Generative AI in Manufacturing

CIRP Journal of Manufacturing Science and Technology 50 (2024) 40–54
Contents lists available at ScienceDirect
CIRP Journal of Manufacturing Science and Technology

journal homepage: www.elsevier.com/locate/cirpj
AI-based optimisation of total machining performance: A review

Katrin Ullrich a, Magnus von Elling b, Kevin Gutzeit c, Martin Dix a, Matthias Weigold b,
Jan C. Aurich c, Rafael Wertheim a, d, I.S. Jawahir e, Hassan Ghadbeigi f, *
a
Fraunhofer Institute for Machine Tools and Forming Technology Chemnitz, Germany
b
Technical University of Darmstadt, Germany
c
RPTU Kaiserslautern-Landau, Institute for Manufacturing Technology and Production Systems, Germany
d
Braude College, Israel
e
University of Kentucky, USA
f
University of Sheffield, Department of Mechanical Engineering, Sheffield, UK
A R T I C L E I N F O A B S T R A C T
Keywords: Advanced modelling and optimisation techniques have been widely used in recent years to enable intelligent
Artificial intelligence manufacturing and digitalisation of manufacturing processes. In this context, the integration of artificial intel
Multi-objective optimisation ligence in machining provides a great opportunity to enhance the efficiency of operations and the quality of
Machining performance
produced components. Machine learning methods have already been applied to optimise various individual
objectives concerning process characteristics, tool wear, or product quality in machining. However, the overall
improvement of the machining process requires multi-objective optimisation approaches, which are rarely
considered and implemented. The state-of-the-art in application of various optimisation and artificial intelligence
methods for process optimisation in machining operations, including milling, turning, drilling, and grinding, is
presented in this paper. The Milling process and deep learning are found to be the most widely researched
operation and implemented machine learning technique, respectively. The surface roughness turns out to be the
most critical quality measure considered. The different optimisation targets in artificial intelligence applications
are elaborated and analysed to highlight the need for a holistic approach that covers all critical aspects of the
machining operations. As a result, the key factors for a successful total machining performance improvement are
identified and discussed in this paper. The AI methods were investigated and analysed in the frame of the
IMPACT project initiated by the CIRP.
1. Introduction to artificial intelligence in machining facilitated their application [3,4]. Predictive models based on ML al
gorithms have been implemented to monitor, regulate, and enhance the
The immense increase of computational power, data storage capac resource and cost efficiency of various machining processes to ensure
ity, and transmission rates together with access to large volumes of data produced parts of high quality.
from a variety of sources provide the basis for the successful imple
mentation of artificial intelligence (AI) technology in a broad spectrum
1.1. AI technologies in machining
of application areas. AI has become a key player in support and
improvement of many everyday life activities and challenges, such as
AI applications in machining, according to the published literature so
navigation, procuring, medical treatment, information management,
far, are mostly focused on using a single evaluation criterion, such as
and communication. It is also an important part of Industry 4.0 which
surface roughness [5–7] or tool wear [8–10], for process assessment and
denotes the digitalisation and automation efforts in manufacturing [1,
optimisation. This is even though machining is a very complex operation
2]. Machining, as one of the key secondary operations to generate
covering several processes defined by the interaction between machine
functional surfaces, has attracted much attention and already benefited
tool, environment, and workpiece. Considering machining as a system of
from AI, machine learning (ML), and optimisation since the rapid
cutting tool, machine tool, and work material, the total machining
technological advancement in information processing technology has
performance (TMP) [11] must be evaluated by several criteria, including
* Corresponding author.
E-mail address: [email protected] (H. Ghadbeigi).
https://doi.org/10.1016/j.cirpj.2024.01.012
Received 30 June 2023; Received in revised form 26 November 2023; Accepted 23 January 2024
Available online 20 February 2024
1755-5817/© 2024 The Authors. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
K. Ullrich et al. CIRP Journal of Manufacturing Science and Technology 50 (2024) 40–54
surface finish, tool wear rate, tool geometry, dimensional and geomet data-driven methods. The conventional machining performance assess
rical accuracy, cutting power, and chip breakability. Therefore, a ment methods are not very efficient for the evaluation of the TMP due to
comprehensive approach to assess and improve the TMP should not only the large number of variables involved. Therefore, a series of fuzzy-set
include process outputs such as cutting forces or tool wear, but also the mathematical models have been developed and implemented to
quality of produced parts, environmental factors, and production costs address this shortcoming [11]. However, the performance of these
to maximise the functional performance and the process sustainability models is bound to the available algorithms and information technolo
via multi-objective optimisation techniques. gies present at the time. Therefore, using the modern data-based toolbox
Various modelling techniques including analytical, numerical, of AI including the methods can significantly enhance the performance
empirical, or physics-based methods [4,7,12–16] as well as data-based of these models with respect to both effectiveness and accuracy.
techniques, such as fuzzy logic and artificial neural networks (ANN),
can be used to model process-related targets as a precondition for single- 1.3. Machine learning in machining
and multi-objective optimisation. However, in the last decade
data-based algorithms from AI, such as ML and evolutionary optimisa ML applications in manufacturing processes have gathered signifi
tion algorithms, have rapidly become an important element of such cant interest among researchers. The existing literature covers either
models to enhance the machining performance and to enable a smart summaries of ML applications in different operations [3,33] or discusses
and sustainable manufacturing process [3,17,18]. specific case studies, such as tool wear or surface quality of the
A considerable number of review papers have already been pub machined parts [34]. Process outputs such as cutting forces and surface
lished in recent years covering various constituents of AI-based process roughness are reportedly predicted using ANNs [35], while sensor fusion
optimisation in machining from general performance assessment mea [36] and fuzzy optimisation [37] have been investigated to establish a
sures and optimisation targets over typical data sources, up to ML and reliable measurement setup for a high-quality data acquisition that is
optimisation algorithms (see Fig. 1). These delve into general discus crucial for optimisation tasks. A review of literature from the late 20th
sions on the successful integration of AI in manufacturing or machining century about the identification of optimal machining parameters pri
[15,19–23] and special considerations on selected machining operations marily using AI methods has been provided in [38]. This review high
including laser beam machining [24], abrasive finishing [25], drilling lights the fact that only a limited number of studies have utilised a
[26,27], or milling [28]. Additionally, application of particular model combined approach involving both ML methods and optimisation al
ling targets with respect to the ML approach and optimisation potential, gorithms as presented in [39,40]. Furthermore, evolutionary and, in
for example, rate of penetration in drilling [4], surface roughness [29], particular, genetic algorithms [12,41–45] were used and evaluated to
cutting forces [16], and tool breakage [30], were analysed. Finally, deep achieve an optimal manufacturing process. Optimisation techniques
learning for tool condition monitoring [31] or evolutionary techniques such as particle swarm optimisation [46], response surface methodology
for machining [32] highlight the particular value of data-based algo [47], weighted grey relational analysis [48,49], or the NSGA-II algo
rithms for manufacturing processes. rithm for single- and multi-view optimisation [50] have also been
implemented in the field of machining.
1.2. Optimisation in machining
1.4. Paper structure
Extensive research has been conducted over the past decades on the
development and application of optimisation algorithms to enhance Starting from recent work on advanced machining in [3] and
machining performance and process monitoring. However, with the considering the exponential development and application of data-based
continuous advancement of algorithms and the rapid progress in infor modelling, the present paper is focused on the state-of-the-art since 2018
mation technologies, there is an urgent need to further develop opti with particular emphasis on the improvement of the machining process
misation methods in machining to fully harness the potential benefits of by integrating ML modelling and data-based optimisation. The paper
Fig. 1. AI-based optimisation in machining includes the identification of performance indices, the selection of optimisation targets and data sources, as well as the
choice of ML and optimisation algorithms.
41
delivers a thorough analysis and systematic presentation of process deep learning (DL) are often used interchangeably and cannot be
parameters and variables as data sources as well as the modelling targets distinguished clearly in the literature, despite being different concepts
process characteristics, tool wear, and workpiece quality. The impor with a decreasing level of generalisation. AI refers to the intelligent
tance of multi-objective optimisation for TMP improvement as well as performance of machines [57,58] enabled through the entirety of al
challenges and preconditions for the successful implementation are gorithms, software, and hardware and, hence, covers the broadest scope.
discussed extensively. According to [59], AI is the ability of a system to correctly interpret
A brief introduction to machining operations in Section 2 is followed external data, to learn from such data, and to use those learnings to
by a brief description of ML and data-based optimisation algorithms achieve specific goals and tasks through their flexible adaptation [60].
applied to the field of machining science. Typical data sources and Whereas, ML refers to the mathematical algorithms which draw con
associated preprocessing, as well as feature extraction methods required clusions from data in the form of predictions, i.e., automatically trans
for high-quality modelling and prediction of machining operations are form experience into knowledge [61]. In this sense, optimisation
discussed in Section 4. The review of the state-of-the-art in ML appli algorithms are both an integral part of AI and are applied to find optimal
cations to machining in Section 5 provides a critical assessment of ML prediction models from ML to solve present learning problems. DL as a
approaches as a precondition for data-based optimisation. Additionally, subgroup of ML refers to the application of artificial neural networks
the most relevant modelling targets together with available information (ANNs) with hidden layers (Section 3.2). Therefore, the ML branches of
sources used as input and output data of the models are presented in ANNs and DL can be regarded practically the same. Modern ANNs are
detail in Section 5. Preconditions and obstacles for the implementation able to autonomously learn the representative features of data which are
of AI and ML towards TMP optimisation are discussed in Section 6 and a needed to solve ambitious learning tasks [62,63]. This specific capa
strategy is proposed for a successful TMP assessment and optimisation. bility of ANNs combined with the increased computing power, storage
Finally, the concluding remarks are described in Section 7. capacity, and data emergence in recent years has led to astonishing re
sults, for example in object detection or speech recognition, however, at
2. The machining process the expense of a limited explainability of the corresponding algorithms
[64].
Machining processes remain among the most frequently used oper An ML model is a function f which assigns an appropriate output or
ations to generate functional surfaces for applications in various sectors label y ∈ Y to every instance x of an input space X (Fig. 2). The input x
including aerospace, automotive, and energy. Cutting is generally is an appropriate feature representation of the instance, e.g., an image,
defined as a form-shaping machining process via chip removal due to an or a sensor measurement and labels are for example real values, classes,
interaction between a workpiece and a tool. Hence, it is commonly or rankings of data sets. If more than one feature sets for the instances
referred to as subtractive manufacturing to be easily differentiated from are available, so-called multi-view learning techniques can be applied
additive manufacturing [51]. Machining processes are often categorised [65]. A predefined training data set is used to generate the predictor
into conventional and non-conventional operations depending on the function using an ML algorithm in the training phase, while the per
source of energy used to remove the material [32,52]. The conventional formance of a model is evaluated in the testing phase with an inde
operations are based on a relative motion between the tool and the work pendent test dataset together with a convenient performance measure.
material, defined by the kinematics of the process, which results in the Supervised, unsupervised, and reinforcement learning (RL) are the three
mechanical chip formation. The operations are divided into cutting (e.g., learning paradigms of ML. Supervised learning refers to scenarios where
turning, drilling, milling) and abrasive processes (e.g., grinding, lapping, labelled examples, i.e., pairs (x, y) from X × Y , are available for
polishing) based on the number of defined geometrical features of the training with the goal of optimally aligning true labels and predicted
cutting edges. The shapes of the cutting edges are known and can be labels in a way that is determined by the algorithm. Regression, the
quantified in the cutting operations, while the features of the cutting prediction of a real-number, or classification, the prediction of two or
edges can only be described statistically in the abrasive processes [53]. more predefined classes, are typical tasks for supervised learning.
Process parameters (i.e., cutting speed, feed, and depth of cut), tool In contrast, in an unsupervised setting only unlabelled instances, i.e.,
properties (e.g., size, coating, edge properties) and the cooling strategy elements of X , are available for training. Clustering and dimensionality
(e.g., flood, minimum quantity lubrication, cryogenic) as well as the reductions are important examples of unsupervised learning tasks. These
workpiece material are among the critical factors required to design a refer to an optimal grouping of instances and or reducing the input
cutting process. dimension with a minimal loss of information, respectively. RL describes
The workpiece quality is often predetermined by technical specifi learning by trial and error, e.g., in playing a game or optimising oper
cations defining surface characteristics, and dimensional accuracy [29]. ations in a production process, whereby a programme or an agent
However, the produced surface integrity, which has an impact on the operates in an environment and aims at maximising its success measured
functional performance of the parts, is characterised by mechanical and in form of cumulative rewards for a sequence of actions [58,60,66].
metallurgical properties within the surface layer, including residual Independent of learning task or paradigm, transfer learning (TL) is an
stresses, microhardness, or the grain structure [54]. Therefore, an ML approach [67] which intends to transfer knowledge from the
optimised cutting process will positively affect the quality of the
generated surfaces [52,55] as well as improve the sustainability of the
machining process with respect to emissions and energy consumption
[47,56].
3. Fundamentals of machine learning and optimisation
In Section 3.1 necessary terms and concepts from ML theory,

