Expert Systems With Applications 249 (2024) 123578

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

Explainable prediction of surface roughness in multi-jet polishing based on

ensemble regression and differential evolution method
Yueyue Wang a, Zongbao He a, b, Shutong Xie a, *, Ruoxin Wang b, Zili Zhang b, Shimin Liu c,
Suiyan Shang b, Pai Zheng b, c, Chunjin Wang b, *
a
School of Computer Engineering, Jimei University, Xiamen 361021, China
b
State Key Laboratory of Ultra-precision Machining Technology, Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom,
Kowloon, Hong Kong, China
c
Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong, China

A R T I C L E I N F O A B S T R A C T

Keywords: Surface roughness is a critical parameter for quantifying the surface quality of a workpiece. Accurate surface
Surface roughness prediction roughness prediction can significantly influence the overall quality of the final components by reducing costs and
Polishing enhancing productivity. Although various prediction methods have been explored in previous research, less
Explainable analysis
attention has been given to explainable prediction methods for surface roughness. This paper proposes a surface
Ensemble learning
Differential evolution algorithm
roughness prediction model based on the Differential Evolution (DE) algorithm with ensemble learning. It aims
Ultra-precision machining to elucidate the intrinsic correlation between processing parameters and surface roughness values in Multi-Jet
Polishing (MJP) using explainable analysis methods. The proposed Ensemble Regression with Differential Evo­
lution (ERDE) algorithm comprises two main modules: data processing and analytics, and ensemble model
prediction. These modules investigate the surface roughness of MJP from the data and model levels, respectively.
To validate the effectiveness of ERDE, MJP experiments were conducted on three-dimensional printed compo­
nents of 316L stainless steel. The experimental data indicated that ERDE outperforms other existing algorithms,
reducing the mean absolute percentage error by approximately 42 %.

intricate surfaces, including lens arrays to achieve nanometer surface

roughness, a critical requirement for numerous high-added-value
1. Introduction
products (Cheung et al., 2018). However, traditional surface rough­
ness inspection methods require a manual, mechanical, and tedious
Surface roughness stands as a pivotal gauge of manufacturing quality
process, while predictive models can effectively reduce time. By accu­
and significantly affects the surface properties of workpieces, such as
rately predicting the surface roughness, the test time during machining
corrosion resistance, life, and fatigue behavior. Surface roughness is
can be effectively reduced, thus improving productivity. Therefore, the
particularly important when it comes to 3D printing technology,
prediction of surface roughness is necessary for the MJP.
prompting the development of various post-processing techniques
Numerous studies have investigated methods for predicting surface
aimed at enhancing the printed surface quality. These techniques
roughness across various processes. Through the research, the surface
include mechanical polishing, electrochemical polishing, laser polish­
roughness prediction studies for polishing processes were categorized
ing, fluid jet polishing, and more. Wang et al. (2017) proposed a multi-
into two types: surface roughness prediction methods for material
jet polishing (MJP) process in 2017 that significantly improves polishing
removal processes and non-material removal processes. In the field of
efficiency. MJP employs multiple pressurized polishing slurry jets that
surface engineering, it is common to use the term “non-material
accelerate abrasive particles to impact the workpiece surface, resulting
removal” to describe surface treatment processes. Unlike physical
in an improved surface finish. It is capable of polishing extensive,

* Corresponding authors at: School of Computer Engineering, Jimei University, Xiamen, Fujian Province 361021, China (S. Xie). State Key Laboratory of Ultra-
precision Machining Technology, Department of Industrial and Systems Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong,
China (C. Wang).
E-mail addresses: [email protected] (Y. Wang), [email protected] (Z. He), [email protected] (S. Xie), [email protected]
(R. Wang), [email protected] (Z. Zhang), [email protected] (S. Liu), [email protected] (S. Shang), [email protected]
(P. Zheng), [email protected] (C. Wang).

https://doi.org/10.1016/j.eswa.2024.123578
Received 14 September 2023; Received in revised form 25 February 2024; Accepted 26 February 2024
Available online 28 February 2024
0957-4174/© 2024 Elsevier Ltd. All rights reserved.
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

Nomenclature yi
̂ The predicted value of the i-th data point (Eq. (3))
xi The individual target vector (Eq. (4))
DE Differential evolution D The dimensionality of the target vector xi (Eq. (4))
MJP Multi-jet polishing N The population size
ERDE Ensemble regression with differential evolution αi The weight of model i
ATPS Anhydrous-based thickened polishing slurry V The mutated individual
CVD Chemical vapor deposition X The initial and final individuals
RSM Response surface methodology vj The j-th mutated gene in V (Eq. (5))
MCF Magnetic compound fluid U The crossover individual
ELGA Ensemble learning by genetic algorithm fitness The fitness function
EDM Electrical Discharge Machining CR The crossover rate
MRR Material removal rate M The total number of models
RF Random forest xj The j-th initial gene in xi (Eq. (6))
ResCNN Residual Convolutional neural network αm The weight assigned to each model m
ANN Artificial neural network y
̂ The predicted surface roughness (Eq. (9))
HPO Hyperparametric optimization Raf The final surface roughness measurement results
RA Regression analysis f Feed rate
SVR Support vector regression P Pressure
XGBoost Extreme gradient boosting TO Tool offset
LASSO Least absolute shrinkage and selection operator d Step distance
ENR Elastic net regression RFE Recursive feature elimination
RR Ridge regression BSS Blind signal separation
LR Linear regression SBS Sequential backward selection
GBR Gradient boosting regression SFS Sequential forward selection
SGDR Stochastic gradient descent regression MSE Mean square error
ETR Extreme random forest MAPE Mean absolute percentage error
Min-Max Minimum-maximum FJP Fluid jet polishing
degree The number of controlling polynomials CoCr Cobalt Chrome
Z The transformed feature vector (Eq. (1)) mj Feature j
NF Number of features after the polynomial feature φj (val) SHAP value of feature j
transformation o Total number of features
fi The i-th individual in the population S Subset of excluded features mj
SHAP Shapley Additive exPlanation
k The number of samples (Eq. (3))

