Application of Artificial Neural Network and Soft Computing Techniques to Engineering World Problems By Dr. K.S. Madhavan , Sr. DGM, Programmable Control Systems, Corporate Research & Development Division, Bharat Heavy Electricals Limited, Hyderabad  500093, India.
Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author.While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server.
Application of Artificial Neural Networkand
Soft Computing Techniques to Engineering World Problems
By
Dr. K.S. Madhavan,
Sr. DGM, Programmable Control Systems,
Corporate Research & Development Division,
Bharat Heavy Electricals Limited,
Hyderabad  500093, India.
Corporate Research & Development, BHEL
2010 – 2020 is a decade of innovation primarily to be steered by Research & Development. In India the play has already begun.
Act I : Develop products based on global R&D collaboration
Act II : Focus on Basic Research
Act III : Build a R&D Powerhouse base within India
Developed economies are facing a problem of declining competitiveness on a global scale. We as developing nations have to capitalize on this aspect to strengthen our R&D base.
R&D is defined as a creative work undertaken on a systematic basis in order to increase the stock of knowledge including knowledge of man, culture, society and the use of this stock of knowledge to devise new applications.
BHEL has invested 2.5% of turnover on R&D. This will shortly be increased to 4%
Neural => neuronA carbon based chip from the brainthe biological computer
4
Brain: A carbonbased computer
Neuron is the Basic Building
block of our Central Nervous
System (CNS)
Human Brain has 1010 neurons.
Human Nervous System has over 1011neurons about the same number as the stars in our galaxy.
Each neuron, on an average,
process about 104inputs
5
Imaging system
Noisy Character Recognized Character
A is represented
As 0 0 1 0 0
0 1 0 1 0
0 1 0 1 0
1 0 0 0 1
1 1 1 1 1
1 0 0 0 1
1 0 0 0 1
M is represented
As 1 0 0 0 1
1 1 0 1 1
1 0 1 0 1
1 0 0 0 1
1 0 0 0 1
1 0 0 0 1
1 0 0 0 1
MATLABProgram
clf;
figure(gcf)
echo on
[alphabet, targets] = prprob;
[R, Q] = size(alphabet);
[S2, Q] = size(targets);
S1 =13; S2 = 26;
net = newff(minmax(alphabet),[S1 S2],{'logsig' 'logsig'}, 'traingdx');
net.LW{2,1} = net.LW{2,1}*0.01;
net.b{2} = net.b{2}*0.01;
noisyL = alphabet(:,1) + randn(35,1)*0.2;
plotchar(noisyL);
pause % strike any key to continue...
net.performFcn = 'sse'; % SumSquared Error performance function
net.trainParam.goal = 0.1; % SumSquared error goal
net.trainParam.show = 20; % Frequency of progress displays (in epochs)
net.trainParam.epochs = 300; % Maximum number of epochs to train
net.trainParam.mc = 0.95; % Momentum constant
% Training begins...please wait..
P = alphabet;
T = targets;
[net,tr] = train(net,P,T);
A2 = sim(net,noisyL);
A2 = compet(A2);
answer = find(compet(A2) == 1);
plotchar(alphabet(:,answer));
Elman Network Program in MATLAB for PHMS
Net = newelm([0 1],[10 1],{‘tansig’,’logsig’});
Net.trainparam.epochs = 300/1000/2000*;
P = […………………………………….];
T = […………………………………….];
Pseq = con2seq (P);
Tseq = con2seq (T);
Net = train (net, Pseq, Tseq);
Y = sim(net, Pseq);
Z = seq2con(Y);
Z{1,1};
Diff1 = T – Z{1,1};
Time Series Prediction of Boiler Drum Level
A time series is a sequence of vectors, x(t), t = 0,1,…where t represents elapsed time.
In practice, for any given physical system, x will be sampled to give a series of discrete data points, equally spaced in time.
