Journal of Innovation & Knowledge 7 (2022) 100276

Journal of Innovation
& Knowledge
ht t p s: // w w w . j our na ls .e l se vi e r .c om /j ou r na l -o f - in no va t i on -a n d- kn owl e dg e

An innovative machine learning model for supply chain management

Haifeng Lina, Ji Lina, Fang Wangb,*
a
College of Information Science and Technology, Nanjing Forestry University, Nanjing 210037, China
b
College of Electronic Engineering, Nanjing XiaoZhuang University, Nanjing 211171, China

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

Article History: Supply chain management (SCM) integrates all links and business processes involved in the supply chain
Received 13 April 2022 through the information management system. Applying artificial intelligence algorithms to the SCM system
Accepted 25 September 2022 can realize the visualization, automation, and intelligent management of all links in the supply chain. This
Available online 10 October 2022
can effectively help enterprises reduce operating costs and improve their ability to respond to market
demands, thereby increasing overall operational efficiency. An effective member selection method is an
Keywords:
important basis for smooth dynamic supply chain operation. To address the problem of high numbers of
Supply chain management (SCM)
decision attributes and low numbers of data samples for decision analysis, this paper proposes a dynamic
Performance optimization
Innovative machine learning
supply chain member selection algorithm based on conditional generative adversarial networks (CGANs). To
Conditional generative adversarial networks ensure that classification performance will not be reduced, the member classification method on the chain
(CGAN) can successfully reduce the data dimension and complexity in the classification process. Furthermore,
Prediction of purchasing and inventory machine learning is used for analyzing and predicting purchase and inventory links in the supply chain. For
Vehicle scheduling the vehicle scheduling module, the path is reasonably planned to improve the operation efficiency. The inte-
grated implementation of the SCM system is finalized using the SSH framework.
© 2022 The Authors. Published by Elsevier España, S.L.U. on behalf of Journal of Innovation & Knowledge. This
is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)

Introduction implementation of manager decisions have become crucial to the

supply chain (Guo et al., 2020). Artificial intelligence provides techni-
Against the backdrop of economic globalization, the supply chain cal potential for the supply chain system to attain these require-
relationship is undergoing dramatic changes. Enterprises develop ments. It can also be used to price, adjust storage capacity, change
and grow effectively only when they constantly adjust their policies transportation plans by predicting the weather, and timely adjusting
in accordance with market changes. Therefore, a smooth, efficient business decisions using external data (Ghorashi et al., 2020; Gholiza-
and cost-saving flow of information, money, logistics, and human deh et al., 2020). A neural network is an adaptive, nonlinear dynamic
resources is crucial to the operation of modern enterprises. This system that consists of hosts of interconnected neurons that enable it
information must be mastered as accurately as possible to ensure to perform complex logical operations. As a data processing system, a
economic success. Accordingly, open access to the “information neural network imitates the structure and function of the human
island” existing in the enterprise supply chain has caused a break- brain. These networks can carry out state responses to continuous or
through for enterprises to improve quality and increase production discontinuous input, and process the dynamic information in nonlin-
capacity. The supply chain is a complete process from order signing ear relationships. A neural network is a manually established
to product arrival in the hands of customers. Other constituent pro- dynamic system that uses a directed graph as its topological struc-
cesses of the supply chain include raw material purchasing, ware- ture. It processes data by making state responses to continuous or
housing, production, inventory, outbound, transportation and discrete inputs. The graph that serves as its topological structure rep-
distribution (Rahmani et al., 2020; Hasani et al., 2021). An efficient resents processing units as nodes and weighted directed arcs (Alavi-
and mature supply chain information management system necessi- doost et al., 2021). Designed for handling search problems, the
tates the functions of on-demand traceability, process visualization, genetic algorithm is applicable to many types of general problems.
and management automation. Therefore, the accurate collection and This algorithm is founded on the parallel search of a chromosome
reception of information in each link and the rapid and efficient group using three types of operations: selections, exchanges and
mutations. This combination makes the genetic algorithm unique.
* Corresponding author. Traditional algorithms are likely to fall into the local optimal solution,
E-mail addresses: [email protected] (H. Lin), [email protected] (J. Lin), as they iterate from a single initial value. The genetic algorithm,
[email protected] (F. Wang).

https://doi.org/10.1016/j.jik.2022.100276
2444-569X/© 2022 The Authors. Published by Elsevier España, S.L.U. on behalf of Journal of Innovation & Knowledge. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/4.0/)
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

however, is efficient for global optimization owing to its large cover- provides a scientific basis for SCM (Xu et al., 2021; Almomani et al.,
age and basis on a string set. It is capable of processing many individ- 2022). However, traditional inventory control theories and methods
uals in a population simultaneously. Because it can evaluate multiple are difficult to adapt to new characteristics and requirements, and
solutions at the same time (Jamrus et al., 2020; Gaurav et al., 2022), enterprises face great risks when adopting traditional management
the genetic algorithm does not risk falling into a local optimal solu- methods. In the past, inventory control was facilitated by the eco-
tion. Furthermore, this algorithm is conductive to parallelization. To nomic purchase quantity management method in microeconomics.
use the genetic algorithm, no knowledge of search space or availabil- When this method was applied to purchase goods, it was found that
ity of auxiliary information is required, except for the value of the fit- inventory quantity has always been in an unreasonable situation, and
ness function, which is needed for individual evaluation. This fitness inventory redundancy and shortage has often occurred as a result
function is not restricted by continuous differentiability, and its defi- (Carrera et al., 2020). BP neural networks have strong non-linear
nition domain can be set arbitrarily, thus significantly expanding the mapping ability, which is especially suitable for solving problems
algorithm’s application range. Furthermore, the genetic algorithm with complex internal mechanisms. Therefore, it is highly feasible
follows probabilistic, rather than deterministic, transition rules in its and reasonable to exercise inventory control according to BP neural
search process thanks to its properties of self-organization, self-adap- network prediction. Traditional prediction methods do not effectively
tion and self-study. When the algorithm executes a search using data solve the problem of accuracy due to the difficulties presented by
derived from the evolutionary process, the individuals with high fit- complexity, dynamic nonlinearity, and uncertainty of the system.
ness exhibit higher survival probability and receive the optimal gene Small changes in the product demand of consumers at the front end
structure for the environment (Goodarzian et al., 2021; Brdesee et al., of the supply chain will lead to progressively greater changes in the
2022). Based on these considerations, this study adopts conditional product demand of companies at the back end of the supply chain.
generative adversarial network (CGAN) to realize the dynamic part- During the overall period, inventory levels in the supply chain exhibit
ner selection method in the supply chain, while utilizing innovative drastic fluctuations. From customers to retailers, distributors, and
machine learning for the analysis of purchasing and inventory links. producers, the fluctuation range always increases, and the fluctuation
The integrated implementation of the supply chain management of product demand has a more obvious impact on enterprises at the
(SCM) system was completed using the SSH framework. top of the supply chain. Thus, enterprises in different stages of the
Establishing strategic partnership is not only the focus of supply supply chain may diverge in market demand, resulting in the collapse
chain strategic management, but also the core of supply chain man- of supply chain coordination. The influencing factors of the chain
agement. The key of supply chain management lies in the connection mainly include demand forecasting, batch ordering, product ration-
and cooperation among the node enterprises of the supply chain, as ing, product pricing, and performance incentive. These factors com-
well as the good coordination among them in design, production, bine and interact in various ways among different supply chains.
and competitive strategy and so on. In this paper, a CGAN method is Therefore, it is important to consider all factors to achieve coordina-
proposed, based on the index data of the relevant factors of the part- tion in the supply chain (Lu & Wang, 2020; Karimian et al., 2020).
ners, the alternative schemes used by the decision makers of the core This study uses demand forecasting as an example to analyze the
enterprises in the supply chain to build the cooperative relationship coordination of a supply chain.
are mined, which improves the accuracy and comprehensiveness of SCM is one of the most important means for enterprises to obtain
the decision-making. This method can cluster the existing candidate international competitive advantages. The use of information tech-
partners of enterprises, and find the types of enterprises to cooperate nology to reconstruct the business process, enable users to cooperate
with, which is very useful for decision makers to make supply chain with suppliers, eliminate wasteful activities in the supply chain, sim-
decisions and logistics planning. Because the judgment of many fac- plify the business process, save time, improve quality, and reduce
tors involved in partner selection is uncertain, this paper get more costs, is an effective means of enhancing global competitiveness of
objective and scientific conclusions through reasonable quantitative enterprises. Extensive academic efforts have been made to research
analysis. This method not only effectively solves the problems of inventory management. Du et al. (Du et al., 2020) developed a system
incompleteness, uncertainty and subjectivity of decision-making model on the basis of the original management mode using a genetic
information, but also effectively solves the problem of dynamic selec- algorithm and BP neural network to improve the efficiency of hospi-
tion of supply chain partners, providing a basis for decision-makers tal drug inventory operations. Teerasoponpong (Teerasoponpong,
to make scientific decisions. 2022) proposed a decision support system (DSS) applicable to sourc-
The rest of the paper is structured as follows. Section 2 discusses ing and inventory management with the goals of helping SMEs in
related studies, and Section 3 describes the design of the SCM system. data compilation and exploitation, and supporting their decisions
Section 4 chiefly presents an integrated implementation of the sys- under business ambiguities. Shen et al. (Shen et al., 2020) set up an
tem framework. The dynamic selection method of supply chain part- integrated model to facilitate the inventory management of farmers’
ners based on the CGAN is proposed in Section 5, and Section 6 fresh agricultural products at various links, including professional
serves as a conclusion, providing a summary and future directions of cooperatives, distribution centers, supermarkets, and other markets.
research. Harifi et al. (Harifi et al., 2021) proposed several models of inventory
control problems. These models were deterministic single-product,
Related Work deterministic multi-product, and stochastic single-product. To solve
the inventory routing problem (IRP), Dosdogru et al. (Dosdogru et al.,
Inventory management plays an important role in the supply 2021) proposed a new hybrid methodology that includes two phases
chain. Through inventory management, enterprises can prevent stock that provide a generic framework for IRP. Wu et al. (Wu et al., 2021)
shortages, maintain the continuity of the operation process, share tackled a multi-period location-inventory-routing problem with time
ordering costs, and quickly meet users’ ordering needs. However, windows and fuel consumption, which optimizes the routing, loca-
because the prediction of future demand has significant uncertainty, tion, and inventory decisions for distribution centers and customers
inventory management is susceptible to issues such as goods hoard- in supply chains of different levels. Badakhshan et al. (Badakhshan et
ing, poor liquidity, capital backlogs, and high inventory management al., 2020) incorporated cash flow modelling into the SD structure of
costs (Saputro et al., 2021). The prediction of demand is one of the the beer distribution game to develop the CFB function. Garg et al.
most important bases for enterprise decision-making and planning, (Garg et al., 2020) presented a multi-objective genetic algorithm opti-
which is important to reduce risks and avoid blindness. This in turn mization framework in the modeling of the bullwhip effect and NSGA
transforms uncertain factors into macro controllable factors, and for centralized and decentralized SCM. Potential improvements of
2
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