including a categorisation of algorithms and learning tasks, are pre
sented. It is followed by an introduction of ML and optimisation algo
rithms most frequently applied in the field of machining.
3.1. Terms and learning tasks Fig. 2. The relation between process input x and output y can be captured with
an ML model or a fitness function and the optimisation model can be used to
The terms artificial intelligence (AI), machine learning (ML), and determine the most suitable input x∗ which minimises the objective function.
42
solution of one learning task to another related task, e.g., by modifying associated with labels and branches correspond to decisions that lead to
an existing ML model to a similar problem and improving training data those labels. At a tree branching the data instances are split according to
set [68]. their feature attributes or thresholds on the feature values. As it is
computationally difficult to calculate the optimal tree, DT algorithms
3.2. Machine learning algorithms are greedy and achieve good results with locally optimal decisions. To
this aim, DT learning starts with the root node and successively grows
ML models are generated to tackle a specified learning problem from the tree deciding which data attribute (feature) to use for the splitting of
practice by inferring the underlying functional relationship between the data subset in each node [60]. Gain measures are used to assess the
input and output of a process. A prior analytical relationship between efficiency of decisions based on to what extent they enhance the purity
input and output quantities is not needed to build an ML model, how of labels in subsequent leaf nodes and, hence, improves the overall
ever, expert knowledge can be greatly beneficial. The performance of predictive performance of the model. DT models are well-known for
models for a given dataset depends on the used ML algorithm and their high degree of explainability and interpretability and although
evaluation measures during training and testing. The majority of ML they have been developed to solve multi-class classification tasks with
algorithms applied in machining-related research in the reviewed discrete-valued attributes, with slight modifications, they can also be
literature can be grouped in (1.) support vector machines (SVMs), (2.) used to solve regression tasks as well. These are originally developed to.
artificial neural networks (ANNs), (3.) decision tree (DT) algorithms, The random forest algorithm (RF) [80] is an ensemble approach using
and (4.) regression analysis (RA). DTs as well as the principles of bagging training data and subsampling
representing features.
3.2.1. Support vector machines and kernel methods
SVMs are linear models arising from a regularised risk minimisation 3.2.4. Regression analysis
approach with beneficial properties of the corresponding loss function, Regression analysis establishes a mathematical model for the statis
originally with the aim to solve supervised classification or regression tical relationship of a real-valued dependent variable on one or more
problems in high-dimensional features spaces [69,70]. The main independent variables for prediction purposes or causality analysis [81].
objective of support vector classification (SVC) is to find a hyperplane, i. Multi-linear regression (MLR) assumes a linear dependency on several
e., a decision boundary, that separates different classes in the inputs input variables, whereas polynomial regression (PR) models the
such that the margin between instances and the hyperplane is maximal dependent variable as a linear combination of the powers of input var
[71]. Additionally, support vector regression (SVR) was developed for iables up to a certain polynomial degree. The method of least squares
regression problems where the optimal hyperplane is determined to best (LS) regression minimises the empirical risk for an MLR problem with a
fit the training data points [72]. SVMs can be transformed into so-called squared loss function that can be solved as a closed formula under
kernel methods [73–76] where a symmetric kernel function is used to specific conditions. Non-parametric regression is conceptually different,
implicitly map the learning instances from the original feature space to as the prediction function does not assess a fixed parameterisation of a
another feature space to enable the solution of a non-linear problem prescribed functional form. Gaussian process regression (GPR) as one
with a linear model. Apart from classification or regression, SVM vari representative uses Bayes’ theorem to find probability distributions over
ations can also be used to tackle other learning problems, such as outlier possible models based on the available dataset.
detection, clustering, or multi-class classification.
3.2.5. Other machine learning modelling approaches
3.2.2. Artificial neural networks Instance-based and generative models from supervised learning have
ANNs are learning algorithms which minimise the expected risk for also been used occasionally to predict machining parameters. Instance-
future predictions with a model inspired by neural networks in biolog based learning does not generate an explicit prediction model but uses
ical organisms [77,78]. An ANN is a connected directional network of the available training data directly for a prediction [60]. The k-nearest
nodes (or neurons) organised in multiple layers or complex structural neighbour (kNN) algorithm, for example, determines the label of new
units, such as convolution layers in convolutional neural networks instances according to the labels of the k closest data points. Generative
(CNNs) [79]. A variety of input data formats, including image, tabular, models, such as Bayesian belief networks [27], take the probability
or vectorial data can be processed by ANNs to solve different learning distribution of input and output variables into account for their pre
tasks. The neurons form the elementary functional units of the network diction. Clustering, dimensionality reduction, and outlier detection are
and receive signals in form of edge weights from preceding neurons, important applications of unsupervised learning. Clustering is the
process the signal, and pass it on to the subsequent neurons. Specific attempt to group data points according to their similarity without the
non-linear activation functions imply a characteristic sensitivity of the awareness of labels. Dimensionality reduction intends to reduce the
neurons and supply ANNs with the capacity to approximate any func feature space dimension, for example, via principal component analysis
tional relationship with arbitrary exactness. In contrast to feedforward (PCA). Outlier detection aims at finding anomalies which significantly
connections, in recurrent neural networks (RNNs) the nodes’ outputs differ from the remaining data.
can affect the input of preceding layers by forming a loop, which make
RNNs capable of predicting data sequence, e.g., via long-short term 3.3. Optimisation algorithms
memory networks (LSTM). The perceptron algorithm displays the basic
functionality of the ANN learning procedure. It implements the succes Optimisation, ML, and AI are closely linked with the notions of
sive adaption of the edge weights to the training inputs via back reasoning and decision making [19,26,82–84] which primarily relate to
propagation (BP). The multi-layer perceptron (MLP) is its generalisation human attributes. The optimisation aims to maximise or minimise an
with hidden layers. Mostly commonly, the gradient descent optimisation output or objective function O by calculating the most appropriate input
algorithm is utilised for the weight adjustment. The explainability and x∗ (Fig. 2). The machining process output y can sometimes be measured
interpretability of ANNs are typically not obvious because of the or calculated directly from the input y and covered with a fitness func
non-linearity of the activations, the non-convexity of the corresponding tion (compare genetic algorithms in Section 3.3.1). Examples of such
optimisation problem during training, and a typically large number of outputs are the measured energy or the calculated production costs. In
parameters. the other case, ML algorithms can be employed to capture the under
lying patterns and correlations from data samples of input x and output
3.2.3. Decision tree learning y. Given specified input values, the aim of single-objective optimisation
A decision tree (DT) is a supervised learning model where leaves are (SOO) is to optimise one objective function with respect to the inputs. In
43
contrast, multi-objective optimisation (MOO) denotes the simultaneous order to formalise fuzzy logic and to allow for conclusions from uncer
consideration of various objective functions and aims to fulfil all targets tain or incomplete information [84]. Fuzzy logic can also be applied in
to their greatest extent [27]. Depending on the preconditions of the process control especially when dealing with input variables that exhibit
optimisation problem, many strategies with corresponding solution instability or fluctuations [26]. The values of fuzzy variables can be
have been developed that involve the calculation of Hessians, gradients, included in quantitative prediction models and optimisation via mem
or values of the focused objective function in different manners. For bership functions or probability distributions [11,98–100]. A special
example, the minimisation of a linear least squares problem can be case is fuzzy clustering, where the assignment of a data object to a
calculated as closed formula whereas Newton’s method is an iterative cluster is not deterministic but fuzzy in the sense of a membership
procedure [88]. The optimisation approaches can be broadly grouped as function [28]. Fuzzy inference systems (FIS) combine a prediction model
deterministic or stochastic algorithms [89–91] based on their and membership functions as well as fuzzy rules (fuzzy logic) to draw
characteristics. conclusions [24,101,102]. Consequently, neuro-fuzzy inference system
In addition to the improvement of the machining process per se, (NFIS) or fuzzy neural networks apply ANNs as prediction model
optimisation is also used to find the best hyperparameters [8,13,24,85, (ANFIS).
86], feature representation, or ML pipeline [87] during training.
Furthermore, optimisation is utilised to determine the best predictor 3.3.3. Further optimisation algorithms
function out of a set of candidates, for example, via sequential minimal Bayesian optimisation (BO), spiral dynamic algorithm (SDA), and the
optimisation (SMO) for SVMs [74] or stochastic gradient descent for Levenberg-Marquardt algorithm (LMA) are also among the methods
ANNs [6]. applied to optimisation in machining. BO is used to globally optimise an
In recent years, the application of data-based optimisation heuristics unknown objective or a function expensive to evaluate by treating it as a
for machining processes are being used to an increasing extent along random function and iteratively updating an objective function until an
with the expansion of AI technology. The most frequently applied data- optimum is reached [103,104]. Similar to biology-inspired algorithms,
based optimisation algorithms in machining are briefly discussed in the the SDA is motivated by spiral phenomena appearing in nature and aims
following sections. Precise applications are presented in Section 5. at a global optimum via diversification and intensification strategies
[105]. The LMA is used to solve non-linear least squares problems [106].
3.3.1. Biology-inspired algorithms
Algorithms inspired by biology mimic natural processes for optimi 4. Data acquisition and preprocessing
sation and improvement [17]. Genetic algorithms (GA), genetic pro
gramming (GP), and particle swarm optimisation (PSO) are prominent Process parameters are defined physical quantities that describe a
representatives of evolutionary algorithms. A subset of domain instances machining process and can take different values to represent various
(population of individuals) is considered in a GA [60,92], i.e., chro operational conditions. Hence, process parameters are targets for pro
mosomes of genes, in a way to satisfy predefined descriptions of indi cess optimisation. In contrast, process variables are continuously
vidual features. The instances run through a cycle of mutation (random monitored during an operation to assess the machining performance.
exchange of genes) and crossover (exchange of chromosome parts be The process variables include all measurable signals, such as forces,
tween instances) for the generation of new instances. A fitness function temperature, or properties that can be observed while the process is
is used to select the most appropriate individuals to enter the next round running. Therefore, they cannot be adjusted or calibrated directly. This
of the cycle. Although most of the conventional GA algorithms target section reports on applied process parameters and measured process
SOO problems, MOO can also be addressed directly using variables with respect to data content, recording, and preparation in the
non-dominated sorting genetic algorithm II (NSGA-II) [93,94]. GPs [28, context of TMP optimisation.
87] follow the principles of evolution as well, but their optimisation
instances are subsequent calculation or preprocessing units called pro 4.1. Process parameters and process variables
grammes whose number is not predetermined, unlike the number of
genes in GAs. In a PSO [28] a population of individuals (swarm of The cutting conditions including cutting speed, depth of cut, and feed
particles) is moving around in the search space in order to find optimal rate very often appear as considered adjustable parameters in AI models.
points with respect to a fitness function. The particles can move either in Other factors such as lubrication, cooling strategies, cutting tool prop
random directions (similar to the mutation step) or according to the erties, such as geometry or coating, and machining tool can also be
movement of the entire swarm, i.e., depending on other particles varied to some extent to achieve an optimised machining performance.
(similar to the crossover step). The flower pollination algorithm (FPA) is Both process parameters and variables are used as model input for ML
another bio-inspired optimisation approach imitating the natural flower prediction models. Process variables as well as process characteristics
pollination process [95,96]. The FPA stochastically combines aspects of like tool and product properties appear typically as ML model outputs.
abiotic self-pollination (local) and biotic cross-pollination (global) using The three most frequent modelling scenarios are shown in Fig. 3.
random parameters and information of both the considered individuals A large variety of sensor types have been used in different AI ap
(flowers) and the remaining population. The navigation behaviour of proaches to improve the TMP. Examples are reported in Table 1. The
birds flying in various groups in search of food has inspired the devel inclusion of acoustic [10,99,107–110], acceleration or vibration [111,
opment of the pigeon optimisation algorithm (POA) [97] which can be 112], force or torque [8,113,114], and displacement sensors [100] is
applied for shortest paths problems in a given population of individuals already well-established and has been studied intensively in the field of
(paths). machining. Their extensive application can be explained with the rela
tive ease of assembly, data capture, and analysis compared to other
3.3.2. Fuzzy optimisation signals, such as machining-induced temperature from thermocouples
Fuzzy optimisation or fuzzy programming refers to optimisation [115] or thermal imaging [116]. Despite technical difficulties involved
approaches under uncertainty conditions, i.e., where data is imprecise or in temperature measurements, it is reported that high-speed and thermal
fuzzy [84]. Often neither the exact values of variables are known nor a imaging combined with modern image processing technologies provides
high-precision outcome is necessary to achieve an acceptable optimi a wealth of information about machining systems that can support the
sation result [98]. A fuzzy set is characterised by its membership func development of relevant DL techniques [117].
tion that indicates for each element the probability of its membership to In general, it is desirable to obtain information from multiple sensors
the respective set. Similar to classical set theory, operations such as to investigate and exploit the potentials of sensor fusion [118]. The
union, intersection, and complement can also be defined for fuzzy sets in simultaneous evaluation and combination of several sensors through
44
Fig. 3. The three typical ML modelling scenarios in machining with their associated (a) modelling targets, (b) input data, and (c) and output data.
Table 1
Important configurations of data acquisition, preprocessing, and feature extraction in ML applications for prediction and optimisation of different modelling targets in
machining.
Data source Reference Data acquisition Preprocessing and feature extraction Modelling target
Acoustic [124] Piezoelectric AE sensor on the workpiece with grease as Amplified and high passed filtered to only observe Tool wear
emission acoustic bond (40MSPS) -> data acquisition card -> voltage- continuous interaction,
(AE) time-signal -> computer -> AE software AE-RMS as input, tool wear as output of a neural
network
[99] AE sensor on dresser holder -> Amplifier module Ratio of power (ROP) from 25-40 kHz band, mean Grinding wheel wear
-> oscilloscope with 2 MHz sampling rate -> PC for storage and standard deviation values of ROP, fuzzy systems
and digital signal processing
Cutting force [125] Kistler 9255B dynamometer ->TEAC DR-FI data recorder Force, feed rate, eccentricity of face cutting and Average cutting force
and torque workpiece geometry, neural network
[100] Dynamometer: force and torque; feed rate, spindle speed and Delamination, thrust force and torque, normalised Delamination, thrust
drill parameter manually recorded data force and torque
[114] Kistler 9257 A dynamometer -> Kistler 5011 signal amplifier Transformation of the signals to numerical format Cutting force
-> data acquisition system in Labview -> data cleaning -> split into training and validation
data
Vibration [126] Endevco 165 3-axis accelerometer to spindle housing and Sampling and FFT -> feature selection -> band pass Class of tool wear
Kistler 8141 A filtering -> backpropagation neural network
accelerometer on workpiece
Merged [127] Dynamometer & accelerometer combined -> amplifier -> data Sampling, new features by combining signals and Tool wear
signals from acquisition system DAP 2400/e6 -> PC their statistical values through addition,
sensor multiplication, and division
fusion [112] Dynamometer & accelerometer combined -> amplifier -> data Feature generation from the power spectral density Roughness, profile
acquisition system from national instruments -> TDMS file of the segmented raw signals -> feature selection deviation, roundness
Vision system [117] Optical imaging of the machined surface Feature extraction, SVR to find relation between Tool flank wear
surface features and tool wear
Motor current [122] Edge device, recording machine data with 500 Hz Data cleaning and sampling, classification model Tool condition
and power (defective tool or not) and regression model (tool monitoring, tool defect
wear) detection
[128] Feedback sensors to record the main drive power Maximum, and minimum value at beginning, middle Flatness deviation
and end of the operation, data averaging
multi-view learning is reported to reveal correlations between input and features are used [19]. Feature extraction techniques can be applied to
output data that would not be available using a single signal [65]. A determine the degree of redundancy and to extract the most relevant
retrofit of machines with additional sensor equipment complicates both features or data representations of the recorded signals (Section 4.3).
the experimental setup and the data analysis, for what reason it is
preferable to use information from already embedded sensors in ma
4.2. Data acquisition and storage
chine tools or obtained from the respective control unit. These can be
extracted using hardware or software solutions, e.g., edge devices or
The acquisition of sufficient data of adequate quality is generally an
gateways, and streaming the data directly to an Internet of Things (IoT)
essential part of ML approaches. For machining problems, the pipeline
platform, databases, or a PC [119]. However, the potential intrinsic
for acquisition, storage, and data transfer depends on the conditions and
redundancy of information is generally lower when only few signals or
associated environment. Table 1 summarises the most important
45
measurement setups, including signals and associated data acquisition as process efficiency, quality measures, or production costs, can be
systems used to obtain required input and output data for ML ap optimised. The results of the applied algorithms must be accessible to
proaches in machining with different optimisation targets. The reviewed the process monitoring as well as the control system to enable the
literature indicates that no standard, such as OPC UA (open platform adjustment of the corresponding parameters [132].
communications unified architecture), is used for data transmission or The reviewed literature revealed that most of the research was
storage in databases. Despite offering flexibility for the data acquisition focussed on ML applications in machining operations with defined
procedure, the lack of standardised pipelines poses a challenge when cutting-edge geometries as they offer more opportunities to monitor and
repeating experiments since details of the measurement chains, such as observe the cutting operation (Fig. 4a). Milling is the most widely
technical specifications and processing steps, are usually not reported or studied process, followed by turning and grinding, while drilling oper
not in sufficient detail. Such critical information, however, provides a ations have received significantly less attention in this context. This
good insight into the state-of-the-art of modern data acquisition and could be due to the complexity of the chip formation conditions,
required preprocessing for ML applications [112]. monitoring of process variables, and surface quality characterisation
during drilling compared with milling and turning. Fig. 4b demonstrates
4.3. Preprocessing and feature extraction that among all ML approaches, ANNs are the most frequently applied
algorithms in machining-related problems, followed by SVM and RF.
The preprocessing of data, including sensor signals or images, is an Relevant examples of the reviewed literature together with the intended
important prerequisite for the training procedure of ML models and outcomes are presented in more detail in Tables 2, 3, and 4.
strongly influences the resulting predictive performance. Preprocessing Recent applications of ML in machining cover various objectives,
steps usually include: typically in one of the fields process characteristics, tool condition
monitoring and tool wear, or surface integrity. The following sections
• Filtering and denoising: reducing the noise level in the gathered provide the state-of-the-art in the respective fields and highlight applied
data using different filters, such as low-pass filters or Gaussian ker algorithms and the main outcomes.
nels [120],
• Replacement of missing values or outliers: implementing 5.1. Process Characteristics
adequate interpolation functions to fill the gaps in the datasets when
the instrumentation fails to capture complete data, Process characteristics comprise all properties of a machining pro
• Normalisation and standardisation: normalising and stand cess which can be observed during operation or inferred afterwards as a
ardising datasets with a large variation in the measured ranges as a direct consequence of the procedure. These include general character
scaling technique to boost the performance of the corresponding istics, such as operation time and costs or CO2 emission, and process-
models, specific characteristics, such as chip formation, cutting forces, or the
• Concept drift detection: detecting a change of statistical properties temperature rise during cutting.
in the datasets (concept drift) using calibration runs [121] to The reviewed literature (Table 2) shows, firstly, that process pa
determine whether a re-training is required to ensure the validity and rameters, such as cutting speed, feed, and depth of cut, are predomi
high performance of the model, nantly used as input variables to predict specific process characteristics.
• Data splitting: splitting available data into training and testing sets This is most probably due to the relatively simple preprocessing pro
to ensure a proper assessment of the model’s performance by inde cedure and their in-process adjustability [116]. Furthermore, cutting
pendent data [112], tool properties, including the nose radius of indexable inserts [113], the
• Data labelling: assigning a label to data instances for supervised tool geometry [133], and the coating [94] have also been utilised as
learning. input parameters. Secondly, cutting forces were considered most
frequently as modelling targets in all studied operations in the literature
A feature representation is a description of data (Section 3.1), for due to their fundamental effect on the productivity, process perfor
example vectors in Rn , images or graphs. Feature extraction can be used mance, and quality of the produced surfaces in machining. Varying
to calculate suitable and informative variables as model input. Statistical cutting forces often cause disturbances of the process. Therefore, a
features such as minimum, maximum, mean or median value, amplitude precise prediction of the cutting forces is fundamentally important for
range, or standard deviation of a time series are examples of time process monitoring and control [134]. For example, different ML
domain features. Typical frequency domain features are characteristic methods, including ANN, SVR and RA models, were applied to predict
frequencies and their magnitudes as well as fast Fourier transformation cutting forces when milling an aluminium alloy for different cutting
(FFT) features or wavelet features (Table 1). Although many signals can parameters [94]. The feed value was found to be the most dominant
be gathered from a machine tool from which again numerous feature factor influencing the predicted forces, followed by the tool type and the
representations can be calculated, various methods such as wrappers, depth of cut. The authors reported that the ANN approach outperformed
filters, and hybrid methods [122,123] are used to extract the most both the SVR and RA models in this specific application. An optimisation
relevant features to enhance the prediction performance and to avoid problem was solved combining the ANN with NSGA-II algorithm to
deteriorating effects. In this context, unsupervised dimensionality minimise the cutting forces as well as the surface roughness depending
reduction techniques, such as PCA, isometric feature mapping, and on the input parameters.
locally linear embedding are employed. Different ML modelling approaches, such as ANN [116], GPR [135],
5. Review on recent ML modelling in machining
Optimised process parameters can maximise the efficiency of the

