ANN Seminar
-
Upload
kiran-hegde -
Category
Documents
-
view
216 -
download
0
Transcript of ANN Seminar
-
8/8/2019 ANN Seminar
1/26
ARTIFICIAL NEURAL NETWORKS FORSTRUCTURAL DAMAGE DETECTION USING
MODAL DATA
Presented By :
KIRAN H
VII Sem. C.E.
-
8/8/2019 ANN Seminar
2/26
Seminar Contents
1. Introduction
2. Modal analysis
3. Damage detection process
3. Design of ANN.
4. Damage detection in cantilever plate
5. Training data for ANN
6. Test/Validation data
7. Damage detection scenarios
24. Applications of ANN
25. Conclusion
Slide no.
-
8/8/2019 ANN Seminar
3/26
Biological Neurons
Neuron : computational unit in the nervous system. It has:
(i) Dendrites (inputs) Neuron receives input from other neurons
(ii) Cell body An electrical pulse that travels from the body, down the
axon (iii) Axon (output) Touch the dendrites or cell body of the next
neuron.
-
8/8/2019 ANN Seminar
4/26
Artificial Neurons
Multiple inputs and a single output.
The neurons can be trained to fire (or not), for particularinput patterns.
Inputs associated with weights. These weights correspond to synaptic efficacy in a
biological neuron.
Each neuron also has a single threshold value.
The weighted sum of the inputs is formed, and the thresholdsubtracted, to compose the activation function.
Note also that weights can be negative, which implies thatthe synapse has an inhibitory rather than excitatory effecton the neuron
-
8/8/2019 ANN Seminar
5/26
Biological andArtificial Neurons
Human neuron A simple neuron with weight
-
8/8/2019 ANN Seminar
6/26
Architectureofneuralnetworks
Feed-forward networks Feedback networks
-
8/8/2019 ANN Seminar
7/26
Structural damage detection
Assess the condition of structures.
Gives improved understanding about structures capacity
and typical performance during its service.
Can be found out by monitoring change in structural
responses, natural frequencies and mode shapes and strain
mode shapes.
These direct methods cant locate and quantify damage.
Modal analysis can give the location and magnitude ofdamage.
-
8/8/2019 ANN Seminar
8/26
Modal analysis
The modal vibration test data is used.
The natural frequencies and mode shapes can be considered
to represent the state of the structure.
Modal parameters depend only on the mechanical
characteristics of the structure and not on the excitation
applied.
Theoretically the structure can be represented by
measurements taken at a single location. Reduction in time and cost of performing damage
monitoring and predictive maintenance.
-
8/8/2019 ANN Seminar
9/26
Damage Detection Process
Every structure has its own natural vibration (or resonance)
frequencies.
Each vibration mode has a different energy distribution.
Any localized damage will affect each mode differently
depending on the location and severity of the damage.
Modal parameters are sensitive to boundary conditions, i.e.,
physical constraints of the structure.
Damage will soften the structure and thus modify itsdynamic characteristics such as frequency and mode shape.
-
8/8/2019 ANN Seminar
10/26
Damage Detection Process
-
8/8/2019 ANN Seminar
11/26
Artificial Neural Network
The feed-forward, multilayered, supervised neural network
with the error back propagation algorithm is used.
Two stages.1st is data feed forward.
The output of each node is defined as
netj =ij Oi + j
Oj = f( netj )
Threshold function is
{1 x>1
f(x) = {x -1x1
{-1 x
-
8/8/2019 ANN Seminar
12/26
Artificial Neural Network
The 2nd stage is error back propagation and weights
adjustment.
The system error is defined as E
E= (1/2P)* p=1P
n=1N (dpn opn)
P - number of instances in the training set,
dpn ,opn - desired and calculated output of the nth output node
for the pth instance, respectively
-
8/8/2019 ANN Seminar
13/26
Damage Detection in Cantilever plate
Cantilever steel plate properties
Cross-section of 50.75 x 6.0 mm2 and mass density of 7670
kg/m3.