Formally this can be stated as: find a function ƒ: RN → R such as to obtain an estimate of x at time t + d, from the N time steps back from time t, so that:
x ( t+ d) = f(x(t),x(t1)… x(tN+1))
Normally d will be one, so that f will be forecasting the next value of x.
a1(k1)
D
P’
LW11
a1(k)
LW11
LW21
a2(k)
b2
b1
Recurrent Neural Networks
a – activation level
P’ – input neurons
LW – Weights
b – bias
D  Delay
The actual/neural network predicted variation of the drum level parameter and the error.
Neural Network Input/Output patterns for Drum Level and error obtained after training
Residual Life Assessment Studies of Transformers using
Artificial Neural Networks
Dissolved Gas Analysis using Artificial Neural Networks
Rogers Fault Diagnosis Table
TableII : Rogers Ratio Codes
TableI : Gas Ratio Codes
TableIII : Rogers Fault Diagnosis Table
TableIV : IEC Ratio Codes
TableV : IEC Fault Diagnosis Table
LIMITATION OF THE RATIO METHODS : THE “NODECISION” PROBLEM
In reallife situations, Ratio Methods lead to codes that are way off from the norms set by the methods. In such a case inference of fault directly from the code pattern becomes difficult and vague.
Interpolations on patterns have to be made which are way beyond normal methods. Sometimes the gas content is undetectable or present as a trace. In such cases assumptions have to be made as to the negligible content of gas present. Ratio Methods fail to give logical diagnosis as to the nature of the fault. No decision can be taken on the diagnostic pattern generated by the Ratio Methods.
Artificial Neural Network (ANN) methods have to be used in order to arrive at a decision regarding the faults.
TRAINING WITH ANN
The neural network used in the present instance is a 3layer Probabilistic Neural Network (PNN) for analysis. There are 4 neurons in the input layer, 12 in the hidden layer, and 12 in the output layer. A smoothing factor of 1.5990609 was decided upon through NETPERFECT for all the links.
The distance metric used is Vanilla Euclidean. PNN networks work by clustering patterns based upon their distance from each other. The Vanilla Euclidean distance metric is recommended for most networks because it works the best. In Vanilla, the output of the network is the square of the distance between the patterns and the weight vector for the neuron; therefore, the winner is the neuron with minimum activation.
TRAINING WITH ANN (Contd.)
The training of ANN made is based on data shown in Tables I to V. When an input is presented, the first layer computes distances from the input vector to the training input vectors, and produces a vector whose elements indicate how close the input is to a training input.
The second layer sums these contributions for each class of inputs to produce as its net output a vector of probabilities.
Finally, a transfer function on the output of the second layer picks the maximum of these probabilities, and produces a 1 for that class and 0 for the other classes.
The Rogers Ratio Table and IEC Ratio Table with known inputs and outputs have been independently stored and trained in the reference ANN model.
Fault Diagnosis of Generator Transformer of Power StationI using ANN Approach
Flowchart of ANNRBS (Artificial Neural Network Rule Based System)
Suggestion : DGA samples from CTs of EHV transformers can also be tested through ANNRBS
Data Input
Neural Network Based Abnormal Detector
RuleBased Abnormal Detector
Y
Both indicate “Normal”?
N
RuleBased fault detector
Neural Network based fault detector
Combined fault diagnosis
Maintenance Action Recommendation
Outputs
Limits of Gases Beyond Which Fault Diagnosis Becomes Necessary
Note : *Before (C57.104) ** Corrected values 1978 IEEE(G) : Generation; IEEE(T) : Transmission
Values in italics are of transformers 67 years old
Unmarked sources are all cited from [Griffin86, Griffin88]
TDCG : Total Dissolved Combustible Gases
Estimation of degree of polymerization and residual age of transformers from Furan concentration dissolved in oil
Architecture of the General Regression Neural Network used
n
P1
ΣYI exp (DI2/22)
I=1
P2
Y1
A
P3
X1
B
Y2
Input Units
(I = 1)
Output Units
Summation Units
(O = 2)
Pn
n
exp (DI2/22)
I=1
Pattern Units
(n=31)
Furfurals are major degradation products of cellulose insulation paper and are found in insulation oils of operating transformers. Furfural analysis is an indirect method to estimate the integrity of cellulose insulation compared to the direct measurement of Degree of Polymerisation of insulating paper.