inventory management in a stochastic environment have also an SC can be designed to solve multi-product, multi-objective, and
attracted significant attention from the academic world. Liu et al. (Liu multi-stage supply chain problems based on real-time data. Because
et al., 2020) focused on a coordinated location-inventory problem in the development of green supply chains has become a trend, Hu et al.
a stochastic supply chain system which is disrupted by supply at ran- (Hu et al., 2020) proposed an effective Tabu search algorithm by
dom. Karimian et al. (Karimian et al., 2020) built a multi-item eco- increasing the sensitivity of governments to environmental issues.
nomic production quantity model with a shortage for a single- This study analyzed the performance of data classification and bal-
vendor, multi-retailer supply chain under vendor-managed inven- ance levels of CGAN, and applied it to partner selection in a dynamic
tory policy in an uncertain environment. Modern supply chains are supply chain, meanwhile, this study completes the integration and
prone to risks due to their complicated networks and wide geograph- implementation of the SCM system framework, so that the system
ical coverage. An integrated planning scheme may adjust the opera- has scalability and portability. Table 1 is the summarize existing
tion of supply chain networks effectively and minimize the impact of work key feature and its limitation.
unknown changes and failures, thus reducing risk from both internal
and external sources. Nezamoddini et al. (Nezamoddini et al., 2020) Supply Chain Management and Machine Learning
constructed a risk-based optimization framework for use in opera-
tional decisions. They proposed a model that handles uncertainties in Supply Chain Network Connection and Collaboration
connection with facility interruptions, lead times, and demands in
supply, production, and distribution channels. Liu and Li (Liu & Li, The essence of supply chain network connection collaboration is
2021) developed a bi-level programming model for joint decision- to realize five connections, namely, connecting customers, connect-
making of order allocation and product configuration, where the ing suppliers, connecting products, connecting production equip-
coordination and protection mechanisms were used to coordinate ment and connecting managers. After the intelligent transformation
between capacity-constrained suppliers and reduce the risks in sup- of the enterprise’s supply chain, these five connections will be con-
ply chains. Han and Zhang (Han & Zhang, 2021) formulated a supply nected and become a supply chain node. A small cycle of connections
chain risk management model based on machine learning. will be generated between customers, suppliers, production equip-
To research SCM system software, industries can integrate MRP Ⅱ ment, products and management. It can also be connected with other
and ERP software interfaces or other specific functional modules, business systems to form a high degree of supply chain network col-
including customer demand forecasting, operation planning and laboration sharing information.
decision analysis ability, enterprise production scheduling, inventory In terms of the integration and transformation of logistics data in
management plan, cooperative enterprise logistics management and the supply chain, the two emerging technologies, machine learning
service, quotation and negotiation decision-making model, transpor- and big data, can be used to realize the real-time transformation of
tation decision, transportation route planning, and other functions. logistics data processing and the integration of supply chain data
The prediction model, decision-making scheme model, path planning using mobile handheld terminal devices. With the continuous devel-
scheme, and scheduling algorithm are integrated through the net- opment of the new generation of information technology, the tradi-
work, database, and front-end technology to provide interactive and tional supply chain management mode has been transformed. The
intuitive options for decision-making at all levels. This approach ena- collaboration of network connection has also brought a new revolu-
bles users to obtain static data and ensure real-time tracking of tion to the industry and realized a wider range of connection and col-
dynamic data, which greatly improves the response of enterprises to laboration.
market changes. The development planning of the enterprise should Supply chain management is a very important link in the com-
proceed from a general perspective to avoid the disadvantages pany’s business operation. At the same time, in the operation link,
caused by the pursuit of local optimization. The integration of pro- the optimization of operation is the core and key to meet the maxi-
duction planning and distribution as the two main functions in sup- mum demand of customers with the minimum cost. The goal of sup-
ply chain (SC) management is a major current problem. Cai et al. (Cai ply chain management is to shorten the company’s capital turnover
et al., 2021) suggested green computing energy management, pro- period, reduce the risks faced by the enterprise, achieve the profit
moted a collaborative distribution optimization framework and algo- growth of the company’s performance, and provide predictable
rithm for an intelligent SCM, and constructed a joint optimization income to the managers.
model of VRP & VFP for logistics distribution. Gharaei and Jolai (Ghar- The traditional supply chain is a logistics network system com-
aei and Jolai, 2021) solved the production scheduling and distribution posed of supply chain, manufacturers, warehouses, logistics centers
problem in a multisite supply chain from the three-agent perspective and channel providers. Different enterprises have different business
of the manufacturer, distributor, and customer. Mousavi et al. nodes, and the supply chain networks formed are different. However,
(Mousavi et al., 2020) investigated the production and air transporta- the core of the supply chain is to manage the company’s logistics,
tion scheduling problem with time windows for the due date. Deng capital flow and information flow around the enterprise. The process
et al. (Deng et al., 2021) studied the carbon emission costs during from the production of products or the purchase of raw materials to
transportation and cold storage, the effect of temperature changes on the final delivery of goods to customers through the company’s sales
the decay rate of fresh products during unloading, and the traffic logistics network.
along the actual distribution route. Khan et al. (Khan et al., 2020) pre- For modern enterprises, better control of enterprise information
sented a hybrid prediction model based on recurrent neural net- flow, better use of artificial intelligence and big data to achieve
works, long short-term memory, gated recurrent units, and a genetic resource sharing, and on this basis, improve the service quality of
algorithm to further optimize the parameters of the hybrid prediction enterprises, shorten logistics costs and time, so as to ultimately
model. Huang and Yang (Huang & Yang, 2020) constructed a multi- achieve a win-win situation between customers and enterprises. It
item joint ordering framework under the global cruise ship supply can be seen that supply chain management is a very important link
chain, including integrated operation of purchases, different order in enterprise operation management. Supply chain management
cycles, delivery, and inventory based on a cruise distribution center. includes all management activities from enterprise strategy to tactics
Shen et al. (Shen et al., 2021) combined big data and intelligent com- formulation and operation management.
puting to study the agricultural product intelligent SC system, and The integration of artificial intelligence and supply chain manage-
used the network equilibrium method to build an SC model that con- ment system can bring synchronization in technical standards to
siders the effort level of multiple producers and retailers. Ali et al. enterprises, without the need to develop other application systems.
(Ali et al., 2021) presented a novel algorithm that demonstrates how It can not only achieve internal management of enterprises, but also
3
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

Table 1.
Existing Work Key Feature and Limitation.

