Cost Estimation of Manufacturing Enterprises based on BP Neural Network
and Big Data Analysis
HUIJUAN MA
Lyceum of the Philippines University Manila Campus,
Manila 1002,
PHILIPPINES
Abstract: - The manufacturing industry is the pillar industry of modern industry, and the cost estimation of
manufacturing enterprises is an important management means of the manufacturing industry. Aiming at the cost
estimation problem of manufacturing enterprises, this research proposes a cost estimation method based on
Back Propagation (Back Propagation) neural network and big data analysis. In the process, the Lambda
architecture was used to construct the big data analysis architecture of manufacturing enterprises, the K-means
clustering algorithm was introduced for data clustering, and then the genetic algorithm was combined with the
Back Propagation neural network to estimate the cost. In the estimation accuracy test, the accuracy of the
research method can reach 94.7% after 240 iterations; in the calculation time test, the calculation time of the
research method is 403 Ks when the data size is 500 Gb in a large-scale data set; in the call data volume test,
the call data volume of the research method is 164 Kb when the research method is carried out to the seventh
step in the small-scale data set; when the application analysis is carried out, the research method completes
accurate cost estimation for 9 target parts. This research method has good model performance and calculation
accuracy, and can effectively estimate manufacturing enterprises’ costs.
Key-Words: - Back Propagation; Lambda architecture; Big data; Cost estimation; Manufacturing; K-means.
Received: April 15, 2023. Revised: October 14, 2023. Accepted: November 6, 2023. Published: November 17, 2023.
1 Introduction
In the manufacturing industry, cost estimation is a
crucial part of enterprise decision-making and
business management. Accurate cost estimation can
help enterprises formulate reasonable pricing
strategies, optimize production processes, and
improve profit margins and competitiveness, [1].
Traditional cost estimation methods usually require
a large amount of data collection and processing,
and these data are often scattered, inconsistent, or
missing, which brings difficulties to cost estimation,
[2], [3]. The models in traditional cost estimation
methods are usually based on simplified
assumptions and empirical formulas, ignoring the
complex manufacturing environment and the
interaction between multiple influencing factors,
resulting in limited accuracy of estimation results,
[4]. A back propagation neural network (BPNN) is a
commonly used artificial neural network model. By
learning training data, it can discover the nonlinear
relationship between input features and costs, and
can automatically adjust the relationship between
neurons through the backpropagation algorithm.
Connection weights between them, so as to achieve
accurate cost estimation, [5]. Big data technology
can help manufacturing companies process and
analyze massive amounts of data, thereby providing
an accurate basis for cost estimation. [6], proposed a
cost estimation method for enterprises that combines
building information models with target value
design. This method can analyze risks and profits,
but its computational efficiency at runtime is
relatively average. Scholars such as Mishra S have
designed an enterprise cost evaluation method using
an ant colony algorithm and resource joint
allocation algorithm. This method can propose
optimization strategies for costs, but the accuracy of
cost estimation is relatively average, [7]. In view of
this, research attempts to design a method based on
the BP neural network to analyze big data and
obtain cost estimation results for manufacturing
enterprises. Utilize BP neural network and genetic
algorithm to achieve efficient cost estimation for
manufacturing enterprises, and additionally
introduce K-means clustering algorithm to improve
the computational accuracy of the research method.
The aim is to integrate the advantages of various
technological means to design a cost estimation
method for manufacturing enterprises with better
performance, providing feasible technical references
for the development of manufacturing enterprises.
The research is mainly carried out in four parts.
The first part discusses and summarizes the relevant
research results of the current cost estimation and
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2567
Volume 20, 2023
BPNN. The second part is mainly to design the cost
estimation method for manufacturing enterprises
based on BPNN and big data analysis. The third part
is the performance test and empirical analysis of the
research method. The last part is the discussion and
summary of the full text.
2 Related works
Cost estimation of manufacturing enterprises can
provide a reference for production planning and
development of enterprises, and many scholars have
conducted related research on cost estimation
methods. Scholars such as Fazeli proposed a
BIM-based estimation method for the cost
estimation of construction projects. In the process,
the material quantity calculation is associated with
the model and expanded. This proposed method can
effectively estimate the construction cost, [8].
Scholars such as Nevliudov proposed an estimation
method using a regression model for the estimation
of material cost in 3D printing. In the process, the
resin consumption and exposure parameters in
printing are correlated, and the correlation
coefficient of the circuit board topology is
calculated. This proposed method has high
computational accuracy, [9]. [10], proposed a
method based on multi-factor analysis for the cost
estimation of automatic mobile phone washing in
public places. During the process, the usage habits
of the hand-washing crowd are analyzed, and the
raw materials are combined for calculation. This
proposed method shows high accuracy. Scholars
such as Leelathanapipat proposed a method based
on multiple linear regression for the cost estimation
of equipment renovation. In the process, three data
related to maintenance were introduced as
independent variables, and the decision coefficient
of the model was adjusted. This proposed method
can effectively estimate, [11]. Scholars such as Rosa
proposed an estimation model based on scale
measurement for the cost estimation of agile
software. In the process, the workload of the project
is provided, and the application domain group is
introduced to improve the accuracy. Experimental
results show that the proposed method has good
performance, [12].
