Professional Documents
Culture Documents
and IEEE Cyber, Physical and Social Computing (CPSCom) and IEEE Smart Data (SmartData)
Abstract— Safety storage in large warehouse is an urgent issue transportation and inventory information of storage objects, with
to be addressed by both the local authorities and businesses, the intelligent decision support for incident situation and storage
especially after the Tianjin explosions. This paper proposes an activities.
inventory management system for a warehousing company. The
system integrates RFID technology and a self-Adaptive distributed We also outline a simulation experiment to validate the
decision support model for inbound and outbound actives, utilisations of our model and the proposed system as well. The
inventory location suggestions and incident handling. The model prototype design is also confirmed by a local warehousing
consists of three major components: environment recognition, company that they estimated it could result in a 10% work
knowledge merging and the decision making. In addition, a ‘self- efficiency improvement to their current work procedures.
adaptive’ feature is adopted for adjusting the knowledge used in
decision making procedure. An experiment is also outlined to The rest of this paper is organized as follows: Section 2
validate the utilisations of our model and the proposed system. describes the approach of our model. Section 3 discussed the
design of our proposed system with the model, and then an
Keywords—Distributed Decision Support; Self-adaptive System; experiments with the data from our previous work is presented.
RFID; Inventory Management
II. SELF-ADAPTIVE DISTRIBUTED DECISION SUPPORT MODEL
I. INTRODUCTION The proposed Self-Adaptive Distributed Decision Support
Businesses incur costs to store, track and secure inventory. Model is a more ‘generic’ approach to provide intelligent
Inventories that are mismanaged can create significant financial support for IoT applications. It is designed with three major
problems for a business, where the mismanagement can cause components: environment recognition, knowledge merging and
an inventory glut or shortage, or even worse, results in deadliest the decision making. A ‘self-adaptive’ mechanism is introduced
disasters. for adjusting the knowledge used for decision making, which
enables the model to give a better ‘fit’ to the deployed scenario
Inventory management of raw materials and products is automatically.
often concerned and attracts a lot of scholars' attentions.
However, with the growing public awareness of security, the The history of the ‘decision support system’ concept can be
safety storage is an urgent issue to be addressed by both the local traced back to 1971 [1], it has been further developed over the
authorities and businesses, especially after the Tianjin past decades. In its research domains, the distributed decision
explosions. support system is the one that has received growing attention
from researchers [2-4]. Most of the researches aims to break
Near midnight on the night of 12th August 2015, a down the decision task to several sub-tasks that are carried out
catastrophic warehouse explosion occurred in Tinajin, near by serval decision agents, the final results are provided from the
Bohai Bay, generating mass casualties, including 165 victims negation of the results, either by the agents, or a central unit[3,
with nearly 800 additional people reportedly injured, and over 5]. Another approach of distributed decision support focus on
$1.5 billion in economic losses. The warehouse was a storage the fusion of heterogeneous information from different source
area for containers of hazardous materials, but inventory [6-8], this approach normally assumes that the outside
negligence led to dangerous chemical storage layout, and finally environment is explicit to the agents and the system[9, 10] with
caused a deadly explosion. a static knowledge base.
An achievable solution for the problem is to adopt the However, under the inventory management scenarios, and
concept of Internet of Things (IoT) using RFID technology to the concept of IoT, the system deployed environments are
track the material, and providing messages or warnings when usually dynamic and complex, thus a new approach of
incorrect behaviours are detected. A novel inventory distributed decision support system is required for the proposed
management system with a self-adaptive distributed decision inventory management system.
support model is proposed in this paper. It aims to provide the
The work procedures of our proposed model are illustrated Action2: Training the artificial neural network: The
in Fig. 1. Briefly, the decision support procedure starts from Network structure and training methods in our model is the
environment recognition by using a neural network approach, Back-Propagation neural network that contains one hidden layer
and the recognition results are used to select the knowledge [14]. It can be replaced by other type of neural network if
segments from several pre-set knowledge bases. A knowledge required. Normally, this network is the same as the network used
merging process is then carried out to generate a local in the feature selection process. The output layer of the network
knowledge base from the selected knowledge segments. After is used to build a matrix as the environment recognition result,
the local knowledge base is created, decisions are then able to be each output node represents an ‘aspect’ of the deployment
made by the node’s own reasoning module. Finally, the decision scenario. This approach enables the ‘free selection’ of the
result is evaluated either by the system or by the users. The knowledge elements by using the result matrix as a benchmark.
evaluation result is used in self-adaption process as a reference The scenario recognition procedure is illustrated in Fig. 2.
for adjusting the benchmark of knowledge selection. Once the
adjustment exceeds a certain threshold, the local knowledge
base is re-created. This process aims to improve the fitness of
the knowledge for nodes’ deployed scenario, and thus, we called
this mode is a ‘self-adaptive’ model.
