)
Sk = Resulting signal/effective input
F = Transfer function
yk = Output signal
Wka & Wkb = respective weights (incoming signals are multiplied with respective weights equivalent to their significance for neuron)
+1 = Bias (additional input to neuron) & W0k= corrosponding weight (allow the neuroan to have an output even input is zero)

 Types of Neural Networks:1) The single-layer Feed-forward network:- There is only one layer with artificial neurons. It is also called as Perceptron. It
uses supervised training to update the connection strengths(weights). An error is the difference between desired output or
target & output calculated by the ANN. Only Linearly separable problems can be solved by this network.
2) The multi-layer Feed-forward Network:- It is also called multi-layer perceptrons (MLPs). it permit the solution of nonlinearly separable tasks. It has at least two layers of neurons, one of which is hidden. MLPs are trained in supervised
manner.

The general structure of two-layered feed-forward neural

network

 Gas Turbine Monitoring:In gas turbine monitoring sub-areas identified are as:1) Condition Monitoring:- process of monitoring one or several parameters of the condition of machinery. A significant

change is indicative of a developing failure.

2) Sensor Validation:- It is the process of monitoring sensor accuracy, detecting & isolating failing sensors & recovering
sensor values.
3) Fault Diagnosis:- it is the process of detecting & isolating faults occurring in the machinery.
Following conditions are typical for Gas turbine Deterioration:- any of these following conditions will cause change in
the performance of a Gas Turbine

The building of mathematical models through physical laws is a complicated process, especially for highly interconnected
systems with losses(turbulence etc.)which are difficult to model in an accurate way. In this aspect neural network provide

the option to built the monitoring models directly from the operational data, thus presenting a cost-effective option for
sensor validation. Neural networks provide the capability of using high dimensional & nonlinear data in an efficient
manner, characteristics which are always hard to simulate with analytical models. Another aspect worth mentioning is that
sensor accuracy etc. is directly included in the neural network. This results in robust models which are insensitive to small
input errors that are common in real measured data.

A case Study

Development & multi-utility of an ANN model for an industrial gas turbine

The power & combined heat & power (CHP) sectors all over the world have gone through major transformations over the
last two decades due to rising competitiveness for deregulation of electricity & more stringent laws for environmental
protection. increasing the Reliability, Availability, & Maintainability of existing plants is a present need to improve the
performance & minimize the environmental impact.
ANN is a tool suitable for simulation & condition monitoring of power & CHP plants.
This case Study was done by Research unit of Thermal Power engineering at Lund University, Sweden. The below Turbine
is installed at one of the Power plant at Gunnesbo, Sweden.
Specification of Gas Turbine under study: Gas Turbine:- GT10B (Siemens Industrial Turbo machinery)
Compressor:- Axial Flow Type (Ten Stage- Overall Pressure Ratio of 13.6:1)
Combustion Chamber:- Annular type- Dual Fuel (Gas & liquid) having 18 Dual fuel air atomized Burners
Compressor Turbine:- Two stage with Internal cooling by convection, Nominal speed of 9770 rpm, Rated air flow of
77.2kg/s
Power Turbine:- Heavy Duty, Nominal Speed of 7700 rpm for Power Generation having capacity of producing 22 MW of
electricity.
Equipped with Heat recovery unit. Heat is extracted in an economizer & exhaust gas condenser. Approximately 35 MW
of heat is extracted in the Heat recovery unit.

Schematic Layout of the CHP plant

Steps for the development of the Gas turbine ANN Model:1) Data collection & screening:- Operational data of one week duration from the plant was obtained as averaged for one
minute.
This data is considered to be the baseline data for definition of the healthy condition of the plant. Data is screened to remove any
transient data. Almost 4000 data points related to enormous sensor signal were removed from the original data set & 7500 data
points remained which were used in the network training & validation process.
2) Selection of Input & output parameters:- Proper selection of input parameters for accurate prediction of a set of output
parameters is important for ANN modelling of any system. Input Parameters for which the accuracy of prediction of output
parameters did not improve was considered redundant & moved from the final ANN model. The complete set of input &
output parameters are given in below table.

3) ANN training parameters:- Training of the ANN was done using commercial software NeuroSolutions Supervised Neural
network with back propagation method was used for training the network. The filtered plant data was divided as 60% for
training, 15% for cross-validation & rest 25% as test data set.
4) Sensitivity analysis:- The purpose of sensitivity analysis is to determine the dependence of output parameters on the
initially selected input parameters. If the prediction accuracy is not affected in spite of removing an input parameter, it proves
that this input parameter is redundant & may be removed from the input set of parameters.

3) Final ANN model & its prediction accuracy:- based on the sensitivity analysis, the final ANN model for the prediction of
gas turbine performance parameters is given below.

The accuracy of prediction of power output with above final trained ANN is shown below.

5) Development of the graphical user interface (GUI):-The idea of developing a GUI for the ANN model was to make it user
friendly so that any person without formal knowledge of either plant or ANN can also use the model to estimate plant
performance. The GUI was created in excel environment. The screen of user interface is shown below.

6) ANN extrapolation capability:- Another utility of the developed model was shown for the extrapolation of predictions even
beyond the range of training data. e.g. The ANN was actually trained for Temperature range of 4.1 to -4.50c. Same ANN was run
for 5 to -140c i.e. beyond the range to examine the extrapolation capacity. The results are as shown below. The prediction
accuracy is good beyond the range of training data

Conclusions:Artificial Neural network is found to be a useful tool for prediction of gas turbine performance if it can be trained properly with
operational data. This is demonstrated by very high prediction accuracy of the developed ANN model. Developed ANN model
may have several utility. It may be used for offline simulation of gas turbine performance or online condition monitoring of the
gas turbine for early detection of faults or degradation. It may also be used for sensor validation purposes.

References:1) Artificial IntelligenceNational Encyclopedia 2010

2) Artificial Intelligence: a modern approach:..Russell SJ, Norvig P. Prentice Hall, 2009
3) Introduction of intelligent tools for gas turbine based small-scale cogeneration:.Arriagada J. Lund University
Sweden 2001
4) Neural networks and statistical techniques: A review of applications:..Paliwal M Kumar UA, 2009.
5) Development & multi-utility of an ANN model for an industrial gas turbine:..Journal of applied energy, Vol.86,
issue 1, pp. 9-47, January 2009

THANK YOU