Some scholars have conducted related research
on BPNN. [13], proposed a method based on BPNN
for the performance prediction of solid oxide fuel
cells. In the process, the support vector machine and
the random forest technology are integrated, and the
model evaluation is performed using multiple
criteria. Experimental results show that the proposed
method has good prediction accuracy. Scholars such
as Li proposed a prediction method based on BPNN
for the early warning of financial risks in business
operations. In the process, the initial financial
problems are analyzed, and the model is reasoned in
the process. This proposed method has a high
prediction accuracy, [14]. Aiming at the problem of
real-time traffic monitoring of roads, scholars such
as Liu proposed a data monitoring system based on
BPNN. In the process, the floating car data is fused
with the fixed detector data, and the genetic
algorithm and ant colony algorithm are introduced
to improve the calculation accuracy of the model.
This proposed method is effective for condition
monitoring, [15]. For the diagnosis of lung cancer,
scholars such as Nanglia proposed a detection image
analysis method based on BPNN. In the process, the
support vector machine is used to simplify the
computational complexity, and the feed-forward and
back-propagation neural network is integrated to
strengthen the features. This proposed method has
good diagnostic accuracy, [16]. [17], proposed an
analysis method based on BPNN for the data
collection of physical components in waste products.
The process combines the physical composition of
solid waste with social factors and uses
hyper-spherical changes to remove constraints. This
proposed method has good data analysis
performance.
To sum up, although BPNN has been
researched and applied in many fields, there is still
little research on the cost estimation of
manufacturing enterprises. In view of this, the study
proposes a manufacturing enterprise cost estimation
method based on BPNN and big data analysis, to
provide more references for the field of
manufacturing cost estimation.
3 Design of Manufacturing
Enterprise Cost Estimation method
based on BPNN and Big Data Analysis
An effective cost estimation method can obtain
accurate product manufacturing cost analysis results.
This section will describe the technical means used
in the manufacturing enterprise cost estimation
method of the research design.
3.1 Big Data Analysis Architecture of
Manufacturing Enterprises based on
Lambda Architecture
Under the trend of industrial manufacturing
intelligence transformation, industrial big data has
become an important information carrier in the
industrial field, and it collects manufacturing
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2568
Volume 20, 2023
information from a comprehensive perspective, [18],
[19].
Data source
Streaming Live view
All data Pre calculation
Batch View
Batch Dataset
Query
Acceleration layer
Batch processing layer
Service layer
Fig. 1: Lambda architecture
Data
transmission
Full data
Batch
processing
framework
Batch
calculation
results
Stream
processing
framework
Flow
calculation
results
Batch View
Stream
Processing
View
Uniform interface
Platform client
Batch processing
layer Stream processing
layer
Service layer
Fig. 2: Multi-mode big data processing architecture
In the application of big data, the use of data
for analysis is the most important link. The Lambda
architecture can perform real-time stream processing
and can be applied to the processing of large-scale
complex data, [20]. The research uses the Lambda
data processing architecture as the foundation to
construct the big data analysis architecture of
manufacturing enterprises. The Lambda architecture
is shown in Figure 1.
As can be seen from Figure 1, the Lambda
architecture includes three main parts: the
acceleration layer, the batch processing layer, and
the service layer. The data source is connected to the
acceleration layer and the batch processing layer,
and the query information is connected to the
acceleration layer and the service layer. The batch
processing layer can perform batch processing
calculations, generate batch processing views, and
transfer data to the service layer for storage. The
batch processing layer can be repeated periodically
when generating batch views to improve data fault
tolerance and is suitable for computing and analysis
on a global scale. The service layer provides support
for the input of query information, accesses the view
with the query conditions contained in the query
information, and calls the real-time view combined
with the batch processing results to give feedback to
the user. Generally, to ensure the simplicity of the
overall system, it is not allowed to Random writes
are performed in the result. The acceleration layer
processes only the latest input data to reduce
processing latency while completing real-time view
generation. The data of manufacturing enterprises
involves many aspects and has many data models.
The Lambda architecture is optimized to establish a
multi-mode manufacturing enterprise big data
processing architecture. The multi-mode big data
processing architecture is shown in Figure 2. From
Figure 2 the multi-mode big data processing
architecture contains three main layers: batch
processing layer, service layer, and stream
processing layer. The service layer is the only
medium connecting the batch processing layer,
stream processing layer, and platform client,
providing the same access interface. During data
transmission, static historical data is directly
imported by the client, and real-time data is
processed by Kafka and the distributed coordinator.