cutting process during machining, e.g., by increasing the material
removal rate [48], minimising tool wear [49], or improving the surface
quality [129]. In this manner, thermally induced defects, progressive
tool wear, or tool vibrations during the process can be prevented [130].
ML combined with optimisation algorithms can be utilised in advanced
machining [130] or manufacturing [131] to autonomously adjust pro Fig. 4. Reviewed literature of ML in machining: proportions of (a) considered
cess parameters. As a result, general or specific output quantities, such machining processes and (b) applied algorithms.
46
Table 2 Table 3
Example applications of ML algorithms and corresponding model input and Example applications of ML algorithms for tool wear prediction and tool con
output for the prediction of process characteristics in machining. dition monitoring with corresponding input parameters and measured variables,
Process ML Model input Model output Year /
respectively.
algorithm Source Process ML algorithm Model input Year/
Source
Turning SVM, GPR Cutting speed, rake Cutting forces, 2020 / Tool wear Tool condition
angle temperature [135] prediction monitoring
ANN, SVM, Cutting speed, feed, Cutting forces 2021 /
GPR depth of cut, nose [113] Turning ANN, SVM, - AE 2021 /
radius RF, DT, KNN [10]
SVM, RF, RA Cutting speed, depth Chatter 2022 / ANN, SVM Cutting speed, - 2018 /
of cut [136] feed, depth of cut [137]
ANN, SVM Cutting speed, feed, Cutting forces 2018 / ANFIS-ANN - Cutting force 2020 /
depth of cut [137] [142]
ANFIS-ANN Cutting speed, feed, Temperature 2018 / ANN Cutting speed, - 2020 /
depth of cut [116] feed, material [9]
Milling ANN, KNN Cutting speed, feed, Cutting forces 2019 / removal
depth of cut, tool [133] ANN Cutting edge - 2018 /
diameter [149]
ANN Feed, depth of cut Cutting forces 2019 / ANN, RF, Depth of cut, AE, cutting force 2021 /
[134] SVM cutting speed, [146]
ANN, SVM, Cutting speed, feed, Cutting forces 2019 / feed
NSGA-II depth of cut, coating [94] ANN - Image processing 2019 /
ANN Cutting forces Chatter 2021 / [150]
[138] ANN - Spindle current 2020 /
SVM, RF, Cutting speed, feed, Cutting forces 2018 / [122]
KNN, PR depth of cut [114] ANN - Cutting force, AE, 2020 /
ANN Cutting speed, depth Chatter 2021 / vibration, [151]
of cut [68] acceleration
ANN Cutting speed, feed, Cutting forces 2021 / CNN - Cutting forces, 2019 /
depth of cut [139] acceleration, [152]
Grinding ANN Cutting speed, feed, Grinding forces 2020 / vibrations
depth per cut, AE [108] CNN Cutting speed, Spindle load 2020 /
GPR Current, voltage Grinding forces 2021 / feed, depth of cut [153]
[140] Milling 1D-CNN Feed AE, vibration, 2021 /
PSO-SVM Cutting speed, depth Grinding forces 2020 / spindle current [154]
of cut [7] RF - Spindle current, 2021 /
GPR, Cutting speed, feed Temperature 2020 / Image processing [155]
Bayesian [141] RF, SVM Cutting speed, AE 2019 /
feed [156]
RNN Cutting speed, - 2022 /
feed, depth of cut [157]
and SVM [135], were also implemented for the prediction of tempera
RF Cutting forces 2020 /
ture rise during machining as technical challenges associated with an [158]
accurate temperature measurement constitute the need for reliable ANN, DT, Cutting speed AE 2021 /
predictive models. In this context, the quality and preprocessing of data KNN [159]
CNN - Cutting force 2018 /
and the applied ML algorithms significantly affect the accuracy of the
[160]
predictions as reported in [116]. It was demonstrated that an adaptive SVM, RF Cutting forces, 2019 /
neuro-fuzzy inference system (ANFIS) model with a feed-forward current [161]
backpropagation multi-layer perceptron (BPMLP) with log-sigmoid RF, SVM Cutting speed, 2021 /
activation function outperforms the temperature rise prediction feed, depth of cut [122]
CNN - Cutting forces 2018 /
compared with an ANN model which was optimised with the
[162]
Levenberg-Marquardt algorithm. ANN - Cutting forces, 2020 /
Table 2 demonstrates that cutting forces are the topic of interest in vibration [163]
the majority of ML applications in the field of process characteristics, Grinding Fuzzy Cutting speed, - 2018 /
feed, tool [99]
perhaps due to their influence on process stability and on TMP in almost
diameter
all of the machining processes. The fact that there are only few studies RF, MLR (RA) Dressing speed, AE 2018 /
on temperature and chatter prediction for the considered operations dresser width [107]
could be linked to the technical difficulties to capture the required high- SVM, kNN- Cutting seed, AE 2019 /
quality data for the generation of ML models with high performance. GA contact force [109]
RF, MLR (RA) - AE, cutting forces, 2019 /
This is even more apparent for the prediction of characteristics such as
acceleration [164]
energy consumption as available data is scarce. RF - Image processing 2021 /
[147]
ANN, SVM, - Image processing 2021 /
5.2. Tool wear prediction and condition monitoring
RF [148]
Drilling ANN, SVM, - Torque 2020 /
A direct tool measurement is impossible in most of the machining RA [14]
operations. Therefore, indirect tool wear characterisation or predictive RF Cutting speed, - 2020 /
models are used to assess the status of the cutting tool. Tool wear pre feed [165]
ANN, SVM, - Image processing 2020 /
diction (TWP) and tool condition monitoring (TCM) denote very closely, SDA [8]
yet different, approaches to estimate the tool life and its performance,
respectively.
TWP prediction is the data-based forecast of wear characteristics and
the respective tool life via ML algorithms in dependence of the process
47
Table 4 design and the parameters chosen. Different combinations of process