stripping means or material grinding techniques, this process can ach­ of processing parameters (e.g., abrasive particle size, polishing pressure,
ieve surface improvement through chemical methods, surface modifi­ speed, etc.) on surface roughness have been systematically investigated
cation, or the addition of new materials. Surface roughness prediction with the help of experimental data and theoretical principles. The pol­
for polished finishes in non-material removal processes requires ishing forces (including normal indentation force and shear force) in the
consideration of models related to chemical reaction kinetics and magnetorheological polishing process have been analyzed and modeled
abrasive deposition, among others. This involves precise modulation of by Peng et al. (2021). On this basis, a final surface roughness model is
surface quality by controlling the chemical composition in the polishing proposed to predict the polishing performance. Also, based on the
solution and the properties of the abrasive particles. The goals of such response surface methodology (RSM), Jian et al. (2022) tried to optimize
processes include improving surface quality, adjusting chemistry, real­ the parameters of the magnetic compound fluid (MCF) polishing process
izing functional coatings, and so on. Min et al. (2021) proposed an based on RSM. The study first identified the key parameters affecting
anhydrous-based thickened polishing slurry (ATPS) and modeled the surface roughness and then designed and conducted experiments using
material removal of ATPS for KDP crystal processing to enable the RSM. However, as a traditional method, the above experimental models
prediction of surface roughness. Wang et al. (Wang, 2023) proposed an are based on trial and error and rely on the experience of domain ex­
improved numerical model by extending the existing laser polishing perts. The process may take a lot of time, effort, and experimental
model and incorporating the effects of mass transfer, solvate-capillary resources.
forces and chemical reactions. This model is used to analyze the effect With the rise of machine learning and its successful application in
of ambient gases on the evolutionary mechanism of laser polishing various fields, more and more studies have adopted machine learning for
surface morphology and to predict the polished surface roughness of 304 surface roughness prediction modeling. They attempted to reveal the
stainless steel based on the model. Laser ablation is demonstrated to be a correlations between process variables and surface roughness by using
competent approach to processing chemical vapor deposition (CVD) machine learning algorithms. Wang et al. (2022) designed a robust
diamonds. By investigating the laser ablation strategy, Chen et al. surface roughness prediction model based on ensemble learning by ge­
(2022) developed an absorption-based laser energy distribution model netic algorithm (ELGA), which can be used to predict the surface
for evaluating surface roughness and ablation depth. roughness of 3D printed 316L stainless steel MJP. Fan et al. (2022)
Comparatively, in the material removal process, research has proposed an online modeling method considering data cleaning based
focused on the interaction of the abrasive with the surface of the on machine learning modeling and used particle swarm optimization
workpiece during the polishing process to chip away the surface mate­ (PSO) algorithm to optimize the polishing parameters. Copper polishing
rial and achieve the desired surface quality. In this context, empirical experiments were then conducted to validate the modeling and opti­
and physical models have become dominant in research, and the effects mization methods. The results show that the model can accurately

2
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

predict the surface roughness of copper samples. Yogesh et al. (2021) selection operator (LASSO) (Tibshirani, 1996), Elastic Net Regression
used the decision tree algorithm and plain Bayesian algorithm to predict (ENR) (Zou and Hastie, 2005), Ridge Regression (RR) (Hilt and Seegrist,
the material removal rate (MRR) and surface roughness for EDM of 1977), SVR (Awad et al., 2015), Linear Regression (LR) (Freedman,
aluminum composites. The model can be used to predict MRR and sur­ 2009), Gradient Boosting Regression (GBR) (Nie et al., 2021), Stochastic
face finish for any combination of process parameters before machining. Gradient Descent Regression (SGDR) (Shalev-Shwartz et al., 2007) and
Zhang et al. (2021) used Random Forest (RF) to obtain process variables Extreme Random Forest (ETR) (Geurts et al., 2006), and optimizes the
(used by the RF algorithm to quantify the importance of features) that weights of these models by using the weight averaging method based on
are important for MRR prediction as inputs to the neural network. In DE. Finally, the proposed method is validated by MJP experiments on 3D
addition, the prediction of MRR with the help of residual convolutional printed component 316L stainless steel. ERDE is compared with other
neural network (ResCNN) is proposed for the first time. Schnecken­ existing methods to validate the efficacy of the proposed method, and
burger et al. (2020) collected data generated by a glass polishing head the feature selection method of ERDE is also compared with alternative
with integrated sensors for training an artificial neural network (ANN) feature selection methods.
model and optimized the model step-by-step using hyperparametric The structure of this paper is outlined as follows: Section 1 describes
optimization (HPO). The results showed that the prediction of the arti­ the relevant background and research significance of this study and
ficial neural network reduced the number of polishing repetitions, thus reviews the related works on surface roughness prediction methods.
shortening the production time. Kanovic et al. (2022) present the results Section 2 introduces the design of the proposed prediction method based
of experiments to investigate ball burning of AISI 4130 alloy steel with a on ensemble regression and DE method (ERDE). Section 3 gives detailed
high-stiffness tool and a ceramic ball. Based on experimental result, information on the experimental setup and dataset. Section 4 compares
surface roughness was modeled using regression analysis (RA), ANN, and discusses the predictive performance of various methods. Section 5
and support vector regression (SVR). The experimental results show that provides a summary of the research presented in this paper and outlines
the application of ANN and SVR is sufficient to simulate the spherical potential avenues for future research.
polishing process and to predict the roughness of the treated surfaces as
far as the possibility of practical application in real industrial conditions 2. Surface roughness prediction method based on DE and
is concerned. ensemble regression
Although machine learning algorithms for surface roughness pre­
diction have been successfully applied in intelligent manufacturing and A surface roughness prediction method of ensemble regression with a
industrial processing, accurate surface roughness prediction may still be differential evolution algorithm (ERDE) is proposed and applied to
a challenge due to the complexity of the surface generation of MJP. The multi-jet polishing, which takes into account the effect of different
correlation between MJP parameters and surface roughness still needs to process variables on surface roughness. Fig. 1 presents the framework of
be extensively studied. On the one hand, for regression prediction, each the ERDE method, which is mainly divided into two aspects: data pro­
model has a diverse assumption space, and the optimal solution may not cessing and analysis and ensemble prediction model. The data process­
be determined by a single model. Therefore, it is usually necessary to ing and analysis are divided into feature engineering module and
combine multiple regression models, i.e., ensemble strategies (Ren et al., Shapley-based explainable analysis module, and the ensemble model
2016). That is, using only simple machine learning models or neural prediction is based on the DE algorithm integrating several base
network models may not be sufficient. We need to use more complex regression models.
models, such as ensemble models to obtain better prediction perfor­ In ERDE, the pre-processed data are first subjected to feature trans­
mance. On the other hand, the analysis and processing of data is a formation and feature selection. Then, further explainable analysis of
prerequisite and key to influencing the performance of the model. data based on Shapley’s theory can help us to identify the key factors
However, many factors affect surface roughness, including depth of cut, impacting surface roughness. Analysis results help researchers to better
feed rate, cutting speed, tip radius, lubrication conditions, cutting understand the significance of the data and thus lay the groundwork for
forces, production time, and cost (Abbas et al., 2016; Abbas et al., 2018; forthcoming experiments and investigations. After data analysis and
Al Bahkali et al., 2016; Yue et al., 2016; Zain et al., 2010). The multitude processing, the transformed samples are sent into several regression
of factors influencing surface roughness makes data analysis prior to models for prediction. The predicted values in the different regression
modeling a significant challenge. This complexity is likely why most models are called response values. The response values obtained from
current surface roughness prediction methods lack a comprehensive the forecast are passed into the DE-based ensemble module. After the
discussion on data analysis. Feature engineering, such as feature selec­ optimal weights of each model are determined by the DE, the algorithm
tion and feature construction, can effectively improve model prediction integrates and calculates the predicted output results to obtain the final
performance. Explainable analysis methods can validate the effective­ surface roughness prediction values.
ness of feature engineering and help researchers to better understand the
data, which can further provide optimization solutions for feature
engineering. 2.1. The data processing and analysis module
In this paper, a novel ensemble regression with differential evolution
(ERDE) method is proposed. The method consists of two main modules: The data processing and analysis module consists of feature trans­
the data processing and analysis and the ensemble model prediction. In formation, feature selection, and explainable analysis. Firstly, Min-Max
the data processing and analysis module, data is processed by feature normalization is chosen to process the data to make the model converge
engineering, and analyzed by interpretable methods. Considering the faster. Besides, given the low feature dimensionality of the MJP dataset,
feature space of the MJP dataset is relatively single, polynomial feature a multi-order feature transformation is used to improve the feature
transformation is used. In addition, we provide an explainable analysis dimensionality. This operation transforms a set of features into a new set
of the results based on the Shapley theory, using visualizations such as while preserving as much information as possible about the original
feature importance plots and sample distribution plots to better analyze features. To further improve the model performance, feature selection is
the impact of features and samples on results. Then DE-based feature used to filter out some redundant features. Finally, the data before and
selection is performed on the new features after transformation to after feature engineering are comprehensively analyzed and compared
enhance the model accuracy. In the ensemble learning prediction by using an explainable method to investigate the inherent connection
module, ERDE combines several machine learning algorithms, including between machining parameters and surface roughness prediction, thus
Random Forest (RF) (Breiman, 2001), Extreme gradient boosting helping researchers to better understand the significance of the data for
(XGBoost) (Chen and Guestrin, 2016), Least absolute shrinkage and subsequent experiments and studies.