The batch processing layer uses the batch
processing framework to perform offline
calculations on the full amount of industrial data in
the distributed database system and outputs the
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2569
Volume 20, 2023
calculation results to the batch view. The historical
results are not retained during calculations to ensure
the timeliness of output data. The stream processing
layer uses the stream processing framework to
process multiple real-time stream data online, and at
the same time sends the latest calculation results to
the stream processing view. The batch processing
framework uses the Hadoop framework, which
includes a distributed storage layer, a resource
scheduling layer, and a batch processing engine. The
distributed storage layer can store and replicate
cluster nodes and store results; the resource
scheduling layer manages and schedules basic
resources; and the batch processing engine for data
calculation. The stream processing framework uses
the Storm framework and uses two modes for the
combined operation to perform strict one-time
processing on the received data. To analyze the data
association in the data, the study introduces the
K-means clustering algorithm for data clustering.
K-means divides the cluster center and calculates
the distance between the data and the cluster center
to divide and update the cluster to achieve the
clustering of the data set. When calculating, the
number of input data is first determined, and then
the initial dataset is set to specify multiple initial
clustering centers. Cluster the data using Euclidean
distance, and then divide the unpartitioned data into
clusters with the same number of initial cluster
centers, as shown in formula (1).
,
, min
x X c C
dist x C x c
(1)
Formula (1),
x
represents the data point;
C
represents the cluster;
c
and represents the cluster
center. After clustering the data, update the
clustering center with the clustering results, as
shown in formula (2).
1
ii
ii
xc
i
cx
NC
(2)
Formula (2),
i
c
represents the new cluster
center of the cluster;
i
NC
represents the number of
data points in the cluster. Set a termination condition
and calculate the sum of squared errors. If the sum
of squared errors is less than the initial threshold,
stop the iteration. The error sum of squares is
calculated as shown in formula (3).
2
1
,
i
k
i p c
E= dist x C
(3)
Formula (3),
E
represents the sum of squared
errors. Introducing parallel coordinates for
dimensionality reduction visualization of
multidimensional data. In a multidimensional space,
if the dataset contains multiple data and each data
contains multiple field attributes, the definition of a
parallel dataset is shown in formula (4).
,1 ,2 ,
, ,..., ,... 1 ,1
m m m m n
D= D D D D m M n N
(4)
Formula (4),
M
it represents the maximum
number of data contained;
N
it represents the
maximum value of the field attribute of the data.
Calculate the relative position value of each data in
the coordinates, as shown in formula (5).
,min
max min
m n n
n
nn
DD
pDD
(5)
Formula (5),
n
p
represents the relative
position value. Draw data in parallel coordinate
systems using relative position values. The
clustering accuracy analysis formula in the
follow-up accuracy analysis is shown in formula (6).
1/
y
A mis n
(6)
Formula (6),
A
represents the accuracy rate;
mis
represents the number of misclassified samples;
y
n
represents the total number of samples. The
calculation formula for subsequent speedup ratio
analysis is shown in formula (7).
s
speedup
r
T
ST
(7)
Formula (7),
speedup
S
represents the speedup
ratio;
s
T
represents the serial execution time on a
single node;
r
T
and represents
r
the parallel
execution time of a computing node.
3.2 Design of Cost Estimation Method based
on Improved BPNN
When carrying out cost estimation based on the big
data of manufacturing enterprises, due to the
different influence of parameters involved in
different products, there are defects in the
characteristics of the calculation, [21], [22]. As a
kind of artificial neural network, BPNN has strong
adaptability when estimating product cost. The
study uses BPNN to estimate costs based on big
data of manufacturing enterprises. The BPNN model
is shown in Figure 3.
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2570
Volume 20, 2023
1 2 3 4
A B C D E
Input layer
Hidden layer
Output layer
Forward propagation
Back Propagation
Fig. 3: BPNN model
It can be seen from Figure 3 that the BPNN
contains an input layer, a hidden layer, and an output
layer. The input layer is the layer that accepts data
input, can normalize the data, and performs data
buffering at the same time. The hidden layer is set
by user requirements, and the number of layers is
determined by repeated combinations of data. If
there is a large discrepancy in the output of the
output layer, the network stops outputting and enters
the backpropagation process, corrects the attribute
values, and then returns to the forward propagation.
In the BPNN, the relationship between neurons is
described by the activation function. During the
training process, the error, threshold, and weight
need to be reduced until they are less than the
present value. In the process of information forward
propagation, the output calculation of a neuron in
the middle layer is shown in the formula (8).
1int , 1,2,...,
mm
z f m q
(8)
Formula (8),
m
z
represents the output of a
neuron in the middle layer;
intm
represents the
information transfer from the input layer to the
output layer;
f
represents the activation function.
The calculation of information transfer is shown in
formula (9).
1
int
n
m im i
i
vx
(9)
Formula (9),
im
v
represents the input layer
information. The output of the last layer of neurons
is shown in formula (10).
