Professional Documents
Culture Documents
TEAMWORK PROJECT
AFNAN KHAN
ORKHAN GURBANLI
i
BUDAPEST UNIVERSITY OF TECHNOLOGY AND ECONOMICS
FACULTY OF MECHANICAL ENGINEERING
DEPARTMENT OF MANUFACTURING ENGINEERING
AFNAN KHAN
ORKHAN GURBANLI
TEAMWORK PROJECT
Deployment of machine learning predictive models in the
micromachining of hardened tool steels.
Budapest, 2022
iii
NOMENCLATURE:
Abbreviations Definitions
AI Artificial Intelligence
AISI American Iron and Steel Institute
ANFIS Adaptive Neuro Fuzzy Inference System
ANN Artificial Neural Network
CNC Computer Numerical Control
CNN Convolutional neural network
FC Fully Connected
GPR Gaussian Process Regression
GR Gaussian Regression
HRB Rockwell Hardness B Scale
HRC Rockwell Hardness C Scale
HV Vickers Hardness
MATLAB Matrix Laboratory
ML Machine Learning
MLP Multi-Layer Perceptron
MMP Multistage Manufacturing Process
RUL Remaining Useful Life
SVM Support Vector Machines
SVR Support Vector Regression
TCM Tool Condition Monitoring
iv
1. INTRODUCTION:
The need for precision machining grows as companies expand around the globe,
necessitating the use of various techniques to optimize manufacturing processes' efficiency.
Today, Industry 4.0 and smart manufacturing are the key production concepts. This study
will discuss thin chip removal techniques like micro-milling and ultra-precision turning, tool
condition monitoring, the use of artificial intelligence (AI) techniques to monitor tool setup
and usage, and the forecasting of thin chip removal process models. The project will also
describe studies that anticipate tool conditions depending on how tools are used. Micro-
milling and hard cutting operations present difficulties in tool condition monitoring because
of uneven wear and chipping of the cutting tools, which result in unexpected tool failure. As
a result, it necessitates a thorough analysis of all relevant studies. This project will include a
thorough description of the micro-milling and ultra-precision turning processes, as well as
many methods based on ANN for predicting tool wear and surface roughness.
2. AIMS:
Firstly, the work's main objective is to investigate tool wear condition monitoring in hard
cutting and build prediction models for the cutting tool surface roughness and lifespan in
the context of Smart Manufacturing and Industry 4.0 phenomena. Second, among other
machining techniques, thin-chip removal, which is our area of interest, will be examined and
summarized. Raw materials, structures, design and optimization methodologies,
manufacturing procedures, performances, and presenting some potential future possibilities
are explored in micro-milling material removal mechanisms. Tool Condition Monitoring
and the use of artificial intelligence in TCM are two further objectives. Finally, the
construction of prediction models using AI is covered. Even though several direct and
indirect approaches for tool wear monitoring have already been examined in various
studies, there is still no consensus on the best option for tool wear monitoring techniques in
micro-milling due to methodological flaws and adverse side effects. As a result, the ANN
application is utilized to forecast the tool life for this project while creating predictive
models. The actions listed below are carried out:
• Information on the cutting parameters (cutting speed, depth of cut, and feed rate) is
gathered.
• Model comparisons.
1
3. LITERATURE REVIEW:
This work has used more than 20 scientific articles to collect comprehensive information
regarding Smart Manufacturing, Thin Chip Removal Processes, AI methods, Tool Condition
Monitoring.
3.1. Micromachining:
Thin chip removal processes are becoming more common machining processes, as the
demands of modern industries shift their trends. The need for thin chip removal usually
occurs in the military industry, space technology, electrical components production, and the
medical field. When we say thin chip removal, we refer to cutting processes where the size
of material removal ranges between 1-999 micrometers (Barnabás Zoltán Balázs, 2021).
Thus, it is about common machining processes such as milling, hard turning, grinding,
broaching, etc. However, in this section of our study, we focus on micro-milling and hard
turning.
Cutting parameters.
There are usually five input-cutting parameters that are used in studies.
• Spindle Speed: It means the rotating speed of the spindle of the CNC machine. It is
usually symbolized by the capital S, simply representing revolutions of the spindle
per minute (rpm).
• Feed: Feed is the material removal rate of the tool per revolution of the tool, often
described as mm/rev and denoted by the letter f.
