Abstract—This study considers a wireless sensor network (WSN) is designed to track specified objects of interest such as bird-calls, insect-images, and so forth. An assumption is made that the sensors in the WSN are capable of analyzing and identifying detected objects and are pre-loaded with the features of the tracked objects before they are deployed. The features associated with the tracked objects are referred to as “model tuples”. When a sensor subsequently detects an object, it extract features from the detected object and then compares it with the tuples stored in its memory in order to determine whether or not the detected object is the tracked object. Since the sensors have only limited memory and storage space, it is impossible to store all the tuples on a single sensor. Furthermore, the sensors are battery operated, and thus the stored tuples are irretrievably lost once the sensor’s energy resources have been consumed. As a result, the network no longer has a complete knowledge of all the tracked information. Accordingly, the present study proposes four tuple dispatching schemes for distributing the tracked information amongst the sensors in such a way as to mitigate the effects of sensor energy depletion, namely Sequential Dispatching, Sequential Dispatching with Overlap, Fixed Distance Dispatching, and Balanced Incomplete Block Dispatching. In addition, an efficient diversity-driven selective forwarding scheme is proposed to resolve the problem where the detected object fails to match the tuples held at the local sensor. In the approach, the local sensor applies the correlation between the sensor identifier and the indexes of the tuples stored at the various sensors to deliver the feature of the object along the paths with the highest diversity. The simulation presents a series of experimental results to benchmark the performance of the proposed forwarding approach for each of the dispatching schemes against that of a blind flooding approach. Index Terms—wireless sensor networks, sequential dispatching, sequential dispatching with overlap, fixed distance dispatching, balanced incomplete block Dispatching, diversity-driven selective forwarding
1. INTRODUCTION
IRELESS sensor networks (WSNs) [5-7],[22] are specialized ad hoc networks composed of numerous low-cost and limited-power sensor nodes designed to work cooperatively in order to monitor a wide range of physical or
environmental conditions, e.g. temperature, sound, vibration, pressure, motion, pollutants, and so forth. Generally speaking, the nodes in such networks either collect all of the specified information of interest and forward it to a centralized sink for further processing, or transmit only the sensed object required to satisfy a specific user query. Depending on the application, the queries submitted to a WSN can be broadly classified as follows [12]:
Continuous versus one-shot queries: Continuous queries refer to queries which result in an extended data flow (e.g. “Report the measured temperature for the next 3 days with a frequency of 1 measurement per hour”). By contrast, one-shot queries are queries which have a simple response (e.g. “Is the current temperature higher than 30
C?”).
Aggregate versus non-aggregate queries: Aggregate queries are queries which require the aggregation of information from multiple sources (e.g. “Report the average temperature of all nodes in region I). Conversely, non-aggregate queries are queries which can be answered by a single node (e.g. “What is the temperature at node i?”).
Complicated versus simple queries: Complicated queries are queries consisting of multiple sub-queries combined by conjunctions or disjunctions in an arbitrary manner (e.g. “What are the values of the following variables: X, Y, and Z?” or “What is the value of (X OR Y) AND Z?”). By contrast, simple queries are queries containing no sub-queries (e.g. “What is the value of variable Z”?).
Queries for replicated data versus queries for unique data: Queries for replicated data are queries in which the response to the query can be provided by multiple nodes (e.g. “Has the target of interest been observed anywhere in the area?”). Conversely, queries for unique data are queries for which the response can only be provided by one node (e.g. “Has the target of interest been observed at node i?”).
The present study considers the sensors deployed in the surveillance area track the specified information of interest in which the goal is to respond to queries for replicated data. An assumption is made that the sensor nodes in the WSN are capable of analyzing and identifying detected objects and are pre-loaded with the features of the objects which are to be tracked before they are deployed. For convenience, this study refers to the features associated with the tracked objects as
An Efficient Diversity-Driven Selective Forwarding Approach for
Replicated Data Queries in Wireless Sensor Networks
Chih-Hung Chao, I-Hui Li,
Chong-Yi Yang, and Jung-Shian Li
W
————————————————
• Chih-Hung Chao is with the Library and Information Center, National University of Kaohsiung, Kaohsiung City 811, Taiwan. E-mail: wilson@nuk.edu.tw.
• I-Hui Li is with Department of Information Networking and System Administration, Ling Tung University, Taiwan. E-mail:
sanity@mail.ltu.edu.tw
• Chong-Yi Yang is with the department of information security unit, Data Communication Business Group, Chunghwa Telecom Company, Taipei City 100, Taiwan. E-mail: cyyang@cht.com.tw.