The tensile strength of the paper decreases corresponding to an increase in the concentrations of the Furfural in the oil. 5Hydroxymethyl2Furfuraldehyde and 2Furfuraldehyde are present in the oil at significantly greater concentrations than any other Furfural components.
Furfural levels range from 0.1 ppm to 10 ppm depending on the age and condition of the transformer insulation. The residual life of the transformer can be predicted by estimating Furfural content in the oil or by the Degree of Polymerisation of cellulose paper taken from lead insulation.
The life assessment can be made faster by estimating furfural from oil which can be collected from the transformer in running condition. The collection of cellulose paper involves cumbersome procedure of shutdown of the transformer and removal of paper from lead insulation after opening the transformer. Hence life assessment by furfural estimation is more popular and rapid method as compared to DP estimation of paper.
Choice of Artificial Neural Networks
The Degree of Polymerisation and Residual Age of transformers are continuous functions of Furfural levels of transformers. Therefore continuous function approximation of multiple outputs through Generalized Regression Neural Networks is one solution for predicting output patterns.
GRNN based on radial units, giving estimates of continuous variables rather than discrete decisions, overcoming the disadvantage of slow training inherent in backpropagation thereby lending itself well to realtime application, is the appropriate choice for our present analysis. Least square method has been used to minimize the error in prediction.
Partial Discharge Classification of 145 kV GIS using
Artificial Neural Networks
INTRODUCTION
GasInsulated Substation (GIS) equipment is being used worldwide in transmission and distribution of electrical energy. However its superior performance severely deteriorates with the presence of foreign particles in the system which causes Partial Discharge, leading to degradation of the insulation.
Reduction in the insulation strength to as low as 80% of the value under clean condition, has been reported. Partial Discharge is one of the effective indications of the defect and degradation of GIS.
The signals obtained from discharges occurring within GIS equipment due to microprotrusion on the enclosure and spacer, voids in solid insulation or due to floating particles are very stochastic in nature.
In the present study, the following three types of faults have been investigated :
Corporate Research & Development, BHEL
Detection of Partial Discharges
In order to determine the extent of partial discharge, a specified voltage was applied from the transformer to the busbar. The discharge signal was picked up from the ungrounded outer enclosure .
The signal is fed to the partial discharge detector and transmitted to the Computerised Discharge Analyser (CDA) for recording and processing the data.
The voltage is slowly increased, from zero, across the test object until corona is noticed in the PD detector. This is the discharge inception voltage. The voltage is continuously increased up to 145/3 ( 83.72) kV until all the three different kinds of partial discharges in the present study are detected.
The voltage is then slowly reduced until the smallest discharge disappears, the discharge extinction voltage. The circuit uses a Straight Detection Method with a potential divider arrangement across the test object.
Floating particles are seen with an applied voltage of 145/3 (83.72) kV, period 5 seconds, noise 15 dB, pressure 2 bar. PDmax (pC) : 172.51, PDavg (pC) : 93.5. Similar recordings are obtained for protrusion and particles sticking on the surface of insulator.
The digital data was recorded by CDA in terms of magnitude (pC) vs phase angle, and number of counts vs phase angle. In order to establish the artificial neural network (ANN) technique, only magnitude vs phase has been considered for simulation. This is considered for all three types of defects simulated for this study.
Prediction of fault probabilities of rotating machinery using GRNN techniques
Table – II Distribution of bit patterns as inputs to GRNN
Fault No. Related Symptoms
(High/Normal/Low)
(Yes/No)
(Faulty/Non faulty)
P1
X1
P2
(Trend in Vibration)
A
X2
P3
Y
X3
Output
Units
B
Xd
Summation
Units
Input Units
Pm
Pattern
Units
Evaluation of Gas Turbine Control Constants
PREDICTION OF GAS TURBINE CONTROL CONSTANTS
Artificial Neural Networks based solution for prediction of only important Site specific tunable control constants has been attempted considering the limitation on the availability of number of data patterns. Data for 8 sites have been taken with 7 sitespecific inputs and 21 output control constants.
