Modelling & simulation of human powered flywheel

IJRET: International Journal of Research in Engineering and Technology eISSN: 2319-1163 | pISSN: 2321-7308
__________________________________________________________________________________________
Volume: 02 Issue: 12 | Dec-2013, Available @ http://www.ijret.org 271
MODELLING & SIMULATION OF HUMAN POWERED FLYWHEEL
MOTOR FOR FIELD DATA IN THE COURSE OF ARTIFICIAL NEURAL
NETWORK – A STEP FORWARD IN THE DEVELOPMENT OF
ARTIFICIAL INTELLIGENCE
A. R. Lende1
, J. P. Modak2
1
Ex. Assistant Professor, Mechanical Engineering, DMIETR, Wardha, Maharashtra, India,
2
Emeritus Professor & Dean (R&D), Mechanical Engineering, PCE, Nagpur, Maharashtra, India
tanuja.chandak04@gmail.com, jpmodak@gmail.com
Abstract
As per geographical survey of India about 65% of human population is living in rural areas where urban resources like electricity,
employment accessibility, etc are very deprived. The country is still combating with fundamental needs of every individual. The
country with immense population living in villages ought to have research in the areas which focuses and utilizes the available human
power. Some Authors of this paper had already developed a pedal operated human powered flywheel motor (HPFM) as an energy
source for process units. The various process units tried so far are mostly rural based such as brick making machine (both
rectangular and keyed cross sectioned), Low head water lifting, Wood turning, Wood strips cutting, electricity generation etc. This
machine system comprises three sub systems namely (i) HPFM (ii) Torsionally Flexible Clutch (TFC) (iii) A Process Unit. Because of
utilization of human power as a source of energy, the process units have to face energy fluctuation during its supply. To evaporate
this rise and fall effect of the energy, the concept of use of HPFM was introduced. During its operation it had been observed that the
productivity has great affection toward the rider and producing enormous effect on quality and quantity of the product. This document
takes a step ahead towards the development of a controller which will reduce system differences in the productivity. This paper
contributes in development of optimal model through artificial neural network which enables to predict experimental results
accurately for seen and unseen data. The paper evaluates ANN modeling technique on HPFM by alteration of various training
parameters and selection of most excellent value of that parameter. The mathematical model of which then could be utilized in design
of a physical controller.
Keywords: Artificial Neural Network, HPFM, MATLAB
-----------------------------------------------------------------------***-----------------------------------------------------------------------
1. OVERVIEW OF HPFM OPERATED PROCECSS
UNIT
1.1 Working of Human Powered Flywheel Motor
Energized Process Unit
This machine system comprises three sub systems namely (i)
HPFM [11] (ii) Torsionally Flexible Clutch (TFC) (iii) A
Process Unit. The process units tried so far are mostly rural
based such as brick making machine [1] [3](both rectangular
and keyed cross sectioned), Low head water lifting, Wood
turning, Wood strips cutting, electricity generation etc.
The Fig-1 shows the schematic arrangement of pedal operated
flywheel motor which comprises of following elements
R= Rider
M = mechanism (01-OA-B-02-01)
BSC = Big Sprocket Chain Drive
Fig -1: Schematics of Human Powered Flywheel Motor
SSC = Small Sprocket Chain Drive
GSR = Gear of Speed Rise
PSR = Pinion of Speed Rise
FW= Flywheel

