Professional Documents
Culture Documents
ABSTRACT Dispatching rules are generally useful for scheduling jobs in flexible manufacturing systems
(FMSs). However, the appropriateness of these rules relies heavily on the condition of the system; thus, there
is no single rule that always outperforms others. In this state of affairs, diverse machine-learning technology
offers an effective approach for dynamic scheduling, as it allows managers to identify the most suitable rule
at each moment. Nonetheless, various machine-learning algorithms may provide diverse recommendations.
The main objective of this study is to implement FMS scheduling using intelligent hybrid learning algorithms
with metaheuristic improvements. The developed model involves three phases: feature extraction, optimal
weighted feature extraction, and prediction. After the benchmark datasets for the FMS are gathered, feature
extraction is performed using t-distributed stochastic neighbor embedding, linear discriminant analysis,
linear square regression, and higher-order statistical features. Further, an optimal weighted feature extraction
method is developed to select the optimal features with less correlation using the improved lion algorithm
(LA), which is called the modified nomadic-based LA (MN-LA). Finally, the optimally selected weighted
features are subjected to a hybrid learning algorithm with the integration of a fuzzy classifier and a deep
belief network (DBN). For improving the prediction model, the membership function of the fuzzy classifier is
optimized using the proposed MN-LA. Moreover, the activation function and the number of hidden neurons
in the DBN are optimized using the MN-LA. The main objective of the optimized hybrid classifier is to
enhance prediction accuracy. The experimental results indicate the effectiveness of the proposed heuristic-
based scheduling method for FMSs.
INDEX TERMS Flexible manufacturing system, hybrid classifier, modified nomadic-based lion algorithm,
optimal weighted feature extraction, rule scheduling, deep belief network, fuzzy classifier, feature extraction,
prediction.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
96088 VOLUME 8, 2020
M. H. Abidi et al.: Optimal Scheduling of Flexible Manufacturing System
FMS using weighted feature extraction; section V describes its derivatives are viable alternative strategies for resolving
the development of the proposed FMS prediction model using the bi-objective consistency of the re-entrant FMS scheduling
a hybrid classifier; section VI presents an improved LA for issue.
enhancing the performance of feature extraction and classifi- Li et al. [27] focused on solving the scheduling prob-
cation; section VII presents the results and discussions; and lem of minimizing the FMS’s total energy consumption
section VIII concludes the paper. based on the FMS’s PN frameworks. On account of the
different energy-consumption rates of the availability of
resources under various operating circumstances, two energy-
II. LITERATURE REVIEW consumption functions were considered. A dynamic pro-
A. RELATED WORKS gramming (DP) system based on PNs was established to
Priore et al. [25] considered and analyzed the fact that resolve the scheduling issues. As the number of states being
generally, for addressing the FMS in scheduling jobs, only examined increases exponentially with the scale of the prob-
the dispatching rules were used. However, the suitability lem, the problem becomes more difficult to solve, rendering
of the dispatching rules was largely dependent on the sys- DP algorithmically impractical. A modified dynamic pro-
tem’s condition. Thus, no specific rule can be considered gramming (MDP) algorithm was applied at an appropriate
as functioning significantly better than another. Therefore, time to achieve an optimal or suboptimal schedule. The sim-
machine-learning methods were used to select a suitable rule ulation results indicated the MDP’s effectiveness in compar-
for the system. Even when machine-learning approaches are ison with traditional models.
used, each approach has different optimal rules. The pro- Luo et al.[14] considered deadlock management problems
posed ensemble methods, which were structured to select the for a general category of the FMS and developed an effective
most appropriate rule over time, clarified this disadvantage. method. The method has three interesting and unique features
A comparison was made among the behavior of boosting, from a software viewpoint. To define and classify the FMS,
stacking, and bagging approaches. Simulation results indi- a PN system was developed to manage all the FMS charac-
cated that with regard to the mean flow and mean tardiness, teristics. The banker algorithm-like deadlock avoidance tech-
the suggested approach was superior to the other traditional nique was recommended. The developed policy was verified
methods. to be polynomial with regard to model size. Furthermore,
Wang et al. [26] suggested an innovative effective estima- experimental results indicated the efficacy of the method and
tion of distribution algorithm (EDA)-based on the place–time its superiority to traditional systems.
PN systems of the FMS to solve FMS problems. A candidate Duan et al. [28] developed an iterative deadlock mitigation
strategy was proposed as an individual method with two strategy dependent on structural analysis for devices with
segments. The first segment contains the information on the simple sequential processes with resources (S3PR), which
route, and the second is a permutation that is replicated in comprised two stages. The first stage was called the ‘‘control
parts. A highly permissive deadlock controller checks and of the siphons.’’ The strict minimal siphons (SMSs) were
ensures individuals’ viability. In an experimental evalua- calculated in an S3PR network, and control places were
tion involving benchmark examples, the proposed algorithm implemented by adding P-invariants combined with extra
outperformed all traditional approaches for the described SMS sets that do not restrict any characteristics of the legal
problem. system. The second stage, which is called non-maxmarked
Rifai et al. [19] proposed a revolutionary strategy siphon control, was performed iteratively or incrementally
called nondominated sorting biogeography-based opti- if the method was not still alive. A mixed-integer linear
mization (NSBBO) for optimizing the complexities of programming (MILP) problem was built in each iteration step
multi-loading FMS scheduling along with the shortcuts to measure a non-maxmarked siphon, and a control place was
infused with re-entrant features. This approach was devel- applied to the siphon’s first-controlled net, resulting in an
oped to identify relatively close-optimal tradeoff solutions enhanced and enlarged network. The iteration was conducted
that can satisfy the two targets of makepan minimization until all new non-maxmarked siphons were produced by a
as well as overall earliness. The goal was to achieve opti- final-augmented net. To prove the superiority of the proposed
mal machine allocation and job sequencing while satisfy- method, its viability was validated and compared with that of
ing all requirements simultaneously. The construction of several traditional approaches.
the NSBBO involved replacing three strategies focused on Dong et al. [29] introduced a sequence of recovery tran-
sinusoidal, quadratic, and trapezoidal systems in the generic sitions to solve deadlock difficulties in dynamic production
linear emigration–immigration method. A series of test issues systems by adopting a PN. In addition to traditional deadlock
were evaluated to determine the effectiveness, complexity, management techniques that mount control places to a web
and reliability of the proposed solutions according to the template, the transitions have been used as legal ones to
specifications of the NSBBO and NSGA-II. The experimen- remove all the deadlock points. First, a series of retrieval
tal results indicated that the NSBBO-trapezoidal framework transitions were obtained from each deadlock marking to be
was successful and competitive with previously reported retrieved, relying on the reachability graph evaluation Addi-
models. Thus, it was argued that the developed NSBBO and tionally, to calculate a recovery change to recover multiple
deadlock markings, a technique of vector intersection was Kamboj et al. [35] have proposed nature-inspired search
introduced. Therefore, an iterative method was developed to algorithm. It consist of two operators namely mutation and
identify a few recovery transitions to restore all the deadlock crossover which are used to maintain the amplitude of the
points. The results indicated that the approach for recovery of search-space boundaries and search-direction matrix, that
the deadlock was efficient and effective. helps to improve the capabilities of the powerful exploration
Bashir et al. [30] proposed offered a novel, innovative and exploitation.
approach for computing the minimum supervisory model, Wang et al. [36] proposed a modified ALO (MALO).It has
which enhances the efficiency of the FMSs in the PN models. a strong exploration ability that is utilized to search for the
The implemented strategy utilized the structural properties of optimal ants and adapt to the discrete optimization problem.
a PN model to avoid the construction of its reachability graph, It consist of two categories supervised and unsupervised. The
which tends to be a state explosion issue. Two main algo- supervised models need to know the class label of each train-
rithms were proposed. The purpose of the first algorithm was ing sample in advance, which results in a better classification
to calculate an efficient uncontrolled transition, and that of result than unsupervised models in most cases.
the second algorithm was to calculate the generalized mutual
exclusion constraints (GMECs) for each stage of the PN FMS B. REVIEW
method. Additionally, a strategy was presented for building Machine learning, big data analysis, Iternet of things (IoT),
control places for each calculated GMEC without resolving etc. are contemporary trends and have applications in almost
the issue of linear integer programming, which significantly every research domain [37], [38]. Even though numerous
reduces the computational overheads. scheduling models for the FMS have been developed, a few
Lee and Shin [31] have introduced a faulty assumptions questions must be answered in future studies. A few of
about the nature of the function. It uses a linear regression as a the important advantages and disadvantages of the existing
learning algorithm, and also the learning would be quick and methods are explained in detail in Table 1. Among them,
easy, but the accuracy of the algorithm would suffer due to support vector machines (SVMs), inductive learning, back-
high bias error that comes, when using a linear regression on propagation neural networks (BPNs), and case-based rea-
nonlinear data. In reducible and variance error the amount of soning (CBR) increase the mean tardiness and mean flow
the function learned from one training data set would change time and improve the dynamic efficiency of the FMS [25].