3
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

Fig. 1. Flow chart of the ERDE prediction method.

2.1.1. Feature transformation relationships.

Feature transformation is converting a set of features into a new set There are only four features in the MJP dataset, but the features can
of features while preserving as much information about the initial fea­ be expanded to NF = 15 features by the second-order polynomial feature
tures as possible, and adding the higher powers and cross terms of the transformation. In the pre-processing module, polynomial feature
original features to the feature set to enhance the ability of machine transformation of varying orders is utilized to project the features into a
learning algorithms to fit nonlinear relationships (Dogan and Birant, high-dimensional space, which helps to find the interrelationship be­
2021). Polynomial feature transformation mainly refers to constructing tween each feature.
more new output features based on the input features according to the
established polynomial rules, where degree in polynomial feature 2.1.2. Feature selection
transformation is the number of controlling polynomials, which can Feature selection refers to the selection of a small number of
affect the complexity and fitting ability of the model. Higher feature important features from many features, thus reducing redundant fea­
dimensions and interrelation terms can be obtained using polynomial tures, modeling training time, and improving model prediction perfor­
features. mance as well. There may still be some redundant features after feature
The raw data contains four machining parameters: feed rate (f), transformation, so a better subset of features can be found by feature
pressure (P), tool offset (TO), and step distance (d). And suppose there selection. Here, the features with high relevance can be effectively
are four input features z1 , z2 , z3 , and z4 (corresponding to the four main selected by using the DE-based feature selection algorithm, which is
machining parameters). The polynomial feature transformation model important for the subsequent modeling process.
can transform them into a combination of polynomial features. Taking Fig. 2 depicts the encoding of the DE-based feature selection algo­
the second-order polynomial as an example, the formula is expressed as rithm. For feature selection based on the DE algorithm, we utilize the DE
follows Eq. (1): algorithm to automatically filter several features from 15 new features.
[ ] Specifically, the population in our designed feature selection method
Z = 1, z1 , z2 , z3 , z4 , z1 2 , z2 2 , z3 2 , z4 2 , z1 z2 , z1 z3 , z1 z4 , z2 z3 , z2 z4 , z3 z4 (1)
based on the DE algorithm contains 15 individuals, each corresponding
to a feature. The dimension of each individual fi is 1, i.e., a binary
where Z is the transformed feature vector containing the combined
number, where ’1′ represents selecting the current bit feature and ’0′
features of the original features z1 , z2 , z3 , and z4 (corresponding to the
represents not selecting the current bit feature. Using the prediction
four main machining parameters feed rate (f), pressure (P), tool offset
error as the fitness function, through the evolution of DE, we can finally
(TO), and step distance (d), respectively). In this way, the linear rela­
obtain a feature selection sequence of length NF (NF = 15), and the
tionship of the original features can be extended into higher-order

4
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

The SHAP method is a typical post-hoc model interpretation

approach, primarily utilized for providing visual explanations of data. In
this study, we use Shapley values for comparative analysis, comparing
the Shapley values of each feature to reveal their respective contribu­
tions to model predictions. Furthermore, we explain the impact of the
newly constructed features, derived from proposed feature engineering,
on the model’s performance, thus assisting in subsequent ensemble
predictions. Lastly, we employ the SHAP library for visualization anal­
ysis. By combining Shapley values with visualizations, we can present a
more intuitive interpretation of the model’s output, enhancing our un­
derstanding of the model’s predictions.