__________________________________________________________________________________________
CH = Chain
CS = Counter Shaft
FS = Flywheel Shaft
1.2 Study of already Available Experimental Data
The various parameters involved [11] in the experimentation
are
Table -1: Independent Variables and their symbols
Sr. No. Independent Variable Symbol
1 Moment of Inertia of Flywheel I
2 Input by the Rider R
3 Time T
4 Mechanical Efficiency ME
5 Gear Ratio G
6 Angular Velocity of Flywheel ω
Table -2: Range of Variation of Independent parameter i.e.
Rider(R)
Range of age 20-25Years
Height 155-170cm
Weight 40-55
Blood Pressure 140-70
Pulse rate 68-80/min.
The observations recorded during the experimentation are as
below
Table -3: Experimental observations
Independent Variables
Dependent
variable
Sr.
No.
Log(I/RT2
) Log(ME) Log(G) Log(ω T)
1 -7.4270 0.00 0.3010 3.6305
↓ ↓ ↓ ↓ ↓
23 -7.1792 0.0662 0.0010 3.5570
↓ ↓ ↓ ↓ ↓
50 -7.2694 0.0600 0.0010 3.5004
↓ ↓ ↓ ↓ ↓
82 -6.1549 0 0.301 3.0767
↓ ↓ ↓ ↓ ↓
141 -5.9717 0 0.301 2.8587
↓ ↓ ↓ ↓ ↓
171 -7.2414 0 0.1139 3.4463
↓ ↓ ↓ ↓ ↓
200 -7.2694 0 0.0569 3.4107
1.3 Mathematical Model [12]
The experimental Independent variables were reduced by
evaluating dimensionless pi terms by Buckingham pi theorem
and a mathematical equation was generated by traditional
method to predict the experimental findings. The equation is
as shown.
ω T = 1.288 ( I/RT2
)-0.46
(ME)-0.87
(G)0.40
2. DETECTION OF DILEMMA AND ELECTED
APPROACH
The plot has been drawn to observe the prediction of
experimental evidences by the traditional empirical model.
The figure 2 evaluates and compares the results. Figure 2
undoubtedly express the inaccuracy in prediction of
experimental evidences by traditional mathematical model.
The existing equation is deficient to predict the desired
experimental findings. Consequently we find the need of
another modelling technique which can predict the evidences
more accurately.
The figure 3 highlights the percentage error in prediction
which is of the order of 20% to 80%. In view of the fact that
independent variables involved are high in numbers, it
becomes a very tedious and inaccurate work to build an
accurate mathematical model. ANN i.e. Artificial Neural
Network has shown its strength in the field of learning and
prediction [6] [7] of the desired results when the input
variables are high in numbers. On the other hand ANN model
can also give more reliable solution with unseen data.
Fig - 2: Prediction of Experimental Findings through
Mathematical Model

__________________________________________________________________________________________
Fig - 3: Percentage Error in Prediction of Experimental
Findings through Empirical Model
3. DECIDING THE SEQUENCE OF EXECUTION
OF PROPOSED ANN MODELLING PROCEDURE
Modelling a system through ANN simulation [9] involves use
of ANN parameters appropriately. A topology is nothing but
the complete architecture of network formed through the use
of ANN parameters. The ANN parameters should be varied
systematically in an attempt to identify best topology for a
specified problem. The number of layers was restricted to two
as the variables involved were high in number. A table for
evaluation of modelling technique is formed [5] as below. The
shaded column indicates the variation of that particular
parameter and shaded row shows the value of that parameter.
Table -3: Sequence of variation of ANN Parameters
Trai
nin
g
Nu
mb
er
Hid
den
lay
er
Siz
e
Type of
Training
Function
Perf
orma
nce
Func
tion
Types of
transfer
function
Type of
Learnin
g
Algorith
m
Layer
1 Layer2
T1 20 trainlm mse tansig purelin learngd
T10 500 trainb mse tansig purelin learngd
T11 500 trainbfg mse tansig purelin learngd
T13 500 trainbr mse tansig purelin learngd
T14 500 traingdm mse tansig purelin learngd
T15 500 traingb mse tansig purelin learngd
T16 500 traincgf mse tansig purelin learngd
T17 500 traincgp mse tansig purelin learngd
T19 500 trainlm mae tansig purelin learngd
T20 500 trainlm sse tansig purelin learngd
T21 500 trainlm mae tansig purelin learngd
T22 500 trainlm mae logsig purelin learngd
T23 500 trainlm mae tansig logsig learngd
T24 500 trainlm mae tansig Purelin
Learnco
n
T25 500 trainlm mae tansig Purelin Learngd
T26 500 trainlm mae tansig Purelin Learnh
T27 500 trainlm mae tansig Purelin Learnk
4. ILUSTRATION OF RESULTS
The graphs for each program are generated which illustrate the
effect of variation of each parameter on prediction of model.
The percentage error in prediction is also plotted to compare
and select the best of the topology amongst these topologies.
Fig - 4: Neural response with 20 Neurons
Fig - 5: Percentage error in predication with 20 Neurons