Example applications of ML algorithms for product quality prediction with parameters and variables have been employed for various ML modelling
corresponding input and output quantities. approaches depending on the kinematics of the studied process and
Process ML algorithm Input Output Year/ cutting tool features as demonstrated in Table 3. The cutting speed, the
Source feed value, and the depth of cut are frequently used input parameters for
Turning ANN Cutting speed Stability 2019 / TWP (Fig. 5), as they can be easily adjusted during a machining process.
behaviour [171] However, it is reported that the accuracy of the predicted outputs
PSO-SVM, Cutting speed, feed, Roughness 2018 / significantly depends on the selected ML algorithm and associated
ANN depth of cut [137]
membership functions [142]. Similar to the cutting force prediction
SVM, RVM, Cutting speed, feed, Roughness 2018 /
PSO depth of cut, [166] (Section 5.1), the ANFIS approach was proven to be a powerful tool in
vibration, power predicting tool wear depending on the cutting speed, feed per tooth, and
ANN, SVM, Cutting speed, feed, Roughness 2019 / depth of cut in milling [143]. The PSO algorithm was used to optimise
GA-GBRT depth of cut [172] the prediction performance of the model.
ANN Cutting speed, feed, Roughness 2020 /
depth of cut, force, [173]
While TWP mainly focusses on the prediction of tool life, TCM uti
vibration, tool wear lises quantities such as cutting forces, temperature, AEs, or vibrations,
ANN Cutting speed, force Roughness, 2021 / which indirectly provide information about the condition of the tool and
Roundness [112] can be used to assess the wear status via ML predictions [144,145]. In
ANN SVM, Cutting speed, feed, Roughness 2021 /
this context, cutting forces [146] and AE signals (Fig. 5) are the pre
CAT (RA), depth of cut, [174]
GBR (RA), DT, vibration dominant input variables in TCM as they can be linked directly to
XGB experimentally determined wear mechanisms.
ANN, SVM, Cutting speed, feed Roughness 2021 / Table 3 shows that AE is the dominant input parameter for TCM in
GPR depth of cut, tool [113] grinding operations, where discontinuous chips are formed and indi
RF Cutting speed, feed, Residual stresses 2019 /
vidual cutting edges cannot be defined geometrically [107]. Direct sig
depth of cut [175]
RF Cutting speed, feed, Residual stresses 2020 / nals from machine tools, such as spindle load [153] and spindle current
depth of cut [176] [154,155], as well as visual characteristics of the produced surfaces,
ANN-POA, Cutting speed, feed, Residual stresses 2021 / have also been used for TCM [147,148]. In this context, image pro
ANN-PSO depth of cut [85]
cessing (IP) is proven to be a powerful tool when the produced surfaces
ANN-POA, Cutting speed, feed, Residual stresses 2021 /
ANN-FPA depth of cut [167] are the primary concern and represent the tool performance [148].
ANN, SVM, RF Force, temperature Microstructural 2021 / In summary, ANN models are widely used for TWP and TCM. In
modification [13] turning operations, CNNs play an important role while other algorithms,
ANN, DT, Temperature Thermal damage 2022 / such as SVM or RF, are also used for tool performance prediction in
ElasticNet [177]
milling, drilling, and grinding.
Milling ANN Tool wear, power Roughness 2018 /
[168]
depth per cut, tool [94] 5.3. Product quality
depth of cut [178] Product quality in machining generally refers to the quality of the
ANN Cutting speed, feed, Roughness 2020 / produced surfaces for what reason the term surface integrity is
depth of cut [6]
frequently used synonymously in this context. The surface can be eval
1D-CNN Cutting speed, feed, Residual stresses 2018 /
depth of cut [179] uated by measures of dimensional and geometrical accuracy (e.g.,
ANN, RF, DT Cutting speed, feed, Barkhausen noise 2021 / roundness), surface finish (e.g., roughness), or mechanical and micro
tool wear [180] structural properties of the surface (Fig. 7). As it can be seen in Fig. 6a,
ANN, SVR, Cutting speed, feed, Microhardness 2020 /
the primary cutting parameters cutting speed, feed, and depth of cut
NSGA-II depth of cut [169]
CNN Tool wear, power Flatness 2021 / were mostly used as input for quality models as they have the most
[128] decisive impact on the machining process along with the tool and
DT, SVM Cutting speed, feed Flatness 2018 / workpiece material [6]. However, also tool wear and the drive power of
[181] the CNC or variables, such as AE and vibrations, can be utilised to
ANN, SMOTE- Residual stresses, Fatigue 2021 /
automatically predict the surface roughness [168]. Additionally, it was
RF, DT, roughness, hardness [170]
Grinding ANN AE Roughness 2020 / observed that measured process variables, such as AE signals [108,110]
[108] and vibrations [86], are predominantly used as input data in ML models
ANN, SVM, Acceleration, power Roughness 2020 / for surface quality prediction in grinding. Fig. 6b shows that surface
GBR [111]
roughness followed by geometrical tolerances are the most widely
BR-ANN, Cutting speed, feed, Roundness 2021 /
SVM, GBR workpiece [182]
studied measures to evaluate the surface quality as an output of ML
geometry models due to the relative ease of measurements compared with other
kNN, SVM, RF AE, power, spindle Grinding burn 2021 / factors of surface integrity such as residual stresses or microstructural
current [110]
kNN, SVM RF Tool, workpiece Grinding burn 2022 /
[183]
LSTM-ANN, Force, AE, vibration Roughness 2021 /
RF [86]
Drilling RF Force Concentricity 2020 /
[184]
RF Spindle current Roundness 2020 /
[185]
Bayesian Cutting speed, feed, Roughness, 2020 /
Network tool coating Roundness [186]
Fig. 5. Reviewed literature of ML in machining: proportions of (a) input pa
rameters for tool wear prediction and (b) input variables for tool condition
monitoring (Table 3).
48
v. investigating optimisation techniques for the training of ML

models, e.g., ANNs [94,116,151].
According to the literature research, the data-based adjustment of