• Jung-Shian Li is with the Department of Electrical Engineering, Institute of Computer and Communication Engineering, National Cheng Kung University, Tainan City 701, Taiwan. E-mail: jsli@mail.ncku.edu.tw.
“model tuples”. When a node subsequently senses an object, it extract features from the detected object and then compares the data with the model tuples stored in its memory in order to determine whether or not the detected object is the target which is to be observed. A typical example of applications of this type is that of a WSN deployed in a wildlife reserve in order to monitor the activities of a particular bird species (e.g. a Bluethroat). In this case, each sensor node is pre-loaded with the model tuples of all the bird-calls heard previously in the reserve (e.g. Formosan Blue Magpies, Common Kestrels, Eurasian Hobbies, and Amur Falcons), and a typical task would be: “Obtain six locations where a Bluethroat’s call has been recorded”.
However, WSN sensors have only limited memory and storage space, and therefore it is impossible to store all of the model tuples required for monitoring and tracking purposes on a single node. Furthermore, WSN sensors are battery operated, and thus all of the information in the sensor memory is lost once the sensor’s energy resources are consumed. As a result, the tracked information stored within the network is incomplete unless the model tuples are stored on multiple nodes. Consequently, effective strategies are required for dispatching the model tuples to all the sensor nodes in such a way as to mitigate the effects of sensor energy depletion. To satisfy this requirement, the present study proposes four tuple dispatching schemes, namely Sequential Dispatching (SD), Sequential Dispatching with Overlap (SDO), Fixed Distance Dispatching (FDD), and Balanced Incomplete Block Dispatching (BIBD).
If the data fails to match any of the model tuples in the sensor memory, it can not be certain whether or not the detected object is the object which is to be monitored or tracked (i.e. the target object). For convenience, this study refers to the data associated with this object as “unknown data”. In theory, the sensor node could simply forward the unknown data to all the other nodes in the network in order to compare it with the information held at each node. However, while such a blind flooding approach guarantees that an object can be definitively classified as either “target object” or “non-target object”, it is highly wasteful of the network’s limited resources. Accordingly, the present study proposes an alternative approach in which the sensor nodes use the correlation between the node identifier and the indexes of the model tuples assigned by the dispatching scheme to forward the unknown data over the paths in the network with the highest diversity. Namely, the collection set of model tuples stored at sensor nodes along the path are the most distinct. The proposed approach not only reduces the total power consumption of the query resolution process, but also improves the classification success rate, i.e. the average success rate in matching the unknown data with the distributed model tuples in the network. The remainder of this paper is organized as follows. Section 2 presents a brief overview of the traditional centralized and distributed data architectures of WSNs, while Section 3 formulates the problem considered in this study and describes the basic assumptions. Section 4 introduces the four tuple dispatching schemes. Section 5 describes the diversity-driven
selective forwarding approach to deal with the propagation the unknown data through the WSN. Section 6 presents the simulation results benchmark the performance of the proposed forwarding approach for each of the dispatching schemes against that of a blind flooding approach. Finally, Section 7 presents some brief concluding remarks.
2. RELATED WORK
The literature contains many query processing models for WSN applications. In early models [6], [24], the query was distributed to every sensor in the network using a blind flooding method, and the relevant collected data was then returned to the sink via the reverse path of the query message. However, such models take no account of the storage constraints and communication costs within the WSN. Accordingly, later models [18-20] attempted to improve the efficiency of the query process by using data-centric storage schemes based on a distributed hash table to store the collected data at different sensor nodes in accordance with the characteristics of the data. Thus, queries relating to a particular type of data can be addressed to a small number of selected nodes, thereby yielding a significant reduction in the communication overhead.
The sensors in a WSN are battery operated and therefore have finite energy resources. As a result, the problem of maximizing the energy utilization efficiency within WSNs has attracted significant attention in the literature. Many previous studies have focused on the so-called in-network processing problem, that is, the pushing of operations, particularly selections and aggregations, into the network in order to reduce communication costs. For example, Bonnet et al. [15] proposed an enhanced centralized architecture designated as Cougar, in which the stored data were represented as relations and the sensed data were represented as time series. The experimental results showed that the proposed system was both flexible and efficient, and made full use of the computing resources available at each sensor node.
Madden et al. [21] developed a distributed query processor for data collection in sensor networks known as TinyDB. The system not only retained many of the features of a traditional query processor, but also incorporated a number of other features in order to minimize the power consumption of the query response process. It was shown that by focusing on the location and costs of the acquired data, TinyDB yielded a significant reduction in the power consumption within the network compared to that of traditional passive query systems.