if a different training data set were used. However, they lack the element of knowledge-based refine-
Chu et al. [32] have developed a modern industrial ment, and for better results, several types of decisions must be
technology, which implements cellular manufacturing sys- implemented in the established FMS. Although the EDA [26]
tems (CMSs) as a solution for complex production. It con- is sufficient for extracting the desired genes and will not
sists of multiple manufacturing units, each of which includes become stuck in the optimum local solution, it has limita-
multiple tasks. In CMS, each worker has one or more skills. tions, e.g., its complexity. Research must be performed to
It can demonstrate cross-training with learning and forgetting improve the developed algorithms. NSBBO [19] can achieve
effect for the worker assignment problem. In this, the pro- bi-objective satisfaction in the FMS scheduling problem and
posed hybrid bacteria foraging algorithm is used to solve the has high levels of diversity and effectiveness. However, in the
joint decision model of worker assignment and production field of tool magazines, which involves the availability of cut-
planning in cellular manufacturing systems. ting tools and material handling systems, further study must
Cavalcante et.al. [33] have introduced the resilient sup- be performed. MDP [27] achieves the ideal or sub-optimal
plier selection that utilizes the advances in data analytics solution within the appropriate time, and the model inher-
when avoiding the two major inconveniences, namely the ently has good performance. However, as the number of
need to estimate the likelihood of disruptions and fore- examined states increases, the size of the problem increases,
casting the performance impacts. The proposed data-driven which makes the problem difficult to solve. This will be
decision-making model for resilient supplier selection can examined in future studies to maximize the total energy
be exploited for the design of risk mitigation strategies in consumption and satisfy the time-related goal. The modi-
supply chain disruption management models, re-designing fied banker’s algorithm [14] has exhibited superiority and
the supplier base or investing in most important and risky higher efficiency compared with other models and appears
suppliers. to be capable of handling large-scale structures. However,
Wang et al. [34] have proposed an improved symmetric it demands the count of the functions to be addressed; no
SCA with adaptive probability selection (SSCA-APS). APS new process can be initiated while executing it. Additionally,
was used to integrate original operators and the proposed it requires the count of resources to stay static; no resource
operators. To combine the new symmetric sine cosine oper- can be reduced for any reason without the likelihood of a
ator with the original sine cosine operator, there are two deadlock. Bio-inspired meta-heuristics are becoming popular
solutions, the first one is that the probability of choosing the for various applications these days [39], [40]. MILP [28] can
new operator and the original operator is fixed. The other is solve various kinds of complex problems; however, it does
to select strategies adaptively according to the success rate of not provide an optimal solution to large problems, and its
offspring. computational time is long. In contrast to other conventional
methods, the iterative intersection approach [29] is effec- necessary to neglect the calculation of the reachability graph
tive and efficient and does not require ILPPs (integer linear of the net model. GMEC [30] does not require the ILPP to
programming problem) to be solved. However, to increase be addressed and minimizes the cost of computing. However,
the efficiency and structural minimality of the system, it is further research must be performed on popular S3PR systems.
without losing relevant or significant information. Features Thus, the extracted t-SNE features are denoted as FS t−SNE .
such as t-SNE, LDA, LSR, and higher-order statistics are LDA [45]: This is an improved feature-extraction and
considered. The descriptions of the features are as follows: dimension-reduction technique that is widely used in the
t-SNE [43]: This is a dimensionality-reduction approach fields of face recognition, speech recognition, and multime-
that reduces high-dimensional information into a low dimen- dia information retrieval. The main objective of LDA is to
sional embedding space for visualization implementations. forecast the optimal transformation according to the high-
Moreover, t-SNE measures the pairwise similarity distri- dimensional information, which is divided into classes. For
bution, attempting to optimize the visualization in a low- solving the problems related to the optimal discrimination
dimensional space by matching the distributions according projection matrix, the within-class and between-class scatter
to the Kullback–Leibler (KL) divergence. Additionally, the matrix is forecasted. The numerical formula for determining
pairwise similarities of t-SNE methods among the points the optimal discrimination projection matrices are as follows:
in both low- and high-dimensional spaces are expressed Mt fv BCclass Mt
as conditional probabilities uj|i and vj|i [44]. For instance, Mtopd = arg max . (8)
mt Mt fv WCclass Mt
the conditional probability uj|i is interpreted as the prob-
ability that a point j is in the neighborhood of point i Here, WCclass and BCclass represent the within- and between-
in the high-dimensional space. The probability of t-SNE class scatter matrices, respectively. To compute WCclass and
methods (as a Gaussian distribution around each data point BCclass , Eqs. (9) and (10) are used, respectively. Moreover,
in the high-dimensional space) is given by Eq. (1). The Eq. (11) gives the eigenvectors of the projection matrix Mt.
lower-dimensional space with the student’s t-distribution L
X
around each data point (for overcoming the overcrowding WCclass = (fvK − µNK ) (fvK − µNK )fv , (9)
issue in the Gaussian distribution) is expressed by Eq. (2). k=1
L
2 X
exp
−kxi −xj k BCclass = JK (µK − α) (µK − α)fv , (10)
2σi2
uj|i = , (1) k=1
2
exp −kxi −x2 k k Mt = eig R−1 BC . (11)
P
K class
2σi
k6=i
2 −1 In the above equations, the term RK is given by RK =
1 +
yi − yj
BCclass + WCclass , and the attributes of the data are denoted
vj|i = P −1 . (2) as fvK . αNK and NK represent the data vector and the data
1 + kyk − yl k2 samples of class K , respectively. Hence, the term FS LDA
k6=l
represents the feature extracted via the LDA method.
Subsequently, t-SNE minimizes the KL divergence among LSR [46]: This is a novel supervised dimensionality-
the distributions, which preserves the local structure of the reduction process. The LSR is used for extracting details
data points across the high- and low-dimensional spaces. from the data. Eq. (12) describes the optimization issue of
Moreover, t-SNE utilizes gradient descent for minimiz- LSR. Here, the class indicator matrix is denoted as Yo =
y1 , y2,··· yn0 , and the matrix with K th columns has a dimen-
ing the KL divergence with the gradient measured using
Eqs. (3) and (4). sionality vector, which is denoted as d ∗ + 1 on the basis of
∂C X Eq. (13). V ∗ represents the optimal transformation matrix.
uij − vij vij yi − yj ,
∗
∗
=4 (3) ∗ fv
The pseudo-inverse of V V is expressed as V V ∗ fv .
∂yi
j
−1 1
fv∗ ∗
2
I (Z ∗ ) = min
Z V − Yo
,
X
W = 1 + kyk − yl k2 , (4) (12)
Z 2 ∗
∗ S
k6=l ∗
Zlsr = V ∗ V fv V ∗ Yofv . (13)
where uij and vij represent the symmetric joint probabilities of
u +u Finally, the features extracted via LSR are expressed
ui|j and uj|i , such that uij = i|j2N j|i . The t-SNE gradient com-
as FS LSR .
putation is reformulated as the N-body simulation difficulty
Higher-Order Statistics: The higher-order statistics, such
by reorganizing the terms into attractive and repulsive forces,
as the skewness, kurtosis, root mean square (RMS), entropy,
as given by Eqs. (5) and (6), respectively. The combined
and energy, are extracted.
attractive and repulsive forces are given by Eq. (7).
X Skewness: This is the degree of distortion from the normal
uij vij W yi − yj , distribution in a set of data. Skewness can be positive, neg-
Atfr = (5)
j∈[1,··· ,N ],j6=i ative, zero, or undefined.
P The 3skewness is calculated using
(xc −x̄)
Eq. (14), where m3 = represents the third-moment
X
v2ij W yi − yj ,
Refr = − (6) n
j∈[1,··· ,N ],j6=i
dataset.
m3
∂C skn1 = 3 (14)
= 4 Atfr + Refr .
(7)
∂yi m22
Kurtosis: This is a statistical measure, which is used to Hence, the weighted feature vector FSi∗ is extracted, in
explain the distribution. It measures extreme values in tail. which the weight function is tuned or optimized by the pro-
Distribution with high kurtosis display tail data exceeds the posed MN-LA for minimizing the correlation between the
tails of the normal distribution. Distributions with less kurto- features. Finally, the optimal weighted feature vector FSi∗ is
sis display tail data, which are usually smaller than the tails subjected to the optimal hybrid classifier to predict the best
of the normal distribution, are given by scheduling rule.
m4
kur4 = 2 . (15) V. DEVELOPMENT OF PROPOSED FMS PREDICTION
m2 MODEL USING HYBRID CLASSIFIER
P
(x −x̄)4
c
P
c (x −x̄)2 A. DBN
Here, m4 = n , and m2 = n ; m4 represents
The DBN consists of several layers, similar to the neural net-
the fourth moment, and m2 represents the variance (σ ).
work (NN) [47]. The visible and hidden neurons are present in
RMS: The RMS is defined as the square root of the mean
the input and output layers, respectively. The DBN employs a
squared, where the mean squared is the arithmetic mean of
Boltzmann network for obtaining the results effectively. The
the squares of numbers. The RMS is also called the quadratic
result from the DBN is denoted as opt, and it has a binary for-
mean and is given by:
mat. Additionally, the result comprises the probability of the
sinusoidal function Bznprb (λ) A, as indicated by Eqs. (22)
r
1 2
trms = t + t22 + · · · + tn2 . (16) and (23). Here, the pseudo-temperature parameter, which is
n1
2 , and denoted as ptmpt, holds the probability’s noise level. Eq. (24)
where the items under observation are represented by tnv
refers to the stochastic model. The Boltzmann system is mod-
the total number of pixels is represented by nv.
eled according to the Boltzmann distribution to accurately
Entropy: This is a statistical measure of uncertainty that
design the input parameters.
accurately reflects the intraset distribution when a set of
opt = 1, with 1 − Bznprb (λ) (22)
patterns is provided. The entropy is given by Eq. (17), where
Qc represents the probability of obtaining the cth value. 0, with Bznprb (λ),
1
n
X Bznprb (λ) = −λ
, (23)
Entropy = − Qc log2 Qc (17) 1 + e ptmpt
c=1 1
lim Bznprb (λ) = lim −λ
Energy: The square root of the ASM (angular second ptmpt→0+ ptmpt→0+ 1 + e ptmpt
moment) texture character is used as the Energy, which is in
0, for λ < 0
the range [0,1]. The sum of the square elements in the GLCM 1
(gray level co-occurrence matrix) is computed via the entropy = , for λ = 0 (24)
2
calculation, according to Eq. (18). 1, for λ > 0.