Gas Turbines are mechanical devices operating on a thermodynamic cycle (Brayton cycle) with air as the working fluid. The air is compressed in a compressor, mixed with fuel and burnt in a combustor with the gas expanded in a turbine to generate power used in driving the compressor and external loads (thrust or shaft power) depending on requirements.
The thermodynamic cycle that represents the common turbomachine is the "open" Brayton cycle.
This hs diagram represents the ideal enthalpy and entropy relationship for the Brayton cycle.
Cycle Processes: 13 Isentropic Compression (q = 0) 34 Isobaric Heat Addition (w = 0) 49 Isentropic Expansion (q = 0)91 Isobaric Heat Rejection (w = 0)
Gas Turbine powerplant performance under ISO conditions (burning a reference fuel, such as natural gas, at 150 C, atmospheric pressure and 60% relative humidity) is information provided by machine manufacturers.
With the increasing utilization of gas turbines in industrial and cogeneration applications, they are taking on a greater role in base load service. Because of their inherent responsiveness, they also offer operating characteristics that can enhance their contribution to utility systems as a generator prime mover.
Detailed discussion were held with RCPuram and were informed that there are about 560 control constants required for this type of machine and about 70 machines have been already supplied by BHEL. Also, information had been obtained that these machines have been supplied to 10 sites in all, since multiple machines have been supplied to the same site.
GT Engg. RCPuram has identified 183 Tunable Control Constants after analysing the individual Big Blocks of Control Sequence Programs (CSPs) and Control Constants. However, from the overall data for 10 sites provided, only 8 sites are found to be useful since the data patterns provided for 2 sites are matching and 1 data set was for a machine with DLN combustor.
In this project, Artificial Neural Network (ANN) techniques are applied to analyze, predict the gas turbine control constants from generic blocks and application specific big blocks of the gas turbine control system. Constants from gas flow calculation, comparator, command state, Fuel Stroke Reference (FSR), Inlet Guide Vane (IGV) fault detection, acceleration control of FSR, Temperature control reference were shortlisted to be predicted by ANN.
A final list of 21 control constants were prepared based on:
1. The data patterns available are enough for training ANN with reasonable accuracy in prediction.
2. Data spread is uniform, without any discontinuities.
7 sitespecific variants were identified as input to ANN:
Machine type
Site Elevation (m)
Site Design Temperature (0C)
Site Relative Humidity (%)
Type of Cycle
Output (MW)
Lower Heating Value of Fuel (Kcal/kg)
21 output variables were identified as output to ANN.
In order to obtain a best possible solution, ANNs have been trained by changing the following variants.
1. Model / Type of ANN (Training algorithm)
2. No. of Epochs
3. Data formats
ANNs have been trained for 2 types of Models and 2 types of data formats by varying the no. of Epochs (Low, Medium & High) for arriving at an optimum solution. The Input/Output data format has been used in engg units as well as normalized for training. The data has been normalised in such a way that the spread is in the range of 1 to +1.The normalised values for the data patterns are computed using the formula (ActualMean) / (MaximumMinimum)
The number of neurons in hidden layer are computed as, (No. of input neurons + No. of output neurons)/2 + Sqrt. (No. of patterns). The number of theseneurons is either 17 or 19, depending upon the type of input data format.
The initial weights of these neurons are assumed to be a low value and are subsequently modified during the training. Selection of very low values for initial weights will result in a longer duration for training and very large values in saturation of the network. Hence, an optimum value of 0.3 has been assumed in this case. The learning rate and momentum are taken to be 0.1 and 0.3 respectively.
The validation of ANN model is the most essential part of identification process. An engineer would never deliver a product without a mention about its accuracy. The approach to the validation of a trained ANN model is to establish the accuracy to which the model is able to predict. A naïve approach to this problem is to validate the trained ANN models with the data that the model was trained with. This is called Naïve Validation. Under Naïve Validation, an ANN model is always bound to predict closely to the trained data, since it was trained on that data. This does not mean that the ANN model is capable of representing the system, but only that the model is able to adjust to the trained data.