Sadagopan et al. [12] proposed an efficient mechanism for obtaining information in sensor networks referred to as ACQUIRE. In the proposed approach, active query packets were injected into the network and followed a random (or possibly pre-determined or guided) trajectory through the network. Upon receiving the active query, each node along the path performed a triggered, on-demand update to obtain information from all its neighbors within a look-ahead distance of d hops, where the value of d was set in such a way as to achieve a tradeoff between the amount of information obtained
(thereby reducing the length of the overall trajectory) and the cost incurred in collecting the information. As the active query propagated through the network, it was progressively resolved into smaller and smaller components until it was completely resolved, at which point it was returned to the querying node in the form of a completed response.
Although ACQUIRE is well-suited to answering one-shot, complex queries for replicated data, it assumes that all of the data tracked by the network is stored at each sensor node. In practice, however, the sensor nodes in a WSN have only limited storage space, and thus it is impossible to store all of the tracked data at each node. In other words, it is necessary to distribute the data amongst all the nodes in the network. As a result, in monitoring and tracking applications, the situation can arise in which the data of an object detected at a particular node can not be matched with the information stored at this node. In this case, the unknown data must be flooded to the other nodes in the network in order to determine whether or not the sensed object is the target object, i.e. the object which is to be tracked by the network. Although using a blind flooding approach can guarantee that an object can be definitively classified as either “target object” or “non-target object”, it is highly wasteful of the network’s limited resources.
3. PROBLEM FORMULATION AND ASSUMPTIONS Consider a WSN with Ns sensor nodes in which each node is
uniquely identified by a node ID ni, where0≤ ≤i Ns−1.
Assume that the network is designed to track specific objects such as bird-calls, insect-images, and so forth. Furthermore, let
0 1 1
{ , , , S }
T = T T …T − be the set of features of S objects (or model
tuples) tracked by the network. Since the sensor nodes have only a limited storage capability, it is impossible to store all of the model tuples at each node. Furthermore, since the sensors are battery operated, all of the tuples stored at any particular node will be lost once the sensor has consumed all of its energy. In other words, some of the tuple information will be lost from the network unless duplicates of this information are held at multiple nodes. Consequently, effective strategies are required for dispatching the model tuples to the various network nodes in such a way as to mitigate the effects of sensor energy depletion. In the WSN monitoring and tracking applications in this study, when a node senses an object, it extract features from the detected object and then compares the data with the tuples held in its memory in order to determine whether or not the detected object is the target object. If the data fails to match any of the tuples in the sensor memory, the data (i.e. the unknown data) should be forwarded to the other nodes in the network in order to obtain a classification decision (i.e. target or non-target). Although the unknown data can be propagated through the network using a blind flooding method, this incurs an excessive communication overhead. As a result, a more efficient flooding strategy to minimize the communication overhead is required whilst simultaneously enhancing the classification success rate.
In developing the tuple dispatching schemes and an efficient method for propagating the unknown data in this study, the following assumptions are made:
1. The WSN has no pre-existing routing protocol. 2. The sensor nodes have no positioning capability (e.g.
GPS).
3. The sensor nodes are randomly deployed and are immobile.
4. Each sensor node has an identical processing and storage capacity.
5. Each sensor node has a limited buffer size and a fixed transmission radius.
6. Each sensor is capable of analyzing and identifying detected objects.
4. TUPLE DISPATCHING SCHEMES
This section develops four schemes for dispatching the model tuples of interest in a WSN to the individual sensor nodes, namely the Sequential Dispatching (SD) scheme, the Sequential Dispatching with Overlap (SDO) scheme, the Fixed Distance Dispatching (FDD) scheme, and the Balanced Incomplete Block Dispatching (BIBD) scheme.
Table 1 summarizes the notations used in the proposed tuple dispatching schemes. The minimum number of sensor nodes required to store a single copy of all the tracked tuples in the network is equal to S N/ e , where S and N are the total e
number of model tuples to be stored and the buffer size required to store one tuple at each sensor, respectively. Parameter N s
(i.e. the number of sensors to be deployed) is also dependent on parameter N (i.e. the number of duplicated copies required of d
each model tuple). In other words, if N duplicated copies of d
the original tuple set are required, the total number of nodes to be deployed is equal to N =s N S Nd / e .
