KE22 FINAL YEAR PROJECT PHASE 3 Modeling and Simulation of Milling Forces SIMTech Project Ryan Soon, Henry Woo, Yong Boon April 9, 2011 Confidential – Internal Only
2 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only AGENDA Objectives Problem Domain Overview System Description Models and Results Benefits to both Organization and Students Demo Q&A
3 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only OBJECTIVES Understand a prognostic problem domain that enables an Hybrid implementation of Knowledge Engineering Techniques Present research effort & implementation result of overall prognostic problem domain Highlight novel prognostic optimization concept and model Challenges and benefits
4 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only PROBLEM DOMAIN OVERVIEW KEY IDEA Optimizing manufacturing asset and predictive maintenance What is Milling? customized material of different shapes and features What to Optimize Predict remaining lifespan of cutter How to Optimize Implementing a Hybrid KE Model using –Hierarchical Clustering (HC) –Adaptive Neural Fuzzy Inferences System (ANFIS) –Resulting in an optimal HC-ANFIS hybrid Why Optimal determine optimal cluster size and automatically produce optimal ANFIS structure
5 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only SYSTEM DESCRIPTION Machine sensors attached to the milling process Cutting force sensor in x, y, z dimension Acoustic emission sensor that measure high frequency stress wave
6 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only SYSTEM DESCRIPTION 6 cutter tools’ data given Over 300+ samples given for each cutter At specific interval –Measure sensors’ readings –Measure tool wear using electronic microscope
7 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only SYSTEM DESCRIPTION ANFIS by itself can solve the prediction problem (Universal Approximator) –But required expert knowledge on rules determination and membership functions –Use HC to determine ANFIS structure and membership parameters How to determine the optimal cluster size of HC –By using cluster balance method Improve overall learning and application performance Coded HC module in.NET C# Coded ANFIS module in Python
8 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only GRID PARTITION WITH HC APPROACH
9 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only GRID PARTITION WITH HC ISSUE Complexity of the ANFIS structure is based on the product of each input’s cluster size Given that p, q, r, s represented the cluster size of the 4 force features ANFIS would generate (p * q * r * s) number of inferences rules For E.g. if p = q = r = s = 10, then number of inferences rules = 10,000! This is computationally intensive and infeasible to implement
10 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only HC-ANFIS APPROACH
11 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only HC-ANFIS APPROACH FINDINGS Lesser rules produced than the previous approach As the features were combined, much lesser ANFIS inferences rules were created thus resulting in a much lesser intensive computation and a practical solution to implement
12 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only HIERARCHICAL CLUSTERING CLUSTER BALANCE
13 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only OVERVIEW OF ANFIS ANFIS architecture Premise ANFIS MF(Bell) Consequence Linear Sugeno Learning Algorithms FWBW PremiseFixedGradient Descent ConsequenceLSEFixed
14 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only BELL MEMBERSHIP FUNCTION C = Cluster Centroid a = Standard Deviation
15 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only HC AND ANFIS ARCHITECTURES
16 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only COMPARISON OF DIFFERENT METHODS Self Training (Single Cutter Tool Training Data) MethodsAccuracyRMSE# of Rules Grid Partition HC-ANFIS SC-ANFIS Generalized Training (Two Cutter Tool Training Data) MethodsAccuracyRMSE# of Rules Grid Partition HC-ANFIS SC-ANFIS Testing (with 3 rd Cutter Tool Production Data) MethodsAccuracyRMSE# of Rules Grid Partition HC-ANFIS SC-ANFIS
17 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only BENEFITS BY ORGANIZATION HC System –Fast and customizable input selection for different application needs –Customized output, to facilitate future seamless integration between HC and other system –Novel cluster balance implementation to determine optimal HC cluster size HC-ANFIS System –Provide an alternative automated tool wear prediction method for SimTech sponsor
18 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only BENEFITS BY STUDENTS Enforce what student learned in course –Knowledge Modeling and Management Use different techniques (i.e. interview, UML diagrams) and CommonKADS to gather and capture user requirements Utilize the knowledge learned in class (i.e. Clustering, Fuzzy Inferences System and Neural Network) to come up with a Hybrid system design and final product –Product Development Understand the underlying principle and math of how Clustering, Fuzzy Inferences System and Neural Network works Explore and innovate new KE techniques Understand the importance and usage of the HC and ANFIS application in real world situation Learned from users on the proper result testing technique –Result must be repeatable and reliable