machining parameters in the sense of (ii.) is mostly realised via data-
based heuristic optimisation techniques such as GAs. Few authors
describe optimisation approaches as mentioned in (i.) based on the
prediction of an upstream ML model [21]. Many AI applications in the
Fig. 6. Reviewed literature of ML in machining: proportions of (a) input and field of machining were found to successfully generate an ML model for
(b) output quantities for product quality prediction in machining (Table 4). the automatic, data-based prediction of relevant process quantities ac
cording to point (iii.) above, without a subsequent optimisation step.
alterations [166]. However, it is important to note, that the last two approaches, i.e., (iv.)
Comparable with the observed trend for TWP (Section 5.2), ANNs and (v.), have only an indirect influence on the process optimisation by
represent the largest group of applied ML algorithms for quality pre finding the most appropriate predictor function for a given learning task.
diction, followed by SVM, RF, and RA, as reported in Table 4. However, Fig. 7 demonstrates the correlation between various process parameters
a broader range of algorithms, including DT, kNN and Bayesian net and variables with the potential targets in ML modelling to define
works, are used for product quality prediction compared with TWP. In objective functions to address the required criteria for TMP. The aim of
addition to the applied data and algorithms, it was shown that the SOO and MOO in the context of TMP optimisation is finally to calculate
training procedures have a strong effect on the prediction performance optimal values for process parameters, such that their feedback in the
[6]. It was demonstrated that a response surface methodology (RSM) actual process results in improved machining results. The solid lines in
model outperforms ANN models trained by various optimisation algo Fig. 7 describe that there is a direct link between the demonstrated
rithms, e.g., quasi-Newton backpropagation, LMA, conjugate gradient or entities. For example, process parameters directly affect the ML targets
resilient BP. Furthermore, hybrid ANN models optimised by different of process characteristics, tool wear and TCM, as well as the product
algorithms, including PSO, the POA, and the FPA were demonstrated to quality, while the broken lines denote an indirect impact. Generally, the
significantly outperform conventional ANNs in the prediction of TMP can be optimised according to the three interdependent criteria of
turning-induced residual stresses [85,167]. In addition to surface product quality, processing time, and processing costs. These criteria are
roughness and residual stresses, other surface integrity measures, strongly correlated with the factors surface finish, dimensional accu
including microstructural alterations [13] and subsurface microhard racy, tool-wear rate, and chip breakability [11].
ness [169], have also been predicted using ANNs and SVMs, respec Product quality can cover one or more aspects of surface integrity,
tively. In turn, these could be used as inputs for a further ML modelling geometric and dimensional accuracy, residual stresses, microhardness,
of fatigue effects at machined parts, e.g., via DTs and partial least and other properties of the machined products (Section 5.1). The pro
squares regression analysis [170]. cessing time can best be assessed via the material removal rate (MRR)
and, hence, the production rate in machining. Although alteration of
6. AI-based total performance optimisation cutting parameters could lead to higher MRR values, this may negatively
affect the surface integrity, i.e., increased roughness or residual stresses,
In the present section the findings of the reviewed literature in the tool life through an increased wear rate, and process sustainability
machining will be summarised and the importance of MOO for suc due to an increased energy consumption and demands for more cool
cessful TMP will be worked out and discussed. ants. The production costs also depend on the MRR, costs associated
with the raw material, and process consumables, including cutting tools.
Depending on the machining process at hand, the three criteria time,
6.1. From ML prediction to process optimisation costs, and quality affect the TMP to varying degrees which should be
captured by ML and optimisation models based on a meaningful
The simultaneous consideration of all parameters and variables database.
affecting TMP is required to achieve the overall optimisation of
machining processes. This directly implies the need for modelling 6.2. Single-objective and multi-objective optimisation in machining
various process-related targets and integrating different MOO methods
for their simultaneous achievement. Targets corresponding to product As explained previously, the aim of SOO is to optimise one objective
quality, tool wear, and process characteristics (Sections 5.1 to 5.3) have function with respect to the inputs. The inclusion of the optimal process
been extensively investigated in recent years. Most of the ML applica parameters via a feedback control system back to the process leads to the
tions in machining can be described as the following approaches: actual desired improvement of the machining process. The cutting
conditions, i.e., cutting speed, depth of cut and feed rate, and to some
i. using an ML algorithm to model the relation between process extent cutting tool materials and characteristics, are the most important
input and output and subsequently selecting appropriate inputs and flexible parameters for process optimisation in machining.
for an optimal output by an appropriate optimisation algorithm Observable and measurable variables such as AE or acceleration signals
[4,187], can be used to predict target quantities, but due to their uncontrollable
ii. directly calculating the optimisation objective (e.g., using a nature they cannot be used for optimisation in the strict sense. If they are
fitness function for GAs) without a need for a preceding ML used as model outputs, they can be optimised themselves depending on
modelling to apply the optimisation algorithm [45], process parameters. As further objectives and potentially contradictory
iii. using an ML model to predict the outcome of a process from input effects of different targets are ignored, SOO limits the optimisation result
parameters without direct exploitation of the result in an opti to a single aspect [47]. MOO, in contrast, considers multiple target
misation algorithm. In this case, the optimisation occurs indi quantities and, therefore, leads to a simultaneously optimised set of
rectly in form of online monitoring [154] or process objectives and corresponding inputs [188]. It can yield SOO as a mar
comprehension from the ML predictions [8], ginal case for the TMP optimisation [43].
iv. applying optimisation techniques for the hyperparameter tuning The modelling of multiple targets in the form of ML predictions can
process to find an optimally parameterised prediction model [7,8, be regarded as a preliminary step or precondition for an MOO. Different
85,174], ANN models are, for example, utilised to predict different aspects of
49
Fig. 7. Overview of TMP optimisation: process parameters and variables as ML model inputs, outputs of ML models as optimisation targets, and the main TMP
optimisation criteria time, costs, and quality.
surface quality [112] or tool wear and roughness [154] from measured SVMs require an appropriate feature representation of data from a
vibrations, forces, sounds, or spindle current signals. Similarly, residual preprocessing procedure, but they are explainable or interpretable by
stresses [85] or surface finish and dimensional deviation [36] are pre design. With respect to the algorithms for process optimisation in
dicted from the cutting conditions using ANNs. Classical ML algorithms machining, the heuristic approaches of evolutionary algorithms, such as
such as kNN, SVMs, DT, and RF [122,135,140] or fuzzy set models [11] GAs, played the dominant role in recent years. The predominance of
are also implemented in multiple regression and classification tasks data-based heuristic methods can be explained with the fact that the
which assess machining processes. objective functions frequently cannot be calculated as closed formula. As
As for SOO, MOO can be performed based on different ML ap no information of function values or derivatives are available, heuristic
proaches and modelling (cases (i.) and (ii.) in Section 6.1). In the first approaches can be employed successfully in compensation. In the future,
case (i.), ML algorithms, such as ANNs, RF, and SVMs, are used for the known properties of the objective function or its derivatives could
prediction of forces, tool edge chipping, and surface quality with and an increasingly be included to the process optimisation approach.
additional optimisation algorithm is utilised to determine the optimal
cutting conditions [94,100,158]. A prediction from analytical, empir 6.3. Challenges and success factors
ical, and numerical model components as fitness function for GA [12]
represents a hybrid approach of (i.) and (ii.). However, MOO is The solution of MOO problems as a mathematical task meets
frequently performed without a preceding ML modelling. Instead, a computational challenges such as, e.g., the search for global optima or
data-based optimisation algorithm is calculated directly with respect to the compromise between Pareto-optimal solutions and optimal values
a measurable or computable objective function. In this context, evolu for single-target objectives. Also the convergence rate [50], overfitting,
tionary algorithms, such as NSGA-II [43,50,189,190] or PSO [46,190], and large datasets for the training of DL algorithms [28] pose problems
as well as the RSM [47,48] and grey relational analysis [48,188] play an for the practical implementation. Other limitations of online-monitoring
important role. and process improvement are real-time predictions [10,192] and the
The presence of conflicting objectives is a general disadvantage of actual implementation of optimisation results at the production plant
MOO which results in so-called Pareto-optimal solutions [43,189,191]. [4].
These are overall best solutions that are potentially suboptimal with To achieve the demanding aim of TMP [11,12] optimisation, the
respect to single optimisation criterion. It is, however, mathematically following key success factors were identified:
feasible to deal with multiple objectives via the introduction of con
straints [44,45] or an SOO approach of a sum or other functional rela • ML prognoses by itself are beneficial for process monitoring or
tionship of individual optimisation targets [41]. related applications. A true improvement of a machining process
Generally, ANNs in their different forms followed by SVMs are the with respect to resources, costs, and product quality can only be
most frequently used ML algorithm in various applications related to achieved if optimal process settings are determined by an optimisa
machining operations. This can be explained by the fact that neural tion procedure and if these settings are used in the actual process.
networks are able to generate the necessary feature representation of • Three classes of objectives along the process chain (refer to Section
data for the learning task at hand as a part of the learning process. 3) have been identified, to which the precise quantities from SOO
Additionally, they can handle multiple, dependent input parameters, as and MOO can be assigned (Fig. 7). In order to ensure a holistic view
it is typical for manufacturing processes like machining. This capacity of on the machining process, the TMP must necessarily be evaluated by
ANNs is at the expense of a low explainability, which inhibits the means of a consideration of all three classes via MOO [11,41,44,46].
deployment of neural networks in industrial applications. In contrast,
50
• The inclusion of the ML and optimisation results into the process is GAs for SOO or NSGA-II for MOO are applied to determine process pa
crucial in practise, either for an immediate regulation or for a future rameters for an optimal machining quality.
process design and planning. Therefore, a pipeline on the basis of MOO has been constituted as precondition for TMP optimisation on
software and hardware must be installed which particularly com the way to advanced machining processes. Future research is required to
prises the database, continuous integration and continuous deploy consider targets related to fair and sustainable machining, in particular,
ment (CI/CD) tools for the industrial implementation [4,26], and a to energy and CO2 reduction in TMP. In this context, continual and
visualisation platform for in-process or online monitoring [126,154]. transfer learning techniques will likewise support the efficient con
• Production plants are typically equipped with multiple sensors sumption of resources and provide a persistent adaption of the AI
which permanently deliver sources of information about the process component.
(either from the beginning or as a result of retrofitting). As the TMP
optimisation exactly requires this variety of information, an effective
sensor fusion procedure for the ML modelling should be integrated Declaration of Competing Interest
[36,118,127,154,192].
The authors declare that they have no known competing financial
Apart from these key factors which are deemed to be necessary for interests or personal relationships that could have appeared to influence
the TMP optimisation, there are further modelling aspects that are not the work reported in this paper.
extensively researched, but significantly contribute to the machining
process improvement. These include a broader observation of cause- References