The SD scheme is based on the simple round-robin approach commonly used for scheduling and data placement purposes in Database Management Systems (DBMSs) [14]. The SDO scheme is an extended version of the SD scheme in which each
TABLE 1
Notations used in tuple dispatching schemes
Notation Description
S Number of model tuples
d
N Number of duplicated copies of the tuple set
s
N Number of sensors to be deployed
e N
Number of tuples that can be stored at each sensor
pair of nodes with sequential IDs shares a common model tuple. In the SD and SDO schemes, the indexes of the model tuples stored in the same sensor node are sequential. However, in the FDD scheme, the indexes of the model tuples stored in the same sensor node are spaced at a constant distance. Finally, in the BIBD scheme, the indexes of the model tuples stored in the same sensor node are spaced at a variable distance, i.e. the model are dispatched in accordance with a symmetric balanced incomplete block allocation method based on an initial cyclic difference set.
4.1 Sequential Dispatching (SD) scheme
The SD scheme sequentially dispatches the model tuples to the sensor nodes using a simple round-robin approach. Let the initial set of tuple indexes be represented by {S S S0, 1, 2,…,Sj},
where 0≤ ≤ −j S 1and S has an integer value. The number of j
elements in the initial set is equal toN and the index e S j
satisfies Eq. (1). Let the tuple set d represent all the model i
tuples stored at node ni and satisfy Eq. (2),
wherei=0,1,…,Ns−1. 1 1 j j S =S − + (1) 0 ( )(mod ) ( )(mod ) { , , } e s j e s i S i N N S i N N d = T + × …T + × (2)
Fig. 1 presents an illustrative example of the SD dispatching scheme. Assume that the original model set comprises seven tuples (i.e. S=7), namely { , ,T T0 1…,T6}, and that each sensor node has sufficient memory space to hold just three tuples (i.e.
Ne=3). If three duplicates of each tuple are to be dispatched to
the sensor nodes (i.e. Nd=3), a minimum of seven sensor nodes
(i.e. Ns=7) are required to store all of the model tuples. In
utilizing the SD scheme to dispatch the model set, the initial set of tuple indexes is given as {0,1, 2} . Therefore, the tuple set d 0
is equivalent to { , ,T T T0 1 2} . That is, model tuples
0, , and 1 2
T T T are stored at node n0. Similarly, tuples
3, 4, and 5
T T T are stored at node n1, model tuples
0, , and 1 6
T T T are stored in node n2, and so forth. 4.2 Sequential Dispatching with Overlap (SDO) scheme
The SDO scheme is similar to the SD scheme above other than the fact that every pair of nodes with sequential node IDs is assigned a common model tuple. Let the initial set of tuple indexes be denoted by {S S S0, 1, 2,…,Sj} , where 0≤ ≤ −j S 1and S has an integer value. The initial set j
comprises N elements, where each element (i.e. index) e
satisfies Eq. (1). In addition, the tuple set d stored at node ni i
satisfies Eq. (3), wherei=0,1,…,Ns−1. 0 ( ( 1))(mod ) ( ( 1))(mod ) { ,.., } e s j e s i S i N N S i N N d = T + × − T + × − (3) The SDO scheme is illustrated schematically in Fig. 2. Assume that there are seven model tuples to be dispatched and that three duplicate copies of each tuple are required (i.e. Nd=3).
Furthermore, assume that each sensor node can only store three tuples in its buffer (i.e. Ne=3). Consequently, a minimum of
seven sensor nodes are required (i.e. Ns=7). Let the initial tuple
set be {0,1, 2} . Therefore, model tuples T0, , and T1 T are 2
stored at node n0, model tuples T2, T3, and T are stored at 4
node n1, model tuples T4, T5, and T are stored at node n6 2, and so forth.
4.3 Fixed Distance Dispatching (FDD) scheme
In the SD and SDO schemes, the indexes of the model tuples dispatched to the same sensor node are sequential. By contrast, in the FDD scheme, the tuple indexes assigned to each node are separated by a constant distance. Let the initial set of tuple indexes be denoted as {S S S0, 1, 2,…,Sj}, where 0≤ ≤ −j S 1. The number of elements in the initial set is equal to N and e
each index S satisfies Eq. (4). Furthermore, the set of model j
tuples stored at node ni is denoted as d and satisfies Eq. (5), i
where i=0,1,…,Ns−1.
1 distance
j j
S =S − + (4) Fig. 2. Illustrative example of model dispatch using SDO scheme Fig. 1. Illustrative example of model dispatch using SD scheme
/ , if / 0 , where distance 1, if / 0 e e e S N S N S N > = = 0 ( )(mod ) ( )(mod ) { , , } s j s i S i N S i N d = T + …T + (5)
Figure 3 presents a schematic illustration of the FDD scheme. Note that the assumptions regarding the total number of model tuples to be dispatched and the buffer size at each sensor node are identical to those described for the SDO scheme. Thus, a minimum of seven sensor nodes (i.e. Ns=7) are again required.