2.2. The integrated model prediction module

2.2.1. Base regression algorithms

In the ensemble module, each base regression model is first modeled
for prediction, and then the DE algorithm is used to assign weights to it.
The above operation will be iterated to find the optimal weight ratio of
Fig. 2. The encoding of DE feature selection. each regression algorithm to get the ensemble prediction. We used ten
regression algorithms including RF, XGBoost, LASSO, ENR, RR, SVR, LR,
feature corresponding to the position with value ’1′ in this sequence is GBR, SGDR, and ETR as the basis for composing the integration module.
the candidate feature. Finally, according to the optimal feature selection The ten regression models mentioned above were first trained, and
result read the feature transformed data to get new data for subsequent hyperparameter tuning was performed for each model to ensure the
modeling use. better performance of each model. Finally, the mean absolute percent­
age error (MAPE) values of each corresponding model were calculated
2.1.3. Shapley-based explainable analysis based on the prediction result and real result by using Eq. (3).
The application of machine learning models to critical areas such as k ⃒ ⃒
biomedicine and autonomous driving has been successful, but the 1∑ ⃒̂y i − yi ⃒
MAPE = ⃒
⃒
⃒ (3)
decision-making process and rationale of the models are often viewed as k i=1 ̂ yi ⃒
a black box and difficult to understand. This puts some limitations on the
application and credibility of the models. To overcome this problem, where k is the number of samples, yi represents the actual value of the ith
interpretable machine learning has become an important area of data point, ̂y i represents the predicted value of the ith data point from
⃒ ⃒
research. The goal of interpretable machine learning is to help re­ the model, denotes the absolute percentage error, ⃒ y i − yi ⃒ which is the
⃒̂ ⃒
̂y i
searchers better understand data and models. Interpretive analytical
absolute value of the relative error between the predicted and actual
studies such as feature significance and correlation can reveal the rela­
values for each data point.
tionship between data and model. Besides, such explainable analysis can
also help to identify the main factors that contribute to model decisions
2.2.2. Ensemble module based on DE algorithm
and gain insights from them to optimize model performance and
The DE is an efficient global optimization technique. It is also a
enhance reliability. Therefore, in this paper, explainable analysis is
population-based heuristic algorithm, which has the advantages of
employed to extract valuable information from both the original and
simple principles, few controlled parameters, and high robustness.
newly constructed features, as well as to validate the robustness of the
Therefore, this module uses DE to process the output of the multi-
newly created ones. Through explainable analysis, the processing pa­
algorithm regression module. In ERDE, a simple coding method was
rameters that have a major impact on surface roughness can be deter­
designed as the coding mechanism of DE, get the optimal weights of
mined, and targeted directions for optimizing the model can be offered.
each model by DE, and integrate the output of each regression module to
Professor Lloyd Shapley (Shapley et al., 1952) introduced Shapley
calculate the ultimate prediction value for surface roughness.
theory to tackle the issue of contribution and benefit distribution in
(1) Individual design
cooperative games. Its core idea involves calculating the contribution of
In ERDE, each individual within the population represents a solution
each participant and ensuring a fair allocation of benefits during the
vector and is also referred to as a target vector xi , which can be expressed
cooperative process. Shapley Additive exPlanation (SHAP) is a “model
by Eq. (4):
interpretation” package based on Python that utilizes Shapley values to
( )
explain the output generated by machine learning algorithms. The xi = xi,1 , xi,2 , ⋯, xi,D i = 1, 2, ⋯, N (4)
central concept of SHAP is to compute the marginal contribution of each
feature to the model’s output, and then calculate the black-box model at where D is the dimensionality of the target vector, representing that one
both global and local levels. Specifically, SHAP values for the jth feature individual contains information of the D dimension, and N means the
mj function φj (val) is defined as follows Eq. (2): population size.
∑ The encoding method of this algorithm is slightly different from
|S|!(o − |S| − 1)! ( ( { }) )
φj (val) = val S ∪ mj − val(S) (2) traditional encoding. As shown in Fig. 3 (a), in the conventional coding
S⊆{m1 ,⋯,mo }\{mj } o!
method, a population contains N individuals. Each individual comprises
where m1, m2, …, mo is the total number of features used in building the D gene, signifying that the algorithm can generate N weight assignment
model, o is the total number of features, and mj is the feature whose schemes for D models. However, this approach may lead to a prolifer­
contribution is to be computed. S is the subset of excluded features mj. ation of individual solutions during the iteration of the algorithm, which
val(S ∪ {mj }) − val(S) is the predicted contribution for the set containing may lead to finding the optimal solution being more difficult and com­
plex. Therefore, to address this problem, DE algorithm with a simple
feature mj minus the predicted contribution for the set excluding feature
encoding mechanism was designed, as shown in Fig. 3(b). Where the
mj, i.e., the marginal contribution of feature mj.
individuals represent the chosen base regression model, the weight of

5
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

Fig. 3. Difference between the traditional coding method of DE and the novel simple coding method in this paper. (a) The traditional encoding method in DE. (b) The
proposed simple encoding method in this paper. (c) Weight of each model.

each individual is 0.05, a total of 20 individuals form a population, and αi = 1) is calculated after the results of the selected models are counted
the population represents a complete weight assignment scheme. On the as shown in Fig. 3 (c).
one hand, this coding method can reduce the dimensionality of the (2) Evolution Operators
search space from D to 1, which, in turn, reduces the complexity of the The DE algorithm mainly uses mutation, crossover, and selection
algorithm and speeds up the search. On the other hand, it can also operations to achieve iterative population updates.
reduce more time and effort to adjust the hyper-parameters, such as the The first step is mutation. Three genes are randomly selected, and the
size of the population, which further improves the efficiency of the mutated individual V = (v1 , v2 , ⋯, vj ) is found by Eq. (5). If each gene vj
algorithm. in V satisfies the constraint, it is crossed with the current individual X;
Fig. 3 (a) and (b) illustrate the distinctions in the encoding method of otherwise, a gene satisfying the condition is randomly generated.
the traditional DE algorithm and the above encoding method. For the ( )
proposed encoding method in this paper, each individual contains only vj = xr1 + F* xr2 − xr3 (5)
one dimension, and the information contained in this dimension is an where vj in the interval [0,9], j, r1 , r2 , r3 are three mutually different
integer in the interval [0, 9]. For each generation, the number of in­ integers chosen randomly from [1, N], and j is different from each of r1 ,r2 ,
dividuals N of its population is defined as 20, and each individual rep­ r3 . xr1 , xr2 , xr3 are three different individual vectors, and F is the scale
resents the probability of each algorithm being selected as 1/20. For factor, xr2 - xr3 is the difference vector.
individual xi , a decimal code [0–9] is provided, representing the base The next step is crossover. The binomial crossover operator is
regressor corresponding to that individual. Based on this individual implemented for X and V. A trial vector U = (u1 , u2 , ⋯, uj ) is generated
design scheme, the final population is a sequence of integers of length
∑ by Eq. (6):
20, i.e., [x1 , x2 , x3 , ⋯, xN− 1 , xN ], where xi = 1. The population can
{
reflect the probability of each underlying regressor model being vj , if randj < CRorj = jrand
uj = (6)
selected. Finally, the weight αi (α ∈ [0, 1] ) of model i (αi represents xj , otherwise
∑
weight assigned to model i within the ensemble framework, where

6
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

where j = 1, 2, 3, ⋯, N, randj a number randomly distributed with uni­

formity within the interval [0, 1], jrand is a randomly chosen integer be­
tween 1 and N, and CR ∈ [0,1] is the so-called cross-control parameter.
Selection is the third stage. Considering a maximization problem, the
individual target X is compared with the crossover individual U by Eq.
(7) based on the objective function fitness. and the one with the better
result will survive to the next generation:
{
U, if fitness(U)⩾fitness(X)
X= (7)
X, otherwise

(3) DE-based ensemble computing

Through the calculation of each individual in the population ac­
cording to Eq. (8), the weight assignment of each regression model is
obtained.
∑20 {
1, INDn = m
αm = (8)
n=1
0, INDn ∕
=m

In this module, predicted values are generated for each data sample
using various regression models, and these values are called response
values. They are used as inputs to this module. In the iterative process,
all the response values are ensemble, and the ensemble surface rough­
ness prediction will be calculated by Eq. (9). In this equation, ̂
y repre­
sents the predicted surface roughness; fm is the response value of each
base model; αm is the weight assigned to each model, and M is the total
number of models.
∑
M ∑
M
y=
̂ αm fm subject αm = 1 (9)
m=1 m=1

