Upload
hilary-hester-daniel
View
216
Download
1
Tags:
Embed Size (px)
Citation preview
RMU_Summer2005_Samanta 1
Applications of Computational Intelligence Techniques in Engineering
B SamantaInternational Visiting Professor
Robert Morris University
RMU_Summer2005_Samanta 2
Presentation Summary Motivation Computational Intelligence Different CI techniques Applications of CI techniques Recent Work Work done at RMU Way forward Conclusions
RMU_Summer2005_Samanta 3
Motivation Use of computers for better understanding and
interpretation of process/system behavior Use of available information to obtain input-
output mapping. Utilization of expert/operator knowledge Ability to use imprecise, uncertain information Integration of knowledge over multiple disciplines Automated machine learning inspired from nature
(neuroscience, genetics, behavioral science) Development of models for optimizing the system
performance satisfying the inherent system/process constraints.
RMU_Summer2005_Samanta 4
Computational Intelligence (CI)
Intelligence built in computer programs
Covers Evolutionary computing Fuzzy computing Neuro-computing
Also known as Soft computing
RMU_Summer2005_Samanta 5
CI Techniques Artificial Intelligence (AI)
Artificial Neural Networks (ANNs) Fuzzy Logic (FL) Support Vector Machines (SVM) Self Organizing Maps (SOM)- unsupervised
Genetic Algorithm (GA) Genetic Programming (GP) Swarm Intelligence/Particle Swarm
Optimization (PSO)
RMU_Summer2005_Samanta 6
CI Techniques (contd.)
ANNs Multi-layer Perceptron (MLP) Radial Basis Function (RBF) Probabilistic Neural Network (PNN)
Fuzzy Logic + ANN Adaptive neuro-fuzzy inference
system (ANFIS)
RMU_Summer2005_Samanta 7
CI Techniques (contd.)
ANN structure Input layer Hidden Layer (s) Output layer Number of nodes in each layer Functions and their parameters
Mostly decided on trial and error basis
RMU_Summer2005_Samanta 8
ANN- a typical example
x1
x2
xN
u1
u2
uQ
y1
y2
yM
.
. ..
.
.
Input layer Hidden layer
RMU_Summer2005_Samanta 9
Fuzzy Logic
Steps involved Fuzzification using membership
functions (MFs)-input Generation of rule base Aggregation Defuzzification using MFs -output
RMU_Summer2005_Samanta 10
Fuzzy Logic (contd.)
Input and output MFs Number Type Parameters
Rule base (experience guided)
RMU_Summer2005_Samanta 11
Neuro-Fuzzy System
Combines the advantages of fuzzy logic (FL) and ANNs
Starts with an initial FL structure Uses ANN for adapting the FL (MF)
parameters and the rule base to the training data
RMU_Summer2005_Samanta 12
Fuzzy Logic – An Example
ANFIS structure for an example system with 2 inputs and 1 output.
RMU_Summer2005_Samanta 14
Genetic Algorithms Construction of genome (individual) Generation of initial population (group of
individuals) Evaluation of individuals Selection of individuals based on criteria Generation of new individuals
Mutation Crossover
Repetition of the process - generation, evaluation, selection
Termination of the process based on max generation no. and/or performance criteria
RMU_Summer2005_Samanta 15
Combinations Combine advantages of GA and other classifiers GA and ANN GA and ANFIS GA and SVM for automatic selection of classifier structure and
parameters ANNs -Number of neurons in hidden layer ANFIS - Number of MFs and their parameters SVM – SVM parameters
Selection of most important system features from a pool Selection of most important sensors (in the context of on-
line condition monitoring and diagnostics)- sensor fusion.
RMU_Summer2005_Samanta 16
Signal Conditioning and Data Acquisition
Feature Extraction
Training Data Set Test Data Set
Training of ANN/ SVM
Is ANN/ SVM Training
Complete ?
No
Yes
ANN / SVM Output
Machine Condition Diagnosis
Trained ANN/ SVM with selected features
Fig. 1. Flow chart of diagnostic procedure
GA based selection of features and parameters
Is GA based selection
over?
Yes
No
Rotating Machine with Sensors
RMU_Summer2005_Samanta 17
Genetic Programming (GP) GP – a branch of GA with a lot of
similarities. Main difference of GP and GA is in the
representation of the solution. In GA, the output is in form of a string of
numbers representing the solution. GP produces a computer program in form
of a tree-based structure relating the inputs (leaves) the mathematical functions (nodes) and the output (root node).
RMU_Summer2005_Samanta 18
GP output –An Example Terminals (leaves): inputs x1, x2 and constant 3 Nodes: Math functions *,+, exp Output: x1*x2+exp(3)
X1X2
times
plus
exp
3
(+ (* (X1 X2))(exp(3))
RMU_Summer2005_Samanta 19
Applications Computer Science
Pattern Recognition (PR) Data Mining Knowledge Discovery/ Machine Learning Feature Extraction and Selection
Mechanical Systems Condition monitoring and diagnostics Multiobjective optimization in design Control System Design
Manufacturing Systems Development of data-driven models Multiobjective optimization of machining parameters
RMU_Summer2005_Samanta 20
Applications (contd.) Engineering Management/IE
Inventory management Project selection Facility layout design Scheduling
Medicine Patient condition monitoring and diagnosis
Social Science Business
Market analysis and forecasting Credit rating
RMU_Summer2005_Samanta 21
Recent Work Machine Condition Monitoring and
Diagnostics using ANNs-MLP, RBF, PNN SVM ANFIS GA-ANN GA-ANFIS GA-SVM GP
Involving signal processing, feature extraction, selection and sensor fusion
RMU_Summer2005_Samanta 22
Recent work (contd.)