Existing work key feature Existing work limitation

Supply chain design The traditional supply chain design problem is to analyze the normal opera- The design of supply chain needs to consider the factors
tion environment. It is usually assumed that the supply chain will operate of unexpected events into the design of supply chain.
according to the plan once it is designed. However, in reality, due to the Taking emergencies into account, the traditional supply
impact of unexpected events, the nodes and their connections of the sup- chain design problems usually change. The objective
ply chain may fail to operate according to the original plan. function of some problems will change.
Supplier selection strategy and The main problem of supplier selection and purchase strategy is the deci- The first is supplier selection, and the second is emer-
procurement strategy sion of supplier combination and order allocation among suppliers. In the gency procurement and backup suppliers. Supplier
event of an emergency, we can try our best to ensure that the supply of selection refers to the selection of the main supplier.
products will not be affected by reasonable selection of suppliers and The backup supplier is the supplier that is activated
allocation of orders. when the main supplier is interrupted. Purchasing from
the backup supplier is backup procurement or emer-
gency procurement.
Inventory and ordering strategy Using inventory and ordering strategies to deal with supply chain disrup- Considering the interruption risk into the traditional
tion risk, especially supply disruption risk, is the focus of many studies. inventory and order models, many models are difficult
The traditional inventory and order model does not consider the possibil- to obtain analytical solutions, but are solved by numeri-
ity of interruption risk. When the interruption risk is considered, the tra- cal analysis and simulation.
ditional inventory and order model needs to be re studied.
Supply chain coordination strategy In the case of decentralized decision-making in the supply chain, due to the Consider supply side risk, demand side risk, transporta-
existence of double marginal effect, the overall optimal profit of the sup- tion risk and other costs
ply chain is often difficult to achieve. In order to mediate conflicts and In the case of risk, analyze the behavior and equilibrium
make better cooperation among supply chain members, it is necessary to conditions of each node.
design a supply chain coordination mechanism.
Demand management strategy Demand management strategies include demand transfer, late delivery Considering the sensitivity of demand time, how to influ-
order and compensation strategy. ence customers’ demand preference, so as to respond to
random supply.

bring improvement to enterprise supply chain management. In order (1) Customer Order Pull
to adapt to the changing enterprise management environment, the
coordination of the supply chain management system can compre- In the supply chain system, the orders issued by customers drive
hensively grasp various management methods required by the enter- the production of enterprises, and the flow of the entire supply chain
prise operation. Dynamic management can implement flexible value flow needs to be pulled by customers’ orders. When no cus-
organization and strengthen internal processes to adapt to external tomer order is generated, the upstream enterprises of the supply
changes. Strengthen planning and organization, and adopt flexible chain do not provide any service. After receiving the customer order,
countermeasures to cope with changes. The coordination of supply the whole supply chain responds quickly and provides services.
chain system can promote the efficiency of production, logistics,
organization and finance.
(2) Punctual and Accurate

Establish an Intelligent Supply Chain Management Mode

Punctuality is a prerequisite for the supply chain system to
achieve internal smoothness and a certain degree of rhythm relaxa-
The goal of the intelligent supply chain management information
tion. All node enterprises in the supply chain must deliver on time
system is that the resources between the business departments of
according to the plan; Accuracy is a necessary condition for the lean
the company can be interconnected. First, the middle layer manage-
supply chain, including accurate customer demand, accurate inven-
ment platform, machine learning algorithm and internal processing
tory, accurate delivery time, etc., which puts forward new require-
flow are introduced to realize the integration. The second step is to
ments for the reasonable layout of suppliers.
standardize the data information format, establish a complete infor-
mation platform, and release the data of various departments in a
hierarchical and standardized manner. The third part is to share the
(3) High Quality and Low Cost
data resources among various departments by using the set permis-
sion specifications, so that the internal personnel can see the relevant
In the supply chain system, each node enterprise greatly reduces
information in a timely, complete and accurate manner, and the data
the production cost and eliminates the waste of equipment, person-
flow between the platforms can also flow smoothly.
nel and process through reasonable resource allocation and efficient
This new mode is one of the ways to realize the transformation of
information exchange. Thus, the main energy is put on quality con-
the supply chain, and the ultimate goal is to realize the information
trol, which ensures the high quality of supply chain services.
sharing among various departments and customers within the com-
pany. The intelligent management mode can also provide electronic
and customized management of supply chain orders, better guide
(4) Emphasis on Coordination
customers to place orders in an electronic way, and internally man-
age these orders in a networked and electronic manner, making
The supply chain system is an organism composed of various
information sharing more convenient and efficient. At the same time,
enterprises and various production processes. If all parts can’t main-
it can also establish a client-side order query mode to facilitate cus-
tain coordination and cooperation, the operation of the supply chain
tomers to timely understand the logistics and inventory situation.
system will face serious problems. The coordination of the supply
chain includes two levels. One is the internal process coordination,
Main Characteristics of Supply Chain Intelligence which emphasizes the information exchange between various links
The main features of supply chain intelligence are as follows: within the production to make each process perfectly connected. The
4
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

second is external communication, which emphasizes the strength- uniform parameters and relationships, and the product can be man-
ening of communication and cooperation among the enterprises at aged with uniform configuration. In addition, this paper uses CGAN
each node of the supply chain and the establishment of a perfect to realize the dynamic selection method of supply chain partners.
information exchange channel. Firstly, the sample data is mapped to Hilbert space, and then the fea-
ture extraction problem is transformed into a linear solution problem
to obtain the feature matrix of the sample data. At the same time, the
(5) Information Networking information entropy and mutual information of each sample data fea-
ture matrix are calculated, and the relative entropy of the target is
A smooth operation of the supply chain system can’t be separated adjusted according to the calculation results. The experimental
from the information exchange between the node enterprises. How- results show that the CGAN data classification is accurate, the data
ever, to make the exchange smooth, we must rely on the powerful balance level is high, and the overall generalization is improved.
Internet system and big data. Only through the Internet platform can Therefore, CGAN is suitable for partner selection in dynamic supply
enterprises in the upstream and downstream of the supply chain chain.
come together to negotiate and deal with the difficulties and prob-
lems encountered in the operation of the supply chain.
Design of SCM System

Function Design of SCM System

(6) Continuous Improvement
Management information system platform modes can be roughly
Compared with other supply chain management, the biggest char-
categorized into four types: host terminal mode, file server mode, cli-
acteristic of intelligent supply chain management is that the
ent/server (C/S) mode, and browser/server (B/S) mode. The first two
improvement activities of intelligent supply chain occur at various
modes cannot fully meet the needs of Internet users due to hardware
nodes and time periods in the operation process of the supply chain.
limitations and the increased number of users and data. Therefore,
It is a dynamic and continuous process. Through the virtuous cycle
modern enterprise management information systems generally imple-
from finding problems to improving problems and then to finding
ment B/S or C/S mode. Our system adopts the current leading B/S
problems, the overall performance of the supply chain is constantly
mode, which is an expansion of the C/S mode. The pre-display function
optimized and improved. Therefore, the first step of implementing
of the original client’s application module is helpful for maintenance
the intelligent supply chain is from improvement.
and upgrading. It is usually composed of presentation, function and
data layers. The architecture of the pattern is shown in Fig. 1.
Integrated Supply Chain Collaborative Management Platform
After the theoretical model and algorithm are established, a prac-
Machine learning is an artificial intelligence technology that dis-
tical application system is needed to verify actual performance to
covers and extracts the hidden information from the data warehouse.
optimize the operation mode and increase management efficiency.
It can automatically analyze the data, conduct inductive reasoning
The specific system function design is shown in Fig. 2.
and association on them, find some internal relationships between
As shown in Fig. 2, the SCM system designed in this study is com-
the data, and find out potential patterns that play a very important
posed of order, purchasing, production, inventory, distribution, and
role in information prediction and decision-making behavior, so as to
transportation modules. Each module includes corresponding busi-
establish new business models to help decision makers make correct
ness. Other interface is possible to work in our model, the innovation
market strategies. The main inputs in the supply chain system are the
of this model lies in:
data from the data warehouse, the converted data of each subordi-
nate department, and the knowledge and experience stored in the
knowledge base of the data mining system. The data selected from (a) First, define the WebService interface. After defining the interface,
the data warehouse is processed in the knowledge base discovery both parties define the transmission rules, transmission fields and
engine, which provides a large number of data extraction algorithms necessary contents. JSCON transmission is used to support multi
to generate auxiliary patterns and relationships. Some discoveries header and table body data and multi table body data.
need to be added to the knowledge base so that subsequent discover-
ies can be extracted and evaluated. The new knowledge obtained by (b) The interface is divided into two categories: down and up (upload
mining will be saved in the mining knowledge base again. and download). It is determined by the information in the header.
The small sample machine learning data classification based on When the interface information receives the content, it will view
the condition generation countermeasure network designed in this the information fixed in the header, such as supplier code,
study has good effect, the data balance processing effect is successful, whether the date is upload or download, and whether the content
and the overall generalization is better than the traditional data proc- in the data table body is empty (this is the pre check content). If it
essing algorithm.
Through the integration of the unified collaboration platform and
various decentralized professional tools, this paper summarizes the
originally scattered data generated in each stage of the whole process
management of the supply chain into the data pool for unified man-
agement, forming a single data source management, ensuring that
the business data can flow smoothly and have traceability at all
stages of the management, and ensuring that customers and manag-
ers can obtain correct and reliable data at any time. At the same time,
the integrated collaborative management platform can be integrated
through a variety of enterprise information systems to open up the
flow of multi professional data in the enterprise. Through the inte-
grated collaborative management platform, a multi-disciplinary and
unified product multi-stage can be formed. Through configuration
management, the product’s multi-disciplinary data can be set with Fig. 1. B/S mode three-tier structure.