19 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only DEMO Show capability of –.NET C# HC program –Grid Partition with HC using Python –HC-ANFIS using Python –Subtractive Clustering (MATLAB)
20 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only THE END Q&A
21 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only BACKUP
22 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only PROBLEM DESCRIPTION 3 set of cutter tool data were given –07, 31, T12 Belong to the same family type but with differences in drill bit shape and knife edges Problem domain requires us to build a hybrid KE system to predict the cutter tool wear Full Microsoft.NET C# implementation of Hybrid KE system Hierarchical Clustering –Derive number of Fuzzy linguistic values for each variable –Derive number of Fuzzy rules ANFIS (Neural Fuzzy System) to learn and predict the tool wear –Generic tool wear prediction model
23 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only ACCURACY & RMSE VS CUSTER # FOR HC-ANFIS
24 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only DATA CORRELATION ANALYSIS – 1 And within each cutter tool data –3 sets of individual tool head data F1, F2, F3 Within each “F” data (315 records) –Acoustic emission data (16 features) –Force (x dimension) data (16 features) –Force (y dimension) data (16 features) Too much features –Use correlation coefficient method and cut down on the features
25 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only DATA CORRELATION ANALYSIS – 2 By using Pearson Correlation Coefficients, the linear dependence between the measured features values and the tool wear values can be calculated AE data is not influencing the tool wear strongly The top influencing features are consistent between the 3 forces AEFxFyFz
26 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only FUZZY SYSTEM IDENTIFICATION
27 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only OVERVIEW OF HIERARCHICAL CLUSTERING Agglomerative HC starts with each object describing a cluster, and then combines them into more inclusive clusters until only one cluster remains. 4 Main Steps –Construct the finest partition –Compute the distance matrix –DO Find the clusters with the closest distance Put those two clusters into one cluster Compute the distances between the new groups and the remaining groups by recalculated distance to obtain a reduced distance matrix –UNTIL all clusters are agglomerated into one group. Ward Methods, minimize ESS (Error Sum-Of-Square)
28 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only OPTIMAL HIERARCHICAL CLUSTERING Determine the numbers of clustering using RSS with penalty. Where, is the penalty factor for addition # of cluster. K’ and K = number of clusters RSS = Residual Sum of Squares Borrow concept from K-means using RSS as goal function.
29 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only HIERARCHICAL CLUSTERING + ANFIS Two Different Approaches for HC + ANFIS –Use HC to determine # of linguistic values for each input features –Use HC to determine # of rules
30 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only OPTIMAL HIERARCHICAL CLUSTERING # OF LINGUISTIC VARIABLES Example on SRE variables, opt # of cluster = 3 Perform HC on selected features on FX Variables Name# of Clusters p2p4 std_fea4 sre3 fstd4
31 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only ANFIS ARCHITECTURES # OF LINGUISTIC VARIABLES ANFIS with 4 inputs variables contains 3~4 linguistics variables generated 192 Rules!
32 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only ANFIS – RESULTS # OF LINGUISTIC VARIABLES ANFIS Predict vs Actual –Train Data with Avg Error 4.84 –Test Data with Avg Error Membership Functions –P2p –Std_fea –Sre –fstd
33 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only OPTIMAL HIERARCHICAL CLUSTERING # OF RULES Build HC on all variables, opt # of cluster = 5
34 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only ANFIS ARCHITECTURES # OF RULES ANFIS with 4 inputs variables contains 5 linguistics variables and 5 rules. Each cluster centre is a fuzzy rules!
35 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only ANFIS – RESULTS # OF RULES ANFIS Predict vs Actual –Train Data with Avg Error 5.75 –Test Data with Avg Error Membership Functions –P2p –Std_fea –Sre –fstd
36 | KE22 FYP, Modeling and Simulation of Milling Forces | Jan | Confidential – Internal Only WHAT’S NEXT? Full.NET C# Implementation Development of Hierarchical Clustering toolset with frontend GUI –Manual range input of number cluster by user –Optimal clustering suggesting the optimal number of cluster Make use of ANFIS model to evaluate –GUI engine for cluster center drawing Development of ANFIS toolset with frontend GUI –Develop the ANFIS Engine which will do the optimization –Develop User Interface for: Display predicted tool-wear result Evaluation of error