The workflow diagram of the DE-based ensemble learning module is

shown in Fig. 4. First, the individuals are randomly generated to
constitute the initialized population. Each individual corresponds to a
regression model whose value indicates that the corresponding weight
of that model is 0.05. After obtaining the weight assignment of each
model, the predicted value of surface roughness after integration of a Fig. 4. Flow chart of DE-based ensemble module.
particular sample is calculated using Eq. (9). Then, the prediction error
of the ensemble model is calculated using the predicted value and target
shown in Fig. 5. Experimental data were obtained for 43 different pol­
value. It serves as the fitness function for DE to obtain the optimal
ishing parameters, including feed rate (f), pressure (P), tool offset (TO),
prediction effect. To ensure the diversity of the population during the
and step distance (d), which are presented in Table 1. The experimental
iteration, crossover, and mutation operations are applied to the current
results are presented in Table 2 (Data from literature (Wang et al.,
population to continuously update the population. In addition, the use of
2022)).
elite retention replicates the best individuals that have emerged
throughout the population’s evolution directly into the next generation,
bypassing genetic manipulation to ensure the overall fitness of the 3.2. Parameter setting of the ERDE
population during the iterative process.
The above process will be repeated until the maximum number of The experimental data mentioned above are randomly separated into
iterations is satisfied. Eventually, the best weights of each model are training, validation, and test datasets with ratios of 80 %, 10 %, and 10
determined, and the ensemble regression model obtained at this %, respectively. After data normalization as per Eq. (1) in the pre-
moment is the best surface roughness prediction model. processing module, the second-order polynomial feature trans­
formation is used to raise the feature dimension of the dataset. Then the
3. Validation experiments training and validation sets are passed into DE for feature selection. New
data is generated by eliminating irrelevant features based on the results
3.1. Experimental setup and data collection of feature selection, while the features are analyzed and interpreted
using Shapley theory. The training dataset is used to train different
Multi-jet polishing was employed for the surface finishing of the 3D- regression models. Five-fold cross-validation determines the optimal
printed 316L stainless steel component. To investigate the correlation parameters for each regression model. After the training process, the
between the MJP parameters and surface roughness of 316L stainless best-trained model is utilized to predict the surface roughness of the test
steel, different polishing parameters were used to polish 3D printed dataset, and the predicted results are referred to as response values.
components of 316L stainless steel with MJP. These polishing parame­ Afterward, these values are passed to the ensemble module. DE performs
ters were determined based on previous research results (Cheung et al., different combinations of response values with the corresponding
2018; Wang et al., 2017; Wang et al., 2023). After polishing, the surface regression model evaluation metrics as fitness functions.
roughness measurements were repeated three times at different regions In the ERDE method, the maximum number of iterations is 150, and
by using the Taylor Hobson Form TalySurf PGI 1240, and the average the size of the population is 15 when DE performs feature selection. The
was taken to determine the final results (Raf). Hence, Raf was used to maximum number of iterations is 50,000, and the size of the population
evaluate the surface roughness. The experimental polishing setup is is 20 when DE performs the weight assignment of models. The best
weight assignment value corresponding to each model and the best

7
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

Fig. 5. Multi-jet polishing experimental setup.

transformation, the second-order polynomial feature transformation

Table 1
had the best outcome. Therefore, the second-order polynomial feature
MJP experimental parameter settings.
transformation is chosen for the subsequent surface roughness
Polishing parameter Range prediction.
Feed rate (f) 10,20,30,40,60,80 mm/min In addition, to verify the advantages of DE-based feature selection
Fluid pressure (P) 4,5,6,7,8,9,10 bar methods in ERDE, five classical feature selection methods were
Tool offset (TO) 2.5,5,7.5,10,12.5,15 mm
compared, which were Recursive Feature Elimination (RFE) (Guyon
Step distance (d) 0.1,0.2,0.3,0.4,0.5,0.6,0.7,0.8 mm
et al., 2002), Random Forest (RF) (Menze et al., 2009), Blind Signal
Separation (BSS) (Cardoso, 1998), Sequential Backward Selection (SBS)
surface roughness prediction value are finally determined by iterations (Reeves and Zhe, 1999), and Sequential Forward Selection (SFS) (Mar­
of calculation. At the same time, the performance of different methods cano-Cedeño et al., 2010). In addition, given that current popular deep
was also tested on the MJP experiments, and MAPE evaluated the al­ learning methods can also automatically extract features, we also
gorithm prediction performance by Eq. (3). selected three deep learning-based feature selection methods for com­
The algorithm is implemented in Python and executed on a computer parison. They are autoencoder-based feature selection methods
featuring an i9-9750H 2.60 GHz CPU. (AutoEncoder), gradient information-based feature selection methods
(Grad), and Lasso-based feature selection methods (Lasso). The predic­
4. Results and discussions tion results still used MAPE as an evaluation metric to assess the pre­
dictive capabilities of the algorithms with different feature selection
4.1. Validation of feature engineering methods. From Fig. 7, it can be found that the prediction results of the
DE-based feature selection method were 14 %, 25 %, 56 %, 30 %, 24 %,
To enhance the predictive performance of the model, feature trans­ 52 %, 4 %, and 51 % lower than those of other feature selection methods,
formation was first performed to obtain more features. As shown in indicating the apparent advantage of the DE-based feature selection
Fig. 6, the differences in model performance under four different feature method.
transformation methods without feature transformation, second-order
polynomial feature transformation, third-order polynomial feature
4.2. Comparison with other surface roughness prediction methods
transformation, and fourth-order polynomial feature transformation
were compared, and MAPE was used as the performance evaluation
To further substantiate the benefits of ERDE, the predictive results of
metric. As shown in Fig. 6, the prediction results after polynomial
ERDE were compared with ten mainstream machine learning algo­
feature transformation were better than those without feature trans­
rithms. The prediction results of different machine learning methods
formation. The model after different polynomial feature transformations
with the ERDE are shown in Table 3. Among the ten classical machine
shows different prediction performances, which indicates that the
learning algorithms, the XGBoost algorithm gave the best results.
feature transformation has some influence on improving the perfor­
However, the MAPE obtained by the ERDE was about 46 % less than the
mance of the model. Among the different orders of polynomial feature
figure of the XGBoost. The MAPE was reduced even more compared to

