Hindawi

Wireless Communications and Mobile Computing

Volume 2022, Article ID 9465433, 13 pages
https://doi.org/10.1155/2022/9465433

Research Article
Visual Detection Application of Lightweight Convolution and
Deep Residual Networks in Wood Production

Zhijie Chen,1 Zhihui He,2 Bao Chao,3 and Hongfei Guo 1,4

1
School of Artificial Intelligence, Jinan University, Zhuhai, China
2
Zhuhai Hengqin Building & Construction Quality Inspection Centre Co., Ltd., Zhuhai, China
3
School of Mechanical Science and Engineering, Huazhong University of Science and Technology, Wuhan, China
4
Institute of Physical Internet, Jinan University, Zhuhai, China

Correspondence should be addressed to Hongfei Guo; ghf-2005@163.com

Received 22 April 2022; Revised 6 May 2022; Accepted 12 May 2022; Published 29 May 2022

Academic Editor: Mohammad Farukh Hashmi

Copyright © 2022 Zhijie Chen et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

In order to improve the production capacity of traditional wood manufacturing industry, efficient wood quality and thickness
detection is a challenging issue. This paper firstly carries out digital twin modeling for a drawer side panel processing line of a
wood company and explores the efficiency problems existing in the links of quality inspection and thickness inspection of
wood by means of value stream mapping. Therefore, we adopted a lightweight convolution neural network MobileNetV2 for
wood quality detection, which realized efficient wood quality identification. In contrast, traditional convolution neural network
has many weighting parameters and large scale of generating detection model, which makes it difficult to apply in situations
with limited computing power and memory. Secondly, due to the stronger robustness and generalization ability of the residual
network, we used ResNet to detect the wood thickness and obtain reliable performance. Finally, we reasonably embedded them
in the whole wood production process and established the simulation model of production line before and after improvement
in FlexSim simulation software. The experimental results show that the improved plan can simplify the workshop production
process, increase the production balance rate by 29.07%, increase the product value-added rate from 0.08% to 0.11%, and
shorten the production cycle by 2 hours. Performance indicators such as product inventory, number of people, and equipment
utilization also improve significantly. Based on the above results, the validity of the production process improvement model
proposed by US based on lightweight convolution and deep residual network is demonstrated.

1. Introduction cal and virtual mapping relationship, and, through continuous

iteration and feedback, make people, machine, material, ring,
Driven by national strategies such as “Made in China 2025” and method really merge together. To be exact, the digital twin
and “Industry 5.0,” many discrete manufacturing enterprises model is a software-based, data-based drive to simulate and
in China have conducted in-depth research in important busi- predict the physical entity’s actual state and future behavior
ness fields such as technology manufacturing and information with relevant improved knowledge [2].
system. In order to make the transition of discrete Value stream mapping is a tool for implementing lean
manufacturing enterprises to digitalization smoothly, we need manufacturing, one of which is designed for physical processes
to construct a digital model for resources first and increase the and is a process analysis tool that starts with suppliers supply-
production flexibility of enterprises [1]. The key to the solu- ing raw materials to warehouse, and then goes into warehouse
tion is correctly using virtual simulation technology to achieve manufacturing, finished goods to warehouse, products to
the goals of multiple experimental iterations at the virtual level warehouse, and until the product reaches the customer. First
and one successful experiment at the physical level. Digital of all, we need to optimize the key parts. Drawing a “future
twin technology uses a mathematical model, historical data state” of the value stream by pulling and smoothing the pro-
reference, and real-time data feedback to construct the physi- cess as the goal of improvement is the main tool for value
2 Wireless Communications and Mobile Computing