Overall absolute deviations (max and min) for different sets of ANN Parameters under Naive validation
Graphs indicating actual/predicted output values of RPCL Kalugurani
Graph indicating the error curve
Three Layer Backpropagation was used as the neural network model and training was done for 2500 epochs with seven input variables and twenty one output variables
It is thus necessary, in order to check the ability of the ANN model to “generalise,” to validate the trained models on an independent set of data, called the Validation data. This final and decisive test for any trained model is a Cross Validation, which involves training with new data sets and making predictions and comparison with the actual data. The predicted data should be identical or within the prescribed uncertainty bounds.
Sets of parameters used for training ANN for cross validation
Set 1:
Data Type : Engineering units
ANN Model : 3layer Back Propagation (BP)
No. of Epochs : 2,500
Set 2:
Data Type : Normalised units
ANN Model : 3layer Back Propagation (BP)
No. of Epochs : 2,500
Set 3:
Data Type : Engineering units
ANN Model : Feed Forward BP using combination of linear, Gaussian, tanh, Gaussian complement and logistic
No. of Epochs : 1,115
Set 4:
Data Type : Normalised units
ANN Model : Feed Forward BP using combination of linear, Gaussian, tanh, Gaussian complement and logistic
No. of Epochs : 1,087
Other Applications
Neural networks are nothing but a mathematical processing unit. It can be used for various applications, for example in:
Dynamic Neural Controller
r
y
: Input signal[Step or square Inputs]
: Neural unit
n
: Control signal for the plant
u
n
y
: Desired output
d
x
T
[
x
x
x
]
are the neural inputs
a
0
1
2
y
: Output of the plant
e
p
: Error between the desired output
and the output of the plant
62
Evolutionary Computation
Evolutionary Algorithms
 Genetic Algorithms
 Evolutionary Programming
 Evolutionary Strategy
 Genetic Programming
 Learning Classifier System
Swarm Intelligence
 Ant Colony Optimization (based on pheromone search)
 Particle Swarm Optimization (based on simulated annealing)
 Stochastic Diffusion Search
Self Organization
 Self Organizing Maps
 Growing Neural Gas
 Competitive Learning
Differential Evolution (based on Genetic Annealing)
 a population based combinatorial algorithm
Artificial Life
 Strong Alife (based on Cellular Automata inspired by Von Neumann)
 Weak Alife (based on Neural Networks)
Harmony Search Algorithms
 Metaheuristic algorithm mimicking the improvisation of musicians
Artificial Immune Systems
Learnable Evolution Model
Cultural Algorithms
 Bridging the gap between Belief Space and Population Space
Studies done on Supercritical Test Rig Facility Data through
Genetic Programming
TEST FACILITY
To find the correlation between Nusselts number (output) and Reynolds number & Prandtl number (inputs).
Does it satisfy the DittusBoelter equation?
Nud = 0.023Red0.8Pr0.4
(for subcritical conditions satisfying heat transfer in fully developed turbulent flow in smooth tubes)
If not find the function linking Nusselts number to Reynolds number & Prandtl number using Genetic Programming.