__________________________________________________________________________________________

__________________________________________________________________________________________
Fig - 21: Percentage error in predication with with 700
Neurons

__________________________________________________________________________________________
Fig - 22: Neural response with training Function “trainb”
Fig - 23: Percentage error in predication with training
Function “trainb”
Fig - 24: Neural response with training Function “trainbfg”
Function “trainbfg”
Fig - 26: Neural response with training Function “trainlm”
Function “trainlm”
Fig - 28: Neural response with training Function “trainbr”
Function “trainbr”

__________________________________________________________________________________________
Fig - 30: Neural response with training Function “traingdm”
Function “traingdm”
Fig - 32: Neural response with training Function “traincgb”
Function “traincgb”
Fig - 34: Neural response with training Function “traincgf”
Function “traincgf”
Fig - 36: Neural response with training Function “traincgp”
Function “traincgp”

__________________________________________________________________________________________
Fig - 38: Neural response with performance Function “mse”
Fig - 39: Percentage error in predication with performance
Function “mse”
Fig - 40: Neural response with performance Function “mae”
Function “mae”
Fig - 42: Neural response with performance Function “sse”
Function “sse”
Fig - 44: Neural response with transfer Function “tansig,
purelin”
Fig - 45: Percentage error in predication with transfer
Function “tansig, purelin”

__________________________________________________________________________________________
Fig - 46: Neural response with transfer Function “logsig,
purelin”
Fig - 47:Percentage error in predication with transfer Function
“logsig, purelin”
Fig - 48: Neural response with transfer Function “tansig,
logsig”
Fig - 49: Percentage error in predication with transfer
Function “tansig, logsig”
Fig - 50: Neural response with learning Function “learncon”
Fig - 51:Percentage error in predication with learning
Function “learncon”
Fig - 52: Neural response with learning Function “learngd”
Fig - 53: Percentage error in predication with learning
Function “learngd”

__________________________________________________________________________________________
Fig - 54: Neural response with learning Function “learnh”
Fig - 55: Percentage error in predication with learning
Function “learnh”
5. DISCUSSION OF RESULTS
5.1 Effect of Variation of Number of Neuron in
Hidden Layer
• Figure 4 to 23 shows that increase of number of
neuron gives better prediction.
• But as the number goes to 700 it takes very large time
to train the network
• Hence Layer size is limited to 500.
5.2 Effect of Variation Training Styles
• Figure 24 to 37 gives the results when training styles
are changed.
• The results are too bad with “trainb” & “trainc”
• As a result it can be concluded that training styles
affect the performance of network to great extent.
• It is been observed that back propagation training
functions are better for fitting function and the
performance seems superior with “trainlm”
5.3 Effect of Variation of Performance Function
• Figure 38 to 43 displays the results with change of
performance function
• The change of performance function has shown little
effect on prediction
• Performance function “mae” has shown better results
for this case.
5.4 Effect of Variation Transfer Function to Hidden
Layer
• Figure 44 to 49 shows the outcome when transfer
functions to hidden layers were changed.
• There are too many transfer functions to use and
network has two hidden layers. Hence various
combinations of these transfer functions were use to
see the effect on performance.
• It is been observed that the outer most layer with
linear transfer function “purelin” gives better results
• The combination of “tansig, purelin” transfer
functions is found best for this case.
5.5 Effect of Variation of Learning Function
• Figure 50 to 55 put view on the results with variation
of learning function.
• It clearly shows the is very mild effect of variation
learning function on performance of network
• Out of those learning functions “learncon” was better
CONCLUSIONS
The optimization methodology adopted is unique and
rigorously derives the most optimum solution for field data
available for Human Powered Flywheel Motor. The effect on
prediction of network is observed very consciously with
variation each ANN parameter.
It could also be concluded from the above results that for a
fitting function type of situation “purelin” is the best transfer
function at the outermost layer of the network. Deciding
numbers of neuron in a layer are also crucial for performance
of the network and this number depends on number of
independent variable counting for the output.
The learning algorithm is also playing a vital role in
performance of the network. Hence precise learning algorithm
should be identified for a particular field problem. If compared
with empirical model prediction ANN is predicting much
better and the error is reduced to 20%.
FUTURE SCOPE
• Prediction of ANN model may be compared with the
available empirical model.
• The best ANN model may be further used to develop
another mathematical model.
• The ANN model could be validated through unseen data.
• This mathematical model then could be utilized to
develop a physical controller.