318
than the corresponding element of the recognition matrix, this This action is carried out after the evaluation is received. The
knowledge segment is then selected. step size is according to the evaluation that is in the format of
percentages. The following empirical equation is used for this
3) Step 3: Knowledge merging action:
The selected knowledge elements may contain confliction
and inconsistencies, and thus they cannot be directly used as the
local knowledge base. A knowledge merging process is then = ∗ − ∗
carried out with a symbolic approach [11, 15, 16]. This approach
is firstly trying to find any possible extra condition attributes (i.e.
Where is the number of the knowledge elements that
the antecedent of a rule) of the inconsistent knowledge elements
and to check it would help for removing inconsistence. Or are in this pre-set knowledge and are also selected in the local
secondly, if the first step attempting fails, a weakening process knowledge base; is an empirical parameter to control the step
for the inconsistent knowledge is then carried out. It identifies size, we set it to 0.05 in our application.
possible conflict, inconsistent or duplicated attributes of The second part of self-adaption is to determining the time
knowledge elements with the help of a relationship table of of local knowledge base regeneration. The regeneration can be
attributes values that created in knowledge selection step. A triggered manually by the users when required, or it can be
meta rule set is associated for controlling the merging process. triggered automatically by measure the changed base weight: if
And after the inconsistence removed, the knowledge elements the total changed base weight in percentage is greater than the
set are then used as the local knowledge base of a node. number of improper used knowledge elements in percentage,
then, the regeneration process is started.
4) Decision Making
Unlike other decision support systems, the decision making
is carried out locally by the nodes individually without any help III. THE SMART INVENTORY MANAGEMENT SYSTEM
from the central unit. The node has its own rule-based system In this section, a smart inventory management system using
and use the local knowledge base to determine the decisions. our proposed model is discussed. The system was original
Normally, the system follows event-drive schema, it could be by designed for a UK local storage company that was handling
using a set of meta rule for event confirmation, or just build into cardboard, metal material and oil. The system was designed for
the application, for example, the detection of RFID tag can be a monitoring the state of the storage warehouse, including
start of a decision making process. However, the determining of environment conditions and inventory information. It also
an event occurrence is an application related design, the details provided decision support during the processes of inbound,
are unable to be clearly provided in model design. outbound and storage. However, the decision support module of
5) Step 5: Self-adaption process the system had been carefully ‘localized’ and ‘customized’ in
The ‘self-adaption’ feature of the model is carried out both order to match its application environment. Thus, it has
locally and globally. It consists of two parts, firstly, modifying difficulties to be used in other scenarios, i.e. by other storage
the base weight of the pre-set knowledge bases of a node, companies. In order to overcome this issues, we outlined an
according to the decision evaluations; and secondly, determining ‘upgraded version’ of the system: the smart inventory
management system with a self-adaptive distributed decision
the time of recreating the local knowledge base.
support model.
The first part of self-adaption requires the evaluation of the
decisions, it is an application related task. But normally, the A. Background
result of evaluation indicates the positive or the negative of the This UK local storage company charges their customers for
effect from the decisions, and thus, the base weight can be storage by volume, weight and material. it has a large old-style
modified according to this evaluation. warehouse for normal goods storage, such as cardboards, metal
The base weight modification step contains two actions: material that does not require special environment, and two
firstly, a ‘reward’ base weight is added to all pre-set knowledge smaller warehouses that are equipped with thermometers,
base that help make out the decision, regardless of the decision hygrometers, air conditioner and motion sensors for goods that
correctness. This action is carried out immediately after a require a special environment. Each smaller warehouse has 12
decision is made out by the following equations: room that the environment can be adjusted respectively.
In the previous design, the smart inventory management
= ∗ + ∗ ⁄ system was mainly developed for tracking purpose by using
RFID equipment either fixed RFID readers or RFID handheld
devices that generated location information of objects. Other
Where is the percentage of the knowledge types of information were inputted by operating staffs, such as
elements that are not selected into the local knowledge base. the type and volume of goods. The pervious design was a
is the number of knowledge elements in the local centralized decision support system, and there is a central server
located in the company headquarter for processing data, making
knowledge base, and is the total number of knowledge decisions and sending back decision results to related devices
elements of the system. that are carried out by operating staff.
Secondly, a ‘reward’ or ‘punishment’ adjusting is applied to
the base weight that regarding the correctness of the decision.