X
Ener = Q (x − y)2 (18) The energy of the Boltzmann model for configuring the
x,y=0 neuron states and the mathematical equation is given by
Eq. (25). Here, the term nse represents the neuron states,
The attained higher-order statics are denoted as FS HIGH . ·
the weights of the neurons are denoted as B , and the biases
The features obtained via the aforementioned techniques, c,d
e.g., t-SNE, LDA, LSR, and higher-order statistics, are pre- of the neurons are denoted as β. The Boltzmann system in the
sented in Eq. (19). The combined form of the extracted fea- DBN encircles the neurons according to Eq. (26).
tures FSi is changed according to Eq. (20). X ·
BZE(nse1 ) = B nsec nsed − βc nsed (25)
FSi = FS t−SNE + FS LDA + FS LSR + FS HIGH (19) c<d
c,d
X ·
FSi = FS1, FS2, · · · FSNF (20) 1BZE (ec ) = nsec B +βc (26)
c,d
d
The merged features FSi are further used for weighted fea-
ture extraction, to alter the scaling of the features, so that each The configuration between the visible and hidden neurons
feature will be less correlative over other by proving unique with respect to the energy is given by Eqs. (27)–(29). In these
information for the prediction. In Eq. (20), NF represents the equations, the terms vsc and hdd refer to the visible and the
total number of features extracted. hidden neurons, respectively. The bias weights are denoted as
vswc and hdwd , respectively. The terms bs∗c and bs∗d represent
C. WEIGHTED FEATURE EXTRACTION the biases considered.
To extend the value of the feature vector at a large scale, +∗ + BZE(vs0 , hd) = −
X ·
B vsc hdd
a weight function WFi is multiplied by the feature vector∗ , c,d
(c,d)
as follows: X X
− vsc bs∗c − hdd hdwd (27)
FSi∗ = FSi × WFi . (21) c d
X
→ ·
1BZE vsc , hd = B hdd + bs∗c (28) The proposed DBN optimizes the activation function and
c,d the number of hidden neurons using the proposed MN-LA.
d
→ X ·
1BZE vs, hdd = B vsc + bs∗c (29)
c,d B. FUZZY CLASSIFIER
d
The major benefit of the fuzzy classifier is that the fuzzy set
The training set of the restricted Boltzmann machine theory can articulate and influence the uncertainty and vague-
(RBM) reduces the considered probabilities during the train- ness [44]. The starting procedure of the traditional fuzzy clas-
ing of the DBN, as the parameter of the weight function sifier is for providing the rules. Here, the actual classification
restricts the probability distribution of input data. Moreover, of the data is provided by the triangular membership function,
it has the ability to allot the probability to each visible and as given by Eq. (33). The membership function of y in C
hidden neuron with the assistance of the energy function, is represented by µC (Y ), and the low, medium and high
as indicated by Eq. (30). Here, the partition function pf is operators are represented as lw, med, and hg, respectively.
given by Eq. (31). The universe of discourse is denoted as Y , and the concerned
element is denoted as y.
1 −D(vs0 ,hd )
RB0 vs0 , hd =
e (30)
pf
0, y ≤ lw
e−D (vs ,hd )
X 0 0
pf = (31) y − lw 0
, lw < y ≤ med
→ → − lw
vs,hd µC (Y ) = medhg − y (33)
, med < y < hg
By measuring the difference in energy between the vis- hg − med
ible and hidden neurons, it is not similar to the RBM, 0, y ≥ hg
the standard Boltzmann machine automatically defines the
visible or hidden neurons. Moreover, the RBM categorizes We assume r̂1 = Eg , F h , where F h represents the output.
the data in an unsupervised fashion. The time consumed by The limiting factor δ1 is given by Eqs. (34) and (35), and
the RBM for obtaining the association via the developed MFLN indicates the number of linguistic variables.
method is long. Therefore, the contrast divergence (CD)
method is selected, and the algorithm related to the CD is as δ1 = max r̂1 − min r̂1
(34)
follows: δ1
(i) Select a training sample of vs and merge it with the δ10 = (35)
MFLN
visible neurons.
(ii) Calculate the term prbhd representing the hidden- The minimum and maximum limits of the linguistic vari-
neuron probability by taking the product of the visible vector ables are given by Eqs. (36)–(38).
and the weight matrix: wmt ∗ prbhd = σ (vs · wmt).
(iii) Sample the hidden states hd from the probabilities Lwmin = min r̂1
prbhd .
Lwmax = min r̂1 + δ10
(36)
(iv) Compute the outer product of the vectors vs
min max
and prbhd . The result is called a positive gradient med = Low + 0.1
φ + = vs · prbD max
= min r̂1 + 2δ10
hd . med (37)
states hd 0 , sample a recreation of
(v) From the hidden Hgmin = min r̂1 + 2δ10 + 0.1
visible states vs0 . Then, resample the hidden states from the
Hgmax = max r̂1
recreation of visible states. (38)
(vi) Calculate the outer product of vs0 and hd 0 . The
result is called the negative gradient (φ − = vs0 · hd 0D ). Because the limits of the membership function play a vital
(vii) The weight updates are computed as the positive role in determining the output degree, the function must be
gradient by subtracting from the negative gradients 1wmt = tuned properly. Hence, in this study, the proposed MN-LA
η φ+ − φ− .
was used to optimize the membership limits, which had a
(viii) Update the weights with the new values: wtc,d ∗ = positive effect on the final membership function.
1wtc,d + wtc,d .
The supervised learning model of the DBN occurs in the C. OPTIMIZED HYBRID CLASSIFIER
NN, and the unsupervised learning model occurs in the RBM. The results obtained from the optimized DBN and the fuzzy
The original output vector and the predicted output vector are classifier (with the proposed MN-LA) are merged via the
denoted as Gv and H v , respectively. The error between the logical AND operation. The AND operation is performed on
original and predicted vectors (given by Eq. (32)) should be each pair of input bits for performing the binary operations,
minimized. which considers two same-length binary depictions and there-
fore assumes the output of both classifiers. A diagram of the
mer1 = Gv − H v (32) optimized hybrid classifier is shown in Fig. 3.
B. SOLUTION ENCODING
As the proposed MN-LA is used in two places (weighted fea-
ture extraction and classification), the objectives tend to solve
the correlation between features and classification accuracy.
The solution encoding of weight optimization in weighted
FIGURE 3. Operation of the hybrid classifier for FMS scheduling.
feature extraction is shown in Fig. 4.
A. OBJECTIVE MODEL
In the proposed FMS scheduling model, the developed MN-
LA optimizes the weight function in weighted feature extrac-
tion, as well as the activation function and number of hidden
neurons in the DBN and the membership limits of the fuzzy
classifier. There are two main objectives for the prediction
model of FMS scheduling.
(a) Minimizing the correlation between features: While
optimizing the weight functions in weighted feature extrac- FIGURE 5. Solution encoding for the optimized DBN.
tion, the first objective of this model is to minimize the corre-
lation between the features. When the correlation between the Moreover, Fig. 5 shows the solution encoding for the opti-
two features is minimized, it is probable that different classes mized DBN in optimizing both the activation function and the
can be distinguished accurately. The correlation between the number of hidden neurons. The activation function AF can be
two features u and v is expressed as follows: sigmoid, ReLU (rectified linear unit), or tanh. The variable
P P P
n uv − u v NHN represents the number of hidden neurons in the DBN,
Correlation = r . which is varied from the number of classes to the total length
P 2 P 2 P 2
n u2 −
P
u n v − v of the features.
(39)
C. CONVENTIONAL LA
Hence, the first objective function, which is given by The LA is inspired by the distinctive social behavior of the
Eq. (40), depends on Eq. (39). lion [48]. There are two special behaviors in the conventional
FR1 = Min (Correlation) (40) LA algorithm: territorial defense and takeover. Moreover, the
{WF1 ,WF2 ,··· ,WFNF } conventional algorithm includes six phases: pride genera-
(b) Maximizing the classification accuracy: The second tion, fertility evaluation, mating, territorial defense, territorial
objective of the proposed prediction model is to optimize takeover, and termination.
In Eq. (53), rand represents a random number. In Eq. (54), FPR, FNR, NPV, FDR, F1 score, and Matthew’s correlation
Rmale refers to the male rate, whose value is 0.15. coefficient (MCC).