5
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

Fig. 2. Function diagram of SCM system.

is unqualified, it will directly return the information to the client relationship, so as to improve the accuracy and comprehensiveness
to upload again. of the decision-making. It can provide data support for formulating
task plans and allocating operation and maintenance forces.
(c) If the information passes the pre check, the infrastructure will be
used to allocate tasks, generate task codes, store them in the data-
base in the temporary table, and wait for the platform system to (1) Reliability requirements:
start the automatic import task.

(d) Register the task type, task type code, name, source data table, tar- (a) Solve the problem of overlapping calls of synchronization tasks,
get table, type, level, time, deletion flag and other information in synchronization tasks are called regularly by scheduled tasks. A
the database. scheme needs to be designed to prevent the same task from over-
lapping, that is, to prevent the previous task from running and the
(e) When each task is started, it first finds the corresponding task next task from starting again (to prevent data inconsistency, dead-
through the transmission information, obtains the storage table lock, etc.).
corresponding to the task through the task, and starts through the
stored procedures in the table. The startup of each stored proce- (b) Solve the problem of data consistency, the data synchronized
dure will generate a unique task number. from the interface table to the platform table or from the platform
table to the interface table must be consistent. It is necessary to
(f) There are three processes in the start-up storage process (includ- avoid transferring inconsistent interface table data to the platform
ing preparing data, verifying data and synchronizing data). These or transferring inconsistent platform data to the interface table.
three processes are associated with a unified number. If the data
is unqualified in the data storage process, the data will be (2) Maintainability requirements:
recorded in the error table.

(g) When the data is synchronized between tables, the number of (a) Data traceability, data written to or read from the interface table
synchronization items and the number of errors are returned after can be traced. Who, when, and what data can be written or read
synchronization so that the foreground personnel can query. by viewing the log? Success or failure? Logging can be performed
according to different
(h) Support simultaneous multitasking.
Synchronize data to set the log level and log switch.
The SCM system designed in this paper standardizes and auto-
mates the operation and maintenance process, simplifies the opera-
tion and maintenance, greatly improves the efficiency of operation (b) Data maintenance, in order to maintain the high performance of
and maintenance, and reduces the maintenance cost. It is a very sta- the system, data (including business data and log data) needs to
ble and well functional SCM management system. In addition, the be archived regularly.
system also combines the CGAN algorithm to analyze the index data.
On the basis of establishing the relevant factors of the partners, it (c) Abnormal pre-warning, in the process of data transmission, it is
mines the alternative schemes that the decision makers of the core necessary to give early warning to relevant personnel for abnor-
enterprises in the supply chain use to build the cooperative mal conditions, so as to manually intervene and handle problems.

6
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

Business Process Design of SCM System Purchasing Business Process Design

According to the order, it is necessary to purchase the detailed list
of materials required for the production of products, put forward the
Order Business Process Design
purchase information to the supplier in the form of purchase order,
This module mainly manages, tracks, and responds to information
and track the purchase process to make sure that the materials reach
requested by customers. Through order management, it commands
the enterprise in the shortest time possible and ensure the normal
the production rhythm and adjusts inventory quantity, thereby
production operation. According to the order number, the system
improving production efficiency and boosting the supply cycle. Links
receives detailed information including material price, performance,
are adjusted in the supply chain according to the order situation,
and model from the supplier through electronic data interchange
thus ensuring rapid response to the market. Once order is signed, the
(EDI), selects the intended company, and signs the purchase order
system automatically generates a unique order number, which can
contract. After signing the contract, the supplier’s logistics informa-
be used to track the flow of orders in all links of the enterprise’s sup-
tion is accessible through the order number until the purchase is
ply chain in real time, and master the overall process of order produc-
completed (Al-Farsi et al., 2021).
tion, packaging, and distribution (Yuan et al., 2020). Furthermore, the
order number is directly scanned and identified from the sensing
layer through an RFID (Radio Frequency Identification Technology) Production Business Process Design
electronic tag, which avoids manual errors. The order business pro- The production module is largely responsible for the management
cess is shown in Fig. 3 and control of production links. Through this module, we can prevent
In the order management, the company shall specify the scope of excessive or insufficient production quantity, accurately understand
abnormal purchase and the approval procedures for abnormal pur- the status of production activities, and improve product quality. The
chase. The system will automatically calculate the quantity and installation equipment can display production information such as
period of purchase for over quantity purchase and near validity drug process flow, material information, production quantity, and yield,
purchase. When creating a purchase order, the over quantity or near and paste the electronic label. Personnel can then manufacture prod-
validity purchase will automatically trigger the approval of abnormal ucts according to the equipment prompt. After the operation, the
purchase. The purchase order can be successfully created only after reader can scan the electronic label and feed back the completion
the approval is passed. The company shall formulate a single procure- information to the SCM system. Finally, managers can understand
ment contract template to clarify the responsibilities and rights of the station production through the system, monitor and adjust the
both parties. The procurement business department shall initiate production process to improve the yield, make production activities
contract approval in accordance with the relevant provisions of con- more refined and flexible, and visualize the production process
tract management and sign the contract after the approval of the (Vegter et al., 2021).
authorized person, so as to ensure that the terms such as procure-
ment price and quantity are consistent with the approved supplier Inventory Business Process Design
selection results and procurement requirements documents, and Inventory business, also known as warehouse management, is
ensure that the contract terms are approved. The procurement busi- largely used to store qualified products temporarily in advance of
ness department shall communicate with the supplier to confirm the orders or according to market forecasts. A temporary transfer station
purchase quantity according to the sales and inventory conditions, in is used for production, supply, and sales. As the integration of logis-
combination with the factors such as order cycle, in transit period, tics and information flow, inventory business is important in improv-
minimum order quantity, service level, etc. the procurement person- ing product delivery speed, market response ability, and production
nel shall create purchase orders in the system according to the pur- efficiency. It is generally composed of receipt, issue, warehouse trans-
chase contract, which shall be implemented after being reviewed by fer, and inventory counting businesses (Asamoah et al., 2021).
the procurement manager. Warehouse transfer orders and tax paid
transfer orders can be created only after they are created and
(1) Warehousing business process design: Paste the electronic tag on
approved.
the storage shelf and mark the storage location information.
When items with an RFID tag pass through the access control, the
reader will automatically collect and match with the system
order, send the information back to the supply chain system, and
simultaneously write it into the database.

(2) Issue business process design: The system first writes the material
information to be delivered out into the RFID electronic tag of the
transport vehicle. When the material passes through the access
control system, the reader displays the material information to be
delivered out in the system. Finally, the system terminal triggers
the delivery business process to complete the delivery of materi-
als.

(3) Business process design of database transfer: The stock transfer

business of inventory goods refers to the transfer of materials
between warehouses or storage locations. The process of ware-
house transfer can accurately identify storage location informa-
tion through an electronic label to ensure the real-time tracking
of goods.