319
The system provided decision support for the inventory devices that carried out by operating staff and 10 fixed Reader
actives in three aspects: inbound and outbound actives, with 40 antennas that deployed in the warehouses. All ‘objects’
inventory location suggestion and incident handling. It aimed to in the system had been labelled with passive RFID tags. The
relief the two members of staff from the increasing workloads ‘objects’ included vehicles, goods and operating staffs that with
who are engaged in arranging the inventory location manually, RFID enabled pass cards.
and to provide solutions for operating staff in unexpected
situation. 1) Pre-set knowledge base
The knowledge used in the application was collected by an
However, the design of the previous work has a weakness: interview with 2 members of management staff and 3 members
the centralized structure requires the decisions have to be made of operating staff. Four of operating handbooks were also
after all related information is received. It could result in examined with the support of the management staff. Finally, 179
incorrect solution and misjudgement of events, due to the raw knowledge elements were collected.
absence of importation data or information.
In addition, the system has difficulties to be used by other A manual procedure was then carried out in order to create
similar companies, it had been carefully ‘localized’ and the pre-set knowledge bases. The raw knowledge elements are
‘customized’ for the company; the knowledge used in decision grouped into 6 pre-set knowledge bases according to their
support module was in a single knowledge base, and the application domains (scenario). During this step, inconsistences,
conflictions and inconsistence were removed manually. Thus, in duplications and conflictions of knowledge elements in the pre-
order to be used for other companies, the decision support set knowledge bases are removed.
module must be re-designed, as well as the knowledge base.
A Boolean matrix was assigned to each rule, and it consisted
Consequently, a new model is developed in order to of 6 elements that indicate the fitness of a rule to a specific
overcome these limitations. It is a more ‘generic’ approach to scenario. In our proposed system, the matrix was used as filter
provide intelligent support for the system with a distributed for knowledge selection from the pre-set knowledge base, and a
schema. The decision procedures are carried out locally at the rule normally had 2-3 of possible application fields.
‘node’ level, rather than in a central decision unit. The nodes of
the model represent the function units of the system, such as In addition, each pre-set knowledge base was assigned a
vehicles, customer sites or a recycling centres etc. Thus, the ‘base weight’ to enable the self-adaption feature. The weights
system is able to make correct decisions without whole system were used in rule selection step and initially set to 0.5. This
information and within a limited amount of time. In addition, weight was changed during the decision evaluation process
with the self-adaption feature, the nodes are able to generate according to the user evaluation, or a fitness function that for
more suitable knowledge base automatically for their deployed minimum the storage cost in terms of total routing length and
environment. actual volume occupied.
2) Scenario Recognition
B. System design The scenario recognition result matrix is generated by a pre-
The layout of the design is briefly illustrated in Fig. 3. Each trained network and describes the current scenario by the beliefs
node has a pre-trained neural network as a scenario recognition level of 6 elements in this result matrix. The deployed
module, and a knowledge-based system as a decision making environment was not clearly classified into a certain category,
module. Thus, a node is able to carry out decision making due to the dynamic and complexity of real world. The
procedure individually. recognition result was used as guidance for selecting knowledge
from pre-set knowledge base, in order to ‘fit’ the real scenario
as much as possible.
In order to train the neural network for scenario recognition,
40 hours of raw data were collected. The data is treated as time
series data and then to be divided into small segments, and the
Sliding Window technology is used. After examination of the
raw data, a frame length of 30 seconds and a windows size of 1
hour is decided in consideration of the sample density.
The possible features used in environment recognition are
computed from the statistical data about objects that were
detected in a window. The raw RFID data such as signal strength
Fig. 3. The system architecture
and reading time are also included. These features are calculated
The nodes communicate to the server with a program that from both time and frequency domains.
running on hand devices or local computers that were deployed The criteria of the accuracy of the network is that the
nearby sensors. A node normally consisted of one or more scenario of the node deployed in the result matrix has the belief
sensors and a computer. The sensor technology used in the that exceeds 0.5. The overall aim of the scenario recognition is
proposed system are RFID technology, motion sensors, to generate the most suitable local knowledge base for nodes.
thermometers and hygrometers. RFID equipment used in this The accuracy is not so important when compared to other pattern
system was from Alien Technology, includes 5 handheld recognition application.
320
TABLE I. SELECTED FEATURES adaption process is to modify the base weight according to the
evaluations, and consequently, in knowledge selection step,
more appropriate knowledge elements could be selected, and
thus, the decision made by the updated knowledge local are
more ‘adaption’ to the scenario. In this proposed system, the
decision receives ‘reward’ if the evaluation is greater than 50%,
otherwise it will be a ‘punishment’. The details of the self-
adaption process are described in Section 2.