Similarly, the nomadic female lion is updated using
Eq. (55), which is based on Eq. (45). B. PERFORMANCE MEASURES
h i In this study, the following 10 performance measures were
∇m = dmfemale + (0.1c2 − 0.05) dmm−nomadic − c1 dmf _cub used to evaluate the performance:
(55) (a) Accuracy: given by Eq. (40)
(b) Sensitivity: measures the number of true positives,
In this work, the parameters have been tuned by the trial which are recognized exactly.
and error method, which is a fundamental method of prob-
trp
lem solving. It is characterized by repeated, varied attempts, Sen = (56)
which were continued until success. The pseudocode of the trp + fan
proposed MN-LA is presented in Algorithm 2. (c) Specificity: measures the number of true negatives,
which are determined precisely.
Algorithm 2 Pseudocode of the Proposed MN-LA
trn
Set the values of Dmale , Dfemale , and Dnomad
1 . Spe = (57)
fap
Determine the values of f (Dmale ), f (Dfemale ), and
f (Dnomad ). (d) Precision: It is the ratio of the number of positive
Initialize f ref = f (Dmale ) and h = 0. observations that are predicted exactly to the total number of
Reserve the values of Dmale and f (Dmale ). observations that are positively predicted.
The fertility evaluation is conducted. trp
The mating is performed, and the cub pool is obtained. Pr e = (58)
Dm_cub and Df _cub are obtained via gender clustering. trp + fap
Set Ucub = 0. (e) FPR: It is computed as the ratio of the number of false-
The cub-growth task is executed. positive predictions to the number of negative predictions.
Update the nomadic male lion using Eq. (54).
Update the nomadic male lion using Eqs. (45) and (55). fap
FPR = (59)
The territorial defense is performed; if the results obtained fap + trn
from the defense are 0, update the values of Dmale and (f) FNR: It is the proportion of positives that yield negative
f (Dmale ). Continue the evaluation. test outcomes.
If Ucub < Umax , the cub growth is executed.
The territorial defense is performed, and the updated values fan
FNR = (60)
of Dmale and Dfemale are obtained. trn + trp
The value of h is increased by 1. (g) NPV: It is the probability that subjects with a negative
If the condition of termination is not satisfied, update the screening test truly do not have the disease.
values of Dmale and f (Dmale ) again. Continue the evaluation
process. fan
NPV = (61)
fan + trn
(h) FDR: It is the number of false positives in all of the
VII. RESULTS AND DISCUSSIONS rejected hypotheses.
A. EXPERIMENTAL SETUP
fap
The developed optimal scheduling for the FMS using the opti- FDR = (62)
mized intelligent model was implemented using MATLAB fap + trp
2018a, and the performance of the model was evaluated. The (i) F1 score: It is defined as the harmonic mean between
population size was fixed as 10, and the number of iterations precision and recall. It is used as a statistical measure to rate
was fixed as 25. The performance of the proposed MN- performance.
LA-based Fuzzy+DBN (FDBN) was compared with that of
traditional algorithms, such as PSO-FDBN [49], grey wolf Sen · Pr e
F1score = (63)
optimizer (GWO)-FDBN [50], whale optimization algorithm Pr e + Sen
(WOA)-FDBN [50], and LA-FDBN [48]. Additionally, the (j) MCC: It is a correlation coefficient computed using four
proposed method was compared with traditional machine- values.
learning algorithms, e.g., the SVM [51], NN [52], DBN [47],
fuzzy classifier [44], and FDBN [44], [47]. The performance trp × trn − fap × fan
MCC = √
was analyzed according to the relevant performance mea- (trp + fap) (trp + fan) (trn + fap) (trn + fan)
sures, e.g., the accuracy, sensitivity, specificity, precision, (64)
FIGURE 6. Performance analysis of the proposed and conventional prediction models using t-SNE+LDA+LSR with the optimized
FDBN: (a) accuracy, (b) sensitivity, (c) precision, (d) FPR, (e) FNR, and (f) F1 score.
C. PERFORMANCE ANALYSIS USING T-SNE+LDA+LSR method is 33.84% greater than PSO-FDBN, 15.38% greater
The performance of the proposed and conventional models than that of GWO-FDBN, 30.76% greater than that of WOA-
using the features t-SNE+LDA+LSR to optimize the DBN FDBN, and 21.53% greater than that of LA-FDBN at a
with respect to the learning percentage is shown in Fig. 6. learning percentage of 65%. The precision of the improved
As shown in Fig. 6(a), the accuracy of the proposed MN-LA- MN-LA-FDBN was determined accurately, and it is shown
FDBN method was 0.2% higher than that of GWO-FDBN, in Fig. 6(c). At a learning percentage of 85%, the precision
0.5% higher than that of WOA-FDBN, and 1% higher than of the proposed MN-LA-FDBN was 35.1% higher than that
that of LA-FDBN at a learning percentage of 35%. From of PSO-FDBN and 42.8% higher than that of WOA-FDBN.
Fig. 6(b), the sensitivity of the proposed MN-LA-FDBN On considering Fig. 6(d), at a learning percentage of 75%,
TABLE 5. Overall performance analysis of proposed and conventional models for optimal scheduling using t-SNE+LDA+LSR with different
Heuristic-Based FDBNs.
the FPR of the proposed MN-LA-FDBN method is 90.6% 94.28% upgraded than WOA-FDBN. In Fig. 7(e), the FNR
better than PSO-FDBN, and 84% better than WOA-FDBN. of the proposed MN-LA-FDBN method is 83.07% better
As shown in Fig. 6(e), the FNR of the improved MN-LA- than PSO-FDBN, 64.61% better than GWO-FDBN, 89.69%
FDBN was 100% higher than those of WOA-FDBN and better than WOA-FDBN, and 76.92% better than LA-FDBN
PSO-FDBN. In Fig. 6(f), the F1 score of the proposed MN- at a learning percentage of 65%. On considering Fig. 7(f)at
LA-FDBN method is 11.76% superior to PSO-FDBN, 8.23% a learning percentage of 85% the F1 score of the proposed
superior to GWO-FDBN, 17.64% superior to WOA-FDBN, MN-LA-FDBN method is 14.11% improved than PSO-
and 29.41% superior to LA-FDBN at a learning percentage FDBN, 16.47% improved than WOA-FDBN, and 15.29%
of 85%. The overall performance of the proposed MN-LA- improved than LA-FDBN. Table 6 presents the overall per-
FDBN and the conventional machine-learning techniques formance of the modified MN-LA-FDBN and the traditional
with t-SNE+LDA+LSR features is presented in Table 5. The optimization-based FDBN methods with higher-order statis-
accuracy of the improved MN-LA-FDBN was 2.2% higher tics. As shown, the accuracy of the proposed MN-LA-FDBN
than that of PSO-FDBN, 0.7% higher than that of GWO- was 6.9% higher than those of PSO-FDBN and WOA-FDBN,
FDBN, 6.2% higher than that of WOA-FDBN, and 1.4% 0.7% higher than that of GWO-FDBN, and 2.9% higher than
higher than that of LA-FDBN. Moreover, the precision of that of LA-FDBN. Additionally, the precision of the proposed
the implemented LA was 21.9% higher than that of PSO- MN-LA-FDBN was higher than those of the other algorithms.
FDBN and 30.7% higher than that of WOA-FDBN. The It was 21.6% higher than those of PSO-FDBN and WOA-
results confirm that the proposed MN-LA-FDBN model was FDBN. Therefore, the proposed MN-LA-FDBN method was
effective for predicting the optimal scheduling in the FMS. superior to the conventional methods for predicting the opti-
mal scheduling in the FMS.
D. PERFORMANCE ANALYSIS USING
HIGHER-ORDER STATISTICS
The performance of the proposed MN-LA-FDBN and the E. PERFORMANCE ANALYSIS FOR
different metaheuristic-based FDBN methods using the T-SNE+LDA+LSR+STATISTICS
higher-order statistical features with respect to the learning The performance of the developed MN-LA-FDBN and
percentage is shown in Fig. 7. The accuracy of the pro- the traditional algorithms with t-SNE+LDA+LSR+higher-
posed MN-LA-FDBN was high for all learning percentages, order statistics with respect to the learning percentage is
as shown in Fig. 7(a). For a learning percentage of 85%, presented in Fig. 8. As shown in Fig. 8(a), the accuracy of
the accuracy of the proposed MN-LA-FDBN was 0.5% the proposed MN-LA-FDBN was high for all the learning
higher than that of LA-FDBN, 4.8% higher than that of percentages. It was 0.5% higher than that of GWO-FDBN,
PSO-FDBN, and 6.1% higher than that of WOA-FDBN. 2.7% higher than that of WOA-FDBN, and 7.5% higher
In Fig. 7(b), the sensitivity of the proposed MN-LA-FDBN than that of PSO-FDBN at a learning percentage of 75%.
method is 38.46% superior to PSO-FDBN, 16.92% supe- From Fig. 8(b) the Sensitivity of the proposed MN-LA-
rior to GWO-FDBN, 33.84% superior to WOA-FDBN, and FDBN method is 24.70% greater than PSO-FDBN, 35.29%
27.69% superior to LA-FDBN at a learning percentage of greater than that of GWO-FDBN, 10.58% greater than that
65%. Additionally, the precision of the proposed MN-LA- of WOA-FDBN, and 41.17% greater than that of LA-FDBN
FDBN was high for all the learning percentages. As shown at a learning percentage of 85%. As shown in Fig. 8(c),
in Fig. 7(c), the precision of the proposed MN-LA-FDBN the precision of the proposed MN-LA-FDBN was high com-
was 11.1% higher than that of PSO-FDBN and 12.3% higher pared with the other models. For a learning percentage of
than that of FDBN. When considering Fig. 7(d),at a learn- 35%, the precision of the proposed MN-LA-FDBN was
ing percentage of 35%, the FPR of the proposed MN-LA- 3% higher than that of WOA-FDBN and 5.2% higher than
FDBN method is 91.42% upgraded than PSO-FDBN, and that of PSO-FDBN. In Fig. 8(d) the FPR of the developed
FIGURE 7. Performance analysis of the proposed and conventional prediction models using higher-order statistics with
the optimized FDBN: (a) accuracy, (b) sensitivity, (c) precision, (d) FPR, (e) FNR, and (f) F1 score.