(4) Inventory business process design: The management personnel

read the information on the electronic tag through the reader,
Fig. 3. Order business flow chart. automatically obtain the details, transmit the information to the
7
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

SCM system through the Internet, compare it with the material and its related applications, and effectively store data to satisfy users’
database, and update the database after confirmation to improve application needs. Database technology is the basis of information
the automation level of inventory counting. systems. The use of information systems is the process of information
processing, storage, and flow, and the success or failure of an infor-
mation system is directly related to the rationality of database design.
The purpose of system database modeling is to enable enterprise
Distribution Business Process Design
leaders, managers, and information technicians to have a unified,
Distribution business is an important link in the supply chain and
general, and complete understanding of the planned information sys-
serves as the embodiment of the relationship between information
tem, so as to formulate the overall scheme scientifically. Whether the
flow and logistics. In the distribution business, the following aspects
database design is reasonable directly affects the difficulty of later
need to be considered: (a) goods must be delivered on time, which
system maintenance, and serves as the basis of system operation sta-
requires the planning and management of vehicle transportation
bility and robustness. Database structure design should also take into
paths. (b) Products must be delivered completely and within the
account the design mode of the existing enterprise database to
agreed time. (c) To reduce the burden of enterprises, delivery costs
ensure smooth integration between different system platforms when
must be minimized. The design process of a distribution business is
implementing the information engineering construction.
as follows: first, classify and package the products to be transported,
and write the information into the RFID electronic tag of the product
packaging box with the terminal. Then, optimize the distribution Database E-R Diagram and Data Table Design
route according to the customer location and vehicle conditions The management system of the supply chain is characterized by a
(Tang and Liu, 2020). complex process that handles immense data shared by multiple mod-
ules. Consequently, it is of vital importance to successfully perform
Transportation Business Process Design data query and management and design a database system that is
In the design of a transportation business process, the goods, con- rationally structured, convenient to use, and highly efficient. The data
tainers, and vehicles are labeled electronically, and the geographical table relationship is shown in Fig. 4.
location is updated through the readers on the delivery point, trans-
fer station, and target warehouse. This ensures identification, anti-
counterfeiting, positioning, and user tracking. Through the electronic Database Features of SCM System
tag and vehicle GIS system, the article and vehicle location informa- SCM in this paper is an open system database server (DB2, Infor-
tion can be uploaded to the Internet, thereby allowing for the accu- mix, SQL server, SQL server, etc.) that integrates multiple platforms;
rate prediction of arrival time and real-time monitoring of the it extends the content management that provides integrated access
transportation process (Srinivasan et al., 2021). channels for various data types of information; Help customers ana-
lyze and extract valuable business information from various data;
Database Design of SCM System Integrate distributed information integration in heterogeneous envi-
ronments.
Database Design Principles SCM in this paper can be transferred to federated systems of dis-
Database design is intended to construct the optimal database tributed servers. More information about data system used in our
mode for a given application environment, establish the database research is as follows:

Fig. 4. Data table design of SCM system.

8
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

(1) Incremental synchronization: because there are many enterprise (3) Distributor cloud: The distributor’s agent in the supply chain plat-
customers and a large amount of data, the system must support form mainly includes the following list of collaborative activities:
incremental modification, and only the modified data is synchro- sales management, purchase management, warehouse manage-
nized each time. ment, transportation management, customer relationship man-
agement, etc.
(2) Support the following data processing: add data / modify data /
seal data / delete data. (4) Retailer cloud: the agent of the retailer in the supply chain plat-
form, which mainly includes the following collaborative activities:
(3) Support setting life cycle: initial account creation / activation / off- Sales management, procurement management, transportation
line / offline activation / refresh. management, warehouse management, etc.

(4) Inbound data integration scenario: before entering the platform (5) Customer cloud: the customer’s agent in the supply chain plat-
table from the interface table, the data should be checked for form.
validity. The system should identify and alert the records that
failed the check. Only the data that successfully passed the check
can enter the distribution platform.
Integrated Implementation of System Framework
(5) Outbound data integration scenario: outbound data has three
integration methods: Modeling of Vehicle Routing Problem

The vehicle routing problem (VRP) is the key to optimize distribu-

(a) The first type is outbound. The data in the interface table is tion and transportation routes, reduce transportation costs, and
accessed by the target system only once. ensure scientific, intelligent, and automatic transportation schedul-
ing. It can also save considerable operational costs for businesses, and
(b) The second is that the data exported to the interface table is is advantageous in SCM and portfolio optimization. Because the VRP
accessed by the target system only twice. relates to delivery time, the VRP with time window (VRPTW) adds
the constraint of delivery time to the basis of VRP. In this type of
(c) The third type is outbound. The data in the interface table may be problem, the earliest and latest arrival times are denoted by Te and Tl,
accessed by multiple (unknown) target systems. respectively. When vehicles distribute materials from one or more
distribution centers to multiple destinations under time constraints,
the vehicle scheduling scheme with the lowest cost is obtained.

(1) Algorithm principle

Cloud Computing Platform of SCM System
In SCM, enterprises can choose to build a cloud computing plat-
form or rent cloud services from cloud based resource service compa- VRPTW is controlled by adding time window T based upon VRP.
nies. By renting basic services, enterprises do not need to spend too The acceptable receiving time limit for customers T is determined by
much hardware cost, and the cost of information sharing will be the earliest receiving time T0 , and the latest receiving time T1 . If the
reduced accordingly. In the case of enterprises’ own construction, the delivery exceeds the time limit, customers cannot accept it.
initial investment is relatively large, but in the long run, the charac-
teristics of cloud computing in low management cost and energy
consumption can make up for the later cost consumption. At the (2) Algorithm description
same time, cloud computing has brought about the development of
1; ieði; jÞ
flat, fuzzy and dynamic collaborative organization results, and collab- Set customer iði 2 vÞ and transport vehicle k; xijk ¼ f .
2ði; jÞ
1; i=
orative resources are also intelligently configured. The flattening of
Define the vehicle load capacity as D and the customer demand as
enterprises can reduce the number of managers and the cost of man-
BBB. The customer di service starts at tbi and ends at toi . When
power; The fuzziness of enterprise boundary can reduce the commu-
xijk ¼ 1, the transport vehicle k is from customer i to customer j, and
nication cost; Enterprise dynamic can increase the technological
yki is the vehicle k serving i. The service time required for i to custom-
advantages of enterprises; Automatic allocation of resources can
ers is tsi ; i.e., tsi ¼ tbi  toi . The driving time from i to j is tij . The wait-
reduce unnecessary investment and make full use of existing resour-
ing time before starting the service is twi .
ces. Based on the change of organizational structure, the cost of man-
power, communication and resources can be reduced, and the cost of
enterprise collaboration can be greatly reduced.
(3) Constraints

P
(1) Supplier cloud: the supplier’s agent in the supply chain platform, d y D 8 k means the total demand of customers ≤ load;
Pi i ki i i
the list of collaborative activities mainly includes warehouse man- k yki ¼ 1; i 2 V, means that each vehicle only serves customers
agement, production management, transportation management, once.
plan management, etc. X X
x ¼ yki ; j 2 V; 8 k;
i ijk
x
j ijk
X X
(2) Manufacturer cloud: the manufacturer’s agent in the supply chain ¼ yki ; j 2 V; 8 k; ji 2 s xijk jSj  1; S  V;
i2s
platform. The list of collaborative activities mainly includes:
demand management, procurement management, transportation Compulsory transportation vehicles form a closed loop from
management, production management, customer relationship company ! customer ! company, where s is the number of custom-
management, plan management, etc. ers visited.
9
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

Fig. 5. SSH integration framework diagram.

toi þ tij tbj ; toi tbj indicates that there is a better transportation Implementation of Each Layer of SSH Integration Framework
path.
tei tbi tli ; i 2 V, the only number of customer service and the only The system that integrates the SSH framework contains four
number of vehicle service customers. layers, each having different responsibilities. They are presentation
layer, business logic layer, data persistence layer, and domain module
layer. This layout enables developers to quickly develop web applica-
(4) Algorithm optimization tions with clear structure, good reusability, and convenient mainte-
nance. The system’s overall infrastructure depends on Struts, which
are responsible for separating MVC. The model part of the Struts
According to the optimization objective of VRPTW, i.e. the mini-
framework controls business jumps and uses the Hibernate frame-
mization of the number of vehicles and total route length, all tasks
work to support the persistence layer. Struts and Hibernate are man-
are completed, and the following equation is obtained:
aged by Spring.
nX X X XX o
minZ ¼ i j
t x ;
k ij ijk j k
xojk ð1Þ
PPP
i j k tij xijk is the total time spent on the planned route carrying
(1) Implementation of SSH Presentation layer
goods.
The vehicle speed is constant, and tij is directly proportional to the The development of SCM systems on the basis of the Internet of
route length. Therefore, the calculated values of the minimum time Things adopts J2EE architecture, provides services in the view layer
and total distance are consistent. of the MVC mode, and realizes the system display layer through JSP.
PP
j k xojk traverses the minimum number of customers. The display layer serves as the interface between the system and
user, as it responds to the user’s requests and displays the results
Integrated SSH Framework returned by the system. Therefore, JSP is used for code development
to realize system response to user requests. The response data is dis-
When developing a SCM system, enterprises should account for played in real time through HTML-XML and other markup languages.
existing system assets and focus on the application of open-source The following problems should be paid attention to in the specific
tools and software to make the system scalable and portable. The development and implementation.
SSH framework is a new technology that inherits the layered archi-
tecture mode of the J2EE framework. The difference between the
frameworks is the implementation method between layers. When (a) Data verification problem. Because the display layer directly inter-
J2EE and SSH frameworks complete the same computing tasks, SSH acts with the user, it is necessary to verify whether user input data
consumes less resources. At the same time, SSH has no special conforms to the format defined by the data layer.
requirements on the platform specificity of business objects. In SSH,
the implementation of business logic can be completed through ordi- (b) Personalized label display needs. Because users may have person-
nary Java objects. alized or customized display needs, the system can define person-
In addition, the SSH composite framework has the following four alized labels through the data layer to realize them.
technical advantages: (a) It separates Java and HTML code, reducing
the requirements for development complexity. (b) The work between (c) User operation problems. The display layer serves as the window
the layers of the system is relatively independent, and the code cou- for users to interact with the system. This interaction is often car-
pling degree is low. (c) Even if it is separated from the AOP mecha- ried out through the page. Therefore, the order of page user input,
nism of spring environment, SSH will not hinder AOP from realizing as well as the realization of list page, paging, and other displays,
business functions. (d) Hibernate and other cross platform technolo- are convenient for use.
gies used with SSH are highly open-source, which has promoted the
framework’s rapid development.
Due to its technical advantages, SSH enables a system developed To fulfill most enterprises’ needs for system application, the SCM
under its framework to retain strong expansibility and portability. system should include core business processing, enterprise daily
Furthermore, the open source model greatly simplifies system devel- application, collaborative office management, personalized require-
opment, thus significantly reducing development time. This study ments, and expandable functional requirements. JSP technology com-
adopts the SSH integration framework to realize the SCM system. The bined with the MVC mode can be used to complete the
framework is illustrated in Fig. 5. implementation of the display layer. Due to the powerful function
10
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