2
• Depth of cut: The depth of the removed material per one tool pass. The unit of the
depth of the cut is mm. It depends on the tool and the material. It is shown by the
letter t.
• Hardness: Hardness is the ability of the material mainly against tool penetration. It
can be measured by the scale of HRC, HRB, or HV.
• Tool nose radius: It is simply the radius of the tooltip, measured in mm (Vishal R.
Balwan, 2022).
The feed rate has the maximum contribution (56 %) for the main cutting force. Next, the
cutting speed contributed around 31 % of the main cutting force. The contribution made by
depth cut is only around 7 %. When it comes to surface finish, the major effect is made by
cutting speed (72 %) followed by feed rate (around 24 %) (Satish KumarP. Tamilselvan,
2022).
3.1.2. Micro-milling:
Micro milling has the main place for producing unconventional miniature patterns and
products because of the flexibility of micro-cutting tools and low cost (Xuewei Zhang *,
2022). Micro products and parts are highly needed for the industries such as electronics,
optics, aerospace, medicine, and biotechnology, nowadays. Many other production methods
have been developed to manufacture such components. However, micro-milling is capable
of machining a wide range of complex products and 3D surfaces which makes this
machining method more popular. The mechanism of chip removal is obviously more diverse
than traditional milling. The factors that make them different from conventional milling are
minimum chip thickness, size effect, elastic recovery, and plowing mechanism, these effects
must be considered. Moreover, micro-cutting tools are vulnerable to excessive deflections,
forces, and vibrations since they are tiny and fragile (Ozcelik, 2014).
Usually, micro-milling can be defined by the diameter of the cutting tool, which ranges
between one micrometer and thousand micrometers. On the contrary, the size of the
conventional milling exceeds a thousand micrometers. However, it is not a precise definition
of micro-milling, since the tolerance of the micro-machining also matters which must be less
than 1 micrometer (Lorcan O’Toole, 2021).
In traditional machining, tool wear is the main side of machining with many discussions
in most of the kinds of literature. However, in the micromachining process, truly little is
known about tool wear and wear mechanisms compared to conventional machining (Aline
Gonçalves dos Santos, 2018). The wear forms presented by the micro-tools are mainly nose
and flank wear, the wear mechanism being more evident is adhesion, evidenced by the
presence of adhered material at the forefront (Aline Gonçalves dos Santos, 2018).
3
3.2. Machine Learning Models:
Machine learning (ML) is an artificial intelligence (AI) technology that allows for the
automatic learning of patterns between inputs and outputs based on training data without
being explicitly programmed. Some machine learning algorithms (in particular, artificial
neural networks (ANN), support vector machines (SVM), and tree learners) can recognize
the complicated relationships among non-linear variables without the use of a physical
model or any specific information, which is a significant advantage over conventional
methods. When compared to traditional simulation approaches, ML methods are frequently
more computationally efficient.
Algorithms typically used to achieve dynamic model selection require feedback from the
system's output, but this feedback is frequently unavailable in real-world applications. A
novel two-step machine learning approach is developed to address this problem by
proposing a dynamic model selection strategy where the selector only needs knowledge
about the system parametrization, not its output. In the first step of this method, nearest
neighbors are used in a selection procedure to choose the best model from a digital library
of eight machine-learning algorithms. The second stage involves making the prediction
using the chosen model. A case study for forecasting the surface roughness of a micro-
machining process, which presents complicated challenges, validates the suggested
approach. The experimental results corroborate the advantages of the proposed method
increasing R2 from 0.892 to 0.915 and decreasing mean absolute percentage error from 19.79
% to 14.63 % when compared to the best individual models’ metrics (Yarens J. Cruz
Marcelino Rivas, 2022).
According to the general image provided by the literature analysis by T. Mikołajczyk, AI-
based models are appropriate for two primary tasks connected to the automation of wear
prediction in turning operations: Image processing is used to evaluate present tool wear and
to anticipate tool wear based on prior behavior of comparable tools. The workpiece is used
in the experiments is C45 with a diameter of 120 mm. The turning process is done by an
uncoated carbide tool.
The comparison of projected and real wear values revealed a good level of accuracy,
particularly in high wear levels, which is essential in industrial settings (T. Mikołajczyk,
2018).
The expected findings matched the experimental comparator readings obtained from the
Equator gauge rather well (Moschos Papananias, 2019).