effect relation chains that support the idea of TMP without limiting a
[1] Shahin M, Chen FF, Bouzary H, Krishnaiyer K. Integration of Lean practices and
complex problem to a single cause-effect relationship [193]. Addition Industry 4.0 technologies: smart manufacturing for next-generation enterprises.
ally, further investigation about the quality and quantity of the used data The International Journal of Advanced Manufacturing Technology 2020;107:
[20], as well as the selection of the best model for given learning tasks in 2927–36.
[2] Oztemel E, Gursev S. Literature review of Industry 4.0 and related technologies.
the context of machining [28] increases the performance of predictive
Journal of Intelligent Manufacturing 2020;31:127–82.
models. Finally, as soon as a prediction and optimisation can be used [3] Kim DH, Kim TJ, Wang X, Kim M, Quan YJ, Oh JW, Min SH, Kim H, Bhandari B,
successfully for the in-process regulation, the question about the reus Yang I, Ahn SH. Smart machining process using machine learning: a review and
perspective on machining industry. International Journal of Precision
ability of models emerges. Transfer learning can be used to adapt ML
Engineering and Manufacturing-Greening Technology 2018;5:555–68.
models for the variation of conditions or other related processes that will [4] Barbosa LFF, Nascimento A, Mathias MH, de Carvalho Jr JA. Machine learning
be the basis for a sustainable, flexible, and robust application of AI in the methods applied to drilling rate of penetration prediction and optimization-a
near future [183,194]. ML models should be continuously retrained or review. Journal of Petroleum Science and Engineering 2019;183:106332.
[5] Correa M, Bielza C, Pamies-Teixeira J. Comparison of Bayesian networks and
adapted using various continual and transfer learning approaches [178] artificial neural networks for quality detection in a machining process. Expert
to avoid the drift in their prediction [181] and performance loss over Systems with Applications 2009;36(3):7270–9.
time due to changes in the input data, the tool, or the environment. To [6] Eser A, Aşkar Ayyıldız E, Ayyıldız M, Kara F. Artificial intelligence-based surface
roughness estimation modelling for milling of AA6061 alloy. Advances in
sum it up, this review paper identifies the issues energy and CO2 Materials Science and Engineering 2021;2021:1–10.
reduction [48,50,188], sustainable machining in general [18], transfer [7] Gu P, Zhu C, Tao Z, Yu Y. A grinding force prediction model for SiCp/Al
learning techniques [157], and effective MOO [93] as the most relevant composite based on single-abrasive-grain grinding. The International Journal of
Advanced Manufacturing Technology 2020;109:1563–81.
future topics in AI-based research on process optimisation in machining. [8] Alajmi MS, Almeshal AM. Predicting the tool wear of a drilling process using
novel machine learning XGBoost-SDA. Materials 2020;13(21):4952.
7. Conclusion [9] Chang WY, Wu SJ, Hsu JW. Investigated iterative convergences of neural network
for prediction turning tool wear. The International Journal of Advanced
Manufacturing Technology 2020;106:2939–48.
In recent years, considerable research activity has been conducted to [10] Ferrando Chacón JL, Fernández de Barrena T, García A, Sáez de Buruaga M,
reliably model parameters and variables in machining. Machine Badiola X, Vicente J. A novel machine learning-based methodology for tool wear
prediction using acoustic emission signals. Sensors 2021;21(17):5984.
learning (ML) and, in a broader context, artificial intelligence (AI) used [11] Fang XD, Jawahir IS. Predicting total machining performance in finish turning
in conjunction with data-based optimisation have become standard tools using integrated fuzzy-set models of the machinability parameters. The
for analysis and monitoring of machining operations. They offer great International Journal of Production Research 1994;32(4):833–49.
[12] Jawahir IS, Wang X. Development of hybrid predictive models and optimization
potentials for a holistic view and improved design of cutting processes.
techniques for machining operations. Journal of Materials Processing Technology
Process characteristics and loads, resulting tool wear, and produced 2007;185(1-3):46–59.
surfaces properties can be predicted via ML depending on cutting con [13] Glatt M, Hotz H, Kölsch P, Mukherjee A, Kirsch B, Aurich JC. Predicting the
ditions, the used machine tool, and other process data for a variety of martensite content of metastable austenitic steels after cryogenic turning using
machine learning. The International Journal of Advanced Manufacturing
machining operations including milling, turning, drilling, and grinding. Technology 2021;115:749–57.
Quantities, such as cutting forces or tool wear are reported to be used [14] Hegab H, Hassan M, Rawat S, Sadek A, Attia H. A smart tool wear prediction
both as input and output for ML models, which is the prerequisite for model in drilling of woven composites. The International Journal of Advanced
Manufacturing Technology 2020;110:2881–92.
prediction and optimisation along the whole process chain with AI. The [15] Aggogeri F, Pellegrini N, Tagliani FL. Recent advances on machine learning
predominant scenarios in the studied ML applications were found to be applications in machining processes. Applied Sciences 2021;11(18):8764.
supervised regression and classification, which refers to learning a [16] Sousa V, Silva FJG, Fecheira JS, Lopes HM, Martinho RP, Casais RB. Accessing the
cutting forces in machining processes: an overview. Procedia Manufacturing
predictor function for a real-valued or discrete process information using 2020;51:787–94.
labelled data. In general, the growing importance of DL due to improved [17] Du Preez A, Oosthuizen GA. Machine learning in cutting processes as enabler for
storage and calculation capacities and the emergence of large data smart sustainable manufacturing. Procedia Manufacturing 2019;33:810–7.
[18] Jawahir IS, Schoop J, Kaynak Y, Balaji AK, Ghosh R, Lu T. Progress toward
volumes with high frequency and dimension is reflected by the modelling and optimization of sustainable machining processes. Journal of
increasing deployment of ANNs within ML applications in machining. Manufacturing Science and Engineering 2020;142(11):110811.
However, the majority of the reviewed literature reports on a successful [19] Arinez JF, Chang Q, Gao RX, Xu C, Zhang J. Artificial intelligence in advanced
manufacturing: current status and future outlook. Journal of Manufacturing
generation of a single prediction model and its evaluation without an
Science and Engineering 2020;142(11):110804.
explicit optimisation step regarding a process-related objective function. [20] Chuo YS, Lee JW, Mun CH, Noh IW, Rezvani S, Kim DC, Lee J, Lee SW, Park SS.
ML predictions are very useful for condition monitoring and predictive Artificial intelligence enabled smart machining and machine tools. Journal of
maintenance. The actual ML models can serve as objective functions, Mechanical Science and Technology 2022;36(1):1–23.
[21] Weichert D, Link P, Stoll A, Rüping S, Ihlenfeldt S, Wrobel S. A review of machine
however, they do not directly lead to a process improvement. To this learning for the optimization of production processes. The International Journal
aim, mainly data-based evolutionary optimisation algorithms such as of Advanced Manufacturing Technology 2019;104(5-8):1889–902.
51
[22] Sarikaya E, Brockhaus B, Fertig A, Ranzau H, Stanula P, Walther J. Data driven [51] Newman ST, Zhu Z, Dhokia V, Shokrani A. Process planning for additive and
production–application fields. Solutions and Benefits 2021. https://doi.org/ subtractive manufacturing technologies. CIRP Annals 2015;64(1):467–70.
10.26083/tuprints-00017874. [52] Liao Z, la Monaca A, Murray J, Speidel A, Ushmaev D, Clare A, Axinte D,
[23] Teti R, Mourtzis D, D’Addona DM, Caggiano A. Process monitoring of machining. M’Saoubi R. Surface integrity in metal machining-Part I: Fundamentals of surface
CIRP Annals 2022;71(2):529–52. characteristics and formation mechanisms. International Journal of Machine
[24] Bakhtiyari AN, Wang Z, Wang L, Zheng H. A review on applications of artificial Tools and Manufacture 2021;162:103687.
intelligence in modeling and optimization of laser beam machining. Optics & [53] Trent EM, Wright PK. Metal cutting operations and terminology. Metal Cutting
Laser Technology 2021;135:106721. 2000:9–20.
[25] Pandiyan V, Shevchik S, Wasmer K, Castagne S, Tjahjowidodo T. Modelling and [54] Aurich JC, Mayer P, Kirsch B, Eifler D, Smaga M, Skorupski R. Characterization of
monitoring of abrasive finishing processes using artificial intelligence techniques: deformation induced surface hardening during cryogenic turning of AISI 347.
a review. Journal of Manufacturing Processes 2020;57:114–35. CIRP Annals 2014;63(1):65–8.
[26] Bello O, Teodoriu C, Yaqoob T, Oppelt J, Holzmann J, Obiwanne A. Application [55] Jawahir IS, Brinksmeier E, M’saoubi R, Aspinwall DK, Outeiro JC, Meyer D,
of artificial intelligence techniques in drilling system design and operations: a Umbrello D, Jayal AD. Surface integrity in material removal processes: Recent
state of the art review and future research pathways. Paper presented at the SPE advances. CIRP Annals 2011;60(2):603–26.
Nigeria Annual International Conference and Exhibition, Lagos, Nigeria, August [56] Balaji AK, Ghosh R, Fang XD, Stevenson R, Jawahir IS. Performance-based
2016.Paper Number: SPE-184320-MS. 2016. https://doi.org/10.2118/184320- predictive models and optimization methods for turning operations and
MSSPE. applications: part 2—assessment of chip forms/chip breakability. Journal of
[27] Noshi, C.I., Schubert, J.J. The role of machine learning in drilling operations; a Manufacturing Processes 2006;8(2):144–58.
review. In SPE/AAPG Eastern regional meeting. OnePetro, 2018. [57] Ertel W. Introduction to artificial intelligence. Springer; 2018.
[28] Torabi AJ, Er MJ, Li X, Lim BS, Zhai L, Oentaryo RJ, Peen GO, Zurada JM. [58] Russell SJ. Artificial intelligence a modern approach. Pearson Education, Inc;
A survey on artificial intelligence-based modelling techniques for high speed 2010.
milling processes. IEEE Systems Journal 2013;9(3):1069–80. [59] Kaplan A, Haenlein M. Siri, Siri, in my hand: Who’s the fairest in the land? On the
[29] Benardos PG, Vosniakos GC. Predicting surface roughness in machining: a review. interpretations, illustrations, and implications of artificial intelligence. Business
International Journal of Machine tools and manufacture 2003;43(8):833–44. Horizons 2019;62(1):15–25.
[30] Xiao W, Huang J, Wang B, Ji H. A systematic review of artificial intelligence in [60] Mitchell TM. Machine learning, 1. New York: McGraw-hill; 2007.
the detection of cutting tool breakage in machining operations. Measurement [61] Shalev-Shwartz S, Ben-David S. Understanding Machine Learning: from Theory to
2022:110748. Algorithms. Cambridge university press; 2014.
[31] Serin G, Sener B, Ozbayoglu AM, Unver HO. Review of tool condition monitoring [62] LeCun Y, Bengio Y, Hinton G. Deep learning. nature 2015;521(7553):436–44.
in machining and opportunities for deep learning. The International Journal of [63] Schulz H, Behnke S. Deep learning: layer-wise learning of feature hierarchies. KI-
Advanced Manufacturing Technology 2020;109:953–74. Künstliche Intelligenz 2012;26:357–63.
[32] Yusup N, Zain AM, Hashim SZM. Evolutionary techniques in optimizing [64] Duval A. Explainable artificial intelligence (XAI). MA4K9 Scholarly Report,
machining parameters: review and recent applications (2007–2011). Expert Mathematics Institute, The University of Warwick 2019:1–53.
Systems with Applications 2012;39(10):9909–27. [65] Xu C, Tao D, Xu C. A survey on multi-view learning. arXiv Preprint arXiv 2013;
[33] Monostori L, Márkus A, Van Brussel H, Westkämpfer E. Machine learning 1304:5634.
approaches to manufacturing. CIRP Annals 1996;45(2):675–712. [66] Kaelbling LP, Littman ML, Moore AW. Reinforcement learning: a survey. Journal
[34] Byrne G, Dornfeld D, Inasaki I, Ketteler G, König W, Teti R. Tool condition of Artificial Intelligence Research 1996;4:237–85.
monitoring (TCM)—the status of research and industrial application. CIRP Annals [67] Pan SJ, Yang Q. A survey on transfer learning. IEEE Transactions on Knowledge
1995;44(2):541–67. and data Engineering 2010;22(10):1345–59.
[35] Yerramareddy S, Lu SC, Arnold KF. Developing empirical models from [68] Unver HO, Sener B. A novel transfer learning framework for chatter detection
observational data using artificial neural networks. Journal of Intelligent using convolutional neural networks. Journal of Intelligent Manufacturing 2021:
Manufacturing 1993;4:33–41. 1–20.
[36] Azouzi R, Guillot M. On-line prediction of surface finish and dimensional [69] Boser BE, Guyon IM, Vapnik VN. A training algorithm for optimal margin
deviation in turning using neural network based sensor fusion. International classifiers. In Proceedings of the fifth Annual Workshop on Computational
Journal of Machine Tools and Manufacture 1997;37(9):1201–17. Learning Theory 1992:144–52.
[37] Karpuschewski B, Wehmeier M, Inasaki I. Grinding monitoring system based on [70] Cortes C, Vapnik V. Support-vector networks. Machine Learning 1995;20:273–97.
power and acoustic emission sensors. CIRP Annals 2000;49(1):235–40. [71] Burges CJ. A tutorial on support vector machines for pattern recognition. Data
[38] Park KS, Kim SH. Artificial intelligence approaches to determination of CNC Mining and Knowledge discovery 1998;2(2):121–67.
machining parameters in manufacturing: a review. Artificial Intelligence in [72] Smola AJ, Schölkopf B. A tutorial on support vector regression. Statistics and
Engineering 1998;12(1-2):127–34. Computing 2004;14:199–222.
[39] Mokhtari Homami R, Fadaei Tehrani A, Mirzadeh H, Movahedi B, Azimifar F. [73] Cristianini N, Shawe-Taylor J. An introduction to support vector machines and