8
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

Table 2 model XGB, ’5′ denotes model ETR, ’6′ denotes model GBR, and ’7′ de­
MJP experimental data(Wang et al., 2022). notes model SVR, ’8′ denotes model LR, and ’9′ denotes model SGDR.
Number f(mm/min) P(bar) TO(mm) d(mm) Raf (μm) Each bit of the sequence indicates the probability of selecting the current
bit of the underlying model as 0.05. By counting the number of occur­
1 10 8 5 0.2 0.021
2 20 8 5 0.2 0.03 rences of each base regression model in the sequence, the result of the
3 30 8 5 0.2 0.032 weight assignment of each base regression model can be obtained. The
4 40 8 5 0.2 0.088 weight of RF model is 8*0.05 = 0.4, the weight of XGBoost model is
5 60 8 5 0.2 0.127 6*0.05 = 0.3, the weight of RR model is 2*0.05 = 0.1, and the rest of
6 80 8 5 0.2 0.144
7 20 4 5 0.2 0.088
them, such as ELN, ETR, LR, and SGDR models, are all assigned 0.05.
8 20 5 5 0.2 0.085 This result shows that our algorithm can effectively select the base
9 20 6 5 0.2 0.083 regression model with better performance, and the final integration
10 20 7 5 0.2 0.039 results are due to the prediction results of each base regression model.
11 20 8 5 0.2 0.039
Fig. 8 compared the performance of four prediction methods, i.e.,
12 20 9 5 0.2 0.037
13 20 10 5 0.2 0.033 Averaging (Liu and Kuo, 2016), Stacking (Ngo et al., 2018), ELGA (Wang
14 20 8 2.5 0.2 0.027 et al., 2022), and ERDE. The histogram visually demonstrates that the
15 20 8 5 0.2 0.04 ERDE method had better prediction results, and its MAPE was reduced
16 20 8 7.5 0.2 0.032 by about 42 %–57 % compared to the results of other methods. This
17 20 8 10 0.2 0.072
18 20 8 12.5 0.2 0.026
demonstrates the significant impact of feature engineering on enhancing
19 20 8 15 0.2 0.036 the model’s prediction accuracy. The efficient global optimization
20 20 8 5 0.1 0.031 capability of the DE algorithm can predict the results accurately.
21 20 8 5 0.2 0.043
22 20 8 5 0.3 0.047
4.3. Explainable analysis
23 20 8 5 0.4 0.085
24 20 8 5 0.5 0.098
25 20 8 5 0.6 0.138 This section interpreted the relationship between processing pa­
26 20 8 5 0.7 0.099 rameters and predicted results by using Shapley theory. The core idea of
27 20 8 5 0.8 0.131
Shapley theory is to calculate the different marginal contributions of a
28 10 5 2.5 0.2 0.044
29 15 5 5 0.4 0.146
feature to all the feature combination scenarios, and then weight the
30 20 5 7.5 0.6 0.183 average of each marginal contribution of the feature based on the
31 25 5 10 0.8 0.229 probability of occurrence of the different scenarios.Finally, we can get
32 10 6 5 0.6 0.076 the feature’s contribution to the prediction result, whose quantitative
33 15 6 2.5 0.8 0.151
value is SHAP Value.
34 20 6 10 0.2 0.065
35 25 6 7.5 0.4 0.165 We also used SHAP values to quantify the contribution of new fea­
36 10 7 7.5 0.8 0.121 tures after feature transformation. Although SHAP values cannot
37 15 7 10 0.6 0.111 directly calculate the collaborative contribution between features, the
38 20 7 2.5 0.4 0.086 collaborative relationship between different features is implied in the
39 25 7 5 0.2 0.066
40 10 8 10 0.4 0.061
weighted result of the marginal contribution. In addition, since the new
41 15 8 7.5 0.2 0.03 features contain the original four processing parameters, we can clarify
42 20 8 5 0.8 0.133 the effect of the interaction of each processing parameter on the pre­
43 25 8 2.5 0.6 0.144 diction results based on the SHAP values calculation results of the new
features, and indirectly based on the potential interaction between the
analyzed processing parameters.
For each feature, the average of the absolute SHAP values before and
after feature engineering was selected as the feature importance metric
and visualized in the form of a bar chart., as shown in Fig. 9(a) and (b).
Fig. 9(a) shows four features of the original data, i.e., the four main
machining parameters feed rate (f), pressure (P), tool offset (TO), and
step distance (d). Fig. 9(b) shows the five newly generated features in
descending order of feature importance after feature engineering.
Comparing Fig. 9(a) and (b), it becomes evident that the processing of
feature engineering can construct more features with higher SHAP
values. Analyzing the features after feature engineering, it can also be
found that three of the five features with SHAP values above 0.01
contain the original feature f. This shows that the feature feed rate (f) has
a greater influence on the predicted surface roughness. It also reflects
from another point of view that feature engineering can retain better
Fig. 6. Comparison of different feature transformation methods. original features and construct new important features from better
original features. In addition, the results in Fig. 9 (b) show that, ac­
the other nine algorithms. Therefore, for the surface roughness predic­ cording to the SHAP method, certain newly constructed features make
tion issue in polishing processes, the ERDE performs better than the no significant contributions to the model.
current mainstream machine learning algorithms, which have a small For each sample, the SHAP value of each feature was plotted, as
prediction error. shown in Fig. 10. The plot provided a better understanding from a
The optimal results of the weight assignments for each of the base comprehensive perspective and may detect outliers. In the figure, each
regression models obtained by the DE algorithm are [0, 4, 4, 4, 5, 1, 0, 0, point represents a sample. The horizontal axis represents samples
9, 0, 1, 0, 8, 4, 0, 3, 4, 0, 0, 4]. Where ’0′ denotes model RF, ’1′ denotes ranked by SHAP value, the vertical axis represents features ranked by
model RR, ’2′ denotes model Lasso, ’3 ’ denotes model ELN, ’4′ denotes importance, and the color depth signifies the magnitude of feature
values (high in red, low in blue). It is important to note that the feature

9
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

Fig. 7. Comparison of different feature selection methods.

Table 3
MAPE comparison of different methods in MJP.
RF RR Lasso ELN XGB ETR GBR SVR LR SGDR ERDE(ours)

MAPE 0.219 0.292 0.497 0.319 0.218 0.231 0.497 0.463 0.402 0.443 0.113

Fig. 8. Comparison of different prediction methods in MJP experiments.

value in this context is distinct from the SHAP value, representing the
actual value of the feature itself.
Fig. 10 showed that feature d was the most important, and this value
may affect the predicted surface roughness when it was too low.
Therefore, in the subsequent experiments, a minimum threshold can be
set for d to ensure that the machining parameter experiments are con­
ducted only when it exceeds this threshold. This approach can effec­
tively reduce the number of experiments. In addition, it can be found
that P and f are also the more dominant influencing factors, but TO Fig. 9. Comparison of SHAP value of each feature before and after feature
minimally impacts surface roughness. This result is consistent with the construction. (a) Mean of SHAP value of each feature before feature engi­
neering. (b) Mean of SHAP value of each feature after feature engineering.
conclusion of the importance analysis of features above, which further
identifies the primary and secondary factors affecting surface roughness.
In the subsequent experiments, the focus can be on optimizing and 4.4. Comparison of different methods on fluid jet polishing (FJP) of 3D-
adjusting these two primary features, giving priority to selecting printed Cobalt Chrome (CoCr)
important features to enhance the model’s predictive performance. The
processing and optimization efforts for less significant features can be To validate the generalization and effectiveness of the proposed
reduced to conserve computational resources and experiment time. By model, we compared the ERDE algorithm with ten mainstream machine
concentrating more efforts and resources on the main influencing fac­ learning algorithms on FJP of 3D-printed CoCr, drawing from a previous
tors, the optimization process can be accelerated, leading to improved study (Xie et al., 2023). In this research, FJP was employed for the post-
experimental efficiency. Explainable analysis based on Shapley values processing finishing of the 3D-printed CoCr alloy. Subsequently,
provides another perspective to validate the effectiveness of feature experimental data was gathered. The prediction results of the ERDE
engineering. algorithm in comparison with different machine learning methods are
presented in Table 4. The experimental findings indicate that the ERDE
algorithm significantly reduces the MAPE compared to the other nine
algorithms. This further substantiates the robust generalization perfor­
mance of the algorithm when handling new data.