4
One of the machining goals is also to deliver an excellent surface finish to the machined
component. A fuzzy logic model is created for forecasting surface roughness while milling
EN31 hardened steel with a TiAlN-coated cutting tool. Surface roughness is predicted by the
proposed model with an average inaccuracy of 3.74%. The hardened EN31 Steel diameter of
50 mm is cut with the tool TiAlN coated cutting tool. MATLAB is used to design the model
(S.Sivarajan, 2022).
Machine learning methods are commonly used in TCM (Tool Condition Monitoring). The
initial phase in TCM is feature extraction. Multiple sensors' gathered input data are
combined to build projections of informative features that accurately reflect the actual data.
It is employed to quote statistical data on the effectiveness of feature grouping.
Consequently, to increase forecast accuracy and reduce computer initiatives. These are the
methods for feature extraction: histogram, wavelet, autoregressive moving average, and
breakdown of variational modes. Feature selections are carefully considered as the
subsequent stage in ML, which excludes irrelevant data and chooses characteristics that
most strongly influence the desired prediction variable. It minimizes complexity in
interpreting the classifier model and lowers training span due to dimensionality and reduces
overfitting to increase overview. The procedures decision trees are used in the feature
selection process (Abhishek D. Patange, 2021).
5
Figure 1: Architectural graph of an MLP network (McCulloch WS, 1943)
Artificial neural networks and deep belief networks are developed to anticipate the wear
condition as well as the remaining useful life (RUL) of the tools to understand and describe
the intricate mechanism of microtool wear and failure. The wear rate was shown to grow as
the tool path radius grew, and the bigger radii could result in catastrophic tool failure. The
neural-network models provided 93-99% accuracy for the wear condition categorization and
RUL prediction deploying ANN, Deep Belief Network, and Hyper Parameter Tuning
(Sumant Bagri, 2021).
3.2.2.2. AlexNet:
Yuqing Zhou et al defined Alex Net as a deeper and broader version of CNN for
training huge datasets introduced by Alex Krizhevsky. AlexNet contains eight layers - five
convolution layers and three fully connected. The first five layers are convolutional layers,
some followed by maximum pooling layers, and the last three layers are fully connected
(FC) layers (Yuqing Zhou, 2022).
7
3.2.2.5. Gaussian Process Regression:
Williams and Rasmussen initially proposed Gaussian process regression (GPR) for machine
learning (Djordje Cica, 2020). In comparison to other kernel-based regression techniques,
such as SVM, GPR is a probabilistic model based on standard Bayesian approaches. GPR is
extremely useful for dealing with complex problems with high dimensionality, nonlinearity,
and a limited number of training parameters. GPR has been widely used in various fields of
engineering in recent years as a result of its excellent performance. In order to monitor the
tool’s flank wear width Guohao Song et al developed a novel tool wear monitoring model
using GPR (Guohao Song, 2021)
8
4. Methods:
S.No V_c (m/min) feed (mm/rev) a_p (mm) R_a (micrometers) T (min)
1 450 0.1 0.4 0.97 18.35
2 500 0.1 0.4 0.8 11.15
3 450 0.15 0.4 1.27 15.16
4 400 0.2 0.4 1.7 25.89
5 450 0.2 0.4 2.01 12.26
6 500 0.2 0.4 1.67 7.45
7 400 0.1 0.8 0.89 30.74
8 450 0.1 0.8 1.11 17.38
9 500 0.1 0.8 1.07 10.26
10 400 0.15 0.8 1.33 26.22
9
11 500 0.15 0.8 1.19 8.34
12 400 0.2 0.8 1.95 21.79
13 450 0.2 0.8 1.89 12.29
14 500 0.2 0.8 1.93 7.03
15 500 0.1 1.2 1.19 9.23
16 400 0.15 1.2 1.24 24.1
17 450 0.15 1.2 1.13 12.13
18 500 0.15 1.2 1.22 7.12
19 400 0.2 1.2 2.14 20.74
20 450 0.2 1.2 1.68 9.54
21 500 0.2 1.2 1.77 5.61
22 450 0.066 0.8 1.12 20.25
23 450 0.234 0.8 2.32 10.93
24 450 0.15 1.47 1.13 10.05
25 425 0.12 0.6 1.12 21.56
26 475 0.12 0.6 1.06 12.79
27 425 0.18 0.6 1.55 17.9
28 475 0.18 0.6 1.41 10.2
29 425 0.12 1 1.03 21.56
30 475 0.12 1 1.15 11.23
31 425 0.18 1 1.6 17.32
32 475 0.18 1 1.44 8.52