stream analysis. Then, we return to the analysis of the status 2. Relevant Work
quo map with the problems found in the future value stream
mapping (VSM) and cycle until a suitable VSM is identified. 2.1. Value Stream Mapping. At present, the application of
Finally, a value flow of the lean production line is designed value stream mapping (VSM) mainly adopts VSM as a
by drawing the future value stream status map, and a working method to identify waste in production. It considers whether
plan is set to eliminate the non-value-added activities in the the production process can be rescheduled in a more effi-
production process as far as possible, so as to reduce costs, cient order and analyzes result to reach a better process pat-
speed up the response to customer needs, and improve enter- tern [3]. However, this does not take into account that the
prise competitiveness, as shown in Figure 1. VSM can only reflect static information and can not present
In addition, the physical process also includes product the problems occurring in workshop production in real time.
inspection and parking. However, because the VSM can And improvement plans based on the VSM can only be
not describe the dynamic behavior of production system in reflected on the future VSM, and we do not know whether
detail, it has limitations in lean production application. these plans match actual production and can not show the
The static limitation of VSM can be overcome by combining complexity and dynamic behavior of the production process
simulation technology. Figure 2 shows the FlexSim simula- [4]. In the current research, Huang et al. [5] used BS-
tion model, which is an object-oriented simulation software, CONWIP (base stock CONWIP control strategy) as the pro-
which can efficiently establish the process of discrete events duction control strategy in combination with VSM and sim-
and provide a highly simulated three-dimensional virtual ulation experiment and used genetic algorithm to optimize
reality environment for material handling and workflow of buffer capacity to improve on-board panel workshop. Yan-
hong et al. [6] combined simulation technology to verify
manufacturing industry, thus achieving virtual simulation
the effectiveness of BS-CONWIP in multiple manufacturing
and dynamic layout of production lines and avoiding the
environments. VSM is mainly used to analyze value streams
limitations of traditional equipment schemes.
within an organization in industrial engineering. Ma et al.
In this study, from the point of view of production line
[7] combined VSM with the Lean Six Sigma method to
transformation and in order to change the static defects of
reflect the impact of lean thinking on cost savings in
previous VSM, we combine VSM and simulation technology
manufacturing. FlexSim software is also used in the technical
and apply VSM to identify the problems existing in the
aspect. Punna Rao et al. [8] combined VSM through method
workshop production process. We use ECRS principles to
and time study and also used FlexSim software to simulate
improve standard operations and FlexSim to quickly and
the time study of current and future state. Aziz et al. [9]
accurately build a three-dimensional model of the equip-
studied the integration of discrete event simulation (DES)
ment layout for the drawer side panel production line. Pro-
and VSM. However, in order to facilitate the study, some
duction simulation was carried out under preset
variables are ignored, and more in-depth studies can be
parameters, and improved simulation of production links
done. Besides, the hybrid DES-VSM method ensures that
was carried out according to lean criteria. Finally, the simu-
an integrated process optimization method is used. Guo
lation test on the improvement plan of the production line
et al. showed how the process of the service system in a
has been completed to meet the increasing customer
pharmaceutical facility could achieve the minimum queue
demand; at the same time, the operating cost and processing
time [10] and listed four scenarios which were simulated
cycle of the production line have been reduced. Overall, our
with FlexSim software to reduce the time and increase the
contributions are as follows:
number of prescriptions provided. In their research [11],
(1) The digital twin model of the wood side panel pro- Jeong and Phillips developed and demonstrated a concept
duction process was established by using VSM and development process (CDP) as a framework to solve process
FlexSim simulation model, which effectively simu- layout optimization problems related to VSM. CDP and
lated the real production environment SCLI are very practice-oriented because they do not require
any complex analytical knowledge. FlexSim software is also
(2) We have simulated and improved the production used for improvement through industrial engineering
links according to the lean principle, which effec- methods, mainly reducing production and environmental
tively improves the production efficiency of wood improvements.
side panels
(3) In addition, we have adopted the computer vision 2.2. Simulation Technology. Visualization simulation tech-
method to complete the tasks of wood quality nology is a technical means to change the digital information
inspection and wood thickness inspection, respec- in simulation into intuitive, graphic, and image representa-
tively, and realized the high automation of the pro- tion, and to present the simulation process changing with
cess flow time and space. With the development of system simulation
and computer technology, it has become an important tech-
The remaining sections of this paper are arranged as fol- nical branch in the field of simulation, which attracts peo-
lows: Section 2 introduces the relevant work of this paper, ple’s attention and develops rapidly [12]. Generally
Section 3 discusses the details of our model concretely, Sec- speaking, the current three-dimensional simulation model-
tion 4 is the experimental data and results, and the summary ing is a one-time realization of modeling, so in order to
and discussion are arranged in Section 5. reduce the operation of redrawing and importing, half of
Wireless Communications and Mobile Computing 3

Draw the Design future VSM

Select physical
current VSM (Improvement Make work plan
process
(Existing problems) target)

Figure 1: VSM design process.

Figure 2: FlexSim simulation model (propaganda map of FlexSim software official website).

the simulation software will be assembled using the model to sort the real wood floor, and the average execution
with component library. At the same time, in order to con- time of this sorting system is 0.4 s, which provides
sider the real time and fidelity of simulation technology, a the technical basis for online detection of real wood
compromise between the two needs to be adopted in the floor [17]. Note that we have adopted the method
design, seeking various help in software and hardware. Sim- of random sampling to improve the random forest
ulation software, in fact, has great guiding significance for model, which effectively improves the generalization
the discrete system manufacturing industry. Schriber and ability of the model
others [13] discussed simulation software such as AutoMod,
(2) Traditional thickness detection methods are very back-
SLX, ExtendSim, Simio, Arena, ProModel, and GPSS/H to
ward. They usually relying on workers to use Vernier
illustrate the entity, resources, control elements and opera-
calipers or micrometers for off-line contact detection
tion, simulation operation, entity status, and list of entities,
of thickness, but their efficiency is low and affected by
and their general management ideas are implemented in
production conditions. Therefore, manufacturers turn
AutoMod, SLX, ExtendSim, and Simio.
their eyes to noncontact detection devices one after
another. At this time, y-ray, laser, optical, ultrasonic,
2.3. Visual Inspection
microwave, infrared, and other methods are commonly
(1) At present, there are three commonly used methods used and mature in the artificial panel detection devices
for single-board UV quality inspection, namely, for thickness detection [18]. These devices enable con-
manual inspection, sensor inspection, and machine tinuous and noncontact thickness measurements with
vision inspection. For examples, Ryu et al. [14] put the advantages of high precision and fast speed. How-
the geometric features of wood surface defects into ever, sensors for noncontact detection are expensive,
multilayer perceptors, and the correct rate of defect and dust and wood chips and thickness greatly affect
classification and recognition reached 84%. Karsulo- detection accuracy during the detection process. Li
vic et al. [15] used ultrasound to identify wood et al. [19] proposed a glass thickness detection algo-
joints, growth rings, textures, etc. The recognition rithm based on the moment invariant algorithm. We
results are consistent with visual judgment results. refer to ResNet here to classify the qualified and
Guo et al. [16] take local threshold segmentation unqualified board thicknesses. ResNet uses the residual
and global visual significance as defect segmentation module to realize the deep network
methods, extract texture features such as gray level
cooccurrence matrix and local binary mode of defect 3. Model Framework
image, and then classify and identify defects using a
support vector machine based on manifold. The final As shown in Figure 3, in view of the problems existing in the
experimental accuracy rate is 91.67%. Hu performs improvement of the side panel production line for drawers,
superpixel segmentation on the real wood floor we first draw a current VSM and analyze the production pro-
image and uses the improved random forest model cess of the side panel. Through the VSM analysis, we found
4 Wireless Communications and Mobile Computing

Start End

Record workshop condition

collect real-time data

Overall improvement for the Flexsim original plan

Draw the current VSM physical workshop modelling

Modify data
Key problem analysis

Simulation experiments
Validation
results analysis

Work our
improvement plan
Modify plan Model constructing
problems
Flexism improved plan
Draw the future VSM
modelling

Figure 3: Lean improvement model for wood drawer side panel processing line optimized by simulation.