#include <math.h>
#include <float.h>
#define TRUNC(x)(((x)>=0) ? floor(x) : ceil(x))
#define C_FPREM (_finite(f[0]/f[1]) ? f[0](TRUNC(f[0]/f[1])*f[1]) : f[0]/f[1])
#define C_F2XM1 (((fabs(f[0])<=1) && (!_isnan(f[0]))) ? (pow(2,f[0])1) : ((!_finite(f[0]) && !_isnan(f[0]) && (f[0]<0)) ? 1 : f[0]))
float DiscipulusCFunction(float v[])
{
long double f[8];
long double tmp = 0;
intcflag = 0;
f[0]=f[1]=f[2]=f[3]=f[4]=f[5]=f[6]=f[7]=0;
double Re=v[0] ;
double Prandtl=v[1] ;
L0: f[0]+=0.002621650695800781f;
L1: f[0]+=Prandtl;
L2: f[2]+=f[0];
L3: f[0]*=f[0];
L4: f[0]*=f[2];
L5: f[0]*=Prandtl;
L6: f[1]+=f[0];
L7: f[0]*=f[2];
L8: f[0]*=pow(2,TRUNC(f[1]));
L9: f[0]*=1.642645597457886f;
L10: f[0]/=Re;
L11: f[0]=fabs(f[0]);
L12: f[0]/=1.238061666488648f;
L13: f[1]+=f[0];
L14: f[0]*=pow(2,TRUNC(f[1]));
L15: f[0]*=f[1];
L16: f[0]+=f[1];
L17: f[0]=f[0];
L18: f[0]+=f[0];
L19: f[0]+=Re;
L20: f[0]=sqrt(f[0]);
L21: f[0]=f[2];
L22: f[0]*=0.7233922481536865f;
L23: if (!_finite(f[0])) f[0]=0; return f[0];
}
float DiscipulusCRegressionFunction(float v [])
{
float ret = DiscipulusCFunction(v) ;
return ret;
}
Boiler Performance Analysis Using Soft Computing Techniques
Presently Boiler performance analysis is being made through standard procedure. But there are deviations in the performance analysed using this procedure.
Presently Boiler Performance is estimated based on indirect method of calculating the cumulative losses from different sources and subtracting it from the theoretical design output.
But this method does not take into account the variations in the coal quality actually fed into the hopper nor does it take into account the variations in design of the boiler based on site requirements.
77
Collaborator’s specifications are based on US grade coal. The deviations arising out of usage of Indian grade coal needs to be studied thoroughly.
For this purpose, the black box technique of neural network is being used to study the inputoutput relationship between various Boiler performance features and relevant inferences are drawn.
78
78
Feature selection encompasses a wide variety of methods for selecting a restricted number of input variables or “features”, which are“relevant” to a problem at hand.
Feature selection is a problem pervasive in all domains of application of machine learning and datamining: engineering applications, robotics and pattern recognition (speech, handwriting, and face recognition), Internet applications (textcategorization), econometrics and marketing applications and medical applications (diagnosis, prognosis, drug discovery).
Restrictingthe input space to a (small) subset of available input variables has obvious economical benefits in terms of data storage, computationalrequirements, and cost of future data collection. It often also provides better data or modelunderstanding and even better predictionperformance.
79
Reasons for performing feature selection include:
81
Burner Tilt was averaged on all the corner points. Tilt expressed in % was converted to degrees using the formula:
((60/100)*x – 30)
Where x is the Tilt in %. Tilt varies from 30 deg. to +30 deg.
MILL COMBINATION
Mill Combination was taken to be the distance from Platen Super heater to the topmost mill fired. This is calculated as follows:
B8 = 18525 + 2*1575 = 21675 mm Where distance
M8 = 18525 + 1*1575 = 20100 mm from topmost mill is 18525mm &
T8 = 18525 mm there are 10 mills in total
81
Excess Air Percentage
(Calculated indirectly from dry O2 percentage in Economizer Outlet by the formula
100 * O2 / (21 – O2))
The average of Left (L) and Right (R) values are taken
LOAD
Load is expressed in MW
REHEATER SPRAY AND SUPERHEATER SPRAY
RH Spray & SH Spray are expressed in tonnes per hour and the sum of L & R values are taken.
82
Block Diagram indicating dynamic program flow
ANN
DEF
File
ANN
DLL
Visual Basic Front End GUI
Predicted Sprays
84
Benefits of the ANN Prediction Tool
The tool will serve as a ready reckoner for operators to predict the quantity of Re Heater Spray and Super Heater Spray, after data entry of other input features like Burner Tilt, Mill Combination, Excess Air and Load.
A good understanding of the Boiler performance and the relationship between different input / output parameters will in due course of time lead to quality improvement in Boiler design and performance and acquiring analytical / design / modelling capabilities for better product design.
The performance analysis techniques used in this project can be extended to different types of Boilers.