and flexible use of MVC, the interactive interface of the display layer Dynamic Selection Method of Supply Chain Partners Based on
developed through JSP technology features personalized characteris- CGAN
tics. It is easy to realize the interface display and the transmission
and application of data through the association with the system busi- In machine learning, traditional algorithms can easily produce
ness through the view layer. unbalanced data, resulting in poor generalizability. To solve this
problem, we propose a dynamic selection method of supply chain
partners based on the CGAN of machine learning.
(2) Implementation of SSH business logic layer
Small Sample Data Processing Based on CGAN

The business logic layer, which is implemented based on com- CGAN is composed of generator and discriminator. The generator
mand calls, integrates a large number of command interfaces. Differ- receives a random noise signal (which can be uniformly or Gaussian
ent functions of the system can call different interactive operations distributed), and generates the corresponding sample. The discrimi-
through the call interface. These include data display, abnormal page nator then receives the real and generated samples. The function of
processing, misoperation prompt, and others. Furthermore, this layer the discriminator is mainly to judge the generated sample’s authen-
provides process organization and management services for priority ticity using a probability value. The generator continuously strength-
management with concurrent users, so as to improve the value of the ens its ability to improve the generated samples, until the
system interaction experience and business logic. discriminator is unable to distinguish between the real and generated
samples. The structural framework of CGAN is shown in Fig. 6.
Below is the CGAN training strategy: randomly initialize generator
(3) Implementation of SSH domain module layer G and discriminator D, and alternatively train G and D until conver-
gence.
In the training of discriminator,
The domain module layer is largely responsible for controlling the
business objects, rules, and logic. It provides access to an interface for
the business logic layer and calls the data persistence layer to com- (1) Sample the authentic samples fx1 ; x2 ; . . . ; xm g of m in the dataset;
plete processing in the database. Therefore, it is a very central layer
that directly determines the correctness, availability, stability, and (2) Sample m vectors fz1 ; z2 ; . . . ; zm g from a distribution (even, nor-
fluency of the supply chain system implementation. Furthermore, it mal);
provides flexibility, maintainability, and expansibility support for the
realization of business. (3) Obtain m synthesized data: f x1 ; x2 ; . . . ; xm g, xi ¼ Gðzi Þ
through G;

(4) Implementation of SSH data persistence layer (4) Update the discrimination parameter ud and maximize the objec-
tive function:
Xm   1 Xm   
~¼ 1
V logD xi þ log 1  D ~x
i
ð2Þ
This layer mainly manages the operation of the database, facili- m i¼1 m i¼1

tates interactive operations with data through the interface, and pro-
vides the method of data operation call to the domain module layer. ud ud þ h r V~ ðud Þ ð3Þ
The data persistence layer provides the implementation method for
all business data processing in the system, as well as the interface for
calling operations. Generally, SQL is used to write specified methods In the training of generator,
and data operations such as addition, deletion, modification, and
query. The layer also enables the combination of these operations.
Therefore, when other layers use the method of the data persistence (1) Sample m vectors fz1 ; z2 ; . . . ; zm g from a distribution (even, nor-
layer, they do not need to understand the mechanism of content mal);
operation. Instead, they only need to provide the data type returned
by the method to ensure the processing of corresponding business (2) Update the generator G’s parameter ug and maximize the objec-
logic. tive function;

Fig. 6. Structural framework of CGAN.

11
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

Xm       1X   
~¼ 1
V log 1  D G zi ð4Þ
_ 1X m m

m i¼1 V¼ logD ci ; xi þ log 1  D ci ; x _ i

m i¼1 m i¼1
  1X m   
ug ug  h r V~ ug ð5Þ þ log 1  D ci ; b
x
i
ð6Þ
m i¼1

The discriminator of CGAN inputs samples (which can be real or ud ud þ h r V ðud Þ ð7Þ
generated) and outputs corresponding scalars (i.e. probability values
belonging to the real samples). At first, because the generator is rela- The difference between this optimization process and the conven-
P
m  
tively weak, the generated samples are easily distinguished; how- i
tional GAN solving process lies in the addition of m1 log 1  Dðci ; b
xÞ .
ever, with continuous strengthening of the generator, the i¼1

discriminator is eventually unable to distinguish the two, and finally In the generator part:
can only output a probability of 0.5 (indicating that the real probabil-
ity is equal to the generated probability). This process is equivalent to
a maximum and minimum game process. The discriminator contin- (6) Sample m noise points fz1 ; z2 . . . zm g from a distribution;
ues to maximize the discrimination into samples, and the generator
also continues to improve its ability to generate samples, eventually (7) Sample m conditions fc1 ; c2 . . . cm g from the dataset;
reaching a Nash equilibrium. Fig. 7 shows the generator of CGAN.
In CGAN, two neural networks confront each other in a zero sum (8) Maximize by updating generator parameter u g :
game. The task of the first network, the generator, is to deceive the _ 1X m    
second network, the discriminator. The generator creates ''fake'' data, V¼ log D G ci ; zi ð8Þ
m i¼1
in which case the two networks can be improved iteratively. These
networks can be considered as black boxes, representing arbitrarily
_  
complex functions, applied to noise or real data. The input of the gen- ug ug þ h r V ug ð9Þ
erator is some random noise that produces a pseudo image. The input
of the discriminator includes a pseudo sample and a sample from the The solution requires not only to train the discriminator in
real set. The training of the two networks requires a loss function, advance or reduce its learning rate compared with the generator, but
and the loss function of each network depends on the other network. also change the number of updates of the generator/discriminator at
To train the network, back propagation is carried out, and the neuron each iteration. It is easy to see when CGAN will merge, because the
weights of other networks are simultaneously frozen. Generally, the stability of the two networks will occur somewhere in the middle.
discriminator and generator networks can take the form of any map- Under the condition of conditional generation countermeasure
ping function. network, the specific steps of small sample machine learning data
Below is the procedure of CGAN algorithm: train the discriminator processing algorithm are as follows
first, and then train the generator in every iteration, in the discrimi-
nator part, Step 1: map the small sample machine learning data into the Hibert
space;
Step 2: transform the data feature extraction problem into a linear
(1) Sample m positive examplesfðc1 ; x1 Þ; ðc2 ; x2 Þ . . . ðcm ; xm Þg from the solution problem, calculate the linear marking of each projection
dataset; direction of the data in the space, obtain the corresponding sam-
ple eigenvalues, and form the eigenmatrix from the correspond-
(2) Samplem noise pointsfz1 ; z2 ; . . . ; zm g from a distribution; ing eigenvectors to obtain the sample data eigenmatrix;
Step 3: in the unbalanced state, adjust the information entropy
(3) Generate data fx _ 1; x _ 2 . . . x _ mg through x _ i ¼ Gðci ; zi Þ; appropriately until the relative entropy reaches the set balance
target, calculate the information entropy and mutual information
1 2 m
(4) Sample m sample points fb
x ;b
x ...b
x g from the dataset; of each sample data characteristic matrix, and adjust the relative
entropy of the target according to the mutual information calcula-
(5) Update the discriminator parameter u d and maximize V _ ; tion results to make the data reach balance;

Fig. 7. Generator of CGAN.