__________________________________________________________________________________________
REFERENCES
[1]. .Modak J. P. and Askhedkar R. D. “Hypothesis for the
extrusion of lime flash sand brick using a manually driven
Brick making machine”, Bulding Research and Information
U.K., V22,NI, Pp 47-54, 1994
[2]. .Modak J. P. and Bapat A. R. “Manually driven flywheel
motor operates wood turning machine”, Contepory
Ergonomics, Proc. Ergonomics Society annual convension13-
16April, Edinburg, Scotland, Pp 352-357, 1993.
[3]. Sohoni V. V., Aware H. V. and Modak J. P. “ Manual
Manufacture of Keyed Bricks”, Building Research and
Information UK, Vol 25, N6, 1997, 354-364.
[4]. Modak J. P.”Design and Development of Manually
Energized Process Machines having Relevance to
Village/Agriculture and other productive operations,
Application of manually energized flywheel motor for cutting
of wood strip”, Human Power, send for Publications.
[5]. H. Schenck Junior “Theory of Engineering
Experimentation”, MC Graw Hill, New York.
[6]. A. R. Lende, “Modelling of pedal driven flywheel motor
by use of ANN”, M. Tech. Thesis, PCE, Nagpur
[7]. S. N. Shvanandam, “Introduction to Neural Network using
Matlab 6.0”, McGraw Hill publisher.
[8]. Stamtios V. Kartaplopoulos , Understanding Neural
Networks and Fuzzy Logics, IEEE Press
[9]. Neural Network Toolbox TM 7 User’s Guide R2010a,
Mathworks.com
[10]. Rudra Pratap, “Getting Started with Matlab7,” Oxford,
First Indian Edition 2006.
[11]. A. R. Bapat, “Experimental Optimization of a manually
driven flywheel motor”, M.E. Thesis, VNIT, Nagpur.
[12]. A. R. Bapat, “Experimentation of Generalized
experimental model for a manually droven flywheel motor”,
PhD Thesis, VNIT, Nagpur.
BIOGRAPHIES
Ms. A. R. Lende, Ex. Asst. Professor, MIT,
Kothrud, Pune, MH, India. Ex. Asst. Professor,
DMIETR, Wardha, MH, India. Qualification: BE.
(Mech. Engg.), M.Tech.(Mech. Engg. Design)
Ph.D. (Pursuing)
Dr. J. P. Modak, Emeritus Professor and Dean
(R&D), PCE, Nagpur, MH, India

Modelling & simulation of human powered flywheel

More Related Content

Viewers also liked (20)

Similar to Modelling & simulation of human powered flywheel (20)

More from eSAT Publishing House (20)

Recently uploaded (20)

Modelling & simulation of human powered flywheel