2
1
, 1, 2,...,
e
n mn m
k
o f w z n q
(10)
Formula (10),
n
o
represents the neuron output
of the last layer;
mn
w
represents the information of
the middle layer;
m
z
and represents the
information of the last layer. The root mean square
error calculation of the forward pass is shown in the
formula (11).
2
1
1
2
l
z n n
n
E y o
(11)
Formula (11),
z
E
represents the root mean
square error of forward transmission;
n
y
which
represents the real output value of the last layer. In
the process of information backpropagation, the
weight is adjusted by solving the partial derivative,
and the root mean square error of the forward
transmission is expanded, as shown in the formula
(12).
2
2
11
1
2
ln
z n im m
nn
E y f w z
(12)
Formula (12),
im
w
it represents the
intermediate layer information in backpropagation.
The intermediate layer information in
backpropagation is shown in formula (13).
im
mn
E
ww
(13)
Formula (13),
represents the learning
efficiency. After adjusting the weights, the root
mean square error is further expanded, as shown in
formula (14).
2
2
11
1int
2
ln
z n im mn m
ni
E y f w f
(14)
However, when only BPNN is used for
calculation, there are problems with small local
poles and insufficient convergence speed. The
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2571
Volume 20, 2023
random search characteristic of the genetic
algorithm is used to improve the convergence speed
of BPNN. The basic flow of the genetic algorithm is
shown in Figure 4.
Create and
initialize
population
Measure and
evaluate
fitness
Select the
most suitable
sample
Butation Cross
production
Is it
The optimal
solution
Output
YesNo
Fig. 4: Basic process of genetic algorithm
Start
Initialize
population
fitness value
Selection,
crossover,
variation
Whether
the number of iterations
is met
Forming a
new species
group
Select the
optimal
individual
Whether
the number of cycles is
met
BP forward
propagation
Is it
less than the global
error
Complete training
and make estimates
Yes
No
Yes Yes
No
No
Fig. 5: Optimization of BPNN validation steps
It can be seen from Figure 4 that when the
basic process of the genetic algorithm is running, it
first needs to create and initialize the population and
encode the parameter characteristics. Then measure
and evaluate the adaptability, select suitable samples
from the evaluation results, and judge whether the
obtained results are the optimal solution after
mutation and crossover operations. If the optimal
solution is not reached, the adaptive measurement
and evaluation are performed again, and then the
loop operation is performed until the obtained result
reaches the optimal solution, and the result is output.
After combining the genetic algorithm with the
BPNN, use the genetic algorithm to optimize the
distribution of the weight thresholds of multiple
targets. When encoding, use the real number
encoding method to encode each individual. First,
the input layer-hidden layer connection weight.
Then encode the hidden layer-output layer
connection weights, encode the neuron threshold of
the hidden layer, and finally encode the neuron
threshold of the output layer. When the population is
initialized, the initial population number is set, and
the initial value of the weight and threshold is
defined as a real number between -1 and 1. The
purpose of training is to make the cost estimate fit
the actual value, and the absolute value of the error
sum of the expected result and the predicted result is
used as the optimization goal. After genetic
manipulation, cost estimation is performed using
optimal weights and thresholds. In the hidden layer
of the neural network, the calculation of the number
of nodes included is shown in the formula (15).
10E E E
N N N t
(15)
Formula (15),
N
represents the number of
nodes;
t
it is an integer between 1 and 10. The
steps of verification according to the optimized
BPNN are shown in Figure 5. It can be seen from
Figure 5 that the optimization BPNN verification
step starts with initializing the population. After
generating the fitness value, if the preset number of
iterations is not reached, the selection, crossover,
and mutation operations are performed to form a
new population, and then repeated iterations until
after the number of iterations reaches the preset
number of times, the optimal individual is selected.
If the preset number of cycles is met, the training is
completed. If not, the forward propagation is
performed until the error is smaller than the global
error.
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2572
Volume 20, 2023
Parameter
selection
Case
Retrieve
Matching
Rules
Search Rules
Sort and
revise cases
in descending
order
Product Library
Is the
class instance similarity
greater than the
threshold
Calculate
directly
BP neural
network for
calculation
Output
Storage and
maintenance
Order entry
Yes
No
Fig. 6: Cost estimation methods for manufacturing enterprises
Integrating the extension matter-element
technology into the cost assessment, the products
involved in the assessment can be searched and
matched quickly and accurately. Finally, the big data
cost estimation method for manufacturing
enterprises constructed is shown in Figure 6.
As can be seen from Figure 6, when
performing cost estimation, firstly, the parameter
selection of the model is performed based on the
input order information, and the threshold value of
the resulting unit is set, and then the instance
retrieval is performed, and the representative class
instances are corrected and then learned. If the
similarity of the retrieved class instances is less than
the threshold, the cases are sorted and corrected in
descending order, and then calculated by the BPNN;
if it is greater than the threshold, the calculation is
performed directly. The calculation results are
output and stored in the product library to expand
the richness of search samples, and the output
results are the cost estimation results.