Monthly shipping data of Side panels of drawer for 2020-2021

Unit: cubic meter
1000.0 933.01
897.5 892.4
900.0 794.1 809.7
780.6 746.5 776.6 748.2
800.0 705.01 713.01 709.7
700.0 633.01 665.01 622.7
615.1
600.0 530.01 569.01
480.01
500.0 437.49
400.0 338.9 340.01
300.0
200.0
100.0
0.0
1 2 3 4 5 6 7 8 9 10 11 12
Year2020
Year2021

Figure 4: Monthly customer demand of drawer panel statistical table for 2020-2021.

out the root causes of the existing production line problems workshop. Through the continuous circulation of such pro-
and formulated improvement measures. Then, we optimize cess, we can finally iterate out a satisfactory optimized
and restructure the production process and draw the future workshop.
state map of the value stream to achieve lean improvement
and reduce waste. At the same time, we established FlexSim 3.1. Pilot Selection. Through on-site inspection of the pro-
models of foundry workshops before and after improvement duction line and according to the process analysis value
according to the current and future state maps of VSM. Next, stream, the improvement of the production line is mainly
we will make a comparison between the before and after vali- based on pull-type. The suitable work-in-process capacity
dation to see if the optimization goal we set at the beginning and supermarket system are determined, and the unreason-
has been achieved. If not, we will check the process in which able equipment and personnel on the production line are
the validation found the problem. If required, the workshop laid out to solve the problems of large material handling vol-
model can be exported. Finally, process optimization can be ume, improper staffing of key processes, and unbalanced
carried out on the actual workshop according to the virtual production.
Wireless Communications and Mobile Computing 5

Production cycle time of major processes before improvement

Unit: second/s
3.5
3.31
3
2.5
2 TT = 1.58s

1.5 1.24 1.37

1.03 0.99
1 0.69 0.68 0.68 0.68
0.5
0

e)
h

es
ng
ng

ng
t

nt
ain
gt
in

rs
sid

fro
hi
di

gi
len

ve
rm

ep

ka
nc
an

(
th

re
Fo

es
d

ac
d
Pu
s

(
bo
xe

Si
ss

sid

es

p
Fi
ne

on

sid

le
th

tic
ick

ng

bo

th

ar
th

hi

bo
V

n
lis
d

io
U

V
xe

Po

ct
U
Fi

e
sp
In
Figure 5: Production cycle time of major processes before improvement (TT = 1:58 is the industry standard, a value below that indicates a
premium level of production speed).

Through field investigation on the production situation daily average customer demand, and TT is the production
of the whole wood processing workshop, it is found that task according to customer demand.
there are a lot of production problems in the drawer side According to the formula, the monthly average customer
panel production line, which affects the processing and demand (MD) reached 614.06 cubic meters, and the average
manufacturing of the whole enterprise. With the increase daily demand (DD) reached 27.9 cubic meters. According to
of customer demand, it is difficult to fulfill the order demand the customer’s daily demand (DD), the production takt time
gradually. Therefore, we selected the side panel processing was calculated to be 1.58 seconds.
line for drawers as the main target for lean improvement.
3.2.2. Overview of Drawer Side Panel Production Workshop.
3.2. Mapping the Current Value Stream The drawer production workshop works 22 days a month,
3.2.1. Analysis of Production Status. Key data are acquired using a double shift system with one shift being 8 hours.
based on customer order information. To ensure data reli- The actual production capacity of the current workshop
ability, interviews are conducted and validated by depart- without overtime is 292.60 cubic meters per month and
ment managers to obtain production plan control and 13.30 cubic meters per day. According to the customer’s
information flow data. These data mainly include customer order demand, without considering customization and
demand of products and relevant information of production according to the customer demand volume from 2020 to
processes. Considering the Sino-US trade war in 2019 and 2021, combined with the average demand calculation for-
the impact of the pandemic from 2020 to the present, this mula, the customer’s monthly average demand (MD) for
paper only takes into account the customer order require- drawer boards has greatly increased, reaching 614.06 cubic
ments of the company in 2021, as shown in Figure 4. meters monthly; the daily demand (DD) reached 27.91 cubic
Shipments from January to June, 2020 are 4432 cubic meters; therefore, the original production is unable to meet
meters and 3085 cubic meters from January to June, 2021, customer demand. Other information requested by the cli-
with a 30.39% drop in shipments; 4299.5 cubic meters from ent is as follows: the order task is shipped once a month,
July to December, 2020, 2920 cubic meters from July to and each transportation is transported in a 40-foot con-
October, 2021, with lost orders from U.S. customers, and tainer, with 54 cubic meters per box. The conventional prod-
there were no orders from November to December 2021. uct specification is 520 length ∗ 92 width ∗ 16 thickness
The formulas for average demand and demand production (mm), that is, the volume of 1306 products is 1 cubic meter.
takt are shown below:
3.2.3. Collect Each Process Information. Figure 5 shows the
YD production cycle of the main processes of the workshop that
MD = , ð1Þ needs to be improved before improvement. The production
12 line of drawer side panels mainly includes 9 main processes
such as thickness-setting sanding and molding. According to
MD
DD = , ð2Þ the definition of value stream, the production activities of
22 the entire workshop are divided into value-added and non-
Daily Working Time value-added activities. It can be divided into the following:
TT = : ð3Þ personnel allocation of each process, single-shift output,
DD
equipment working time, equipment quantity, and other
MD is the monthly average customer demand, DD is the primary data. In order to draw the VSM, it is necessary to
6 Wireless Communications and Mobile Computing