12
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

Step 4: input the processed data into the condition generation coun-
termeasure network, take the data processing target as the objec-
tive function, the discriminator reaches the optimal result, output
the optimal result under the condition of given the network opti-
mal discriminator, and complete the small sample machine learn-
ing data processing

Experimental Analysis of Small Sample Data Classification and Data

Balance Based on CGAN

The generator of CGAN takes noise and labels as inputs, and the
discriminator will generate objects, and the original objects and
labels as inputs. At the same time, CGAN optimizes the judgment cri-
teria of the discriminator, which judges false data + arbitrary labels
and true data + wrong labels as false, and judges’ true data + correct
labels as true. In this way, the user-defined label can output the
desired data.
Fig. 8. Artificial imbalance data.
CGAN (Conditional GAN) uses tags to train generators and dis-
criminators. The generator learns to generate realistic samples for
distribution was obtained through linear affine transformations, and
each tag in the training dataset, while the discriminator learns to dis-
the two-dimensional standard Gaussian distribution data, namely
tinguish between true sample tag pairs and false sample tag pairs.
x » Mð0; 1Þ, y » Mð0; 1Þ, were obtained through a one-dimensional
The generator of CGAN synthesizes a pseudo sample using the
standard normal distribution. The data pairs used in the experiment
noise vector and the label. The purpose of this pseudo sample is to
meet the two-dimensional standard Gaussian distribution, and the
make the discriminator think that it is the real sample of a given tag
distribution of the obtained data on the two-dimensional plane is
as much as possible. The discriminator of the CGAN accepts the
shown in Fig. 8.
labeled real sample or the labeled pseudo sample. The discriminator
It can be seen in Fig. 8 that the imbalance between the positive
learns how to identify real data and match pairs, and also learns to
and negative sample data in the artificial unbalanced data set is obvi-
identify false sample tag pairs and distinguish them from real sample
ous, and the distribution of the data in the two-dimensional space
tag pairs.
appears to be oval, indicating that the experimental data set was
CGAN training process:
transformed from the standard Gaussian distribution. In prior studies,
the unbalanced data had a negative impact on the data processing
(1) Training a discriminator to randomly take a small batch of labeled algorithm, thus deteriorating the processing effect. In this experi-
real samples and their labels (x, y); Calculate d ((x, y)) of a given ment, the proportion of positive and negative samples was con-
small batch and back propagate the binary loss update to mini- trolled, the generated data set was studied, and the ROC curve
mize the loss; Randomly take a small batch of random noise vec- comparison results were obtained using CGAN, BPNN, and PSO, as
tors and class labels, and generate a small batch of pseudo shown in Fig. 9.
samples; Calculate small batch and back propagate binary loss The abscissa FP-rate and ordinate TP-rate shown in Fig. 9 respec-
updates to minimize losses. tively represent the proportion of negative and positive samples,
with classification errors in the total samples. By observing the ROC
(2) A training generator that randomly takes a small batch of random curve, it can be seen that among the three groups of experimental
noise and category labels to generate a small batch of pseudo sam- results, the area enclosed by the CGAN and FP rate axes is the largest,
ples; Calculate and back propagate the binary loss update for a and the area enclosed by the BP neural network and FP rate axes is
given small batch to maximize the losses.

In the experimental study of small sample machine learning data

processing, the car evaluation data set was selected from the public
data set UCI, with the number of samples being 96. To illustrate the
global optimality of CGAN, we compared it with the BP neural net-
work (BPNN) and particle swarm optimization (PSO), and analyzed
the generalizability of each data processing algorithm through an
AUC curve.
In the experiment, a two-dimensional Gaussian model was used to
process experimental data into unbalanced data. All positive samples
   
6 3 2
were sampled at M ; , and all negative samples were
7 2 4
   
5 4 2
sampled at M ; . The proportion of positive and negative
4 2 3
samples of the generated experimental data was 1:10, with the entire
sample set divided into training and testing sets. We used a Gaussian
distribution under non-degenerate linear transformation to ensure
the unity of data. Beginning with the two-dimensional standard nor-
   
1 1 0
mal distribution M ; , any two-dimensional Gaussian Fig. 9. Comparison results of data classification.
1 0 1
13
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

Fig. 12. Data balance test results of CGAN.

Fig. 10. Data balance test results of BPNN.

the smallest. After calculation, the AUC value of PSO is 0.9152, that of transformed into a linear solution problem to obtain the sample data
the BP neural network is 0.8533, and that of CGAN is 0.9627. In data feature matrix. At the same time, the information entropy and mutual
classification, a higher AUC value denotes a better classification information of each sample data feature matrix were calculated, and
effect. the relative entropy of the target was adjusted according to the calcu-
It can be seen from the results that CGAN exhibits a better classifi- lation results. The experimental results show that the data classifica-
cation and processing effect for small sample unbalanced data. Based tion of CGAN is accurate, the data balance level is high, and overall
on this information, the data processed by the algorithm is displayed generalizability is improved. Therefore, CGAN is suitable for partner
through visualization technology. The comparison results of the data selection in dynamic supply chains.
balance experiment are shown in Figs. 10-12.
As can be seen in the figures, although the negative sample seg-
Conclusion and Future Work
mentation in Fig. 10 is obvious, the positive sample data is distributed
around the negative sample, and the data are cross-fused together
This study developed an overall design of the SCM system, includ-
and obviously unbalanced. The experimental results in Fig. 11 show
ing six modules: order management, purchasing management, pro-
that negative and positive sample data are mixed together, there is
duction management, inventory management, distribution
no obvious boundary, and the data balance is poor. The experimental
management, and transportation management. Then we expounded
results in Fig. 12 show that positive and negative samples exhibit
these six modules in detail, designed the system database, and con-
obvious segmentation, obvious cluster form, and good data balance.
structed an entity ER diagram. To solve the problem of more decision
Overall, the small sample machine learning data classification effect
attributes and less data samples for decision analysis, we proposed a
based on conditional generation countermeasure network designed
dynamic supply chain member selection algorithm based on CGAN.
in this study is good, the data balance processing effect is successful,
Furthermore, we used innovative machine learning to solve the prob-
and the overall generalizability is better than that of the traditional
lems of procurement, inventory, and route planning in the supply
data processing algorithm.
chain. We also employed SSH to integrate the system framework in
In this study, CGAN was used to realize the dynamic selection
the presentation, business layer, data persistence, and domain mod-
method of supply chain partners. First, the sample data was mapped
ule layers. The focus of future research is further optimization and
in Hilbert space, and then the data feature extraction problem was
improvement of certain modules involved in this study. We also aim
to determine how the system can process massive data effectively,
mine more effective resources from existing data, and apply these
resources to the SCM system.

Acknowledgements

This work was supported by Key Research and Development plan

of Jiangsu Province (Grant No.BE2021716), The National Natural Sci-
ence Foundation of China (32101535), Jiangsu Modern Agricultural
machinery equipment and technology demonstration and promotion
project (NJ2021-19) and Jiangsu Postdoctoral Research Foundation
(2021K112B).

References

Alavidoost, M H, Jafarnejad, A, & Babazadeh, H (2021). A Novel Fuzzy Mathematical

Model for an Integrated Supply Chain Planning Using Multi-Objective Evolutionary
Algorithm. Soft Computing, 25(3), 1777–1801 FEB.
Al-Farsi, S, Rathore, M M, & Bakiras, S (2021). Security of Blockchain-Based Supply
Chain Management Systems: Challenges and Opportunities. Applied Sciences-Basel,
Fig. 11. Data balance test results of PSO. 11(12) JUN.