From Eq. (4), the distance between the indexes of the tuples stored at each node is equal to 2 (i.e. 7 / 3 ). Therefore, the initial set is given by{0, 2, 4} . In other words, the tuple set d is 0
equal to { ,T T T0 2, 4} . Consequently, model tuples
0, 2, and 4
T T T are stored at node n0, model tuples
1, 3, and 5
T T T are stored at node n1, model tuples
2, 4, and 6
T T T are stored at node n2, and so on.
4.4 Balanced Incomplete Block Dispatching (BIBD) scheme In the BIBD scheme, the tuples are dispatched to the sensor nodes using a symmetric balanced incomplete block allocation scheme, and thus the distance between the tuple indexes at the same node varies. In general, the balanced incomplete block (BIB) allocation scheme [2], [23] partitions v treatments belonging to a set Ω ={1, 2,..., }v into b blocks (subsets of Ω) in such a way that
(i) Each block contains (k <v) distinct treatments; (ii) Each treatment appears in r blocks,
(iii) Every pair of treatments appears together in λ blocks. Here, the integers v b r k, , , , and λ are the BIB allocation parameters and are related as follows:
vr=bk (6) (v 1) r k( 1)
λ − = − (7) For a BIB allocation process with parameters v=b, r=k, λ, any two blocks have exactly λ treatments in common, and thus the solution is referred to as a symmetric BIB (v,k,λ) allocation. If λ=1, the solution sets which satisfy Eqs. (6) and (7) are known as cyclic difference sets and parameters v and k satisfy
Eqs. (8) and (9). Table 2 illustrates the initial block sets for the first six (v,k,1) designs.
2 1
v=n + +n (8) , where n is a prime number or a power of a prime number k=n+1 (9)
In the BIBD scheme proposed in this study, the model duples are regarded as treatments and the sensor nodes are regarded as blocks. Let the initial cyclic difference set be defined as
0 1 2
{S S S, , ,…,Sj} for 0≤ ≤ −j S 1, where
j
S is an integer
and referred to as an index of the tuple in the model set. The number of elements in the initial set is equal toN . Furthermore, e
the set of model tuples stored at node ni is denoted as d and i
satisfies Eq. (10), where i=0,1,…,Ns−1.
0 ( )(mod ) ( )(mod ) { , , } s j s i S i N S i N d = T + …T + (10)
The BIBD scheme is illustrated schematically in Fig. 4. The same assumptions as those made for the SDO and FDD schemes TABLE 2
Correlation between prime numbers, treatments and initial cyclic difference sets
n v Initial cyclic difference set 2 7 {0,1,3} 3 13 {0,1,3,9} 2 2 21 {0,1,4,14,16} 5 31 {0,1,3,8,12,18} 7 57 {0,1,3,13,32,36,43,52} 3 2 73 {0,1,3,7,15,31,36,54,63}
Fig. 3. Illustrative example of model dispatch using FDD scheme
described above are also applied here, and thus a minimum of seven sensor nodes (i.e. Ns=7) are required. Let the initial cyclic
difference set be defined as {0,1, 3} . Thus, tuple set d is equal 0
to { , ,T T T0 1 3}, tuple set d is equal to 1 { ,T T T1 2, 4}, tuple set d 2
is equal to { ,T T T2 3, 5}, and so on.
4.5 Analyzing the union set between two senor nodes To estimate the distribution characteristics of the proposed dispatching schemes, we use the mentioned above example to calculate the expected and variance value of the tuple union size between two senor nodes. The tuple union size between two sensor nodes is the number of all distinct model tuples stored in two sensor nodes and we assume that X is a discrete random variable to represent the tuple union size.
In the SD dispatching scheme, there are 7 cases for the union size of 6, 7 cases for the union size of 5, and 7 cases for the union size of 4. Hence, the expected value µSD and the variance value σ2SD of the union size for the SD scheme are
calculated as follows: 7 7 7 2 2 2 2 2 2 2 7 7 7 2 2 2 ( ) ( 6) * ( 6) ( 5) * ( 5) ( 4) * ( 4) 7 7 7 = * 6 *5 * 4 5 7 7 7 2 ( ) * (6 5) * (5 5) * (4 5) 3 SD SD X P X X P X X P X X C C C X C C C µ σ = = = + = = + = = + + = = − + − + − =
For the SDO and FDD dispatching scheme, there are also 7 cases for the union size of 6, 7 cases for the union size of 5, and 7 cases for the union size of 4. Thus, the expected values µSDO and µFDD and the variance values σ2SDO and σ2FDDare the same as the values of the SD dispatching scheme. Nevertheless, since the BIBD dispatching scheme is based on the cyclic difference sets as described in subsection 4.4, the union size of the scheme is a constant and equal to 2*N -1. Thus, in the e
example, the expected value µBIBDand the variance value
2 BIBD
σ are equivalent to 5 and 0, respectively.