Figure 6: Current VSM before improvement (the figure shows the specific details of each step and the time they take, and the arrows
indicate the sequence of each step).

use the collected primary data to calculate the secondary thickness), improving employee operation posture, and reduc-
data through formulas (4)–(6). The secondary data repre- ing bending motion. By analyzing the value of the whole pro-
sents the mapping parameters required by the model. duction line, the link of benefit waste is analyzed, and the
The mapping parameters required to draw a VSM following improvement scheme is put forward in detail.
mainly include the equipment utilization rate, value-added
time, and non-value-added time. The specific calculation 3.3.1. Bottleneck Process Improvement. For the two processes
formula is as follows: of forming and lengthening, the equipment used for forming
is four-sided planing, which is mainly used for widening and
Actual Production Time
ð4Þ grooving of sanded wood with fixed thickness. The equip-
Ut = , ment used for fixing the length is double-end milling, with
Total available Time
the main purpose being setting the length at both ends. Four
n and six people are needed for the original two processes,
VAT = 〠 CTi , ð5Þ respectively, and only one person in each process is respon-
i=1 sible for production, i.e., value-added activities and the non-
n

 value-added activities for intermediate feeding need 0.56 hrs.
i :
PLT = 〠 CTi + CT ð6Þ According to the principle of merging improvement in
i=1 ECRS principle, we merge two processes, two four-sided
planers plus feeding device, and adopt double-end milling
Among them, CTi refers to the production time of the i
 i refers to the interval time between the i and i connection in the end feeding. Through this improvement,
process, CT only three people are needed in the two processes.
+ 1 processes. After data collection, it can be concluded that For the drilling procedure, we have improved the origi-
the value-added time of the production line is 12.72 seconds, nal single-row drilling equipment. Single-row drilling equip-
and the sum of the value-added time and non-value-added ment is responsible for perforating 3 rows of holes in the
time of the product is 4 hours. product. Half of the 20 cubic meters of single shift output
3.2.4. Drawing the Current VSM. According to the informa- need to be perforated repeatedly, which will lead to inven-
tion of each process, the VSM is drawn as shown in Figure 6. tory accumulation and low equipment utilization. According
Considering that the purpose of status map is to fully visual- to the principle of simplified equipment improvement in the
ize the waste in the current production situation by drawing ECRS principle, we adopted automatic multirow drilling,
and calculating and the supply of transport lines and filler which changed the original equipment that had to repeat
lines is sufficient, we selected 8 processes in the production perforation many times into automatic feeding and one-
line that have a greater impact on the value-added time step forming and adjusted the staffing from 4 to 2.
and carried out VSM analysis. According to the above For the side painting process, the original equipment was
VSM, it can be seen that the equipment of the entire produc- the paint line, requiring 15 employees to perform the follow-
tion line is backward and the production efficiency is low. ing processes: spray primer-air dry-grind-spray primer-air
dry-grind-spray finish-air dry-shelf. These complex proce-
3.3. Improvement of ECRS Principles. Based on the analysis of dures result in inconsistency between the upstream and down-
the key data above, we utilize the principle of ECRS (cancel, stream procedures. According to the eliminate and compile
merge, adjust, and simplify) and analyze the following principles of ECRS, the vacuum UV line is used to improve
schemes to improve the production line: process improvement the original multipass painting process to one-step forming
of production line, elimination of bottleneck process (check after vacuum UV process, which is not only environmentally
wood quality), updating equipment, improving tool quality friendly but also cost-effective. Because of the reduced process,
to meet customer personalized customization (adopting wood the original 15 people were also reduced to 3.
Wireless Communications and Mobile Computing 7

Manufacturing Customer
orders object expectations for Customer
features Demand orders
Price generation Quality
Manufacturing Quality Price
order Delivery data QFD Delivery data
Service level Service level
Cost

Contract

Manufacturing
Manufacturing Internal order Delivery
decision production
Delivery
Order

Virtual
enterprises
Out sourcing
Contract contract

Figure 7: Production planning and order operation system model based on supply chain management.