33 400 0.1 0.4 0.77 32.66
34 400 0.15 0.4 1.33 28.05
35 500 0.15 0.4 1.05 8.26
36 450 0.15 0.8 1.28 14.43
37 400 0.1 1.20 1.11 28.43
38 450 0.1 1.2 1.08 15.18
39 366 0.15 0.8 1.37 34.46
40 534 0.15 0.8 1.31 6.12
41 450 0.15 0.13 1.17 12.18
Table 1
Data training was done in MATLAB software using the above data in which the data
collected from the first 31 experiments was used to train the model while the remaining data
was used to test the model. During data training different combinations of “Training
Function” and “Transfer Function” were examined in order to figure out the model which
provided the more accurate output results. For training the data, a two layered feed-forward
backprop ANN model was used with three different combinations of training functions
namely TRAINLM, TRAINSCG and TRAINGDM. TANSIG and PURELIN were used as the
transfer functions in the two layers. The first hidden layer comprised of 10 nodes while 2
nodes were used in the second hidden layer. A general layout of the used ANN network is
presented below:
10
Figure 4: Layout of the ANN network
Using the above settings, a combination of 12 models were training during this project
using different combinations of transfer and training functions and the model which
according to the results gave the most accurate output was the one that was trained using
TRAINLM training function and TANSIG transfer function was used in both the layers. The
comparison of results obtained from all our models is presented below:
11
5. RESULTS:
After training 12 different models in MATLAB using the “nntool” module, the model which
outperformed the others in predicting the output parameters namely surface roughness and
tool lifetime is presented below:
TANSIG transfer function was used for both the layers and the training function used in
this model was TRAINLM. The performance of the trained model is represented below:
12
Figure 6
The training state for our best performing model is given below:
Figure 7
13
Figure 8
The output results obtained after training the model are presented in the table below:
After training the model, the remaining data was used to simulate the model and the
results obtained are:
R_a (experimental) T (experimental) R_a (Test) T (Test)
1.44 8.52 1.582034 9.159912
0.77 32.66 1.032182 28.48693
1.33 28.05 1.395324 25.95248
1.05 8.26 1.403721 8.649895
1.28 14.43 1.37454 15.77121
1.11 28.43 0.982218 25.42697
1.08 15.18 0.986417 16.77568
1.37 34.46 1.367487 30.30539
1.31 6.12 1.381594 1.237039
1.17 12.18 1.416385 18.33392
Table 4
15
The graphical comparison between the experimental and trained as well as simulated data
is presented below:
2
1.5
1
0.5
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Experiment
30
25
20
15
10
5
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31
Experiments
T (Experiment) T (Trained)
16
Experimental vs Simulated Data
2
Surface Roughness [µm] 1.5
1
0.5
0
1 2 3 4 5 6 7 8 9 10
Experiments
R_a (Experimental) R_a (Simulated)
17
Experimental vs Simulated Data
40
Tool Lifetime [min] 30
20
10
0
1 2 3 4 5 6 7 8 9 10
Experiments
T (Experimental) T (Simulated)
x = input;
t = target;
18
net.divideParam.testRatio = 10/100;
% Plots
% Uncomment these lines to enable various plots.
%figure, plotperform(tr)
%figure, plottrainstate(tr)
%figure, ploterrhist(e)
%figure, plotregression(t,y)
%figure, plotfit(net,x,t)
% Deployment
% Change the (false) values to (true) to enable the following code blocks.
% See the help for each generation function for more information.
if (false)
% Generate MATLAB function for neural network for application
% deployment in MATLAB scripts or with MATLAB Compiler and Builder
% tools, or simply to examine the calculations your trained neural
% network performs.
genFunction(net,'myNeuralNetworkFunction');
y = myNeuralNetworkFunction(x);
end
if (false)
% Generate a matrix-only MATLAB function for neural network code
% generation with MATLAB Coder tools.
genFunction(net,'myNeuralNetworkFunction','MatrixOnly','yes');
y = myNeuralNetworkFunction(x);
end
if (false)
% Generate a Simulink diagram for simulation or deployment with.