10
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

Fig. 10. Distribution of SHAP value of each sample.

Table 4
MAPE comparison of different methods on FJP of 3D-printed CoCr.
RF RR Lasso ELN XGB ETR GBR SVR LR SGDR ERDE(ours)

MAPE 0.521 0.962 0.85 0.964 0.764 0.503 0.85 1.108 0.902 0.82 0.343

5. Conclusion and future research performance. Secondly, this method incorporates feature engineering,
which includes polynomial feature transformation and DE-based feature
This paper introduces an ensemble regression method predicated on selection. However, the polynomial feature transformation often results
differential evolution (ERDE) for predicting the surface roughness of 3D- in redundant features. Consequently, future work could consider
printed 316L stainless steel components in multi-jet polishing. The ERDE employing alternative feature transformation methods to identify more
method is specifically designed to address challenges at both the data effective features. In summary, future enhancements could be
and modeling levels. At the data level, new features are generated approached in two ways. Firstly, to validate the method’s generality,
through a feature engineering method, wherein the data are processed experiments could be conducted using different processing techniques,
using polynomial feature construction and DE-based feature selection. allowing for methodological improvements. Secondly, optimizing the
Concurrently, Shapley’s theory is employed to analyze the influence of ensemble learning module could further enhance the prediction per­
features (processing parameters) and sample distribution on the output formance of surface roughness.
(surface roughness). This approach aids researchers in better under­
standing the data, optimizing the model, and achieving more accurate CRediT authorship contribution statement
predictions. At the modeling level, the initial step involves constructing
a predictive model that links surface roughness with processing pa­ Yueyue Wang: Conceptualization, Methodology, Software, Data
rameters. The new data is subsequently input into multiple basic curation, Validation, Writing – original draft. Zongbao He: Software,
regression algorithms for training and prediction. Following this, an Resources, Writing – original draft, Writing – review & editing. Shutong
ensemble learning method is utilized to integrate multiple basic Xie: Conceptualization, Methodology, Writing – review & editing.
regression models into an ensemble framework. During the ensemble Ruoxin Wang: Methodology, Writing – review & editing. Zili Zhang:
process, the DE is employed to optimize weight allocation, thereby Methodology. Shimin Liu: Writing – review & editing. Suiyan Shang:
obtaining weights for each basic regression model. Ultimately, based on Writing – review & editing. Pai Zheng: Funding acquisition, Writing –
these weights, the weighted average prediction results are calculated to review & editing. Chunjin Wang: Supervision, Conceptualization,
obtain the final prediction for surface roughness. Funding acquisition, Writing – review & editing.
To validate the effectiveness of the ERDE method, it was compared
with numerous existing algorithms using the MJP dataset. The results
underscore the efficacy of the ERDE method in predicting surface Declaration of competing interest
roughness, outperforming both the ELGA algorithm proposed in a pre­
vious paper and the current mainstream machine learning algorithms. In The authors declare that they have no known competing financial
the MJP experiments, the mean percentage prediction error results of interests or personal relationships that could have appeared to influence
the ERDE method (MAPE = 0.113) surpassed the results of the ELGA the work reported in this paper.
method (MAPE = 0.195) and the results of the classical machine
learning algorithm (MAPE = 0.218). The explainable analysis results not Data availability
only demonstrate the effectiveness of feature engineering but also help
uncover key information for reference in subsequent experiments. Data will be made available on request.
Despite the promising results, this study acknowledges two areas for
potential improvement. Firstly, the ERDE method predicts surface Acknowledgments
roughness by utilizing DE to optimize the weights of multiple regressors
within a relatively singular ensemble framework. Therefore, exploring The work described in this paper was mainly supported by a
more advanced ensemble methods could enhance prediction Shenzhen-Hong Kong-Macau Technology Research funding from the
Shenzhen Municipal Science and Technology Innovation Commission