products produced per minute is the same, i.e., the synchro-

Table 1: Description of network structure parameters for
nization of production lines. At the same time, the inspec-
MobileNetv2.
tion of raw materials before the first procedure, without
Input Operator T C N S relevant equipment in the workshop, requires manual
inspection. Through a comprehensive and innovative plan-
224 ∗ 224 ∗ 3 Conv2d 32 1 2
ning of the supply process, this part of the treatment is
112 ∗ 112 ∗ 32 Bottleneck 1 16 1 1 passed on to the supplier. With the supplier processing, the
112 ∗ 112 ∗ 16 Bottleneck 6 24 2 2 production activities can, therefore, be quickly organized
56 ∗ 56 ∗ 24 Bottleneck 6 32 3 2 based on the actual dynamic demand of the current mar-
ket [20].
28 ∗ 28 ∗ 32 Bottleneck 6 64 4 2
Considering that the processing speed of punching in the
14 ∗ 14 ∗ 64 Bottleneck 6 96 3 1 later process is still slower than that of the first three pro-
14 ∗ 14 ∗ 96 Bottleneck 6 160 3 2 cesses, after these three processes, tray code package is car-
7 ∗ 7 ∗ 160 Bottleneck 6 320 1 1
ried out to establish a semifinished supermarket. Priority is
given to the storage of some semifinished products. After
7 ∗ 7 ∗ 320 Conv2d 1 ∗ 1 — 1280 1 1 start-up, the UV line can be started as long as the preheating
7 ∗ 7 ∗ 1280 Avgpool 1 ∗ 7 — — 1 — is completed without waiting for the upstream process to
1 ∗ 1 ∗ 1280 Conv2d 1 ∗ 1 — k — — pack and deliver the semifinished products, thus shortening
the inventory level of the upstream process and improving
the order cycle [21]. At the same time, the production plan
For two-sided polishing, two sets of sanders were origi- and order operation system model associated with the sup-
nally used for fine polishing on both sides of the product plier, customer, and order is established, forming pull-type
to facilitate the next UV process. Replacing one-sided production oriented to customer demand and realizing flex-
sanders with two-sided sanders, according to the compile ible restriction of production, production schedule adjust-
principle in ECRS principle, saves one sander and support- ment, and production capacity restriction. From user order
ing operators. input to order completion, all departments revolve around
order operation, demonstrating the integrated enterprise
3.3.2. Non-Value-Added Time Improvement. We analyzed thought of vertical and horizontal integration: horizontal
the temporary storage area of the production line according order-driven pull-type production and vertical resource-
to the current VSM and found that the temporary storage constrained production control. The specific model is shown
area was set from fixed thickness sander, four-sided planing, in Figure 7.
and double-end milling. However, in practice, there is the
possibility of realizing process linkage in these three pro- 3.4. Quality and Thickness Detection Model of Board. Wood
cesses. Our goal is to ensure that the number of semifinished quality inspections are two important link in side panel
8 Wireless Communications and Mobile Computing

(a) (b)

Figure 8: Unqualified (a) and qualified (b) UV surfaces.

(a) (b)

Figure 9: Qualified thickness (a). Unqualified thickness (b).

production, but manual screening is easy to misjudge and a deep network using the residual module, which is
time-consuming. In order to realize automatic inspection more robust than traditional deep networks. As the
of wood surface quality and wood thickness, we established neural network gets deeper, the correlation between
in-depth learning models for both. the returned gradients gets worse and worse and
finally approaches the white noise, making the depth
(1) In order to detect wood surface quality, we intro- of the neural network meaningless. Since images are
duced convolution neural network for identification locally correlated, it can be assumed that gradients
processing. Here, we use MobileNetv2, which has should have similar correlations, so updating gradi-
the advantage of highly movable model and high ents with residual modules makes sense. ResNet
performance. Table 1 illustrates the network struc- learns from residuals and passes features prior to
ture settings of our model. Input refers to the charac- the previous activation function through an identical
teristic input scale of the layer, operator refers to the connection to the output of this calculation, with
operation of the layer, T refers to the expansion their sum as the output of the cost layer. Then, the
index, C refers to the output dimension, N refers to whole network is followed by a Softmax classifier,
the number of repetitions, and S refers to the step which classifies it into normal thickness and abnor-
mal thickness, completing an effective classification.
We randomly divided 1000 datasets collected from the Specifically, our ResNet structure settings are shown
factory into training, test, and validation sets at 8 : 1 : 1. First, in Table 2
we used image enhancement on the dataset, as shown in
Figure 8. Then, we use MobileNetv2 to capture the picture
features. Finally, we use the Softmax classifier to process 4. Experiment
the acquired features to get the probability of image classifi-
cation. After cross-validation, the current accuracy rate is 4.1. Data Sources. In this study, the quality control in wood
96.875%, which can meet the requirements of the task. processing is mainly the quality detection of wood thickness
and brightness. We take pictures on the spot according to
(2) In order to detect the thickness of wood and achieve the workshop and select more shooting angles to ensure
automatic tool selection, we refer to ResNet here to more accurate results of training detection. It is divided into
classify the qualified and unqualified wood board self-made CNN datasets made after evaluation, mainly wood
thickness, as shown in Figure 9. ResNet implements thickness datasets, including 500 eligible thickness picture
Wireless Communications and Mobile Computing 9