From the above simple analyses, the BIBD dispatching scheme has more uniformity characteristics than other dispatching schemes.
5. DIVERSITY-DRIVEN SELECTIVE FORWARDING SEARCH APPROACH
The four tuple dispatching schemes described in the previous section yield a correlation between the node identifiers and the indexes of the corresponding model tuples. In this section, this correlation is utilized to develop an efficient diversity-driven forwarding method for the propagation of unknown data through the WSN. In the proposed approach, the diversities of the propagation paths in the network are calculated by collecting the two-hop node identifiers associated with each node. The node identifiers are collected using a simple node identifier discovery protocol based on the following messages:
HELLO message
NEIGHBOR_LIST_REQUEST message
NEIGHBOR_LIST_REPLY message
Once the sensor nodes have been deployed in the surveillance region, they broadcast HELLO messages to all the nodes within their transmission range. Each sensor node receiving the HELLO message updates its list of neighboring nodes accordingly. The sensor nodes then flood a NEIGHBOR_LIST_REQUEST message with a time-to-live (TTL) parameter setting of 1 to all their neighboring nodes. When a node receives this message, it responds with a NEIGHBOR_LIST_REPLY message which contains both its own ID and those of all its neighboring nodes. The source node then uses this information, together with that received from all its other neighboring nodes, to construct a complete list of node identifiers for all its two-hop neighbors.
Once a sensor node has compiled its two-hop neighboring node ID list, it uses the corresponding knowledge of the tuple dispatching scheme to construct one bit map to record the model tuples stored at its one-hop and two-hop neighbors. In the bit map, a value of 1 in the k-th bit indicates the existence of the k-th model tuple, while a value of 0 indicates the nonexistence of the k-th model tuple. Let the notation D denote a bit map ij
calculated by node ni and showing the model tuples stored at
node nj and node nj’s neighbors except for node ni. Furthermore,
let D be the bit map calculated by node nii i showing the model
tuples stored within its own memory. Node ni can then calculate
the model tuple diversity of neighboring node nj using
conventional bitwise OR and exclusive OR operations. Let the bitwise OR operator be represented by “|” and the bitwise exclusive OR operator be represented by⊕. The diversity (i,j), i.e. a quantity of bits summation for a bit map, is then defined as shown in Eq. (11), where the bit addition operation is denoted by “ ∑ ”.
( , ) (( ii| ij) ii)
Diversity i j ≡∑ D D ⊕D (11) Fig. 5. Model tuples dispatched to each sensor buffer by BIBD scheme
In the diversity-driven selective forwarding method proposed in this study, a sensor with unknown data sends this information to all its neighboring nodes with TTL=1 except for the top-N diversity neighboring nodes with TTL≥2. In other words, the forwarding process uses the TTL and top=N diversity parameter to limit the total number of hops and the search scope over which the query passes. Whenever a sensor node receives the query (i.e. the unknown data), it decrements the TTL value by one. If the act of decrementing the TTL parameter reduces the TTL parameter to zero, the sensor node takes no further action. However, if a sensor node receives a query with TTL>2, it decrements the TTL parameter and then broadcast the query to all its neighboring nodes with TTL=1 except for its top-N diversity neighbors with the decremented TTL value. Finally, sensor nodes receiving a query with TTL=2, decrement the TTL value and then broadcast the query to all their immediate neighbors. In addition, whenever the query is matched with the information stored at one node, the node will no longer forward the query to its neighbors irrespective of the TTL value.
The forwarding process uses a sequence-number-controller (SNC) to prevent the sensor nodes within the network from delivering the same unknown data. Specifically, the source sensor node enters its ID and a search sequence number into the query for the unknown data, and then sends this query to its neighboring nodes. Each sensor node maintains a list of all the source node IDs and sequence numbers it has previously received and forwarded. Thus, when a node receives a query, it uses this information to check whether or not it has already processed the same query from another node. If so, the query is simply dropped; otherwise it is processed as described in the previous paragraph.
