This ppt was presented to defend my thesis for the completion of Master of Engineering in Production Engineering.

1. APPLICATION OF STATISTICAL LEARNING TECHNIQUES AS PREDICTIVE TOOLS
FOR MACHINING PROCESSES
by
Shibaprasad Bhattacharya
ROLL: 001911702013
Under the guidance of
Prof. Shankar Chakraborty
Department of Production Engineering

2. ACKNOWLEDGEM
ENT
Eternally grateful to Prof. Shankar Chakraborty

3. Contents
Objectives
Importance
Introduction
Statistical learning techniques as predictive tools
Conclusion
Future scope of Work

4. Objectives
Integration of Statistical Learning Techniques for Machining Processes
Incorporating non-parametric methods to bypass the rigid assumptions of
parametric methods
Laying down a framework for selecting the right model keeping in mind
different tradeoffs
Selection of right values of parameters for different models

5. Importance
Better analysis
Close control
Understanding of the process parameters
Competitive edge
Satisfying customer demands

6. Introduction
What is Statistical Learning?

8. Difference between Statistical Learning and Machine
Learning
Different buzz words: Statistics/Operations
Research/Mathematics vs Computer Science
Size of the dataset
Number of variables
Why and What vs How

9. Steps involved in Statistical Learning
Selecting the dataset
Dividing it for training and testing
Training the model with the training data
Testing it with the testing data
Validating the testing results with error estimators
Selection/Rejection of the used model

11. Statistical Learning Techniques
Supervised Unsupervised Reinforcement
Linear Regression
KNN
Trees
Naïve-Bayes
Logistic Regression
Clustering
PCA
LDA
Apriori
FP-growth
Hidden Markov
model

12. Fitting the data

13. Underfit
When the model fails to generalize the data
Performs poorly for both the training and testing set
Can be recognized by looking the training error
Switching to a more complex model will help solve it
Adding more features will also help

14. Overfit
When the model memorises the data too closely
It fails to identify the pattern
Performs well on Training set
Performs poorly on Testing set
Reducing the number of features will help
Adding more data can also solve this problem

15. Trade-off in Statistical Learning
Trade-off between the Bias and Variance
Trade-off between Prediction accuracy and Interpretability

16. Bias-Variance Trade-off
Bias
Introduced because of the of
the over simplification of the
model assumption
Difference between average
prediction of the build model
and the actual value it is trying
to predict
Variance
Error induced by the
randomness of the training
data
Fails to generalize the model
and hence gives higher testing
error

18. Trade-off between Predictive accuracy and Model interpretability

19. Error estimators
Mean absolute percentage error or MAPE:
Root mean squared percentage error or RMSPE:
Root mean squared logarithmic error:

20. Error estimators
Correlation coefficient or R:
Root relative squared error or RRSE:
A= Actual value
P= Predicted value
𝐴= Mean of the actual values
𝑃= Mean of the predicted values

21. Statistical Learning Techniques as Predictive tool
for Machining Processes
1. Prediction of Responses in a Dry Turning Operation: A
Comparative Analysis
2. Prediction of Responses in a CNC Milling Operation using
Random Forest Regressor
3. Predicting Responses for Turning High Strength Steel
Grade H with XGBoost

22. Prediction of Responses in a Dry Turning Operation: A Comparative Analysis

23. Experimental details
Dry turning operation using a heavy duty lathe
Input parameters: Cutting speed, Feed rate and Depth of cut
Output responses: Surface roughness, Cutting force and Material removal rate
Number of training points: 18
Number of testing points: 9
Statistical learning techniques used:
Multivariate regression analysis
Artificial neural network (ANN)
Fuzzy logic
Adaptive neuro-fuzzy inference system (ANFIS)

24. Turning parameters with their operating levels

25. Multivariate regression analysis
It generally follows the following form:
Where β0 is the Y-intercept coefficient, β1-βn are the main effect coefficients and βij is the interaction
coefficient
These coefficients are initially unknown and they are computed by fitting the data.