MN-LA-FDBN was 100% better than those of WOA-FDBN existing models for t-SNE+LDA+LSR+Statistics is pre-
and PSO-FDBN. In Fig. 8(e) the FNR of the proposed sented in Table 7. The accuracy of the proposed MN-LA-
MN-LA-FDBN method is 58.82% better than PSO-FDBN, FDBN was 6.9% higher than that of PSO-FDBN, 0.7% higher
50.58% better than GWO-FDBN, 76.47% better than WOA- than that of GWO-FDBN, and 2.9% higher than that of WOA-
FDBN, and 42.35% better than LA-FDBN at a learning FDBN. The precision of the proposed MN-LA-FDBN was
percentage of 85%. On considering Fig. 8(f) at a learning per- 27.5% higher than that of PSO-FDBN and 19% higher than
centage of 75% the F1 score of the proposed MN-LA-FDBN that of WOA-FDBN. The results indicate that the imple-
method is 4% improved than PSO-FDBN, 13.33% improved mented MN-LA-FDBN method with the combined features
than WOA-FDBN, and 10.66% improved than WOA-FDBN. was suitable for predicting the optimal scheduling in the FMS
The performance of the proposed MN-LA-FDBN and the and was superior to the other models.
TABLE 6. Overall performance analysis of proposed and conventional models for optimal scheduling using Higher-Order statistics with different
Heuristic-Based FDBNs.
TABLE 7. Overall performance analysis of proposed and conventional models for optimal scheduling using t-SNE+LDA+LSR+Statistics with different
Heuristic-Based FDBNs.
F. PERFORMANCE ANALYSIS FOR OPTIMAL with optimized weighted features is presented in Table 8.
WEIGHTED FEATURES As shown, the accuracy of the improved MN-LA-FDBN was
The performance of the improved MN-LA-FDBN and the tra- 14.1% higher than that of PSO-FDBN, 6.6% higher than that
ditional FDBN with the optimal weighted feature according of GWO-FDBN, 13.2% higher than that of WOA-FDBN, and
to the learning percentage is shown in Fig. 9. The accuracy 5% higher than that of LA-FDBN. Moreover, the precision of
of the proposed MN-LA-FDBN was high compared with the the proposed MN-LA-FDBN was 28.9% higher than that of
other algorithms, as shown in Fig. 9(a). It was 1.4% higher PSO-FDBN and 30% higher than that of WOA-FDBN. Thus,
than that of PSO-FDBN, 1.5% higher than that of WOA- the proposed MN-LA-FDBN exhibited good performance for
FDBN, 2% higher than that of LA-FDBN, and 2.6% higher optimal scheduling with the optimized weighted features.
than that of PSO-FDBN for a learning percentage of 35%.
From Fig. 9(b), the sensitivity of the proposed MN-LA- G. EFFECT OF T-SNE+LDA+LSR ON MACHINE LEARNING
FDBN method is 67.27% greater than PSO-FDBN, 54.54% The performance analysis of the developed MN-LA-FDBN
greater than that of GWO-FDBN, 56.36% greater than that and the conventional classifiers with t-SNE+LDA+LSR
of WOA-FDBN, and 49.09% greater than that of LA-FDBN with respect to the learning percentage is presented in Fig. 10.
at a learning percentage of 55%. In Fig. 9(c), the precision As shown in Fig. 10(a), the accuracy of the improved
of the proposed MN-LA-FDBN was 4.7% higher than that MN-LA-FDBN has exactly defined the positive observations
of WOA-FDBN and 5.3% higher than that of PSO-FDBN from the whole observations for all the learning percentages.
for the learning percentage of 35%, as shown in Fig. 9(d). It was 1% higher than those of the FDBN and DBN, 6.5%
In Fig. 9(e), the FNR of the proposed MN-LA-FDBN method higher than those of the fuzzy classifier and the NN, and
is 85.71% better than PSO-FDBN, 82.85% better than GWO- 3.1% higher than that of the SVM for a learning percent-
FDBN, 88.57% better than WOA-FDBN, and 71.42% better age of 35%. As shown in Fig. 10(b) at a learning percent-
than LA-FDBN at a learning percentage of 35%. As indicated age of 85%, the Sensitivity of the proposed MN-LA-FDBN
by Fig. 9(f), the FNR of the proposed MN-LA-FDBN is is 5.88% higher than those of FDBN and DBN, 27.05%
correctly defined the false-negative observations from all the higher than fuzzy classifier, 23.52% higher than NN, and
observations for all the learning percentages. The FNR of 12.94% higher than SVM. For all the learning percentages,
the proposed MN-LA-FDBN was 37.5% higher than that of the precision of the implemented MN-LA-FDBN high and
WOA-FDBN, 52.3% higher than that of GWO-FDBN, and it is correctly defined the true values from all the values,
60% higher than that of LA-FDBN. The overall performance as shown in Fig. 10(c). For a learning percentage of 85%,
of the proposed MN-LA-FDBN and conventional models the precision of the proposed MN-LA-FDBN was 42.8%
FIGURE 8. Performance analysis of the proposed and conventional methods using t-SNE+LDA+LSR+Statistics with the
optimized FDBN: (a) accuracy, (b) sensitivity, (c) precision, (d) FPR, (e) FNR, and (f) F1 score.
higher than those of FDBN and DBN, 100% higher than that MN-LA-FDBN method is 4% superior to SVM, 6.66% supe-
of the fuzzy classifier, 81.8% higher than that of NN, and rior to NN, 5.33% superior to DBN, 1.33% superior tofuzzy
66.6% higher than that of the SVM. On considering Fig. 10(d) classifier, 8% superior to FDBN. The overall classification
at a learning percentage of 55% the FPR of the proposed MN- analysis of the proposed MN-LA-FDBN and the conventional
LA-FDBN method is 72.72% improved than SVM, 74.54% classifiers with t-SNE+LDA+LSR is presented in Table 9.
improved than NN, 90.90% improved than DBN, 78.18% As shown, the accuracy of the proposed MN-LA-FDBN was
improved than fuzzy classifier, and 94.54% improved than 8.8% higher than that of the SVM, 5.4% higher than that of
FDBN. In Fig. 10(e), the FNR of the proposed MN-LA- the NN, 3.8% higher than that of the DBN, 12.5% higher
FDBN method is 80% better than DBN, and fuzzy classifier than that of the fuzzy classifier, and 3% higher than that of
at a learning percentage of 35%. When considering Fig. 10(f) the FDBN. Moreover, the precision of the proposed MN-LA-
at a learning percentage of 75% the F1 score of the proposed FDBN was 52% higher than that of the SVM, 44% higher
FIGURE 9. Performance analysis of the proposed and conventional methods using optimal weighted features with the
optimized FDBN: (a) accuracy, (b) sensitivity, (c) precision, (d) FPR, (e) FNR, and (f) F1 score.
than that of the NN, 28.5% higher than that of the DBN, 60% higher-order statistics for different learning percentages,
higher than that of the fuzzy classifier, and 26.1% higher than as shown in Fig. 11. The accuracy of the proposed MN-LA-
that of the FDBN. The results indicate that the proposed MN- FDBN was high for all the learning percentages, as shown
LA-FDBN outperformed the conventional algorithms and is in Fig. 11(a). It was 4.9% higher than those of the FDBN and
reliable for optimal scheduling. DBN, 28.7% higher than those of the fuzzy classifier and the
NN, and 21.4% higher than that of the SVM for a learning
H. EFFECT OF HIGHER-ORDER STATISTICS percentage of 85%. On considering Fig. 11(b) at a learning
ON MACHINE LEARNING percentage of 55% the Sensitivity of the proposed MN-LA-
We analyzed the proposed MN-LA-FDBN and the con- FDBN method is 81.81% upgraded than SVM,NN and fuzzy
ventional machine-learning algorithms with the use of classifier, 54.54% upgraded than DBN,and FDBN. When
TABLE 8. Overall performance analysis of proposed and conventional models for optimal scheduling using weighted feature with different
Heuristic-Based FDBNs.
TABLE 9. Overall performance analysis of proposed and conventional machine-learning models for predicting optimal scheduling in FMS with
t-SNE+LDA+LSR.
TABLE 10. Overall performance analysis of proposed and conventional machine-learning models for predicting optimal scheduling in FMS with
Higher-Order statistics.
considering Fig. 11(C) at a learning percentage of 85% the accuracy of the proposed MN-LA-FDBN was 11.2% higher
precision of the proposed MN-LA-FDBN method is 47.05% than that of the SVM, 17.8% higher than that of the NN,
superior to SVM, 49.41% superior to NN,and fuzzy classifier 8.6% higher than that of the DBN, 13.1% higher than that of
23.52% superior to DBN, and FDBN. From Fig. 11(d) the the fuzzy classifier, and 7.8% higher than that of the FDBN.