Table 2: Description of ResNet’s structure parameters. 4.2. Evaluation Index

Layer name Outputs size 34-layers 4.2.1. Product Value-Added Rate. Product value-added rate is
Conv1 112 ∗ 112 7 ∗ 7, 64 stride 2 an important indicator to measure the processing efficiency of
production line. By comparing product value-added time with
3 ∗ 3 max pool, stride 2
( non-value-added time, we can find relevant data to reflect the
Conv2.x 56 ∗ 56 3 ∗ 3, 64 production mode and time utilization rate of production line;
3×
3 ∗ 3, 64 map the value generated for the layout of production line, the
( configuration of personnel, and the storage of material filling
3 ∗ 3, 128 temporary storage area; and identify the waste problem. The
Conv3.x 14 ∗ 14 4×
3 ∗ 3, 128 formula for calculating the product value-added rate is as fol-
( lows:
3 ∗ 3, 256
Conv4.x 7∗7 6×
3 ∗ 3, 256 VAT
( PCE = × 100%: ð7Þ
3 ∗ 3, 512 PLT
Conv5.x 1∗1 3×
3 ∗ 3, 512 According to the current VSM, the product value-added
rate is 0.08%, which indicates that the actual value-added time
Table 3: Testing results of wood quality. used in the production line is very short. Due to the long prep-
aration time in wood side panel processing, improper material
Ours Model A Model B filling and handling arrangements lead to long waiting time of
AUC 97.78% 84% 91.67% equipment and large storage of line materials. Although UV
Recall 99.62% — — line production needs to be warmed up for more than one
F1-score 98.2% — —
hour in advance, the processing speed is far ahead of the first
6 procedures, which will cause inconsistency between down-
Precision 96.875% — —
stream production and downstream production. At the same
time, most personnel arrange non-value-added links such as
feeding, discharging, and handling materials. This will result
Table 4: Testing results of wood quality. in many downstream processes being on standby when the
material does not arrive, leading to a lot of waste.
Ours Model A Model B
AUC 99.61% 74% 98% 4.2.2. Balance Rate of Production Line. As an index to mea-
Recall 99.24% — — sure the leanness and performance of production line, the
F1-score 99.61% — — production balance rate reflects whether the time load of
Precision 100% — — each process is uniform in the whole process. The higher
the balance rate of production line is, the waiting time
between processes is less and the non-value-added time in
the process is less. The formula for calculating the balance
Table 5: Model performance after improving training strategy. ratio of the production line is as follows:
Parameters Accuracy rate ∑ni=1 CT i
Baseline 99.61% LOB = : ð8Þ
n × CT max n
+cosine decay 99.77%
+label smoothing 98.79% The total value-added time of the product is 12.72 sec-
+mixup 98.76% onds, and there are 9 processes in total. Among them, two
process exceeds the takt, and the bottleneck process is the
datasets and 500 unqualified thickness picture datasets. side paint. Not only the production cycle is long but also
Label is also used for image labeling of thickness datasets, 15 employees are required to carry out the following pro-
which can be used as additional data to detect the position cesses in this process: primer spraying-drying-polishing-
of the board in the training of a neural network. Another primer spraying-drying-polishing-topcoat spraying-drying-
dataset is the wood paint surface dataset, which includes shelf, and the weather factor should also be considered.
500 qualified picture datasets and 500 unqualified picture According to the formula, the balance rate of the produc-
datasets. Both datasets were acquired by the participants tion line is only 42.76%. The production cycle of each process
during field shooting in the workshop. Of the two datasets, is too different, and the imbalance problem is obvious. At the
80% were used for training and 20% for testing. The number same time, due to the problems of the equipment itself, it takes
of experiments was the same and was performed under the 1-2 hours to adjust the machine for production when the cus-
same experimental conditions. Personal PC (Intel(R) Cor- tomer needs to develop new samples. In addition, the bottle-
e(TM) i5-8300H CPU @ 2.30 GHz) repeats the experiment neck process needs to be improved urgently due to the long
30 times and takes the average value. waiting time of materials and a serious waste of personnel.
10 Wireless Communications and Mobile Computing

Figure 10: FlexSim model of the production process of drawer side panel workshop before improvement.

4.2.3. Image Classification Error Rate. The AUC calculation (1) The learning rate decay strategy uses cosine function.
is shown in The diagram of learning rate change with epoch is
similar to cos, as shown in
∑insi ∈Pclass rank insi − ðM ∗ ðM + 1ÞÞ/2
AUC = , ð9Þ
M∗N   
epoch
newlr = etamin + 0:5 ∗ ðinitiallr − etamin Þ ∗ 1 + cos ∗π
where the serial number representing the i sample is the T max
rank insi ; M, N are number of positive samples and negative ð10Þ
samples, respectively; ∑insi ∈ Pclass indicates that only the
serial numbers positive sample are added together.
4.3. Experimental Results. Model A is a laser-based method (2) Label smoothing softens the label of one-hot type
to calibrate the distance between upper and lower laser sen- that was commonly used in the past, so that it can
sors by static and dynamic noncontact measurements of the reduce overfitting to some extent when calculating
thickness of manmade panel using two laser sensors. Model the loss value
B is a thickness detection method based on machine vision.
Through image gray level processing, perspective correction, (3) Mixup is a data enhancement method. If mixup
sliding plate image filtering, and sliding plate image training is used, then two input images are read at
enhancement processing, image preprocessing is completed a time; assuming that they are represented by ðxi , yi
for sliding plate images taken in laboratory environment. Þ and ðx j , y j Þ, then a new image can be synthesized
An image tracking method is proposed for edge detection by the following two formulas, and this new image
and morphological open-close operation of sliding plate can be used for training as follows:
image to extract and locate the edge of sliding plate and cal-
culate the minimum thickness of sliding plate.
Model A is a thickness detection method based on laser
with two laser sensors being used to carry out static and (i) x = λxi + ð1 − λÞx j where xi , x j are the raw input
dynamic noncontact measurement experiments on an artifi- vectors,
cial panel thickness. A calibration method for distance (ii) y = λyi + ð1 − λÞy j where yi , y j are one-hot label
between the upper and lower laser sensors is proposed.
encodings
Model B is a thickness detection method based on
machine vision. Through gray processing, perspective cor- Note that all the experiments, except the above improve-
rection, skateboard image filtering, and skateboard image ments, remain the same settings and are trained on NVIDIA
enhancement processing of skateboard pictures taken in V100. Table 5 shows the model performance after improving
the laboratory environment, image preprocessing is com- training strategy.
pleted. It is proposed that after edge detection and morpho-
logical on-off operation of skateboard image, image tracking 5. Summary and Discussion
is used to extract and locate skateboard edges, and the min-
imum skateboard thickness is calculated. Tables 3 and 4 5.1. Modeling Simulation Lean Model. According to the
shows the testing results of wood quality. improvement scheme, the FlexSim simulation model of the
Through the experimental results, we can find that our side panel workshop of the drawer before improvement
model achieves more than 90% performance on multiple which is shown in Figure 10 (unit time in seconds) should
indicators and has higher accuracy than Model A and Model be improved. In the process of modeling, the required mod-
B, which indicates the validity of our experimental model. In ules are selected first, and the raw materials are supplied to
addition, to verify the validity of our model parameters, we the required feeding process with the generator. Key pro-
also explored the effect of different training strategies on cesses in the production line are completed with a processor.
the convergence speed of the wood thickness detection Considering the production bottleneck in the original pro-
model. cess, the temporary storage area is set before length fixing,
Wireless Communications and Mobile Computing 11

Figure 11: Improved FlexSim model of the production process of drawer side panel workshop.