Optimization of turning process using artificial intelligence technology. The other kernel-based learning methods. Cambridge university press; 2000.
International Journal of Advanced Manufacturing Technology 2014;70:1205–17. [74] Platt, J., 1998. Sequential minimal optimization: A fast algorithm for training
[40] Sivanaga Malleswara Rao S, Venkata Rao K, Hemachandra Reddy K, support vector machines. https://www.microsoft.com/en-us/research/publi
Parameswara Rao CV. Prediction and optimization of process parameters in wire cation/sequential-minimal-optimization-a-fast-algorithm-for-training-support-ve
cut electric discharge machining for High-speed steel (HSS). International Journal ctor-machines/1998.
of Computers and Applications 2017;39(3):140–7. [75] Schölkopf B, Smola AJ. Learning with kernels: support vector machines,
[41] Wang X, Jawahir IS. Optimization of multi-pass turning operations using genetic regularization, optimization, and beyond. MIT press; 2002.
algorithms for the selection of cutting conditions and cutting tools with tool-wear [76] Hastie T, Tibshirani R, Friedman JH, Friedman JH. The elements of statistical
effect. International Journal of Production Research 2005;43(17):3543–59. learning: data mining, inference, and prediction, 2. New York: springer; 2009.
[42] Wang X, Da ZJ, Balaji AK, Jawahir IS. Performance-based predictive models and p. 1–758.
optimization methods for turning operations and applications: Part 3—optimum [77] Grossi E, Buscema M. Introduction to artificial neural networks. European
cutting conditions and selection of cutting tools. Journal of Manufacturing Journal of gastroenterology & hepatology 2007;19(12):1046–54.
Processes 2007;9(1):61–74. [78] Krenker A, Bešter J, Kos A. Introduction to the artificial neural networks.
[43] Deb K, Datta R. Hybrid evolutionary multi-objective optimization and analysis of Artificial Neural Networks: Methodological Advances and Biomedical
machining operations. Engineering Optimization 2012;44(6):685–706. Applications. InTech,; 2011. p. 1–18.
[44] Sardinas RQ, Santana MR, Brindis EA. Genetic algorithm-based multi-objective [79] O’Shea K, Nash R. An introduction to convolutional Neural Networks arXiv
optimization of cutting parameters in turning processes. Engineering Applications Preprint arXiv 2015;1511:08458.
of Artificial Intelligence 2006;19(2):127–33. [80] Breiman L. Random forests. Machine Learning 2001;45:5–32.
[45] D’addona DM, Teti R. Genetic algorithm-based optimization of cutting [81] Cherkassky V, Mulier FM. Learning from data: concepts, theory, and methods.
parameters in turning processes. Procedia CIRP 2013;7:323–8. John Wiley & Sons; 2007.
[46] Rao RV, Pawar PJ, Shankar R. Multi-objective optimization of electrochemical [82] Phillips-Wren, G., Jain, L.. Artificial intelligence for decision making. In
machining process parameters using a particle swarm optimization algorithm. Knowledge-Based Intelligent Information and Engineering Systems: 10th
Proceedings of the Institution of Mechanical Engineers, Part B: Journal of International Conference, KES 2006, Bournemouth, UK, October 9–11, 2006.
Engineering Manufacture 2008;222(8):949–58. Proceedings, Part II 10 (pp. 531–536). Springer Berlin Heidelberg; 2006.
[47] Bagaber SA, Yusoff AR. Multi-objective optimization of cutting parameters to [83] Pomerol JC. Artificial intelligence and human decision making. European Journal
minimize power consumption in dry turning of stainless steel 316. Journal of of Operational Research 1997;99(1):3–25.
Cleaner Production 2017;157:30–46. [84] Carlsson C, Fullér R, Carlsson C, Fullér R. Fuzzy Reasoning for Fuzzy
[48] Yan J, Li L. Multi-objective optimization of milling parameters–the trade-offs Optimization. Fuzzy Reasoning in Decision Making and Optimization 2002:
between energy, production rate and cutting quality. Journal of Cleaner 157–205.
Production 2013;52:462–71. [85] Elsheikh AH, Muthuramalingam T, Shanmugan S, Ibrahim AMM, Ramesh B,
[49] Kant G, Sangwan KS. Prediction and optimization of machining parameters for Khoshaim AB, Moustafa EB, Bedairi B, Panchal H, Sathyamurthy R. Fine-tuned
minimizing power consumption and surface roughness in machining. Journal of artificial intelligence model using pigeon optimizer for prediction of residual
Cleaner Production 2014;83:151–64. stresses during turning of Inconel 718. Journal of Materials Research and
[50] Wang Q, Liu F, Wang X. Multi-objective optimization of machining parameters Technology 2021;15:3622–34.
considering energy consumption. The International Journal of Advanced [86] Guo W, Wu C, Ding Z, Zhou Q. Prediction of surface roughness based on a hybrid
Manufacturing Technology 2014;71:1133–42. feature selection method and long short-term memory network in grinding. The
52
International Journal of Advanced Manufacturing Technology 2021;112: response surface methodology for the use in soft sensors. Procedia CIRP 2022;
2853–71. 108:270–5.
[87] Nag K, Pal NR. Genetic programming for classification and feature selection. [116] Masoudi S, Sima M, Tolouei-Rad M. Comparative study of ANN and ANFIS models
Evolutionary and swarm Intelligence algorithms 2019:119–41. for predicting temperature in machining. Journal of Engineering Science and
[88] Nocedal, J., Wright, S.J. eds., 1999. Numerical optimization. New York, NY: Technology 2018;13(1):211–25.
Springer New York. [117] Dutta S, Pal SK, Sen R. On-machine tool prediction of flank wear from machined
[89] ÖZKARACA O. A review on usage of optimization methods in geothermal power surface images using texture analyses and support vector regression. Precision
generation. Mugla Journal of Science and Technology 2018;4(1):130–6. Engineering 2016;43:34–42.
[90] Cavazzuti, M. Deterministic optimization. in Optimization Methods: From Theory [118] Albertelli P, Goletti M, Torta M, Salehi M, Monno M. Model-based broadband
to Design Scientific and Technological Aspects in Mechanics, pp. 77–102; 2013. estimation of cutting forces and tool vibration in milling through in-process
[91] Kunche P, Reddy KVVS. Metaheuristic applications to speech enhancement. indirect multiple-sensors measurements. The International Journal of Advanced
Springer International Publishing; 2016. p. 7–16. Manufacturing Technology 2016;82:779–96.
[92] Thede SM. An introduction to genetic algorithms. Journal of Computing Sciences [119] Fertig A, Bauerdick C, Weigold M. In-process quality monitoring during turning
in Colleges 2004;20(1):115–23. based on high frequency machine data. In Proceedings of the Machining
[93] Verma S, Pant M, Snasel V. A comprehensive review on NSGA-II for multi- Innovations Conference (Micilor) 2020.
objective combinatorial optimization problems. IEEE Access 2021;9:57757–91. [120] Liang YC, Wang S, Li WD, Lu X. Data-driven anomaly diagnosis for machining
[94] Yeganefar A, Niknam SA, Asadi R. The use of support vector machine, neural processes. Engineering 2019;5(4):646–52.
network, and regression analysis to predict and optimize surface roughness and [121] Altendeitering M, Dübler S. Scalable detection of concept drift: a learning
cutting forces in milling. The International Journal of Advanced Manufacturing technique based on support vector machines. Procedia Manufacturing 2020;51:
Technology 2019;105:951–65. 400–7.
[95] Altintas Y, Yang J, Kilic ZM. Virtual prediction and constraint of contour errors [122] Fertig A, Grau L, Altmannsberger M, Weigold M. Tool condition monitoring and
induced by cutting force disturbances on multi-axis CNC machine tools. CIRP tool defect detection for end mills based on high-frequency machine tool data.
Annals 2019;68(1):377–80. MM Science Journal 2021:5160–6.
[96] Yang, X.S.Flower pollination algorithm for global optimization. In [123] Wang J, Xie J, Zhao R, Zhang L, Duan L. Multisensory fusion based virtual tool
Unconventional Computation and Natural Computation: 11th International wear sensing for ubiquitous manufacturing. Robotics and computer-Integrated
Conference, UCNC 2012, Orléan, France, September 3–7, 2012. Proceedings 11 Manufacturing 2017;45:47–58.
(pp. 240–249). Springer Berlin Heidelberg, 2012. [124] Giriraj B, Raja VP, Gandhinadhan R, Ganeshkumar R. Prediction of tool wear in
[97] Goel, S. September. Pigeon optimization algorithm: A novel approach for solving high speed machining using acoustic emission technique and neural network.
optimization problems. In 2014 International Conference on Data Mining and Indian Journal of Engineering Materials Science 2006;13:275.
Intelligent Computing (ICDMIC) (pp. 1–5). IEEE; 2014. [125] Lin SC, Lin RJ. Tool wear monitoring in face milling using force signals. Wear
[98] Loucks DP. Fuzzy optimization. Public Systems Modelling: Methods for 1996;198(1-2):136–42.
Identifying and Evaluating Alternative Plans and Policies. Cham: Springer [126] Hsieh WH, Lu MC, Chiou SJ. Application of backpropagation neural network for
International Publishing; 2022. p. 223–35. spindle vibration-based tool wear monitoring in micro-milling. The International
[99] Alexandre FA, Lopes WN, Lofrano Dotto FR, Ferreira FI, Aguiar PR, Bianchi EC, Journal of Advanced Manufacturing Technology 2012;61:53–61.
Lopes JC. Tool condition monitoring of aluminium oxide grinding wheel using AE [127] Chen SL, Jen YW. Data fusion neural network for tool condition monitoring in
and fuzzy model. The International Journal of Advanced Manufacturing CNC milling machining. International Journal of Machine tools and manufacture
Technology 2018;96:67–79. 2000;40(3):381–400.
[100] Anand G, Alagumurthi N, Elansezhian R, Palanikumar K, Venkateshwaran N. [128] Bustillo A, Pimenov DY, Mia M, Kapłonek W. Machine-learning for automatic
Investigation of drilling parameters on hybrid polymer composites using grey prediction of flatness deviation considering the wear of the face mill teeth.
relational analysis, regression, fuzzy logic, and ANN models. Journal of the Journal of Intelligent Manufacturing 2021;32(3):895–912.
Brazilian Society of Mechanical Sciences and Engineering 2018;40:1–20. [129] Zhang H, Deng Z, Fu Y, Lv L, Yan C. A process parameters optimization method of
[101] Wu J, Su Y, Cheng Y, Shao X, Deng C, Liu C. Multi-sensor information fusion for multi-pass dry milling for high efficiency, low energy and low carbon emissions.
remaining useful life prediction of machining tools by adaptive network based Journal of Cleaner Production 2017;148:174–84.
fuzzy inference system. Applied Soft Computing 2018;68:13–23. [130] Mekid S, Pruschek P, Hernandez J. Beyond intelligent manufacturing: a new
[102] Yang LD, Chen JC, Chow HM, Lin CT. Fuzzy-nets-based in-process surface generation of flexible intelligent NC machines. Mechanism and Machine Theory
roughness adaptive control system in end-milling operations. The International 2009;44(2):466–76.
Journal of Advanced Manufacturing Technology 2006;28:236–48. [131] Wuest T, Weimer D, Irgens C, Thoben KD. Machine learning in manufacturing:
[103] Močkus J. On Bayesian methods for seeking the extremum. Optimization advantages, challenges, and applications. Production & Manufacturing Research
Techniques IFIP Technical Conference: Novosibirsk, July 1–7, 1974. Springer 2016;4(1):23–45.
Berlin Heidelberg; 1975. p. 400–4. [132] Arnold F, Hänel A, Nestler A, Brosius A. New approaches for the determination of
[104] Shahriari B, Swersky K, Wang Z, Adams RP, De Freitas N. Taking the human out specific values for process models in machining using artificial neural networks.
of the loop: a review of Bayesian optimization. Proceedings of the IEEE 2015;104 Procedia Manufacturing 2017;11:1463–70.
(1):148–75. [133] Wenkler E, Arnold F, Hänel A, Nestler A, Brosius A. Intelligent characteristic
[105] Tamura K, Yasuda K. Spiral dynamics inspired optimization. Journal of Advanced value determination for cutting processes based on machine learning. Procedia
Computational Intelligence and Intelligent Informatics 2011;15(8):1116–22. CIRP 2019;79:9–14.
[106] Moré, J.J. The Levenberg-Marquardt algorithm: implementation and theory. In [134] Vaishnav S, Agarwal A, Desai KA. Machine learning-based instantaneous cutting
Numerical Analysis: Proceedings of the Biennial Conference Held at Dundee, June force model for end milling operation. Journal of Intelligent Manufacturing 2020;
28–July 1, 1977 (pp. 105–116). Berlin, Heidelberg: Springer Berlin Heidelberg, 31:1353–66.
2006. [135] Hashemitaheri M, Mekarthy SMR, Cherukuri H. Prediction of specific cutting
[107] Chen J, Chen H, Xu J, Wang J, Zhang X, Chen X. Acoustic signal-based tool forces and maximum tool temperatures in orthogonal machining by support
condition monitoring in belt grinding of nickel-based superalloys using RF vector and Gaussian process regression methods. Procedia Manufacturing 2020;
classifier and MLR algorithm. The International Journal of Advanced 48:1000–8.
Manufacturing Technology 2018;98:859–72. [136] Yesilli MC, Khasawneh FA, Otto A. Topological feature vectors for chatter
[108] Mirifar S, Kadivar M, Azarhoushang B. First steps through intelligent grinding detection in turning processes. The International Journal of Advanced
using machine learning via integrated acoustic emission sensors. Journal of Manufacturing Technology 2022:1–27.
Manufacturing and Materials Processing 2020;4(2):35. [137] Jurkovic Z, Cukor G, Brezocnik M, Brajkovic T. A comparison of machine learning
[109] Pandiyan V, Caesarendra W, Tjahjowidodo T, Tan HH. In-process tool condition methods for cutting parameters prediction in high speed turning process. Journal
monitoring in compliant abrasive belt grinding process using support vector of Intelligent Manufacturing 2018;29:1683–93.
machine and genetic algorithm. Journal of Manufacturing Processes 2018;31: [138] Saadallah A, Finkeldey F, Morik K, Wiederkehr P. Stability prediction in milling