85
Backpropagation was tried to start with for 70 sets of power plant data classified as Training set & Test set. Data is sorted in terms of Burner Tilt, Excess Air, Mill combination and Load individually. Then it is trained with different learning rates, momentum, epochs & events since minimum average error > 40,000. With 90% Training Data,10% Test Data, epochs varying from 2019 to 4765 correlation coefficient of RH Spray is found to vary from 0.9531 to 0.9684, correlation coefficient of SH Spray is found to vary from 0.9879 to 0.9919.
GRNN proved better with correlation coefficient of RH Spray varying from 0.9505 to 0.9947, correlation coefficient of SH Spray varying from 0.9925 to 0.9970 with an optimum smoothing factor of approx. 0.0255 which was generated through genetic algorithm. 80% Training data and 20% Test data was used.
Cross Validation (3rd Data Pattern):
Tilt MC Excess Air Load RH Spray SH Spray
22 20100 27.65 503.91 56.01 11.71 (Actual)
51.55 11.35 (Cross Validation)
7.96% 3.07% (Error) 86
Generation of smoothing factor using Genetic
Algorithm
Training using GRNN
Training using Backpropagation
Soft Computing Components Methodology used
Boiler Performance Inputs
Algorithm
Training using Backpropagation
Training using GRNN
87
(General Regression Neural Network)
89
¥
y
(
X,y)
dy
¦

¥
E[y
X] =
¥
¦
(
X
,y)
dy

¥
Multiplying the measured value of the output with the
appropriate probability density function of the Euclidean
distance (
Di) of any input variable X from other input
variables occurring in the attribute space and averaging
yields the estimated value of the predicted output
n
Yi exp(D2/22)
I=1
n
exp(D2/22)
I=1
Y (X) =
Many spheres of influence will be formed for various points.
The appropriate sphere of influence is defined as the one that
produces the smallest mean square error between the actual
and predicted output values. Determination of this appropriate
sphere of influence i.e. (smoothing factor) is where learning
takes place in GRNN.
In order to find out the optimum smoothing factor, Genetic Algorithms are used. A genetic breeding pool size of 20 is used here.
Search/Optimization technique by choosing survival of the fittest
through chromosome crossover or mutation. Unit of GA is allele.
92
How Genetic Algorithm Works? (Contd.)
After dozens or even hundreds of generations, a population eventually emerges wherein the individuals will solve the problem very well. In fact the most fit (elite) individual will be an optimum or close to optimum solution.
The processes of selection, mutation and crossover are called genetic operators. Genetic Algorithm includes additional genetic operators. One is called diversity operator.
Genetic Algorithm for Boiler Performance Analysis
In the project “Boiler Performance Analysis using Soft Computing Techniques”, the twin objectives of minimizing ReHeater Spray as well as SuperHeater Spray become essential.
SuperHeater Spray and ReHeater Spray are functions of burner tilt in our curve fits generated. The other parameters like Mill Combination, Excess Air, Load can become part of the Chromosome (continuous type) input into the Genetic Algorithm software.
Therefore we now have fitness functions as well as chromosomes to achieve multiple objectives of minimizing ReHeater Spray as well as SuperHeater Spray.
The Curves are shown in the following slides:
Power Station  I
Power Station II
HYBRID OPTIMIZATION USING GENETIC ALGORITHMS
The coefficients of the trend line equation, centroid of the excess air values after grouping, centroid of the Mass Flow values after grouping, different values of tilt after grouping were all taken as continuous chromosomes to be further treated by Genetic Algorithm (GA).
The sum of the squares of the error function between calculated values from the trend line equation and the actual RH, SH Spray were taken as two fitness functions.
The multiple objectives were to minimize both the fitness functions for RH as well as SH Spray.
The steam temperature corresponding to the data sets were taken as constraints and these were not to exceed 545 deg. C with a tolerance of 5 deg. C. The ranges of all the input chromosomes were defined. The GA was run for 650 generations before it converged on an optimum solution set.