FPR of the proposed MN-LA-FDBN method is 72.30% better As indicated by Table 10, the precision of the proposed MN-
than SVM, and fuzzy classifier, 61.53% better than NN, LA-FDBN was 52% higher than that of the SVM, 44% higher
92.30% better than DBN, and 95.38% better than FDBN at than that of the NN, 30% higher than that of the DBN, 55.9%
a learning percentage of 65%. In Fig. 11(e) the FNR of the higher than that of the fuzzy classifier, and 28.2% higher than
proposed MN-LA-FDBN method is 80% better than DBN, that of the FDBN. The results confirm that the proposed MN-
and FDBN at a learning percentage of 35%. For the learning LA-FDBN performed well and is appropriate for optimal
percentage of 85%, in Fig. 11(f) the F1 score of the proposed scheduling.
MN-LA-FDBN was 4.1% higher than those of the FDBN and
DBN, 15.3% higher than those of the fuzzy classifier and I. EFFECT OF T-SNE+LDA+LSR+STATISTICS
the NN, and 11.9% higher than that of the SVM, as shown ON MACHINE LEARNING
in Fig. 11(i). The overall classification performance with the The classification performance of the improved MN-LA-
use of higher-order statistics for the proposed MN-LA-FDBN FDBN and the traditional classifiers with t-SNE+LDA+
and the traditional algorithms is presented in Table 10. The LSR+Statistics according to the learning percentage is shown
FIGURE 10. Performance analysis of the proposed and conventional machine-learning algorithms using t-SNE+LDA+LSR
with the optimized FDBN: (a) accuracy, (b) sensitivity, (c) precision, (d) FPR, (e) FNR, and (f) F1 score.
in Fig. 12. For a learning percentage of 85%, the accuracy MN-LA-FDBN was 3% higher than that of the FDBN, 21.9%
of the proposed MN-LA-FDBN was 1.2% higher than those higher than that of the fuzzy classifier, 4.1% higher than
of the FDBN and DBN, 72% higher than that of the fuzzy that of the DBN, 23.4% higher than that of the NN, and
classifier, 26.1% higher than that of the NN, and 17.1% 21.9% higher than that of the SVM. From Fig. 12(d) the
higher than that of the SVM, as shown in Fig. 12(a). When FPR of the proposed MN-LA-FDBN method is 72.72% bet-
considering Fig. 12(b) at a learning percentage of 65%, ter than SVM, 69.09% better than NN, 90.90% better than
the sensitivity of the proposed MN-LA-FDBN method is DBN, and FDBN and 70.90% better than fuzzy classifier
53.84% superior to SVM, NN,and fuzzy classifier 38.46% at a learning percentage of 55%. In Fig. 12(e), the FNR
superior to DBN, and FDBN. As indicated by Fig. 12(c), for of the proposed MN-LA-FDBN method is 80% better than
a learning percentage of 35%, the precision of the proposed DBN, and FDBN at a learning percentage of 35%. When
FIGURE 11. Performance analysis of the proposed and conventional machine-learning algorithms using higher-order
statistics with the optimized FDBN: (a) accuracy, (b) sensitivity, (c) precision, (d) FPR, (e) FNR, and (f) F1 score.
considering Fig. 12(f), at a learning percentage of 85% the proposed MN-LA-FDBN was 60% higher than that of the
F1 score of the proposed MN-LA-FDBN method is 24.70% SVM, 44% higher than that of the NN, 37.8% higher than that
superior to SVM, 25.88% superior to NN, 23.52% superior of the DBN, 49.9% higher than that of the fuzzy classifier,
to DBN, and FDBN and 17.64% superior to fuzzy classi- and 27% higher than that of the FDBN. The results indicate
fier. The overall classification analysis of the proposed and that the proposed MN-LA-FDBN outperformed conventional
conventional models with t-SNE+LDA+LSR+Statistics is classifiers.
presented in Table 11. As shown, the accuracy of the proposed
MN-LA-FDBN was 15% higher than that of the SVM, 7.8% J. EFFECT OF OPTIMAL WEIGHTED FEATURES ON
higher than that of the NN, 12.1% higher than that of the MACHINE LEARNING
DBN, 10.4% higher than that of the fuzzy classifier, and 8.6% A classification analysis of the proposed MN-LA-FDBN and
higher than that of the FDBN. Moreover, the precision of the the classical methods with weighted features according to the
FIGURE 12. Performance analysis of the proposed and conventional machine-learning algorithms using
t-SNE+LDA+LSR+Statistics with the optimized FDBN: (a) accuracy, (b) sensitivity, (c) precision, (d) FPR, (e) FNR, and
(f) F1 score.
learning percentage was performed, as shown in Fig. 13. The learning percentage of 35%, the precision of the proposed
accuracy of the proposed MN-LA-FDBN is high for all the MN-LA-FDBN was 5.2% higher than that of the FDBN,
learning percentages, as shown in Fig. 13(a). For a learning 21.9% higher than that of the fuzzy classifier, 6.3% higher
percentage of 35%, the accuracy of the proposed MN-LA- than that of the DBN, 17.6% higher than that of the NN,
FDBN was 3.1% higher than those of the FDBN and DBN, and 20.4% higher than that of the SVM. From Fig. 13(d)
7.6% higher than that of the fuzzy classifier, 4.2% higher the FPR of the proposed MN-LA-FDBN method is 73.33%
than that of the NN, and 5.3% higher than that of the SVM. better than SVM, 66.66% better than NN, and fuzzy classifier,
On considering Fig. 13(b) at a learning percentage of 55% 86.66% better than DBN, and 89.33% better than FDBN
the Sensitivity of the proposed MN-LA-FDBN method is at a learning percentage of 75%. In Fig. 13(e) the FNR of
81.81% upgraded than SVM,NN and fuzzy classifier, 54.54% the proposed MN-LA-FDBN method is 77.33% better than
upgraded than DBN,and FDBN. As shown in Fig. 13(c) at the DBN, and FDBN at a learning percentage of 75%. When
FIGURE 13. Performance analysis of the proposed and conventional machine-learning algorithms using optimized
weighted features with the optimized FDBN: (a) accuracy, (b) sensitivity, (c) precision, (d) FPR, (e) FNR, and (f) F1 score.
considering Fig. 13(f) at a learning percentage of 65% the 39.9% higher than that of the SVM, 54% higher than those
F1 score of the proposed MN-LA-FDBN method is 23.07% of the fuzzy classifier and the NN, 24.3% higher than that of
superior to SVM, 24.61% superior to NN, 29.23% superior the DBN, and 19.5% higher than that of the FDBN. Thus,
to DBN, and FDBN and 20% superior to fuzzy classifier. the proposed MN-LA-FDBN was more effective than the
Table 12 presents the overall performance of the improved conventional methods and was suitable for optimal schedul-
MN-LA-FDBN and conventional models using weighted fea- ing in the FMS.
tures. As shown, the accuracy of the proposed MN-LA-
FDBN was 11.5% higher than that of the SVM, 17.8% higher K. ANALYSIS BASED ON COMPUTATIONAL TIME
than those of the fuzzy classifier and the NN, 10.6% higher The computational time of the proposed and existing meth-
than that of the DBN, and 9% higher than that of the FDBN. ods is analyzed and listed in Table 13. On considering the
Similarly, the precision of the proposed MN-LA-FDBN was computational time, the proposed algorithm is better than
TABLE 11. Overall performance analysis of proposed and conventional machine-learning models for predicting optimal scheduling in FMS with
t-SNE+LDA+LSR+Higher-Order statistics.
TABLE 12. Overall performance analysis of proposed and conventional Machine-Learning models for predicting optimal scheduling in FMS with
optimized weighted features.
TABLE 13. The computational time of the proposed and existing method. pertaining to the FMS, features (e.g., t-SNE, LDA, LSR, and
higher-order statistical features) are extracted. Next, optimal
weight feature extraction is performed for selecting the opti-
mal features with low correlation via the developed MN-LA.
These features are applied to a hybrid classifier. The activa-
tion function and the number of hidden neurons in the DBN,
as well as the membership function in the fuzzy classifier, are
optimized by the proposed MN-LA. According to test results,
with the use of the optimized weighted features, the accuracy
of the proposed MN-LA-FDBN was 11.5% higher than that
of the SVM, 17.8% higher than those of the fuzzy classifier
the existing methods, such as PSO, GWO, WOA, and LA. and the NN, 10.6% higher than that of the DBN, and 9%
The computational time of the proposed MNLA is 17.43% higher than that of the FDBN. Hence, the proposed MN-LA
better than PSO, 35.60% better than GWO, 11.75% better is suitable for predicting the optimal scheduling in FMSs.
than WOA, and 73.49% better than LA, respectively. The advantages of the proposed method include, reduced
Time Complexity of the Proposed Method Using Big O manufacturing cost, increased labor productivity, increased
Notation: machine efficiency, improved product quality, increased sys-
Big O notation is the most common metric for calculating tem reliability, reduced parts inventory, shorter lead times,
time complexity. Big O specifically describes the worst-case and increased production rate. The disadvantages include
scenario, and can be used to describe the execution time a high initial set up cost, requirements of skilled workers,
required or the space used by an algorithm. The time com- and a more complicated system. Moreover, the proposed
plexity of the proposed MN-LA is O (p∗ n ∧ 2), where, n is MN-LA has not been tested to solve many complex problems.
the initial population size and p is the no. of iterations. In future, a 3D animation module for will be developed for
better visualization of the system and results.