19
% Simulink Coder tools.
gensim(net);
end
To enhance our work, we used Minitab data analytics tool and investigated by which
cutting parameters, the surface roughness and tool life is affected the most. The Pareto charts
shows the results in the Figure 13 and Figure 14. As it can be seen, the surface roughness is
highly influenced by feed rate, while the tool life is mostly affected by cutting speed.
20
Figure 14: The effect of different cutting parameters on tool lifetime
6. Conclusions:
To sum up, our work provided literature survey on micromachining techniques and various
machine learning models. Furthermore, we introduced sample experimental data of cutting
parameters and used the data to predict the tool life and surface roughness of a sample
workpiece with the deployment of MATLAB and Minitab tools.
7. References:
Abhishek D. Patange, R. J. (2021). Review on tool condition classification in milling: A machine learning.
Materials Today: Proceedings, 46, 1106-1115.
Aline Gonçalves dos Santos, M. B. (2018). Tungsten carbide micro-tool wear when micro milling UNS S32205
duplex stainless steel. Wear, 414-415, 109-117.
Barnabás Zoltán Balázs, N. G. (2021). Analysis of cutting force and vibration at micro-milling of a hardened
steel. Procedia CIRP, 99, 177-182.
Djordje Cica, B. S. (2020). Predictive modeling of turning operations under different cooling/lubricating
conditions for sustainable manufacturing with machine learning techniques. Applied Computing and
Informatics Emerald Publishing Limited.
21
Guohao Song, J. Z. (2021). Tool wear monitoring based on multi-kernel Gaussian process regression and
Stacked Multilayer Denoising Auto Encoders. Mechanical Systems and Signal Processing, 186.
Jianmin Wanga, Y. L. (2021). An accurate tool wear prediction method under different cutting conditions
based on network architecture search. Procedia Manufacturing, 274-278.
Kamil Işık, S. E. (2022). A benchmark comparison of Gaussian process regression, support vector machines,
and ANFIS for man-hour prediction in power transformers manufacturing. Procedia Computer
Scieence, 207, 2567-2577.
Lorcan O’Toole, C.-W. K.-Z. (2021). Precision micro-milling process: state of the art. Advances in
Manufacturing, 9, 173-205.
Mahsa Valizadeh, & Sarah Jeannette Wolff. (2022). Convolutional Neural Network applications in additive
manufacturing. Advances in Industrial and Manufacturing Engineering, 4.
McCulloch WS, P. W. (1943). A logical calculus of the ideas immanent in nervous activity. Bull Math Biol, 5,
115-33.
Moschos Papananias, T. E. (2019). An Intelligent Metrology Informatics System based on Neural Networks
for Multistage Manufacturing Processes. Procedia CIRP , 82, 444-449.
Papanania Moschos, M. T. (2019). An Intelligent Metrology Informatics System based on Neural Networks
for Multistage Manufacturing Processes. Procedia CIRP, 82, 444-449.
S.Sivarajan, M. M. (2022). Prediction of surface roughness in hard machining of EN31 steel with TiAlN
coated cutting tool using fuzzy logic. materialstoday, 65, 35-41.
Satish KumarP. Tamilselvan, M. F. (2022). Hard turning of AISI D2 steel with cubic boron nitride cutting
inserts. Materials today.
Sumant Bagri, A. M. (2021). Tool wear and remaining useful life prediction in micro-milling along complex
tool paths using neural networks. Journal of Manufacturing processes, 71, 679-698.
T. Mikołajczyk, K. N. (2018). Predicting tool life in turning operations using neural networks and image
processing. Mechanical Systems and Signal Processing, 104, 503-513.
Tim Raffin, Tobias Reichenstein, Jonas Werner, Alexander Kühl, & Jörg Franke. (2022). A reference
architecture for the operationalization of machine learning models in manufacturing. Procedia CIRP,
115, 130-135.
Vishal R. Balwan, B. D. (2022). Influence of hard turning parameters on cutting forces of EN 353 steel.
Materialstoday;proceedings, 63, 149-156.
Xuewei Zhang *, T. Y. (2022). An intelligent sustainability evaluation system of micro milling. Robotics and
Computer-Integrated Manufacturing, 73.
Yarens J. Cruz Marcelino Rivas, R. Q. (2022). A two-step machine learning approach for dynamic model
selection: A case study on a micro milling process. Computers in Industry, 143.
22
Yuqing Zhou, G. Z. (2022). A new tool wear condition monitoring method based on deep learning.
Measurement, 1-11.
23