Vibration Based FuzzyNeural System for Structural Health Monitoring

1
Vibration Based Fuzzy-Neural System for
Structural Health Monitoring

• Lakshmanan Meyyappan (Laks)

2
Vibration Based Fuzzy-Neural System for
Structural Health Monitoring

• Objective
• Overall System
• Experimental Setup
• Damage Detection
• Fuzzy Logic Decision System
• Neural Network Prediction System
• Research Findings
• Future Work

3
Objectives

• The main goal is to develop a practical
  real-time structural health monitoring system
  using smart systems engineering concepts and
  tools.

4
2. Overall System
5
3. Experimental Setup

• Data Type
• Data Acquisition Method
• Experiment
• Data Analysis

6
3.1 Data Type

• Acoustic or ultrasonic
• X-radiography
• Eddy-currents
• Microwaves
• Magnetic fields
• Thermal fields
• Vibration Signatures
• Many more

7
3.1.1 Vibration Signatures

• It must be a non-destructive evaluation (NDE)
  technique
• It should be able to monitor the structure on a
  global basis
• It must not cause any disruption to the normal
  operation of the structure
• The equipment used for collecting data should not
  be bulky, must be easy to replace and must be
  reliable
• Initial and operating costs should not be very
  high
• It must be sensitive and must be able to detect
  small damages

8
3.1.1 Vibration Signatures

• The basic premise of vibration based structural
  health monitoring is that damage in the structure
  or change in its physical properties (i.e.,
  stiffness, mass and/or damping) will, in turn,
  alter the dynamic response of the system

9
3.1.1 Vibration Signatures

• Advantages
• NDE Technique
• Global Analysis
• Normal Operation of the Structure
• Small
• Reliable
• Less Expensive (both initial and operating costs)
• Sensitive

• Disadvantages
• Unsupervised Learning Mode
• Data Accuracy (Potential problem with any type of
  data)

10
3.2 Data Acquisition Methods

• The Structure has to be excited by some means to
  collect vibration data
• Measured Input Excitation Methods
• Step Relaxations
• Shakers
• Ambient Excitation Methods
• Test Vehicle Method

11
3.2.1 Test Vehicle Method

• Advantages
• Realistic
• Easy to collect (not as tedious as the measured
  input excitation methods)
• Does not cause damage to the structure
• Normal operation of structure is possible 3.1.1
  Vibration Signatures
• Disadvantages
• The collected data depends on the type of test
  vehicle, traffic on the bridge, speed of the
  vehicle

12
3.3 Experiment

• Teardrop
• Bridge

13
3.4 Data Analysis

• Pre-Processing
• Feature Selection
• Signal processing
• Data Cleansing

14
3.4.1 Pre-Processing
15
3.4.2 Feature Selection

• Feature selection involves condensing the data
  and finding out a distinguished feature that can
  be used as an input to a tool, which would
  analyze the data and detect damage.
• Peak
• Mean
• Median
• Mode

16
3.4.3 Signal Processing

• Processing the raw data to get some useful
  information (Also reduces the dimension of the
  data)
• Fourier Transform
• Sampling and Quantization
• Power Spectrum

17
3.4.3 Signal Processing
18
3.4.4 Data Cleansing

• Data cleansing is the process of selectively
  choosing data to accept for, or reject from the
  feature selection process.
• Power Spectrum Peak Values

19
4. Damage Detection

• Interesting Facts on Extensive Analysis
• Vibration Value ( Power Spectrum Peaks) vary with
  speed and mass of the vehicle, and few other
  parameters like the temperature.
• All these relations are non linear
• Even though vibration varies with a number of
  parameters, the pattern, that is, its relation
  with a different member remains the same.

20
4. Damage Detection

• For simplicity of explanation the data collected
  with the sensors attached to the above five
  locations are used.

21
4. Damage Detection
22
4. Damage Detection

• Relationship between the members remains
  the same that is member 3 has the highest power
  spectrum value in all of the above cases followed
  by member 1, 5, 4 and 2 respectively

23
4. Damage Detection

• At all different conditions the relationship
  between the members remains the same that is
  member 3 has the highest power spectrum value in
  all of the above cases followed by member 1, 5, 4
  and 2 respectively
• This is true for any number of members
• Advantage
• Independent of all other conditions like speed,
  mass, type of the vehicle.

24
Why Fuzzy Logic Neural Networks

• Fuzzy logic and neural networks are used to make
  the predictions because of their inherent
  robustness and their abilities to handle
  nonlinearities and uncertainties in structural
  behavior
• Conventional decision making systems and control
  systems rely on the accuracy of the modeling of
  the system dynamics. A fuzzy system does not
  require accurate dynamic modeling

25
5. Fuzzy Logic Decision System

• Goal To take power spectrum values of various
  members as input and predict a possible damage
• Method Fuzzy Ranking System

26
5.1 Fuzzy Ranking System

• 101 Membership functions
• Triangular Membership functions

27
5.1 Fuzzy Ranking System

• Zadeh Operators
• AND gt use the minimum of the options.
• OR gt use the maximum of the options
• NOT gt use 1-option
• Zadeh Implications
• AND For combining two different input classes.
• OR When choosing the membership value that
  belongs to an output class.
• NOT While combining two incompatible input
  classes.

28
5.1 Fuzzy Ranking System

• Fuzzy Ranking based on Fuzzy Integral values
  calculated using the formula
• where a, b, c are the vertices of the triangular
  membership functions
• Alpha is the index of optimism and it varies
  between 0 and 1

29
5. Fuzzy Logic Decision System

• Output
• The bridge is perfect
• The members 2, 3 may be damaged
• This output is fed to the neural networks.

30
6. Neural Network Prediction System

• Goal To make the final prediction on the
  condition of the bridge
• Inputs
• Fuzzy logic system output
• Speed of the vehicle ( Speed Gun output)

31
6. Neural Network Prediction System

• Input 100 Data Points (speed)
• Target 100 Data Points (Power
  Spectrum Peak Value)
• Algorithm Back Propagation (LM Method)
• Layers 2 Layers 15 1
• Transfer
• Functions Tansig Purelin
• Error Rate 1e-8
• Max Epochs 1500

32
6. Neural Network Prediction System
33
6. Neural Network Output

• The Member N is Fine
• The Member N has Small Damage
• The Member N has Medium Damage
• The Member N has Large Damage

34
7. Research Findings

• Real-time Health Monitoring
• Global Monitoring
• NDE Technique
• Simple, Low Cost and Reliable

35
8. Future Work

• Finite Element Analysis for simulating damaged
  member data
• Fuzzy Clustering instead of Fuzzy Ranking