VIII. CONCLUSIONS
This paper introduced an effective prediction model using REFERENCES
intelligent hybrid learning for FMS scheduling. It comprises
[1] J. Wang, Y. Ma, L. Zhang, R. X. Gao, and D. Wu, ‘‘Deep learning for
feature extraction, optimal weighted feature extraction, and smart manufacturing: Methods and applications,’’ J. Manuf. Syst., vol. 48,
prediction. Once the data are collected from standard datasets pp. 144–156, Jul. 2018.
[2] A. Hayes. (2019). Flexible Manufacturing System (FMS). [23] S. Burnwal and S. Deb, ‘‘Scheduling optimization of flexible manufac-
Accessed: Feb. 27, 2020. [Online]. Available: https://www.investopedia. turing system using cuckoo search-based approach,’’ Int. J. Adv. Manuf.
com/terms/f/flexible-manufacturing-system.asp Technol., vol. 64, nos. 5–8, pp. 951–959, Feb. 2013.
[3] R. Rendall, I. Castillo, B. Lu, B. Colegrove, M. Broadway, L. H. Chiang, [24] P. Priore, D. De La Fuente, A. Gomez, and J. Puente, ‘‘A review of machine
and M. S. Reis, ‘‘Image-based manufacturing analytics: Improving the learning in dynamic scheduling of flexible manufacturing systems,’’ Artif.
accuracy of an industrial pellet classification system using deep neural Intell. Eng. Des., Anal. Manuf., vol. 15, no. 3, pp. 251–263, Jun. 2001.
networks,’’ Chemometric Intell. Lab. Syst., vol. 180, pp. 26–35, Sep. 2018. [25] P. Priore, B. Ponte, J. Puente, and A. Gómez, ‘‘Learning-based scheduling
[4] S. Wang, Y. C. Liang, W. D. Li, and X. T. Cai, ‘‘Big data enabled intel- of flexible manufacturing systems using ensemble methods,’’ Comput. Ind.
ligent immune system for energy efficient manufacturing management,’’ Eng., vol. 126, pp. 282–291, Dec. 2018.
J. Cleaner Prod., vol. 195, pp. 507–520, Sep. 2018. [26] X. Wang, K. Xing, X. Li, and J. Luo, ‘‘An estimation of distribution
[5] B. Wally, J. Vyskocil, P. Novak, C. Huemer, R. Sindelar, P. Kadera, algorithm for scheduling problem of flexible manufacturing systems using
A. Mazak, and M. Wimmer, ‘‘Flexible production systems: Automated Petri nets,’’ Appl. Math. Model., vol. 55, pp. 776–788, Mar. 2018.
generation of operations plans based on ISA-95 and PDDL,’’ IEEE Robot. [27] X. Li, K. Xing, M. Zhou, X. Wang, and Y. Wu, ‘‘Modified dynamic
Autom. Lett., vol. 4, no. 4, pp. 4062–4069, Oct. 2019. programming algorithm for optimization of total energy consumption in
[6] A. M. Farid, ‘‘Static resilience of large flexible engineering systems: flexible manufacturing systems,’’ IEEE Trans. Autom. Sci. Eng., vol. 16,
Axiomatic design model and measures,’’ IEEE Syst. J., vol. 11, no. 4, no. 2, pp. 691–705, Apr. 2019.
pp. 2006–2017, Dec. 2017. [28] W. Duan, C. Zhong, X. Wang, A. U. Rehman, U. Umer, and N. Wu,
[7] K. S. Kul’ga and A. A. Kitaev, ‘‘Simulation of flexible manufacturing ‘‘A deadlock prevention policy for flexible manufacturing systems mod-
systems at the preliminary design stage,’’ Russian Eng. Res., vol. 37, no. 11, eled with Petri nets using structural analysis,’’ IEEE Access, vol. 7,
pp. 974–982, Nov. 2017. pp. 49362–49376, 2019.
[8] P. Wenzelburger and F. Allgöwer, ‘‘A novel optimal online scheduling [29] Y. Dong, Y. Chen, S. Li, M. A. El-Meligy, and M. Sharaf, ‘‘An efficient
scheme for flexible manufacturing systems,’’ IFAC-PapersOnLine, vol. 52, deadlock recovery policy for flexible manufacturing systems modeled with
no. 10, pp. 1–6, Jan. 2019. Petri nets,’’ IEEE Access, vol. 7, pp. 11785–11795, 2019.
[9] X. Gang and Q. Quan, ‘‘Research on planning scheduling of flexible man- [30] M. Bashir, Z. Li, M. Uzam, A. Al-Ahmari, N. Wu, D. Liu, and T. Qu,
ufacturing system based on multi-level list algorithm,’’ Procedia CIRP, ‘‘A minimal supervisory structure to optimally enforce liveness on Petri
vol. 56, pp. 569–573, Jan. 2016. net models for flexible manufacturing systems,’’ IEEE Access, vol. 5,
[10] A. M. El-Tamimi, M. H. Abidi, S. H. Mian, and J. Aalam, ‘‘Analysis of pp. 15731–15749, 2017.
performance measures of flexible manufacturing system,’’ J. King Saud [31] I. Lee and Y. J. Shin, ‘‘Machine learning for enterprises: Applications,
Univ.-Eng. Sci., vol. 24, pp. 115–129, Jul. 2012. algorithm selection, and challenges,’’ Bus. Horizons, vol. 63, no. 2,
[11] P. Kostal, A. Mudrikova, and D. Michal, ‘‘Possibilities of intelligent flex- pp. 157–170, Mar. 2020.
ible manufacturing systems,’’ in Proc. IOP Conf. Ser., Mater. Sci. Eng.,
[32] X. Chu, D. Gao, S. Cheng, L. Wu, J. Chen, Y. Shi, and Q. Qin, ‘‘Worker
vol. 659, Oct. 2019, Art. no. 012035.
assignment with learning-forgetting effect in cellular manufacturing sys-
[12] S. Wang, C. Wang, and Y. Yu, ‘‘Comments on ‘Siphon-based deadlock tem using adaptive memetic differential search algorithm,’’ Comput. Ind.
prevention policy for flexible manufacturing systems,’’’ IEEE Trans. Syst., Eng., vol. 136, pp. 381–396, Oct. 2019.
Man, Cybern. A, Syst. Humans, vol. 41, no. 2, pp. 338–340, Mar. 2011.
[33] I. M. Cavalcante, E. M. Frazzon, F. A. Forcellini, and D. Ivanov, ‘‘A super-
[13] Z. Li and G. Liu, ‘‘Comments on ‘An iterative synthesis approach to Petri
vised machine learning approach to data-driven simulation of resilient
net based deadlock prevention policy for flexible manufacturing systems,’’’
supplier selection in digital manufacturing,’’ Int. J. Inf. Manage., vol. 49,
IEEE Trans. Syst., Man, Cybern. A, Syst. Humans, vol. 39, no. 3, p. 692,
pp. 86–97, Dec. 2019.
May 2009.
[34] B. Wang, T. Xiang, N. Li, W. He, W. Li, and X. Hei, ‘‘A symmetric sine
[14] J. Luo, Z. Liu, and M. Zhou, ‘‘A Petri net based deadlock avoidance
cosine algorithm with adaptive probability selection,’’ IEEE Access, vol. 8,
policy for flexible manufacturing systems with assembly operations and
pp. 25272–25285, 2020.
multiple resource acquisition,’’ IEEE Trans. Ind. Informat., vol. 15, no. 6,
pp. 3379–3387, Jun. 2019. [35] V. K. Kamboj, A. Nandi, A. Bhadoria, and S. Sehgal, ‘‘An intensify Harris
hawks optimizer for numerical and engineering optimization problems,’’
[15] K. Barkaoui, A. Chaoui, and B. Zouari, ‘‘Supervisory control of discrete
Appl. Soft Comput., vol. 89, Apr. 2020, Art. no. 106018.
event systems based on structure theory of Petri nets,’’ in Proc. IEEE Int.
Conf. Syst., Man, Cybern. Comput. Cybern. Simulation, vol. 4, Oct. 1997, [36] M. Wang, C. Wu, L. Wang, D. Xiang, and X. Huang, ‘‘A feature selection
pp. 3750–3755. approach for hyperspectral image based on modified ant lion optimizer,’’
[16] H. Liu, K. Xing, W. Wu, M. Zhou, and H. Zou, ‘‘Deadlock prevention Knowl.-Based Syst., vol. 168, pp. 39–48, Mar. 2019.
for flexible manufacturing systems via controllable siphon basis of Petri [37] G. T. Reddy, M. P. K. Reddy, K. Lakshmanna, R. Kaluri, D. S. Rajput,
nets,’’ IEEE Trans. Syst., Man, Cybern., Syst., vol. 45, no. 3, pp. 519–529, G. Srivastava, and T. Baker, ‘‘Analysis of dimensionality reduction tech-
Mar. 2015. niques on big data,’’ IEEE Access, vol. 8, pp. 54776–54788, 2020.