26. ANN

27. ANN
Designed to imitate the human behavior
The processing units/nodes/neurons are the building blocks of ANN
Type of neural network considered: Feedforward neural network
Algorithm used: Backpropagation (Levenberg-Marquardt)

28. Fuzzy Logic
Deals with imprecise information to arrive at logical conclusions
Input values are generally converted to linguistic terms like (High-Low)
A fuzzy logic unit contains a fuzzifier, membership functions, a fuzzy rule base, an inference
engine and a defuzzifier
Fuzzy logic is generally based on if-then rules like this:
Rule 1: If x1 is A1 and x2 is B1 and x3 is C1 and x4 is D1, Then output (O) is E1, else
Rule 2: If x1 is A2 and x2 is B2 and x3 is C2 and x4 is D2, Then output (O) is E2, else
Rule n: If x1 is An and x2 is Bn and x3 is Cn and x4 is Dn, Then output (O) is En.

29. Fuzzy Logic Rule Viewer

30. ANFIS

31. ANFIS Architecture

32. Comparison of different membership functions

33. Rule viewer for ANFIS

34. Comparison of actual and predicted response values for Ra

35. Comparison of actual and predicted response values for Fc

36. Comparison of actual and predicted response values for MRR

37. Performance indices for different prediction tools

38. Prediction of Reponses in a CNC Milling Operation using Random Forest
Regressor

39. Experimental details
CNC Milling Process
Input parameters: Cutting speed, Feed rate, Depth of cut and Width of cut
Output responses: Surface roughness, Material removal rate and Active energy consumption
Number of training points: 21
Number of testing points: 6
Statistical learning techniques used: Random Forest

40. Milling parameters with their operating levels

41. Random Forest
Bagging technique : Parallel ensemble method
Multiple weak learners are combined
This brings more stability
RF is made by combining multiple decision trees
The final output is yielded after taking into account of all the trees

42. Schematic diagram of a random forest

43. Default values of different parameters

44. MRR
Ra
AEC

46. A sample decision tree for MRR

47. A sample decision tree for Ra

48. A sample decision tree for AEC

50. Predicting Responses for Turning High Strength Steel Grade-H with
XGBoost

51. Experimental details
CNC Turning Process
Input parameters: Cutting speed, Feed rate, and Depth of cut
Output responses: Surface roughness and Material removal rate
Number of training points: 100
Number of testing points: 25
Statistical learning techniques used: XGBoost

52. Input parameters
Output responses

54. XGBoost
Boosting technique : Sequential ensemble method
Another tree based method
Like Random Forest, weak learners are combined
Instead of parallel method, a sequential method is used

56. Process Parameters
nrounds: The number of trees in the model.
eta: The learning rate. Range: 0 to 1.
max_depth: The greatest depth to which a tree can grow.
early_stopping_rounds: After how many rounds should
the model stop if there is no improvement in the
predictions?

57. Hyperparameter tuning

58. Parameter
Lowest
value
Highest
value
Step size
Number of
combinatio
ns
nrounds 1 1000 10 100
eta 0.2 1 0.02 41
max_depth 1 6 1 6
The total number of possible combinations from the above configuration
is: 100×41×6 = 24600
Parameter MRR Ra
nrounds 61 71
eta 0.64 0.2
max_depth 3 6

59. Sample trees for Ra

60. Sample trees for MRR

63. Statistical Index Ra MRR
MAPE 2.34 13.88
RMSPE 3 18.33
RMSLE 0.014 0.16
R 0.99 0.99
RRSE 0.035 0.115

64. Conclusions
Selecting the best technique is important by comparing
Hybrid learning approaches outclasses the counterparts
Parametric methods are flexible to model
Parametric methods work well even with smaller dataset
Non-parametric methods bypasses rigid assumptions of parametric methods
Non-parametric methods can easily accommodate all kinds of variables
Selection of the right parameter values are important for non-parametric methods

65. Future Scope of Work
Extending the current work by incorporating different types of variables
Example: Coolant type, Tool diameter etc.
Including several conditions as input variables
Example: Skill level of the worker, tool quality etc.
Include budget constraint as output variable to predict the expected cost

66. Thank You!