14
H. Lin, J. Lin and F. Wang Journal of Innovation & Knowledge 7 (2022) 100276

Ali, S I, Ali, A, AlKilabi, M, & Christie, M (2021). Optimal Supply Chain Design with Prod- Hu, H, Li, J, Li, X, & Shang, C J (2020). Modeling and Solving a Multi-Period Inventory
uct Family: A Cloud-Based Framework with Real-Time Data Consideration. Com- Fulfilling and Routing Problem for Hazardous Materials. Journal of Systems Science
puters and Operations Research, 126 FEB. and Complexity, 33(3), 760–782 JUN.
Almomani, A, Alauthman, M, Shatnawi, M T, et al. (2022). Phishing Website Detection Huang, L L, & Yang, J Q (2020). An Improved Swarm Intelligence Algorithm for Multi-
with Semantic Features Based on Machine Learning Classifiers: A Comparative Item Joint Ordering Strategy of Cruise Ship Supply. Mathematical Problems in Engi-
Study. International Journal on Semantic Web and Information Systems (IJSWIS), 18 neering, 2020 MAR.
(1), 1–24. Jamrus, T, Wang, H K, & Chien, C F (2020). Dynamic Coordinated Scheduling for Supply
Asamoah, D, Agyei-Owusu, B, Andoh-Baidoo, F K, & Ayaburi, E (2021). Inter-Organi- Chain Under Uncertain Production Time to Empower Smart Production for Indus-
zational Systems Use and Supply Chain Performance: Mediating Role of Supply try 3.5. Computers and Industrial Engineering, 142 APR.
Chain Management Capabilities. International Journal of Information Manage- Karimian, Y, Mirzazadeh, A, Pasandideh, S H, & Namakshenas, M (2020). A Geometric
ment, 58 JUN. Programming Approach for A Vendor Managed Inventory of A Multiretailer Multi-
Badakhshan, E, Humphreys, P, Maguire, L, & McIvor, R (2020). Using Simulation-Based Item EPQ Model. Rairo-Operations Research, 54(5), 1401–1418 JUL.
System Dynamics and Genetic Algorithms to Reduce the Cash Flow Bullwhip in the Khan, P W, Byun, Y C, & Park, N (2020). IoT-Blockchain Enabled Optimized Provenance
Supply Chain. International Journal of Production Research, 58(17), 5253–5279 SEP. System for Food Industry 4.0 Using Advanced Deep Learning. Sensors, 20(10) MAY.
Brdesee, H S, Alsaggaf, W, Aljohani, N, et al. (2022). Predictive Model Using a Machine Liu, D Z, & Li, Z K (2021). Joint Decision-Making of Product Family Configuration and
Learning Approach for Enhancing the Retention Rate of Students At-Risk. Interna- Order Allocation By Coordinating Suppliers Under Disruption Risks. Journal of Engi-
tional Journal on Semantic Web and Information Systems (IJSWIS), 18(1), 1–21. neering Design, 32(5), 213–246 MAY.
Cai, L, Yan, Y C, Tang, Z M, & Liu, A J (2021). Collaborative Distribution Optimization Liu, Y H, Dehghani, E, Jabalameli, M S, Diabat, A, & Lu, C C (2020). A Coordinated Loca-
Model and Algorithm for an Intelligent Supply Chain Based on Green Computing tion-Inventory Problem With Supply Disruptions: A Two-Phase Queuing Theory-
Energy Management. Computing AUG. Optimization Model Approach. Computers and Industrial Engineering, 142 APR.
Carrera, D A, Mayorga, R V, & Peng, W (2020). A Soft Computing Approach for Group Lu, Z G, & Wang, H (2020). An Event-Based Supply Chain Partnership Integration Using
Decision Making: A Supply Chain Management Application. Applied Soft Comput- a Hybrid Particle Swarm Optimization and Ant Colony Optimization Approach.
ing, 91 JUN. Applied Sciences-Basel, 10(1) JAN.
Deng, H X, Wang, M, Hu, Y, Ouyang, J Z, & Li, B R (2021). An Improved Distribution Cost Mousavi, M, Hajiaghaei-Keshteli, M, & Tavakkoli-Moghaddam, R (2020). Two Cali-
Model Considering Various Temperatures and Random Demands: A Case Study of brated Meta-Heuristics to Solve an Integrated Scheduling Problem of Production
Harbin Cold-Chain Logistics. IEEE Access, 9, 105521–105531. and Air Transportation With the Interval Due Date. Soft Computing, 24(21), 16383–
Dosdogru, A T, Ipek, A B, & Gocken, M (2021). A Novel Hybrid Artificial Intelligence- 16411 NOV.
Based Decision Support Framework to Predict Lead Time. International Journal of Nezamoddini, N, Gholami, A, & Aqlan, F (2020). A Risk-Based Optimization Framework
Logistics-Research and Applications, 24(3), 261–279 MAY. for Integrated Supply Chains Using Genetic Algorithm and Artificial Neural Net-
Du, M, Luo, J W, Wang, S P, & Liu, S (2020). Genetic Algorithm Combined with BP Neural works. International Journal of Production Economics, 225 JUL.
Network in Hospital Drug Inventory Management System. Neural Computing and Rahmani, D, Abadi, M Q H, & Hosseininezhad, S J (2020). Joint Decision on Product
Applications, 32(7), 1981–1994 APR. Greenness Strategies and Pricing in A Dual-Channel Supply Chain: A Robust Possi-
Garg, A, Singh, S, Gao, L, Xu, M J, & Tan, C P (2020). Multi-Objective Optimization bilistic Approach. Journal of Cleaner Production, 256 MAY.
Framework of Genetic Programming for Investigation of Bullwhip Effect and Net Saputro, T E, Figueira, G, & Almada-Lobo, B (2021). Integrating Supplier Selection With
Stock Amplification for Three-Stage Supply Chain Systems. International Journal of Inventory Management Under Supply Disruptions. International Journal of Produc-
Bio-Inspired Computation, 16(4), 241–251. tion Research, 59(11), 3304–3322 JUN.
Gaurav, A, Gupta, B B, & Panigrahi, P K (2022). A Comprehensive Survey on Machine Shen, J C, Xu, C J, & Ying, Y (2021). Construction of Intelligent Supply Chain System of
Learning Approaches for Malware Detection in IoT-based Enterprise Information Agricultural Products Based on Big Data. ACTA Agriculturae Scandinavica Section B-
System. Enterprise Information Systems, 1–25. Soil and Plant Science DEC.
Gharaei, A, & Jolai, F (2021). An ERNSGA-III Algorithm for the Production and Distribu- Shen, L X, Li, F C, Li, C C, Wang, Y M, Qian, X Q, Feng, T, & Wang, C (2020). Inventory
tion Planning Problem in the Multiagent Supply Chain. International Transactions Optimization of Fresh Agricultural Products Supply Chain Based on Agricultural
in Operational Research, 28(4), 2139–2168 JUL. Superdocking. Journal of Advanced Transportation, 2020 JAN.
Gholizadeh, H, Tajdin, A, & Javadian, N (2020). A Closed-Loop Supply Chain Robust Srinivasan, M, Hamdani, M, & Ma, S Q (2021). Four Supply Chain Management Systems:
Optimization for Disposable Appliances. Neural Computing and Applications, 32(8), From Supply Chain Strategies to Human Resource Management. Business Horizons,
3967–3985 APR. 64(2), 249–260 Mar.
Ghorashi, S B, Hamedi, M, & Sadeghian, R (2020). Modeling and Optimization of A Reli- Tang Y F, Liu Y J. “Information Management System and Supply Chain Management
able Blood Supply Chain Network in Crisis Considering Blood Compatibility Using (SCM)”, International Conference on Cyber Security Intelligence and Analytics
MOGWO. Neural Computing and Applications, 32(16), 12173–12200 AUG. (CSIA), Cyber Security Intelligence and Analytics, vol. 928, pp.1421-1426, 2020.
Goodarzian, F, Shishebori, D, Nasseri, H, & Dadvar, F (2021). A Bi-Objective Production- Teerasoponpong, S (2022). Decision Support System for Adaptive Sourcing and Inven-
Distribution Problem in A Supply Chain Network Under Grey Flexible Conditions. tory Management in Small- and Medium-sized Enterprises. Sopadang A, Robotics
Rairo-Operations Research, 55, S1287–S1316 MAR. and Computer-Integrated Manufacturing, 73 FEB.
Guo, J Q, Yu, H L, & Gen, M S (2020). Research on Green Closed-Loop Supply Chain with Vegter, D, Hillegersberg, J, & Olthaar, M (2021). Performance Measurement Systems for
the Consideration of Double Subsidy in E-Commerce Environment. Computers and Circular Supply Chain Management: Current State of Development. Sustainability,
Industrial Engineering, 149 NOV. 13(21) NOV.
Han, C L, & Zhang, Q (2021). Optimization of Supply Chain Efficiency Management Wu, W T, Zhou, W, Lin, Y, Xie, Y Q, & Jin, W Z (2021). A Hybrid Metaheuristic Algorithm
Based on Machine Learning and Neural Network. Neural Computing and Applica- for Location Inventory Routing Problem With Time Windows And Fuel Consump-
tions, 33(5), 1419–1433 MAR. tion. Expert Systems with Applications, 166 MAR 15.
Harifi, S, Khalilian, M, Mohammadzadeh, J, & Ebrahimnejad, S (2021). Optimization in Xu, X, Lee, S D, Kim, H S, & You, S S (2021). Management and Optimisation of Chaotic
Solving Inventory Control Problem Using Nature Inspired Emperor Penguins Col- Supply Chain System Using Adaptive Sliding Mode Control Algorithm. International
ony Algorithm. Journal of Intelligent Manufacturing, 32(5), 1361–1375 JUN. Journal of Production Research, 59(9), 2571–2587 MAY.
Hasani, A, Mokhtari, H, & Fattahi, M (2021). A Multi-Objective Optimization Approach Yuan, H Q, Qiu, H B, Bi, Y, Chang, S H, & Lam, A (2020). Analysis of Coordination Mecha-
for Green and Resilient Supply Chain Network Design: A Real-Life Case Study. Jour- nism of Supply Chain Management Information System from The Perspective of
nal of Cleaner Production, 278 JAN. Block Chain. Information Systems and E-Business Management, 18(4), 681–703 DEC.

15