4 Effectiveness Analysis of
Manufacturing Enterprise Cost
Estimation Method based on BPNN
and Big Data Analysis
Manufacturing enterprise cost estimation can bring
data reference for enterprise decision-making. This
section will conduct performance tests and
application analyses of the research method to
determine the effectiveness of the research method.
4.1 Performance Test of Cost Estimation
Method based on BPNN and Big Data
Analysis
To analyze the effectiveness of the cost estimation
method based on BPNN and big data analysis
designed by the research in estimating
manufacturing enterprises, the research first tests the
performance of the designed method. The data set
used in the test is a composite data set formed by
mixing the historical cost data set and the external
data set, and the data set is divided into two
sub-blocks. The decision tree random forest
algorithm is a method that allows for online model
updates. The decision tree model provides a clear
decision path, allowing users to understand how the
model makes predictions; The random forest
algorithm can perform data analysis without the
need for a large amount of preprocessing and is a
high-performance enterprise data analysis method.
The support vector machine deep neural network
algorithm has strong generalization ability and can
adapt to the complex data distribution of
manufacturing enterprises; The support vector
machine algorithm can provide certain model
interpretability, help users understand content, and
has good performance in manufacturing-related data
analysis. In this context, the study compared the
decision tree random forest algorithm with the
support vector machine deep neural network
algorithm. Test the loss value during the training of
the research method, as shown in Figure 7.
It can be seen from Figure 7 that during training, the
loss values of the three different methods all
gradually decrease in the early stage, and tend to
stabilize after reaching the lowest interval.
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2573
Volume 20, 2023
Epoch
050 100 150 200 250 300 350
0
0.2
0.4
0.6
0.8
1
Val loss
(a)DT-RF
Epoch
0
0.2
0.4
0.6
0.8
1
Val loss
(b)SVM-DNN
Epoch
0
0.2
0.4
0.6
0.8
1
Val loss
(c)Research method
050 100 150 200 250 300 350 0 50 100 150 200 250 300 350
Fig. 7: Training loss value test
Accuracy(%)
40 80 120 160 200
0
10
20
30
40
50
60
70
80
90
100
Epoch
240
Accuracy(%)
30
40
50
60
70
80
90
Epoch
(a)Training set (b)Validation set
Research method
SVM-DNN
DT-RF
040 80 120 160 200 2400
Research method
SVM-DNN
DT-RF
Fig. 8: Estimating accuracy testing
The loss value of the decision tree-random
forest algorithm reaches the lowest interval after 77
iterations, and the curve has obvious fluctuations in
the process of descending, and the lowest value of
the interval is 0.17. The loss value of the support
vector machine-deep neural network algorithm
reaches the lowest interval after 191 iterations, and
the curve has obvious fluctuations during the
decline process, and the lowest value of the interval
is 0.19.
The loss value of the research method reaches
the lowest interval after 61 iterations, and there is a
very small fluctuation in the lowest interval, and the
lowest value of the interval is 0.12. It shows that the
research method has faster training speed and better
training results. The estimation accuracy of the
research method was tested, as shown in Figure 8.
It can be seen from Figure 8 that in both the
training set and the verification set, the estimation
accuracy of the three methods increases with the
number of iterations, and tends to be stable after
reaching the highest interval. In the training set, the
estimation accuracy of the decision tree-random
forest algorithm increased rapidly during the first
138 iterations, and the estimation accuracy was
90.8% when the number of iterations reached 240.
The estimation accuracy of the support vector
machine-deep neural network algorithm increased
rapidly during the first 129 iterations, and the
estimation accuracy was 84.1% when the iteration
number reached 240. The estimation accuracy of the
research method increased rapidly during the first
120 iterations, and the estimation accuracy was
94.7% when the number of iterations reached 240.
In the verification set, the estimation accuracy of the
decision tree-random forest algorithm increases
rapidly during the first 141 iterations, and the
estimation accuracy is 81.7% when the number of
iterations reaches 240. The estimation accuracy of
the support vector machine-deep neural network
algorithm increases rapidly in the first 148 iterations,
and the estimation accuracy is 77.6% when the
iteration number reaches 240. The estimation
accuracy of the research method increased rapidly
during the first 120 iterations, and the estimation
accuracy was 83.2% when the iteration number
reached 240. It shows that the research method has
better estimation accuracy. The calculation time of
the research method in different data scales is tested,
as shown in Figure 9.