Production cycle time major procss after improvement

Unit: second/s TT = 1.58s
1.6 1.38
1.4 1.19 1.19
1.2 1.03 1.10
1
0.8 0.68 0.68 0.68
0.6
0.4
0.2
0

)
h

es
ng

ng

ng
nt

se
ain
gt

sid

er
or
di

hi

i
len

ck
ep

ev
an

nc

(F
h

pa
(R
ot
d

d
Pu
s

es
xe

Si
ss

cle
b

sid

es
g fi
ne

ng

ti
sid
ick

th

ar
hi
in

bo

th
lis

n
rm
th

io
bo
Po

V
Fo
d

ct
xe

e
sp
Fi

In
Figure 12: Production cycle time of main process after improvement.

Figure 13: Future VSM with improvement.

punching, and side painting process. Finally, the corre- 5.2. Drawing the Future VSM. According to the improve-
sponding means of transport is selected, or connecting each ment plan and FlexSim simulation model made by specific
process directly. To simplify the model and reduce runtime, analysis, the future VSM is shown in Figure 12.
operators are not allocated in the model. According to the above simulation model and improve-
The FlexSim simulation model of the production process ment plan, we draw the future of VSM. After improvement,
of the drawer side panel workshop, which is built according 37 employees were reduced, 4 single-row drills are changed
to the improved scheme, is shown in Figure 11 (unit time in to 1 automatic multirow drills, and 2 single-sided sanders
seconds). to 1 double-sided sander. For the improvement of the
12 Wireless Communications and Mobile Computing

Four sided Automatic multi

Sander Four sided planing1 planing2 Sander Molding machine row drill

31.11% 18.73% 18.73% 73.72% 85.18% 97.78%

Double end milling Single row drill Paint line

Vacuum UV line Double sided sander UV line front

29.91% 31.12% 99.99% 78.74% 48.67% 48.67%

Polishing sander UV line front UV line reverse Uv line reverse Packing

41.38% 20.54% 20.54%

48.67% 48.67%

Packing
Idle
Processing
20.84%
Blocked

Idle
Processing
(a) Before improvement (b) After improvement

Figure 14: Pie chart of utilization ratio of main process equipment in drawer side panel production.

bottleneck process, the production time of each process is that is closest to the production pace. After forming and
made less than the takt time, and the balance rate of produc- length synthesis, the utilization ratio of overall equipment
tion line changes from 42.76% to 71.83%. The product has been improved a lot, and after two Sanders are combined
value-added rate has been increased from 0.08% to 0.11%. into one double-sided sander, the utilization ratio of equip-
The monthly production volume of products has changed ment has also been improved. Besides, after replacing the
from 292.60 cubic meters/month to 703.10 square/month. vacuum UV line of equipment, the overall equipment utili-
The production cycle of the improved main processes is zation of the original high-load paint line is eased down.
shown in Figure 13. The utilization ratio of adjacent production equipment is
similar, which improves the stability and continuity of the
5.3. Simulation Lean Model Analysis. The FlexSim model is whole production process, reduces blockage, and smoothens
run to compare the equipment utilization rate of production interval of the whole production line.
line before improvement. The results are shown in the fig-
ure. It can be seen from Figure 14(a) that before improve- 5.4. Summary and Discussion. In this paper, the drawer side
ment, the overall workshop production balance is low and panel processing workshop is selected, and the digital twin
the main reasons are the bottleneck process and equipment technology is introduced to carry out the lean improvement
failure. The highest utilization of paint line equipment as a and simulation analysis research of this production line.
bottleneck process also means that equipment burden is Firstly, the basic product information of the workshop is col-
high and are prone to equipment failures. Secondly, lected and the current VSM is drawn. Through the value
although the processing cycle of single-disc drilling exceeds flow chart and the ECRS principle, the existing waste prob-
the production takt, four single-row drills are running due lems are analyzed, and the improvement plan is put forward
to the production process considerations in the workshop, in light of the key problems. In order to verify the objectivity
which leads to the low utilization of the overall equipment, and reliability of the improvement plan, a simulation model
and one equipment is basically idle. Other processes are of the drawer side panel processing workshop based on Flex-
affected by bottleneck processes and equipment utilization Sim software is established. Finally, the improvement plan is
is low for a long time. In addition, the equipment is idle determined according to the operation results of the simula-
due to the low product value-added rate caused by manual tion model and the future VSM was drawn. The research
transportation and packing as well as preheating of the shows that the combination of VSM and simulation model
equipment. As shown in Figure 14(b), the overall utilization technology based on digital twin technology boasts certain
of the improved equipment has been increased, with the guiding significance and reference value for improvement
highest utilization being achieved in the perforation process of the manufacturing industry.
Wireless Communications and Mobile Computing 13