Development of Artificial Neural Networks (ANN) based Prediction Model for NOx Emissions from Utility Boilers
101
Thermal NOx and Fuel NOx may be controlled but Prompt NOx remains uncontrolled and is quantitatively less compared to the former two.
NOx can be controlled in two phases:
During Combustion Process
Post Combustion
Combustion Controls:
1. Flue Gas Recirculation
2. Over Fire Air
3. Low NOx burners
4. Reburn
Post Combustion Controls:
Selective NonCatalytic Reduction (SNCR)
Selective Catalytic Reduction (SCR)
Thermal NOx can be controlled by bringing Temperature down. Also bringing concentration of O2 or N2 down. Thermal NOx rises slowly.
Fuel NOx can be controlled by delayed mixing of fuel and air to form N2 rather than NOx. Fuel NOx rises rapidly.
Thermal, Prompt and Fuel NOx are controlled by modifying:
Combustion Gas Temperature
Residence Time
Turbulence
(The Three T’s)
Alternative Techniques for reducing NOx
Burner Out Of Service (BOOS): Stopping fuel flow from individual Burners. Air flow is maintained through idle burners.
Biased Firing (BF): Injecting more fuel to some burners and reducing the amount of fuel to other burners.
Close Coupled Over Fire Air (CCOFA): In the Main Wind Box
Separated Over Fire Air (SOFA): Installed above main Wind Box using separate ducting.
SNCR: By injecting Ammonia in the flue gas. Ammonia is a pollutant and can react with SOx in the flue gas to produce Ammonium salts which can deposit in downstream equipment such as air heaters.
SCR: By injecting Ammonia in the presence of catalysts like Platinum or Palladium / Vanadium or Titanium / Zeolites (crystalline ammoniosilicate compounds)
Reduce Excess Air
The Inputs and Output to the Neural Network
NOx Prediction Tool (15 features)
Feature Selection was tried out by
Feature Selection based on Fdistribution and its corresponding p value
Ten dominant features are extracted from 15 features
F Value
p Value
Application of Artificial Neural Networks for NOx Prediction
(General Regression Neural Network Model)
Network type : GRNN, genetic adaptive
Patterns processed : 49
Smoothing factor : 0.1342353
Output : NOx(in ppm)
R squared : 0.9718
r squared : 0.9739
Mean squared error : 175.507
Mean absolute error : 6.557
Min. absolute error : 0
Max. absolute error : 48.004
Correlation coefficient r :0.9868
Percent within 5% :79.592
Percent within 5  10% : 12.245
Percent within 10  20% : 8.163
Percent within 20  30% : 0
Percent over 30% : 0
NOx
NOx Prediction Tool (10 features)
Some Suggested Projects !!!
An Assessment of Back Propagation Neural Networks for Weather Forecasting
This research project looks at using local weather conditions and a Back Propagation neural network to predict the following day's barometric pressure.
An Assessment of Brain State in a Box Neural Networks for Database Assessment
This research project looks at using Brain State in a Box Neural Networks to assess the contents of a database and find data according to partial patterns provided.
SURVEY OF RURAL STUDENT PROSPECTIVE ANALYSIS
1. Course of engineering is most usable in daytoday life.
2. Engineering course takes background of 12th standard / preuniversity course.
3. It is risky to get admission for engineering during recession.
4. Jobs / opportunities are more in engineering stream.
5. Innovation / discovery opportunity is more in engineering.
6. Relatives taking admission in engineering is an influencer.
7. 12th standard / preuniversity teachers inspire engineering study.
8. Foreign opportunity is more after BE , BTech.
9. Family members suggest to go for engineering study.
10. Admission into any stream of engineering is OK.
11. Government providing loan for engineering is an enabler for engineering education.
12. Engineering is better than other courses in science, math, arts, fine arts.
13. Entrance exam for engineering is a hectic task.
14. Criteria for admission affects student behaviour.
We can introduce student behaviour system towards Engineering and Research with the help of artificial neural network model in Microsoft Excel. Excel is a popular and low cost technical education modelling technique with decision analysis.
Automobile
Agriculture
Aerospace
Forestry
Construction
113
Milling / Drilling
Robotics