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2574
Volume 20, 2023
100 200 300 400 500
0
100 200 300 400 500
200
300
400
500
600
700
800
900
100
200
300
400
500
600
700
800
900
1000
Time(Ks)
(b)Dataset B Subblock
Data scale(Mb)
Time(s)
(a)Dataset A Subblock
Research method
SVM-DNN
Data scale(Gb)
DT-RF
Research method
SVM-DNN
DT-RF
Fig. 9: Calculation time test
1 3 5 6 7
0
400
800
1200
1600
2000
2400
2800
Computational procedure
Call Data Volume(Kb)
(a)Dataset A Subblock
2 4 1 3 5 6 7
2000
4000
6000
8000
10000
12000
14000
16000
Computational procedure
Call Data Volume(Kb)
(b)Dataset B Subblock
2 4
Research method
SVM-DNN
DT-RF
Research method
SVM-DNN
DT-RF
Fig. 10: Call data volume
It can be seen from Figure 9 that the
calculation time of the three methods increases with
the increase of the data size. In the A sub-block with
a smaller data size, the decision tree random forest
algorithm has a computation time of 181 seconds
when the data size is 100 Mb and 998 seconds when
the data size increases to 500 Mb. The calculation
time of the support vector machine-deep neural
network algorithm is 134 s when the data size is 100
Mb, and the calculation time is 946 s when the data
size increases to 500 Mb. The calculation time of
the research method is 69 s when the data size is
100 Mb, and the calculation time is 431 s when the
data size increases to 500 Mb. In the B sub-block
with a large data size, the calculation time of the
decision tree-random forest algorithm is 289 Ks
when the data size is 100 Gb, and the calculation
time is 721 Ks when the data size increases to 500
Gb. The calculation time of the support vector
machine deep neural network algorithm is 376s
when the data size is 100Mb, and 830Ks when the
data size increases to 500Gb. The calculation time
of the research method is 244 Ks when the data
scale is 100 Gb, and the calculation time is 403 Ks
when the data scale increases to 500 Gb. It shows
that the research method has better calculation speed.
The recalled data volumes of the research method at
different computational steps are tested, as shown in
Figure 10. It can be seen from Figure 10 that the call
data volumes of the three methods decrease
continuously as the calculation proceeds. In the A
sub-block with a smaller data size, the decision tree
random forest algorithm has a data call volume of
2711Kb at step 1 and 762Kb at step 7. The data call
volume of the support vector machine deep neural
network algorithm in step 1 is 1522Kb, and the call
data volume in step 7 is 1241Kb. The data call
volume in step 1 of the research method is 2203Kb,
and the call data volume in step 7 is 164Kb. In
sub-block B with large data size, the calling data
volume of the decision tree-random forest algorithm
in the first step is 15213 Kb and the calling data
volume in the seventh step is 8123 Kb. The call data
volume of the support vector machine-deep neural
network algorithm in the first step is 15653 Kb, and
the call data volume in the seventh step is 6167 Kb.
In the research method, the calling data volume in
the first step is 1442 Kb, and the calling data
volume in the seventh step is 2481 Kb. It shows that
the research method has better data retrieval
simplicity.
4.2 Application Analysis of Cost Estimation
Method based on BPNN and Big Data
Analysis
In the application analysis of the research method,
an elevator parts manufacturing enterprise is taken
as the analysis object. First, analyze the processor
usage of the system when performing cost
estimation, as shown in Figure 11.
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2575
Volume 20, 2023
0 100 200 300 400
0
20
40
60
80
100
Run time(s)
CPU usage ratio(%)
0 100 200 300 400
0
20
40
60
80
100
Run time(s)
CPU usage ratio(%)
(a)SVM-DNN (b)Research method
Fig. 11: Processor occupancy
12345
0
200
400
600
800
1000
1200
1400
Part Number
Cost value(Yuan)
SVM-DNN
Research method
Actual cost
6 7 8 9
Fig. 12: Cost estimation results
From Figure 11, it can be seen that during cost
estimation, there is a certain degree of fluctuation in
the processor occupancy of different methods during
a total runtime of 400s. The maximum processor
utilization ratio of the support vector machine deep
neural network algorithm reached 97%, and there
have been multiple instances during this period
where it approached the maximum processor
utilization ratio; The average processor usage during
this period reached 61%. The minimum processor
occupancy ratio is the case of algorithm pauses or
system protection, and does not have a reference
value. The maximum processor utilization ratio of
the research method is 55%, and the average
processor utilization ratio during the period is 25%;
Due to the relatively small maximum processor
utilization ratio, the fluctuation in processor
utilization ratio during the time period is also
relatively small. It shows that the research method
brings less burden to the processor during actual
operation, and has lower requirements on hardware.
The cost estimation results of the research methods
are compared and analyzed, as shown in Figure 12.
It can be seen from Figure 12 that the cost of 9
target parts has been successfully estimated using
the support vector machine-deep neural network
algorithm and research method. The actual
processing costs of these nine parts are all below
1,400 yuan. The minimum difference between the
estimated results of the support vector machine-deep
neural network algorithm and the actual processing
costs is 22 yuan, and the maximum difference
reaches 117 yuan; among these 9 target parts, there
were 8 instances where there was a significant
deviation from the actual processing cost. the
minimum difference with the actual processing cost
is 4 yuan, and the maximum difference is 19 yuan;
among these 9 target parts, there was no significant
difference between them and the actual processing
cost. It shows that the research method can
effectively and accurately estimate the cost of
manufacturing enterprises. Sensitivity analysis was
conducted on cost indicators using research methods,
and seven common implicit costs were set as
indicators. The results are shown in Table 1.