11
Y. Wang et al. Expert Systems With Applications 249 (2024) 123578

(Project No: SGDX20220530110804030), and the Research and Inno­ Menze, B. H., Kelm, B. M., Masuch, R., Himmelreich, U., Bachert, P., Petrich, W., &
Hamprecht, F. A. (2009). A comparison of random forest and its gini importance
vation Office of The Hong Kong Polytechnic University (Project code:
with standard chemometric methods for the feature selection and classification of
BBR2, BD9B and BBXL). spectral data. BMC Bioinformatics, 10, 1–16.
Min, L., Minghui, L., Riemer, O., Fangzeng, S., & Binghai, L. (2021). Anhydrous based
References shear-thickening polishing of KDP crystal. Chinese Journal of Aeronautics, 34(6),
90–99. https://doi.org/10.1016/j.cja.2020.09.019
Ngo, T. D., Kashani, A., Imbalzano, G., Nguyen, K. T., & Hui, D. (2018). Additive
Abbas, A. T., Ragab, A. E., Bahkali, A.l., & El Danaf, E. A. (2016). Optimizing cutting manufacturing (3D printing): A review of materials, methods, applications and
conditions for minimum surface roughness in face milling of high strength steel challenges. Composites Part B: Engineering, 143, 172–196. https://doi.org/10.1016/j.
using carbide inserts. Advances in Materials Science and Engineering, 2016, 7372132. compositesb.2018.02.012
https://doi.org/10.1155/2016/7372132 Nie, P., Roccotelli, M., Fanti, M. P., Ming, Z., & Li, Z. (2021). Prediction of home energy
Abbas, A. T., Ragab, A. E., El-Danaf, E. A., & Al Bahkali, E. A. (2018). Effect of equal- consumption based on gradient boosting regression tree. Energy Reports, 7,
channel angular pressing on the surface roughness of commercial purity aluminum 1246–1255. https://doi.org/10.1016/j.egyr.2021.02.006
during turning operation. Proceedings of the Institution of Mechanical Engineers, Part B: Peng, Z., Song, W.-L., Ye, C.-L., Shi, P., & Choi, S. B. (2021). Model establishment of
Journal of Engineering Manufacture, 232, 1006–1995. https://doi.org/10.1177/ surface roughness and experimental investigation on magnetorheological finishing
0954405416662083 for polishing the internal surface of titanium alloy tubes. Journal of Intelligent
Al Bahkali, E. A., Ragab, A. E., El Danaf, E. A., & Abbas, A. T. (2016). An investigation of Material Systems and Structures, 32(12), 1278–1289. https://doi.org/10.1177/
optimum cutting conditions in turning nodular cast iron using carbide inserts with 1045389X20930095
different nose radius. Proceedings of the Institution of Mechanical Engineers, Part B: Reeves, S. J., & Zhe, Z. (1999). Sequential algorithms for observation selection. IEEE
Journal of Engineering Manufacture, 230. https://doi.org/10.1177/ Transactions on Signal Processing, 47(1), 123–132.
0954405416662085 Ren, Y., Zhang, L., & Suganthan, P. N. (2016). Ensemble classification and regression-
Awad, M., Khanna, R., Awad, M., & Khanna, R. (2015). Support vector regression. recent developments, applications and future directions. IEEE Computational
Efficient learning machines: Theories, concepts, and applications for engineers and system Intelligence Magazine, 11(1), 41–53. https://doi.org/10.1109/MCI.2015.2471235
designers, 67–80. Schneckenburger, M., Garcia-Barth, L., & Börret, R. (2020). Machine learning model for
Breiman, L. (2001). Random forests. Machine Learning, 45(1), 5–32. https://doi.org/ robot polishing cell. Seventh European Seminar on Precision Optics Manufacturing.
10.1023/A:1010933404324 SPIE, 2020(11478), 145-154. https://doi.org/10.1117/12.2564633.
Cardoso, J. F. (1998). Blind signal separation: Statistical principles. Proceedings of the Shalev-Shwartz, S., Singer, Y., & Srebro, N. (2007). Pegasos: Primal estimated sub-gradient
IEEE, 86(10), 2009–2025. solver for svm. In Proceedings of the 24th international conference on Machine learning,
Chen, N., Wang, R., Nagarajan, B., Yan, B., Wu, Y., He, N., & Castagne, S. (2022). 807-814. Corvalis, Oregon, USA.
Investigation of metal-coating-assisted IR nanosecond pulsed laser ablation of CVD Shapley, L., Artin, E., & Morse, M. (1952). Quota solutions op n-person games1. RAND
diamond. Journal of Materials Research and Technology, 18, 4114–4129. https://doi. Corporation.
org/10.1016/j.jmrt.2022.04.072 Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the
Chen, T., Guestrin, C., 2016. Xgboost: A scalable tree boosting system. In Proceedings of Royal Statistical Society Series B, 58(1), 268–288. https://doi.org/10.2307/41262671
the 22nd acm sigkdd international conference on knowledge discovery and data mining, Wang, C. J., Cheung, C. F., Ho, L. T., Liu, M. Y., & Lee, W. B. (2017). A novel multi-jet
San Francisco, California, USA, 785–794. polishing process and tool for high-efficiency polishing. International Journal of
Cheung, C. F., Wang, C., Ho, L. T., & Chen, J. (2018). Curvature-adaptive multi-jet Machine Tools and Manufacture, 115, 60–73. https://doi.org/10.1016/j.
polishing of freeform surfaces. CIRP Annals, 67(1), 357–360. https://doi.org/ ijmachtools.2016.12.006
10.1016/j.cirp.2018.04.072 Wang, C., Loh, Y. M., Cheung, C. F., Liang, X., Zhang, Z., & Ho, L. T. (2023). Post
Dogan, A., & Birant, D. (2021). Machine learning and data mining in manufacturing. processing of additively manufactured 316L stainless steel by multi-jet polishing
Expert Systems with Applications, 166. https://doi.org/10.1016/j.eswa.2020.114060 method. Journal of Materials Research and Technology, 23, 530–550. https://doi.org/
Fan, J., Ren, X., Pan, R., Wang, P., & Tao, H. (2022). Model and optimize the magnetic 10.1016/j.jmrt.2023.01.054
composite fluid (MCF) polishing process with machine learning modeling and Wang, R., Cheng, M. N., Loh, Y. M., Wang, C., & Cheung, C. F. (2022). Ensemble learning
intelligent optimization algorithm. International Journal of Precision Engineering and with a genetic algorithm for surface roughness prediction in multi-jet polishing.
Manufacturing, 1–11. https://doi.org/10.1007/s12541-022-00668-5 Expert Systems with Applications, 207, Article 118024. https://doi.org/10.1016/j.
Freedman, D. A. (2009). Statistical models: Theory and practice. Cambridge University eswa.2022.118024
Press. Wang, W., Zou, P., Xu, J., & Ehmann, K. F. (2023). Surface morphology evolution
Geurts, P., Ernst, D., & Wehenkel, L. (2006). Extremely randomized trees. Machine mechanisms of laser polishing in ambient gas. International Journal of Mechanical
Learning, 63(1), 3–42. https://doi.org/10.1007/s10994-006-6226-1 Sciences, 250, Article 108302. https://doi.org/10.1016/j.ijmecsci.2023.108302
Guyon, I., Weston, J., Barnhill, S., & Vapnik, V. (2002). Gene selection for cancer Xie, S., He, Z., Wang, C., Liu, C., & Ke, X. (2023). A generic evolutionary ensemble
classification using support vector machines. Machine Learning, 46, 389–422. learning framework for surface roughness prediction in manufacturing. International
Hilt, D. E., & Seegrist, D. W. (1977). Ridge, A Computer Program for Calculating Ridge Journal of Computer Integrated Manufacturing, 1–23. https://doi.org/10.1080/
Regression Estimates. Department of Agriculture, Forest Service, Northeastern Forest 0951192X.2023.2204486
Experiment Station. Yue, C., Wang, L., Liu, J., & Hao, S. (2016). Multi-objective optimization of machined
Jian, Y., Ri, P., Fan, J., Wang, Z., Zhang, D., & Ji, S. (2022). Optimization of magnetic surface integrity for hard turning process. International Journal of Smart Home, 10,
composite fluid polishing process based on response surface method. Journal of the 71–76. https://doi.org/10.14257/ijsh.2016.10.6.08
Chinese Institute of Engineers, 45, 1–7. https://doi.org/10.1080/ Zain, A. M., Haron, H., & Sharif, S. (2010). Application of GA to optimize cutting
02533839.2021.1983460 conditions for minimizing surface roughness in end milling machining process.
Kanovic, Z., Vukelic, D., Simunovic, K., Prica, M., Saric, T., Tadic, B., & Simunovic, G. Expert Systems with Applications, 37(6), 4650–4659. https://doi.org/10.1016/j.
(2022). The modelling of surface roughness after the ball burnishing process with a eswa.2009.12.043
high-stiffness tool by using regression analysis, artificial neural networks, and Zhang, J., Jiang, Y., Luo, H., & Yin, S. (2021). Prediction of material removal rate in
support vector regression. Metals, 12(2), 320. https://doi.org/10.3390/ chemical mechanical polishing via residual convolutional neural network. Control
met12020320 Engineering Practice, 107, Article 104673. https://doi.org/10.1016/j.
Liu, C. A., & Kuo, B. S. (2016). Model averaging in predictive regressions. The conengprac.2020.104673
Econometrics Journal, 19(2), 203–231. https://doi.org/10.1111/ectj.12063 Zou, H., & Hastie, T. (2005). Regularization and variable selection via the elastic net.
Marcano-Cedeño, A., Quintanilla-Domínguez, J., Cortina-Januchs, M. G., & Andina, D. Journal of the Royal Statistical Society Series B: Statistical Methodology, 67(2), 301–320.
(2010). Feature selection using sequential forward selection and classification https://doi.org/10.1111/j.1467-9868.2005.00527.x
applying artificial metaplasticity neural network. In IECON 2010–36th annual
conference on IEEE industrial electronics society (pp. 2845–2850).

12