The proposed diversity-driven selective forwarding approach is demonstrated by means of a simple illustrative example. Assume that a total of 52 tuples are to be dispatched and that each sensor node can store only four tuples in its buffer. Therefore, 13 sensor nodes are required in the surveillance region. Utilizing the BIBD dispatching scheme, the initial cyclic difference set is defined as {0,1,3,9}. Thus, model tuples
0, , 1 3, and 9
T T T T are dispatched to node n0,
TABLE 4
Tuple bit maps for neighboring nodes of node n7
0 1 2 3 4 5 6 7 8 9 10 11 12 Diversity 7 0 D 1 1 1 1 0 0 1 1 0 1 0 0 0 5 7 2 D 0 0 1 1 0 1 0 0 0 0 0 1 0 3 7 9 D 1 1 0 1 0 1 0 0 0 1 1 0 0 5 7 10 D 1 1 1 0 1 1 1 1 1 0 1 1 1 8 7 11 D 0 1 0 1 1 1 1 1 1 0 1 1 1 6 TABLE 3
Neighboring node information for node n7
One-hop Neighboring nodes
Two-hop Neighboring nodes
n0 n6
n2 Null
n9 n6
n10 n1,n4
n11 n3,n5
1, 2, 4, and 10
T T T T are dispatched to node n1, T2, T3, T5, and T11,
are dispatched to node n2, and so forth (see Fig. 5).
Assume that a top-2 diversity-driven forwarding approach is implemented in the WSN (i.e. N=2). Assume also that the 13 sensor nodes are randomly deployed in the surveillance region and have the network topology shown in Fig. 6. In the initialization phase, each sensor node uses the node ID discovery protocol to find its one-hop and two-hop neighboring node IDs. Table 3 lists the neighboring node IDs for node n7.
Note that node n7 obtains its two-hop neighboring node ID
information from its one-hop neighboring nodes. For example, the two-hop neighboring node IDs n1 and n4 are obtained from
neighboring node n10. Having acquired the one-hop and
two-hop neighboring node IDs, each node constructs one bit map indicating the model tuples stored at each of their one-hop and two-hop neighbors. Table 4 lists the tuple index bit map for each of node n7’s neighbors, i.e. D , 07
7 2
D , D , 97 D , and 107 D . 117
Since model tuples T3, T7, T8, and T are stored at node n10 7, bit map D77 has the form (0001000110100)2. Therefore, Diversity(7,0) is equal to 5 (i.e. ∑(D77|D07)⊕D77). Similarly, Diversity(7,2), Diversity(7,9), Diversity(7,10), and Diversity(7,11) are equal to 3, 5, 8, and 6, respectively.
When node n7 senses an object whose features do not match that stored within its buffer, this unknown data is forwarded both to its local neighboring nodes and along the top-2 diversity paths of these neighboring nodes. Thus, as shown in Fig. 7, and assuming the value of the TTL parameter to be equal to 4, the unknown information is sent initially to all of the one-hop neighbors of node n7, and is then forwarded by nodes n10 and n11 (i.e. the two nodes with the highest diversity values) to their own neighboring nodes in the event that they too lack the information required to classify the sensed object.
The diversity-driven selective forwarding approach avoids the need to flood the unknown data to the entire network. Rather, the nodes forward the data only along those paths with a higher diversity. Thus, the number of nodes traversed by the unknown data is reduced while the classification success rate (i.e. the
probability of classifying the sensed object) is increased. In other words, a robust network performance is obtained with a minimal power consumption.
6. SIMULATION RESULTS AND PERFORMANCE ANALYSIS 6.1 Simulation Model
In this section, the performance of the proposed diversity-driven selective forwarding method is evaluated for different tuple dispatching schemes and different TTLs and diversity settings. The default system parameter settings are as follows:
Model set size: 1600, 5000 tuples
Tuple size: 12 Kbytes (RGB 64*64 Pixels, 256 colors )
Deployment area: 3000*3000, 10000*10000 m2
Number of sensor nodes: 1450, 4858
Flash size of each sensor node: 512, 840 Kbytes
Number of tuples stored at each sensor node: 42, 70
Number of model duplicates: 38, 68
Transmission range of sensor nodes: 100 m
Average neighbor degree of sensor network: 10
Tuple dispatching schemes: SD, SDO, FDD and BIBD
In performing the evaluations, the simulation considers different system operating conditions under various parameter settings, namely, Scenario 1 and 2. In Scenario 1, the sensor nodes in the WSN have the form of MICAZ sensors and are equipped with a camera to capture the sensed objects. The captured images have a size of 12 Kbytes (RGB 64*64 Pixels, 256 colors). Meanwhile, each sensor has a FLASH size of 512 Kbytes and a transmission range of 100 m. It is assumed that the original WSN model set comprises 1600 tuples (images) and that 38 duplicates of each tuple are required to be dispatched. Each sensor node is capable of storing 42 tuples, and thus 1450 sensors are deployed in the surveillance region. The sensor nodes are deployed in a 3000*3000 m2 surveillance region and form a connected network in which the average degree of each sensor node is equal to 10.