Table 1. Sensitivity Analysis of Implicit Cost
Indicators
Index
Sensitivity
value
Policy guarantee capability
-0.514
Market growth degree
-0.708
Ecological environment
-0.330
Supporting service industry
development environment
-0.130
Government work efficiency
-0.218
Investment and financing environment
-0.526
Science and technology innovation
ability
-0.617
From Table 1, it can be seen that the research
method has obtained analysis results on the seven
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2576
Volume 20, 2023
cost sensitivities of manufacturing enterprises. The
absolute sensitivity value of the market development
level is above 0.7, which is the highest among the 7
items, indicating that the cost of manufacturing
enterprises is greatly influenced by market
development level factors and has a strong
sensitivity to changes in the market development
level. The absolute sensitivity value of the
development environment of the supporting service
industry is 0.130, which is the smallest of the seven,
indicating that the cost of manufacturing enterprises
is less affected by changes in the development
environment of the supporting service industry.
5 Conclusion
The cost estimation of manufacturing enterprises
can provide a decision-making reference for the
production planning of manufacturing enterprises.
Based on big data analysis, the research proposes a
cost estimation method for manufacturing
enterprises using BPNN. Firstly, the construction of
the big data analysis architecture is completed based
on the Lambda architecture, and then the data
association analysis is carried out through the
clustering algorithm, and the cost is estimated using
the optimal weight and threshold, and finally, the
effectiveness of the research method is tested. The
experimental results show that in the training loss
value test, the loss value of the research method
reaches the lowest interval after 61 iterations and the
lowest reaches 0.12; in the estimation accuracy test,
the research method in the verification set reaches
240 iterations. The estimation accuracy is 83.2%; in
the calculation time test, the calculation time of the
research method is 69 s when the data size in the
small-scale data set is 100 Mb; in the analysis of
processor occupation, the maximum processor
occupation of the research method is 400 s. The
ratio is 55%; the maximum difference between the
cost estimation results of the nine target parts and
the actual value by the research method is only 19
yuan. The results show that the research method has
better computational efficiency and accuracy of
results when estimating the cost of manufacturing
enterprises, and the burden on the hardware is
smaller. In the future, research methods can be
applied to manufacturing enterprises with intelligent
data collection equipment. The data collection
equipment monitors and collects data on the
production line, calculates and analyzes the
collected production data through research methods,
and obtains corresponding analysis results.
Enterprise management personnel refer to the
analysis results to adjust the production plan of the
enterprise and make decisions on the development
direction of the enterprise. However, research
methods are more focused on designing for
mechanical manufacturing enterprises, and data
from mechanical manufacturing enterprises is also
used for application analysis. Currently, it is
uncertain how effective the application will be in
areas where there are few intelligent data collection
devices in the light textile industry and handicrafts
industry. In the future, application analysis will be
conducted for manufacturing categories with
relatively small amounts of intelligent data
collection to enrich experimental results and
optimize research methods.
References:
[1] L. Mauler, F. Duffner, W. G. Zeier, J. Leker.
"Battery cost forecasting: a review of methods
and results with an outlook to 2050," Energy
Environ. Sci., vol. 14, no. 9, pp. 4712-4739,
2021.
[2] C. Diagne, B. Leroy, A. C. Vaissière, R. E.
Gozlan, D. Roiz, I. Jarić, J. Salles, C.
Bradshaw, F. Courchamp. "High and rising
economic costs of biological invasions
worldwide," Nature, vol. 592, no. 7855, pp.
571-576, 2021.
[3] X. R. Zhang, X. Sun, W. Sun, T. Xu, P. P.
Wang, S. K. Jha. "Deformation expression of
soft tissue based on BP neural network," Int. J.
Autom. Soft Comput., vol. 32, no. 2, pp.
1041-1053, 2022.
[4] K. Schwarze, J. Buchanan, J. M. Fermont, H.
Dreau, M. W. Tilley, J. M. Taylor, P. Antoniou,
S. Knight, C. Camps, M. Pentony, E. Kvikstad,
S. Harris, N. Popitsch, A. Pagnamenta, A.
Schuh, J. Taylor, S. Wordsworth. "The
complete costs of genome sequencing: a
microcosting study in cancer and rare diseases
from a single center in the United Kingdom,"
Genet. Med., vol. 22, no. 1, pp. 85-94, 2020.
[5] B. Khan, W. Khan, M. Arshad, N. Jan.
"Software cost estimation: algorithmic and
non-algorithmic approaches," Int. J. Data Sci.
Adv. Analytics, vol. 2, no. 2, pp. 1-5, 2020.