[17] B. Huang, H. Zhu, G. Zhang, and X. Lu, ‘‘On further reduction of con- [38] S. Bhattacharya, S. R. K. S, P. K. R. Maddikunta, R. Kaluri, S. Singh,
straints in ‘nonpure Petri net supervisors for optimal deadlock control of T. R. Gadekallu, M. Alazab, and U. Tariq, ‘‘A novel PCA-firefly based
flexible manufacturing systems,’’’ IEEE Trans. Syst., Man, Cybern., Syst., XGBoost classification model for intrusion detection in networks using
vol. 45, no. 3, pp. 542–543, Mar. 2015. GPU,’’ Electronics, vol. 9, no. 2, p. 219, Jan. 2020.
[18] D. Herrero-Perez and H. Martinez-Barbera, ‘‘Modeling distributed trans- [39] T. R. Gadekallu, N. Khare, S. Bhattacharya, S. Singh, P. K. R. Maddikunta,
portation systems composed of flexible automated guided vehicles in I.-H. Ra, and M. Alazab, ‘‘Early detection of diabetic retinopathy using
flexible manufacturing systems,’’ IEEE Trans. Ind. Informat., vol. 6, no. 2, PCA-firefly based deep learning model,’’ Electronics, vol. 9, no. 2, p. 274,
pp. 166–180, May 2010. Feb. 2020.
[19] A. P. Rifai, H.-T. Nguyen, H. Aoyama, S. Z. M. Dawal, and [40] C. Iwendi, P. K. R. Maddikunta, T. R. Gadekallu, K. Lakshmanna,
N. A. Masruroh, ‘‘Non-dominated sorting biogeography-based optimiza- A. K. Bashir, and M. J. Piran, ‘‘A Metaheuristic optimization approach for
tion for bi-objective reentrant flexible manufacturing system scheduling,’’ energy efficiency in the IoT networks,’’ Softw., Pract. Exp., to be published,
Appl. Soft Comput., vol. 62, pp. 187–202, Jan. 2018. doi: 10.1002/spe.2797.
[20] M. G. Filho, C. F. Barco, and R. F. T. Neto, ‘‘Using genetic algorithms [41] D.-C. Li, C. Wu, and F. M. Chang, ‘‘Using data-fuzzification technology in
to solve scheduling problems on flexible manufacturing systems (FMS): small data set learning to improve FMS scheduling accuracy,’’ Int. J. Adv.
A literature survey, classification and analysis,’’ Flexible Services Manuf. Manuf. Technol., vol. 27, nos. 3–4, pp. 321–328, Dec. 2005.
J., vol. 26, no. 3, pp. 408–431, Sep. 2014. [42] M. Srinivasan and Y. B. Moon, ‘‘Aframework fora goal-drivenapproachto
[21] K. Kim, G. Yamazaki, L. Lin, and M. Gen, ‘‘Network-based hybrid genetic grouptechnology applications using conceptual clustering,’’ Int. J. Prod.
algorithm for scheduling in FMS environments,’’ Artif. Life Robot., vol. 8, Res., vol. 35, no. 3, pp. 847–866, Mar. 1997.
no. 1, pp. 67–76, Sep. 2004. [43] D. M. Chan, R. Rao, F. Huang, and J. F. Canny, ‘‘t-SNE-CUDA: GPU-
[22] A. Prakash, F. T. S. Chan, and S. G. Deshmukh, ‘‘FMS scheduling with accelerated t-SNE and its applications to modern data,’’ presented at the
knowledge based genetic algorithm approach,’’ Expert Syst. Appl., vol. 38, 30th Int. Symp. Comput. Archit. High Perform. Comput. (SBAC-PAD),
no. 4, pp. 3161–3171, Apr. 2011. Lyon, France, 2018.
[44] M. J. Sousa, A. Moutinho, and M. Almeida, ‘‘Classification of potential MUNEER KHAN MOHAMMED received
fire outbreaks: A fuzzy modeling approach based on thermal images,’’ the master’s degree in industrial engineering
Expert Syst. Appl., vol. 129, pp. 216–232, Sep. 2019. from King Saud University. He has grad-
[45] E. Hidayat, N. A. Fajrian, A. K. Muda, C. Y. Huoy, and S. Ahmad, ‘‘A uated from Osmania University, Hyderabad,
comparative study of feature extraction using PCA and LDA for face India. He is currently a Researcher with the
recognition,’’ in Proc. 7th Int. Conf. Inf. Assurance Secur. (IAS), Melaka, Advanced Manufacturing Institute, College of
Malaysia, Dec. 2011, pp. 354–359. Engineering, King Saud University. His research
[46] H. Zhao, Z. Wang, and F. Nie, ‘‘Orthogonal least squares regression for
interests include laser micro-machining, man-
feature extraction,’’ Neurocomputing, vol. 216, pp. 200–207, Dec. 2016.
ufacturing systems, additive manufacturing,
[47] Y. Liu, H. Zhou, F. Tsung, and S. Zhang, ‘‘Real-time quality monitoring
and diagnosis for manufacturing process profiles based on deep belief non-conventional machining processes, and so on.
networks,’’ Comput. Ind. Eng., vol. 136, pp. 494–503, Oct. 2019. He has published several research articles in international journals and
[48] R. Boothalingam, ‘‘Optimization using lion algorithm: A biological inspi- conferences of repute.
ration from Lion’s social behavior,’’ Evol. Intell., vol. 11, nos. 1–2,
pp. 31–52, Oct. 2018.
[49] M. E. H. Pedersen and A. J. Chipperfield, ‘‘Simplifying particle swarm
optimization,’’ Appl. Soft Comput., vol. 10, no. 2, pp. 618–628, Mar. 2010.
[50] S. Mirjalili and A. Lewis, ‘‘The whale optimization algorithm,’’ Adv. Eng.
Softw., vol. 95, pp. 51–67, May 2016.
[51] S. Yu, K. K. Tan, B. L. Sng, S. Li, and A. T. H. Sia, ‘‘Lumbar ultrasound
USAMA UMER received the B.E. and M.S.
image feature extraction and classification with support vector machine,’’
degrees in mechanical engineering from the
Ultrasound Med. Biol., vol. 41, no. 10, pp. 2677–2689, Oct. 2015.
[52] F. Fernández-Navarro, M. Carbonero-Ruz, D. Becerra Alonso, and NED University of Engineering and Technology,
M. Torres-Jimenez, ‘‘Global sensitivity estimates for neural network Karachi, Pakistan, in 1998 and 2002, respec-
classifiers,’’ IEEE Trans. Neural Netw. Learn. Syst., vol. 28, no. 11, tively, and the Ph.D. degree in manufacturing and
pp. 2592–2604, Nov. 2017. automation from the Beijing Institute of Technol-
[53] S. Mirjalili, S. M. Mirjalili, and A. Lewis, ‘‘Grey wolf optimizer,’’ Adv. ogy, Beijing, China, in 2007. He worked as a Qual-
Eng. Softw., vol. 69, pp. 46–61, Mar. 2014. ity Assurance Officer at Philips/Whirlpool, from
October 1998 to April 2002. He started teaching
career in November 2002 as an Instructor with the
MUSTUFA HAIDER ABIDI received the mas- National University of Science in Technology (NUST), Pakistan. After Ph.D.
ter’s degree in industrial engineering from King he rejoined NUST as Assistant Professor, in July 2007. Later he joined
Saud University. He has graduated from Jamia FARCMT at King Saud University, in September 2012. He is currently
Millia Islamia, New Delhi, India. He is currently working as Associate Professor with the Advanced Manufacturing Institute,
a Researcher with the Advanced Manufacturing King Saud University Riyadh, Saudi Arabia. His research interests include
Institute, College of Engineering, King Saud Uni- high-speed machining, modeling and simulation of manufacturing opera-
versity. He has received a Gold Medal from the tions, laser beam machining, and optimization methods in manufacturing.
Faculty of Engineering and Technology, Jamia
Millia Islamia. The application of virtual reality
techniques for sustainable product development is
the major focus of his research. His other research interests include but
not limited to, human–computer interaction (HCI), artificial intelligence
(AI), reverse engineering, micro-manufacturing, and additive manufacturing.
He obtained Black Belt for Lean Six Sigma, trained in Project Management.
JABER E. ABU QUDEIRI received the B.Sc.
He is a Certified Supply Chain Manager. He has published several research
degree in mechanical engineering from the Uni-
articles in international journals and conferences of repute.
versity of Jordan, in 1992, and the M.Sc. and
Ph.D. degrees in manufacturing engineering from
HISHAM ALKHALEFAH received the master’s Gifu University, Japan, in 2005 and 2008, respec-
and Ph.D. degrees from the Industrial Engineering tively. He joined UAEU as an Associate Professor,
Department, King Saud University. He is currently in 2017. He was a Letter’s Patent registered at the
serving as a Supervisor of the Advanced Manu- Ministry of Industrial and Tried under No. P/1775.
facturing Institute, College of Engineering, King His current research interests include modeling
Saud University. His research interests include and optimization of manufacturing systems, opti-
advanced manufacturing technologies especially mization of sequence of operations in CNC, predict springback in sheet metal
additive manufacturing, digital manufacturing, bending process. He received the Scholarship (Monbukagakusho) for his
manufacturing systems, non-conventional machin- M.Sc. and Ph.D. degrees, in 2002. He received the Best Paper Award at
ing to name a few. He has published several ICACTE 2014.
research articles in internationals journal and conferences of repute.