Materials ANN based estimation of fatigue life Modeling of material properties in
terms of heat treatment parameters Rotordynamics Control System Design
RMU_Summer2005_Samanta 23
Work done at RMU Intelligent Manufacturing Systems Development of Tool Wear Model
ANFIS and GA-ANFIS Genetic Programming (GP)
Development of machined surface roughness model ANFIS and GA Genetic Programming (GP)
Mutliobjective optimization of machining parameters Minimization of machining cost Minimization of surface roughness Minimization of production time Subject to constraints on
Operating parameters –speed, feed, depth of cut Cutting Force Power consumption
Tested on 5 different data sets Involves different machining operations
Milling, turning and Turning of hard material (>Rc 65)
RMU_Summer2005_Samanta 24
Tool Wear Model Mapping of Inputs and Outputs
Inputs Tool type- geometry, material Work piece Cutting speed (V) Feed rate (f) Depth of cut (d) Vibration (Vx, Vy, Vz) Forces (Fx, Fy, Fz) Cutting Time (t)
Outputs Tool wear Remaining Tool Life
GA/GP based selection of characteristic inputs
RMU_Summer2005_Samanta 25
ANFIS based Tool Wear Model – An Example Input pool
Spindle speed (x1) Feed rate (x2) Machining time (x3) Ratio of forces in 2 directions: Fx (feed)/ Fz (tangential) (x4)
Output – Tool wear level Data set
Training – 25 Test - 38
Number of MFs - 2 Performance –
Training Root Mean Square Error (RMSE) 1.30% Test data set RMSE : 8.52% Training time 0.34 s
RMU_Summer2005_Samanta 26
0 5 10 15 20 25-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
index i
Nor
mal
ized
Too
l Life
Fig. 1. Results of training data set
Actual
PredictedPrediction error
RMU_Summer2005_Samanta 27
0 5 10 15 20 25 30 35 40-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1.2
index i
Nor
mal
ized
Too
l Life
Fig. 2. Results of test data set
Actual
PredictedPrediction error
RMU_Summer2005_Samanta 28
GA-ANFIS based roughness model – An Example Input pool
Spindle speed (x1) Feed rate (x2) Depth of cut (x3) Vibration in 3 directions
x (radial) (x4) y (tangential) (x5) z (feed) (x6)
Output – surface roughness Data set
Training – 36 Test - 24
GA based selection of best 3 features: x2, x1, x5 Number of optimum MFs - 2 Performance –
Training Root Mean Square Error (RMSE) 2.60% Test data set RMSE : 6.65% Training time 263.2 s
RMU_Summer2005_Samanta 30
0 5 10 15 20 25 30 35 40-0.1
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
index i
Nor
mal
ized
too
l fla
nk w
ear
Fig. 1. Results of training data set
Actual
PredictedPrediction error
RMU_Summer2005_Samanta 31
0 5 10 15 20 25-0.2
0
0.2
0.4
0.6
0.8
1
1.2
index i
Nor
mal
ized
too
l fla
nk w
ear
Fig. 2. Results of test data set
Actual
PredictedPrediction error
RMU_Summer2005_Samanta 32
GP model for surface roughness
GP was used for same data sets Training – 36 Test set – 24
Performance Training RMSE: 3.79% Test RMSE : 6.90% Training time: 463.7 s
RMU_Summer2005_Samanta 33
X2
X2
X3
X2
X1
X2
step
asin
plus
power
X4
X3
step
asin
plus
powerX4
divide
log10
tanh avg
X2
acos
log
X3
X2
X4
X3
step
asin
plus
powerX3
divide
log10
tanh
avg
power
asin power
exp
sqrt
power
exp
sqrt
power
GP output tree for Roughness model
RMU_Summer2005_Samanta 34
Publications Planned Predictive modeling of tool wear in
turning using adaptive neuro-fuzzy inference system
Modeling and prediction of tool wear in turning using genetic programming
Predictive modeling of surface roughness in turning using adaptive neuro-fuzzy inference system and genetic algorithms
RMU_Summer2005_Samanta 35
Publications Planned (contd.) Modeling and prediction of surface
roughness in turning using genetic programming
Predictive modeling of surface roughness in milling using adaptive neuro-fuzzy inference system and genetic algorithms
Multiobjective evolutionary optimization of a machining process
RMU_Summer2005_Samanta 36
Conferences/Journals North American Manufacturing Research
Conference (NAMRC 34 ), NAMRI/SME, May 23-26, 2006, Milawukee, WI, USA.
Flexible Automation and Intelligent Manufacturing (FAIM) June 26-28, 2006, Univ of Limerick, Ireland.
IFAC Symposium on Information Control in Manufacturing (INCOM) May17-19, 2006, France.
Journal of Manufacturing Systems/SME International Journal of Machine Tools &
Manufacture
RMU_Summer2005_Samanta 37
Industry-RMU collaboration Potential Interest in RMU-EOC research
collaboration in the area of Laser machining.
Development of machining models using CI Multiobjective constrained optimization of
machining/laser system parameters Sensor fusion
Interest in RMU-ExOne research collaboration in the areas of 3D printing
process system Design optimization
RMU_Summer2005_Samanta 38
Way Forward
Scope for further collaboration with RMU
Teaching – Development of new elective or short courses in consultation with Faculty
Research – Joint supervision of projects/theses at Senior, MS and PhD levels
Collaborative work with Faculty Outreach- Industry and Government
supported research projects/contracts
RMU_Summer2005_Samanta 39
Conclusions
Increasing popularity of CI techniques Integrating capability over multiple
disciplines Capability of incorporating imprecision
and uncertainty Suitability for hard-to-model processes
/systems Better alternatives to traditional hard
computing scenario