However, this paper only uses the concept of digital twin [7] J. Ma, H. Chen, Y. Zhang et al., “A digital twin-driven produc-
for reference, and the establishment of the model does not tion management system for production workshop,” Interna-
fully meet the requirements and level of digital twin technol- tional journal of Advanced manufacturing technology,
ogy. At the same time, there is still room for research on the vol. 110, no. 5-6, pp. 1385–1397, 2020.
optimization of workshop logistics handling route and on- [8] G. V. Punna Rao, S. Nallusamy, and P. Raman, “Enhancement
site customization improvement. Relevant intelligent algo- of production in subassembly line of a medium scale industry
rithms can be considered to verify the simulation optimiza- using different lean tools and FlexSim simulation software,”
tion, which is also the optional direction in the future. International Journal of Engineering Research in Africa,
vol. 44, pp. 229–239, 2019.
[9] Z. Aziz, R. M. Qasim, and S. Wajdi, “Improving productivity of
Data Availability road surfacing operations using value stream mapping and
discrete event simulation,” Construction Innovation, vol. 17,
The data used to support the findings of this study are avail- no. 3, pp. 294–323, 2016.
able from the corresponding author upon request. [10] H. Guo, L. Zhang, Y. Zhang et al., “A differential game-based
approach for school-enterprise collaborative R&D strategy
Conflicts of Interest on digital twin technology,” IEEE Access, vol. 8, pp. 188688–
188698, 2020.
The authors declare no conflicts of interest. [11] K. Y. Jeong and D. T. Phillips, “Application of a concept devel-
opment process to evaluate process layout designs using value
stream mapping and simulation,” Journal of Industrial Engi-
Acknowledgments neering and Management (JIEM), vol. 4, no. 2, pp. 206–230,
2011.
This work is supported by the Guangdong Academic Degree
and Graduate Education Reform Research Project (No. [12] F. J. Kang, X. Hua, H. H. Li, B. Zhang, and M. Su, “Overview of
development of visualization simulation technology,” Journal
2019JGXM15), Guangdong Province Higher Education
of System Simulation, vol. 21, no. 17, pp. 5310–5313, 2009.
Teaching Research and Reform Project (No. 2020059),
[13] T. J. Schriber, D. T. Brunner, and J. S. Smith, “Inside discrete-
Guangdong Graduate Education Innovation Project (No.
event simulation software: how it works and why it matters,”
82620516), Jinan University Off-Campus Practice Teaching in 2013 Winter Simulations Conference (WSC), pp. 424–438,
Base Construction Project (No. 55691207), Guangdong Washington, DC, USA, 2013.
Higher Education Association “14th Five-Year Plan” Higher
[14] C. Ryu, S. G. Kong, and H. Kim, “Enhancement of feature
Education Research Topic (No. 21GZD09), Guangdong extraction for low-quality fingerprint images using stochastic
Province “Quality Engineering” Construction Project (No. resonance,” Pattern Recognition Letters, vol. 32, no. 2,
210308), Jinan University Graduate Education Teaching pp. 107–113, 2011.
Achievement Cultivation Project (No. 2021YPY010), and [15] J. T. Karsulovic, L. A. León, and L. Gaete, “Ultrasonic detection
Jinan University Teaching Reform Research Project (No. of knots and annual ring orientation in Pinusradiata Iumber,”
JG2022086). Wood and Fiber Science, vol. 32, no. 3, pp. 278–286, 2020.
[16] H. Guo, Y. Zhu, Y. Zhang, Y. Ren, M. Chen, and R. Zhang, “A
References digital twin-based layout optimization method for discrete
manufacturing workshop,” The International Journal of
[1] L. Tian, G. Liu, and S. C. Liu, “Digital twin and production line Advanced Manufacturing Technology, vol. 112, no. 5-6,
simulation technology,” Journal of Graphics, vol. 42, no. 3, pp. 1307–1318, 2021.
pp. 349–358, 2021. [17] J. F. Hu, Research on Technology of Solid Wood Floor Sorting
[2] Siemes, “Digital twin - driving business value throughout the Based on Machine Vision, Northeast Forestry University, 2015.
building life cycle,” [2020-12-28].https://new.siemens.com [18] G. Wang and D. Li, “Development of automatic detection
/global/en/products/buildings/contact/digital-twin. html. technology for artificial wood panel thickness,” Forestry
[3] H. Guo, X. Lian, Y. Zhang et al., “Analysis of environmental Machinery &Woodworking Equipment, vol. 4, pp. 4–6, 1999.
policy's impact on remanufacturing decision under the effect [19] J. X. Li, H. L. Lin, B. Yu, W. Chen, C. H. Lin, and S. T. Huang,
of green network using differential game model,” IEEE Access, “Machine vision-based non-congeneric product detection
vol. 8, pp. 115251–115262, 2020. algorithm for vehicle quarter glass,” Journal of Fujian Univer-
[4] Y. H. Lian and H. V. Landeghem, “Analysing the effects of lean sity of Technology, vol. 19, no. 3, pp. 223–229, 2021.
manufacturing using a value stream mapping-based simula- [20] H. Guo, M. Chen, K. Mohamed, T. Qu, S. Wang, and J. Li, “A
tion generator,” International Journal of Production Research, digital twin-based flexible cellular manufacturing for optimi-
vol. 45, no. 13, pp. 3037–3058, 2007. zation of air conditioner line,” Journal of Manufacturing Sys-
[5] X. B. Huang, L. Y. Gong, and L. X. Yang, “Study on automotive tems, vol. 58, pp. 65–78, 2021.
back door panel injection molding process simulation and
[21] H. Guo, M. Chen, R. Zhang et al., “Research on improvement
process parameter optimization,” Manufacturing Engineering,
of truck vibration based on systematic G8D method,” Shock
vol. 2021, article 9996423, pp. 1–15, 2021.
and Vibration, vol. 2019, Article ID 1416340, 11 pages, 2019.
[6] W. Yanhong, H. Junxia, T. Yuanyuan, T. Zhongda, and Z. Jun,
“A modified CONWIP production control strategy for lamp
production line,” in 2018Chinese Control And Decision Confer-
ence (CCDC), pp. 3916–3921, Shenyang, China, 2018.