[6] F. Elghaish, S. Abrishami, M. R. Hosseini, S.
Abu-Samra. "Revolutionising cost structure
for integrated project delivery: a BIM-based
solution," Eng. Constr. Archit. Manage., vol.
28, no. 4, pp. 1214-1240, 2021.
[7] S. Mishra, M. N. Sahoo, A. Kumar Sangaiah,
S. Bakshi. "Nature-inspired cost optimisation
for enterprise cloud systems using joint
allocation of resources," Enterprise Inf. Syst.,
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2577
Volume 20, 2023
vol. 15, no. 2, pp. 174-196, 2021.
[8] A. Fazeli, M. S. Dashti, F. Jalaei, M.
Khanzadi. "An integrated BIM-based
approach for cost estimation in construction
projects," Eng. Constr. Archit. Manage., vol.
28, no. 9, pp. 2828-2854, 2021.
[9] I. Nevliudov, I. Razumov-Fryziuk, V.
Yevsieiev, D. Nikitin, D. Blyzniuk, R. Strelets.
"Cost estimation of photopolymer resin for
3D exposure of circuit boards," Technol. Audit
Prod. Reserves, vol. 2, no. 2 (64), pp. 43-49,
2022.
[10] P. Karar, M. Chatterjee, S. Deshi, M. Das, P.
Das, S. Pal. "Cost estimation and fabrication
of automatic hand sanitizing machine," Int. J.
Res. Eng. Sci. Manage., vol. 4, no. 5, pp.
24-27, 2021.
[11] J. Leelathanapipat, P. Paichit. "Cost estimate
of repairing refurbished equipment using
multiple regression model," Eng. J. Res. Dev.,
vol. 31, no. 2, pp. 127-138, 2020.
[12] W. Rosa, B. K. Clark, R. Madachy, B. W.
Boehm. "Empirical effort and schedule
estimation models for agile processes in the
US DoD," IEEE Trans. Software Eng., vol. 48,
no. 8, pp. 3117-3130, 2021.
[13] S. Song, X. Xiong, X. Wu, Z. Xue. "Modeling
the SOFC by BP neural network algorithm,"
Int. J. Hydrogen Energy, vol. 46, no. 38, pp.
20065-20077, 2021.
[14] X. Li, J. Wang, C. Yang. "Risk prediction in
financial management of listed companies
based on optimized BP neural network under
digital economy," Neural Comput. Appl., vol.
35, no. 3, pp. 2045-2058, 2023.
[15] J. Liu, J. Huang, R. Sun, H. Yu, R. Xiao.
"Data fusion for multi-source sensors using
GA-PSO-BP neural network," IEEE Trans.
Intell. Transp. Syst., vol. 22, no. 10, pp.
6583-6598, 2020.
[16] P. Nanglia, S. Kumar, A. N. Mahajan, P. Singh,
D. Rathee. "A hybrid algorithm for lung
cancer classification using SVM and Neural
Networks," ICT Express, vol. 7, no. 3, pp.
335-341, 2021.
[17] S. Ma, C. Zhou, C. Chi, Y. Liu, G. Yang.
"Estimating physical composition of
municipal solid waste in China by applying
artificial neural network method," Environ.
Sci. Technol., vol. 54, no. 15, pp. 9609-9617,
2020.
[18] L. Chen, V. Jagota, A. Kumar. "Retracted
article: research on optimization of scientific
research performance management based on
BP neural network," Int. J. Syst. Assur. Eng.
Manage., vol. 14, no. 1, pp. 489-489, 2023.
[19] X. Rong, Y. Liu, P. Chen, X. Lv, C. Shen, B.
Yao. "Prediction of creep of recycled
aggregate concrete using back‐propagation
neural network and support vector machine,"
Struct. Concr., vol. 24, no. 2, pp. 2229-2244,
2023.
[20] S. Nimrah, S. Saifullah. "Context-free word
importance scores for attacking neural
networks," J. Comput. Cognit. Eng., vol. 1, no.
4, pp. 187-192, 2022.
[21] X. B. Zhang, G. D. Cheng, J. L. Zhao, W. Dai,
Y. Tao. "Research on classification and
evaluation of Chang 3 reservoir in H area
based on BP neural network technology,"
Prog. Geophys., vol. 38, no. 3, pp. 1272-1281,
2023.
[22] Z. Cui, L. Wang, Q. Li, K. Wang. "A
comprehensive review on the state of charge
estimation for lithium-ion battery based on
neural network," Int. J. Energy Res., vol. 46,
no. 5, pp. 5423-5440, 2022.
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
Huijian Ma is the solely author who conducted the
methodology, writing and revision of this
manuscript.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
No funding was received for conducting this study.
Conflict of Interest
The authors have no conflict of interest to declare.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on BUSINESS and ECONOMICS
DOI: 10.37394/23207.2023.20.219
Huijuan Ma
E-ISSN: 2224-2899
2578
Volume 20, 2023