199–213. processes using a simulation-based Machine Learning approach. Procedia CIRP
[110] Sauter E, Sarikaya E, Winter M, Wegener K. In-process detection of grinding burn 2018;72:1493–8.
using machine learning. The International Journal of Advanced Manufacturing [139] Wang J, Zou B, Liu M, Li Y, Ding H, Xue K. Milling force prediction model based
Technology 2021;115:2281–97. on transfer learning and neural network. Journal of Intelligent Manufacturing
[111] Mahata S, Shakya P, Babu NR, Prakasam PK. In-process characterization of 2021;32:947–56.
surface finish in cylindrical grinding process using vibration and power signals. [140] Dörr M, Ott L, Matthiesen S, Gwosch T. Prediction of tool forces in manual
Procedia CIRP 2020;88:335–40. grinding using consumer-grade sensors and machine learning. Sensors 2021;21
[112] Du C, Ho CL, Kaminski J. Prediction of product roughness, profile, and roundness (21):7147.
using machine learning techniques for a hard turning process. Advances in [141] Maier M, Rupenyan A, Bobst C, Wegener K. Self-optimizing grinding machines
Manufacturing 2021;9:206–15. using Gaussian process models and constrained Bayesian optimization. The
[113] Alajmi MS, Almeshal AM. Modeling of cutting force in the turning of AISI 4340 International Journal of Advanced Manufacturing Technology 2020;108:539–52.
using Gaussian process regression algorithm. Applied Sciences 2021;11(9):4055. [142] Marani M, Zeinali M, Kouam J, Songmene V, Mechefske CK. Prediction of cutting
[114] Charalampous P. Prediction of cutting forces in milling using machine learning tool wear during a turning process using artificial intelligence techniques. The
algorithms and finite element analysis. Journal of Materials Engineering and International Journal of Advanced Manufacturing Technology 2020;111:505–15.
Performance 2021;30:2002–13. [143] Xu L, Huang C, Li C, Wang J, Liu H, Wang X. Estimation of tool wear and
[115] Grossmann F, Basten S, Kirsch B, Ankener W, Smaga M, Beck T, Uebel J, Seewig J, optimization of cutting parameters based on novel ANFIS-PSO method toward
Aurich JC. Predictive modelling of cryogenic hard turning of AISI 52100 based on intelligent machining. Journal of Intelligent Manufacturing 2021;32:77–90.
53
[144] Oraby SE, Hayhurst DR. Tool life determination based on the measurement of [170] Holmberg J, Wretland A, Hammersberg P, Berglund J, Suárez A, Beno T. Surface
wear and tool force ratio variation. International Journal of Machine Tools and integrity investigations for prediction of fatigue properties after machining of
Manufacture 2004;44(12-13):1261–9. alloy 718. International Journal of Fatigue 2021;144:106059.
[145] de Oliveira JG, Dornfeld DA. Application of AE contact sensing in reliable [171] Cherukuri H, Perez-Bernabeu E, Selles MA, Schmitz TL. A neural network
grinding monitoring. CIRP Annals 2001;50(1):217–20. approach for chatter prediction in turning. Procedia Manufacturing 2019;34:
[146] Shen Y, Yang F, Habibullah MS, Ahmed J, Das AK, Zhou Y, Ho CL. Predicting tool 885–92.
wear size across multi-cutting conditions using advanced machine learning [172] Zhou T, He L, Wu J, Du F, Zou Z. Prediction of surface roughness of 304 stainless
techniques. Journal of Intelligent Manufacturing 2021;32:1753–66. steel and multi-objective optimization of cutting parameters based on GA-GBRT.
[147] Wang N, Zhang G, Ren L, Pang W, Li Y. Novel monitoring method for belt wear Applied Sciences 2019;9(18):3684.
state based on machine vision and image processing under grinding parameter [173] Vasanth XA, Paul PS, Varadarajan AS. A neural network model to predict surface
variation. The International Journal of Advanced Manufacturing Technology roughness during turning of hardened SS410 steel. International Journal of
2022;122(1):87–101. System Assurance Engineering and Management 2020;11:704–15.
[148] Lee ET, Fan Z, Sencer B. Estimation of cBN grinding wheel condition using image [174] Nguyen VH, Le TT, Truong HS, Le MV, Ngo VL, Nguyen AT, Nguyen HQ. Applying
sensor. Procedia Manufacturing 2021;53:286–92. bayesian optimization for machine learning models in predicting the surface
[149] Mikołajczyk T, Nowicki K, Bustillo A, Pimenov DY. Predicting tool life in turning roughness in single-point diamond turning polycarbonate. Mathematical
operations using neural networks and image processing. Mechanical Systems and Problems in Engineering 2021;2021:1–16.
Signal Processing 2018;104:503–13. [175] Dong P, Peng H, Cheng X, Xing Y, Tang W, Zhou X. Semi-empirical prediction of
[150] Lutz B, Kisskalt D, Regulin D, Reisch R, Schiffler A, Franke J. Evaluation of deep residual stress profiles in machining IN718 alloy using bimodal Gaussian curve.
learning for semantic image segmentation in tool condition monitoring. 2019 Materials 2019;12(23):3864.
18th IEEE international conference on machine learning and applications [176] Peng H, Dong P, Cheng X, Zhang C, Tang W, Xing Y, Zhou X. Semi-empirical
(ICMLA). IEEE; 2019. p. 2008–13. prediction of residual stress distributions introduced by turning Inconel 718 alloy
[151] Segreto T, D’Addona D, Teti R. Tool wear estimation in turning of Inconel 718 based on Lorentz function. Materials 2020;13(19):4341.
based on wavelet sensor signal analysis and machine learning paradigms. [177] Ouerhani N, Loehr B, Rizzotti-Kaddouri A, De Pinho Santo, D, Limat A,
Production Engineering 2020;14:693–705. Schinderholz P. Data-Driven Thermal Deviation Prediction in Turning Machine-
[152] Twardowski P, Wiciak-Pikuła M. Prediction of tool wear using artificial neural Tool-A Comparative Analysis of Machine Learning Algorithms. Procedia
networks during turning of hardened steel. Materials 2019;12(19):3091. Computer Science 2022;200:185–93.
[153] de Farias A, de Almeida SLR, Delijaicov S, Seriacopi V, Bordinassi EC. Simple [178] Möhring HC, Eschelbacher S, Georgi P. Machine learning approaches for real-time
machine learning allied with data-driven methods for monitoring tool wear in monitoring and evaluation of surface roughness using a sensory milling tool.
machining processes. The International Journal of Advanced Manufacturing Procedia CIRP 2021;102:264–9.
Technology 2020;109(9-12):2491–501. [179] Reimer A, Luo X. Prediction of residual stress in precision milling of AISI H13
[154] Huang PM, Lee CH. Estimation of tool wear and surface roughness development steel. Procedia Cirp 2018;71:329–34.
using deep learning and sensors fusion. Sensors 2021;21(16):5338. [180] Uhlmann E, Holznagel T, Schehl P, Bode Y. Machine learning of surface layer
[155] Yuan J, Liu L, Yang Z, Bo J, Zhang Y. Tool wear condition monitoring by property prediction for milling operations. Journal of Manufacturing and
combining spindle motor current signal analysis and machined surface image Materials Processing 2021;5(4):104.
processing. The International Journal of Advanced Manufacturing Technology [181] Codjo L, Jaafar M, Makich H, Knittel D, Nouari M. Honeycomb Core Milling
2021;116:2697–709. Diagnosis using Machine Learning in the Industry 4.0 Framework. 2018 IEEE
[156] Li Z, Liu R, Wu D. Data-driven smart manufacturing: tool wear monitoring with 23rd International Conference on Emerging Technologies and Factory
audio signals and machine learning. Journal of Manufacturing Processes 2019;48: Automation (ETFA), Vol. 1. IEEE; 2018. p. 1251–8.
66–76. [182] Safarzadeh H, Leonesio M, Bianchi G, Monno M. Roundness prediction in
[157] Hua J, Li Y, Mou W, Liu C. An accurate cutting tool wear prediction method under centreless grinding using physics-enhanced machine learning techniques. The
different cutting conditions based on continual learning. Proceedings of the International Journal of Advanced Manufacturing Technology 2021;112:
Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture 1051–63.
2022;236(1-2):123–31. [183] Sauter E, Winter M, Wegener K. Analysis of robustness and transferability in
[158] Bleicher F, Ramsauer CM, Oswald R, Leder N, Schoerghofer P. Method for feature-based grinding burn detection. The International Journal of Advanced
determining edge chipping in milling based on tool holder vibration Manufacturing Technology 2022;120(3-4):2587–602.
measurements. CIRP Annals 2020;69(1):101–4. [184] Schorr S, Möller M, Heib J, Bähre D. Quality prediction of drilled and reamed
[159] Twardowski P, Tabaszewski M, Wiciak–Pikuła M, Felusiak-Czyryca A. bores based on torque measurements and the machine learning method of
Identification of tool wear using acoustic emission signal and machine learning random forest. Procedia Manufacturing 2020;48:894–901.
methods. Precision Engineering 2021;72:738–44. [185] Ziegenbein, A., Fertig, A., Metternich, J., Weigold, M. Data-based process analysis
[160] Gouarir A, Martínez-Arellano G, Terrazas G, Benardos P, Ratchev SJPC. In-process in machining production: Case study for quality determination in a drilling
tool wear prediction system based on machine learning techniques and force process, 2020; 93, p. 1472.
analysis. Procedia CIRP 2018;77:501–4. [186] Wang, X., Eisseler, R., Moehring, H.-C.Prediction and optimization of machining
[161] Schwenzer M, Miura K, Bergs T. Machine learning for tool wear classification in results and parameters in drilling by using Bayesian networks, 2020. 14, p. 373.
milling based on force and current sensors. IOP conference series: materials [187] Stoll A, Pierschel N, Wenzel K, Langer T. Process Control in a Press Hardening
science and engineering, Vol. 520. IOP Publishing; 2019, 012009. Production Line with Numerous Process Variables and Quality Criteria. Machine
[162] Terrazas G, Martínez-Arellano G, Benardos P, Ratchev S. Online tool wear Learning for Cyber Physical Systems. Berlin, Heidelberg: Springer Berlin
classification during dry machining using real time cutting force measurements Heidelberg; 2019. p. 77.
and a CNN approach. Journal of Manufacturing and Materials Processing 2018;2 [188] Pawanr S, Garg GK, Routroy S. Multi-objective optimization of machining
(4):72. parameters to minimize surface roughness and power consumption using TOPSIS.
[163] Zhang X, Han C, Luo M, Zhang D. Tool wear monitoring for complex part milling Procedia CIRP 2019;86:116–20.
based on deep learning. Applied Sciences 2020;10(19):6916. [189] Dhabale R, Jatti VS, Singh TP. Multi-objective optimization of turning process
[164] Oo H, Wang W, Liu Z. Tool wear monitoring system in belt grinding based on during machining of AlMg1SiCu using non-dominated sorted genetic algorithm.
image-processing techniques. The International Journal of Advanced Procedia Materials Science 2014;6:961–6.
Manufacturing Technology 2020;111:2215–29. [190] Rao RV, Rai DP, Balic J. Multi-objective optimization of machining and micro-
[165] Oberlé R, Schorr S, Yi L, Glatt M, Bähre D, Aurich JC. A use case to implement machining processes using non-dominated sorting teaching–learning-based
machine learning for life time prediction of manufacturing tools. Procedia CIRP optimization algorithm. Journal of Intelligent Manufacturing 2018;29:1715–37.
2020;93:1484–9. [191] Ghahramani M, Qiao Y, Zhou M, Wu N. An AI-based Multi-objective Optimization
[166] Xie N, Zhou J, Zheng B. An energy-based modeling and prediction approach for Approach for Monitoring Manufacturing Processes. 2021 International
surface roughness in turning. The International Journal of Advanced Conference on Cyber-Physical Social Intelligence (ICCSI). IEEE; 2021. p. 1–5.
Manufacturing Technology 2018;96:2293–306. [192] Beyca OF, Rao PK, Kong Z, Bukkapatnam ST, Komanduri R. Heterogeneous sensor
[167] Khoshaim AB, Elsheikh AH, Moustafa EB, Basha M, Mosleh AO. Prediction of data fusion approach for real-time monitoring in ultraprecision machining (UPM)
residual stresses in turning of pure iron using artificial intelligence-based process using non-parametric Bayesian clustering and evidence theory. IEEE
methods. Journal of Materials Research and Technology 2021;11:2181–94. Transactions on Automation Science and Engineering 2015;13(2):1033–44.
[168] Pimenov DY, Bustillo A, Mikolajczyk T. Artificial intelligence for automatic [193] Junejo F, Amin I, Hassan M, Ahmed A, Hameed S. The application of artificial
prediction of required surface roughness by monitoring wear on face mill teeth. intelligence in grinding operation using sensor fusion. Geomate Journal 2017;12
Journal of Intelligent Manufacturing 2018;29:1045–61. (30):11–8.
[169] Tampu C, Chirita B, Cristea I, Zichil V, Schnakovszky C, Herghelegiu E, Carausu C. [194] Fertig A, Kohn O, Brockhaus B, Weigold M. Consistent contextualisation of
Influence of cutting parameters on surface hardness in milling of AL6061T6. IOP process and quality information for machining processes. Production at the
Conference Series: Materials Science and Engineering, Vol. 916. IOP Publishing; Leading Edge of Technology: Proceedings of the 11th Congress of the German
2020, 012118. Academic Association for Production Technology (WGP), Dresden, September
2021. Cham: Springer International Publishing; 2021. p. 195–202.
54

Generative AI in Manufacturing

Uploaded by

Copyright:

Available Formats

Generative AI in Manufacturing

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Generative AI in Manufacturing

Uploaded by

Copyright:

Available Formats

CIRP Journal of Manufacturing Science and Technology 50 (2024) 40–54

Contents lists available at ScienceDirect

CIRP Journal of Manufacturing Science and Technology

AI-based optimisation of total machining performance: A review

3. Fundamentals of machine learning and optimisation

In Section 3.1 necessary terms and concepts from ML theory,

5. Review on recent ML modelling in machining

Optimised process parameters can maximise the efficiency of the

Table 4 design and the parameters chosen. Different combinations of process

v. investigating optimisation techniques for the training of ML

According to the literature research, the data-based adjustment of

You might also like