321
provided by the company after a carefully examination of our [3] O. López-Ortega, and M.-A. Rosales, “An agent-
design. oriented decision support system combining fuzzy
They confirmed the design has the following advantages: clustering and the AHP,” Expert Systems with
firstly, the tracking feature is the most important feature. Applications, vol. 38, no. 7, pp. 8275-8284, 2011.
Secondly, the instruction and incident handling feature are [4] B. Ramesh, and K. Mohan, "Integrating group
useful for their new employee. However, they pointed out there decision and negotiation support systems with work
is cost issue about the investment of the layout, the handle device processes." p. 9 pp.
is expensive and they prefer using square code with Android [5] J. Qin, and X. Liu, “Multi-attribute group decision
phone or cheaper RFID reader. Finally, they confirmed that the making using combined ranking value under interval
system would improve at least 10% their work efficiency, in type-2 fuzzy environment,” Information Sciences, vol.
terms of storage process and incident handling time. 297, no. 0, pp. 293-315, 2015.
[6] D. R. Ball, A. Deshmukh, and N. Kapadia, “An
V. CONCLUSION options-based approach to coordinating distributed
decision systems,” European Journal of Operational
In this paper, a novel inventory management system
Research, vol. 240, no. 3, pp. 706-717, 2015.
designed for storage company is described. It uses the RFID
[7] T. Bui, and J. Lee, “An agent-based framework for
technology for tracking the movement of goods, and a proposed
self-adaptive distributed decision support model is introduced to building decision support systems,” Decision Support
enable the system automatically ‘fit’ to its deployed Systems, vol. 25, no. 3, pp. 225-237, 1999.
environment by adjusting of the nodes’ knowledge base. [8] M. V. Sokolova, and A. Fernández-Caballero,
“Modeling and implementing an agent-based
The proposed model provides a more generic approach for environmental health impact decision support system,”
decision support in inventory management. It uses a distributed Expert Systems with Applications, vol. 36, no. 2, Part
schema and all decision makings are carried out by the nodes 2, pp. 2603-2614, 2009.
individually to avoid any possible delay caused by network [9] Ç. Elmas, and Y. Sönmez, “A data fusion framework
communications. The model uses the result of scenario
with novel hybrid algorithm for multi-agent Decision
recognition as a benchmark for knowledge selection to create a
Support System for Forest Fire,” Expert Systems with
local knowledge base that used by node individually to carry out
the decisions with its own rule-based system. In addition, a self- Applications, vol. 38, no. 8, pp. 9225-9236, 2011.
adaption step is introduced to further modify the local [10] D. Fogli, and G. Guida, “Knowledge-centered design
knowledge base for better adaption to the scenario where the of decision support systems for emergency
node deployed. management,” Decision Support Systems, vol. 55, no.
1, pp. 336-347, 2013.
A simulation experiment is then carried out in this paper, in [11] S. Konieczny, and R. P. Perez, “Propositional belief
order to prove the utilization of the proposed inventory base merging or how to merge beliefs/goals coming
management system, as well as the self-adaptive model. This from several sources and some links with social choice
design is also confirmed by another chemical storage company, theory,” European Journal of Operational Research,
and they estimated the design could bring 10% work efficiency
vol. 160, no. 3, pp. 785-802, Feb, 2005.
improvement to their current work procedures.
[12] I. Kononenko, and M. Kukar, Machine learning and
data mining : introduction to principles and
ACKNOWLEDGMENT algorithms, Chichester: Horwood Publishing, 2007.
This work was supported by the Applied Basic Research [13] W. Wang, P. Jones, and D. Partridge, “Assessing the
Programs of Science and Technology Department of Sichuan impact of input features in a feedforward neural
Province of P.R. China [grant numbers 2015JY0228]; the network,” Neural Computing & Applications, vol. 9,
Project Sponsored by the Scientific Research Foundation for the no. 2, pp. 101-112, 2000, 2000.
Returned Overseas Chinese Scholars, State Education Ministry; [14] E. Alpaydin, Introduction to machine learning, Third
and the Fundamental Research Funds for the Central edition. ed., Cambridge, Massachusetts: The MIT
Universities [grant numbers ZYGX2015J063]. Press, 2014.
[15] R. Q. Lu, and Z. Jin, “Formal ontology: Foundation of
domain knowledge sharing and reusing,” Journal of
REFERENCES Computer Science and Technology, vol. 17, no. 5, pp.
[1] S. Morton.M.S, Management decision systems, 535-548, Sep, 2002.
Division of Research, Graduate School of Business [16] J. Masters, and I. Isif, Structured Knowledge Source
Administration, Harvard University, 1971. Integration and its applications to information fusion,
[2] C. W. Holsapple, “Decision Support Multi-Participant 2002.
Decision Making,” Journal of Computer Information
System, pp. 37-45, 1991.
322