The first 6 natural frequencies for the intact and damaged
configurations for this cantilever plate are taken from
literature.
The damage configuration was introduced by saw cuts at
the clamped end (Cut-1) and the mid-span (Cut-2) of theplate.
The depth (d) of the cut is 20 mm.
-
8/8/2019 ANN Seminar
14/26
Damage Detection in Cantilever plate
Damage made by cuts at locations 1 and 5 for the cantilever
plate
Division into 9 components for simulation of damage in
FEA
-
8/8/2019 ANN Seminar
15/26
Measured and computed (FEA) frequencies.
Frequencies, Hz
Case Mode 1 Mode 2 Mode 3 Mode 4 Mode 5 Mode 6
Measured(undamaged)
5.67 35.58 100.33 196.38 319.50 485.71
Measured
(damaged)
5.11 31.60 94.44 179.97 308.89 452.71
FEA(undamaged)
5.68 35.61 99.70 195.36 322.90 482.27
-
8/8/2019 ANN Seminar
16/26
The Neural Network Architecture
A feed forward back-propagation neural network has been
used.
In input stage, 6 processing elements in the input layer
representing a vector of the first 6 natural frequencies of thebeam.
In output stage,10 PEs for the output layer, where the first
node consists of damage magnitude and the remaining 9
nodes represent the damage indicators at each of 9 locations. The damage indicators '0' and '1' represent the undamaged
and fully damaged conditions, respectively.
-
8/8/2019 ANN Seminar
17/26
The Neural Network Architecture
In the hidden layer design ,18 PEs in single hidden layer
network was chosen as it results in less RMS error.
-
8/8/2019 ANN Seminar
18/26
Training Data
To obtain training data finite element (FE) model has been
generated for the cantilever plate with 36 finite elements.
The damage was defined as the fractional loss of the second
moment of area over one location.
The data was generated for a wide range of damage
magnitudes of 1%, 5%, 10%, 15%, 20%, 25% and 30%.
Hence, the database consists of training sets corresponding
to 7 damage states in each of 9 locations of the cantileverbeam. This results in 63 training sets.
-
8/8/2019 ANN Seminar
19/26
Training Data For multiple damage scenario, namely, simultaneous
damage in two locations, training data has been generated
for 1%, 10% and 20% damage magnitudes in each of two
damage locations simultaneously which resulted in 144training sets.
-
8/8/2019 ANN Seminar
20/26
Test /Validation Data For validating the network, a test data set has been
generated for3 damage magnitudes of3%, 13% and 23%.
This test database consist of data pertaining to 3 damage
levels for each of 9 damage locations which resulted in 27test sets for single damage scenario.
For validating multiple damage state, test data set
corresponding to the above three damage magnitudes in
three selected pairs of damage locations, namely, at (1,3),(1,5) and (1,7), have been generated which resulted in 9 test
sets.
-
8/8/2019 ANN Seminar
21/26
Neural Network Training
The learning rate parameter () for the network is chosen as
0.01 and the error tolerance is chosen as 0.01 after trials.
Fractional change in resonant frequencies used to get
accurate results.
zi = (fui fdi ) / fui
zi is the fractional change in the i-th mode,
fu and fd are the frequencies of undamaged and damaged
structure, respectively.
-
8/8/2019 ANN Seminar
22/26
Damage detection using ANN
-
8/8/2019 ANN Seminar
23/26
Multiple damage detection
-
8/8/2019 ANN Seminar
24/26
Applications for Neural Networks
Detection of medical phenomena
Stock market prediction
Credit assignment
Monitoring the condition of machinery. Engine management
In marketing
Facial recognition
Robotics
-
8/8/2019 ANN Seminar
25/26
Conclusion
The detection and identification of structural damage is a
vital part of the monitoring and servicing of structural
systems during their lifetime.
It was observed that prediction capability of ANN is betterby using fractional difference of frequencies instead of
using the frequencies.
The results show that the ANN method is capable of
predicting the location and magnitude of damage with goodaccuracy for single and multiple damage scenarios.
-
8/8/2019 ANN Seminar
26/26