In Scenario 2, each sensor has a FLASH size of 840 Kbytes and a transmission range of 100 m. The original WSN model set comprises 5000 tuples (images) and that 68 duplicates of each tuple are required to be dispatched. Each sensor node is capable of storing 70 tuples, and thus 4858 sensors are deployed in the surveillance region. The sensor nodes are deployed in a 10000*10000 m2 surveillance region and form a connected network in which the average degree of each sensor node is also equal to 10.
In the initialization phase irrespective of the manner in which the deployment area, the number of sensor nodes, or the flash size of sensor nodes, each sensor node collects the IDs of its one-hop and two-hop neighbors using the node ID discovery protocol and then constructs the corresponding bit maps and calculates the path diversities. In the simulations, it is assumed that objects appear randomly in the sensor node monitoring region and Fig. 7. Use of top-2 diversity-driven selective forwarding approach by
cannot be classified using the model tuples stored at the local sensor node.
In each scenario, the performance of the WSN is quantified using two metrics, namely the classification success rate and the number of traversed nodes crossed by the unknown data. The classification success rate is defined here as the average success rate of the proposed diversity-driven selective forwarding scheme in classifying the sensed object. For each tuple dispatching scheme, once the unknown data has been successfully matched, the average number of traversed nodes crossed by the unknown data is evaluated in order to quantify the power consumption of the corresponding scheme. The simulation results also compare the performance gain of the proposed diversity-driven selective forwarding search algorithm with a blind flooding algorithm. In the blind flooding algorithm, a sensor with unknown data send this information to all its neighboring nodes and only uses the TTL parameter to limit the search scope over which the query process irrespective of the top-N value.
6.2 Simulation results and analysis
Fig. 8 illustrates the variation of the classification success rate with the degree of path diversity when using the SD, SDO, FDD and BIBD dispatching schemes and specifying the time-to-live parameter as TTL=3 under Scenario 1. In the top-1 diversity case (i.e. N=1), the unknown data is broadcasted to all the local neighboring nodes and is then forwarded by the neighboring node with the highest diversity. As shown, the corresponding
classification success rate is less than 20% for each of the dispatching schemes. However, the success rate of the blind flooding is up to 97% since the unknown data is broadcasted to a great number of sensor nodes only limited by the TTL=3 setting irrespective of the degree of path diversity. When the value of N is increased, the unknown data is forwarded to a greater number of sensor nodes, and thus the classification success rate increases for each of the dispatching schemes. From inspection, the highest classification success rate (80%) is obtained using the BIBD dispatching scheme and a diversity setting of N=3. The classification success rate of the BIBD dispatching scheme is approaching to the success rate of the blind flooding scheme. Fig. 9 compares the number of traversed nodes crossed by the unknown data when using the SD, SDO, FDD and BIBD dispatching methods and a time-to-live parameter setting of TTL=3. It can be seen that the number of traversed nodes is virtually the same for each dispatching scheme. However, the BIBD scheme provides a slight performance advantage as the value of N increases. Moreover, although a higher classification success rate can be obtained by using a blind flooding scheme, the number of traversed nodes when using the scheme is far more than that when using the proposed forwarding scheme.
To evaluate the effect of the TTL parameter on the system performance, Figs. 10 and 11 show the variation of the classification success rate as a function of the path diversity for TTL=4 and TTL=5, respectively. As the TTL value increases, the length of the path over which the unknown data is delivered
TTL=3 0.00% 20.00% 40.00% 60.00% 80.00% 1 2 3 top-N diversity C la ss if ic at io n su cc es s ra te SD SDO FDD BIBD Flooding
Fig. 8. Classification success rate for various diversity paths and TTL=3 in Scenario 1 TTL=3 20 70 120 170 1 2 3 top-N diversity T ra ve rs ed n od es SD SDO FDD BIBD Flooding
Fig. 9. Number of nodes traversed by unknown data for various diversity paths and TTL=3 in Scenario 1
TTL=4 0% 20% 40% 60% 80% 100% 1 2 3 top-N diversity C la ss if ic at io n su cc es s ra te SD SDO FDD BIBD Flooding
Fig. 10. Classification success rate for various diversity paths and TTL=4 in Scenario 1 TTL=5 0% 20% 40% 60% 80% 100% 1 2 3 top-N diversity C la ss if ic at io n su ce ss r at e SD SDO FDD BIBD Flooding
Fig. 11. Classification success rate for various diversity paths and TTL=5 in Scenario 1
also increases. As a result, a greater number of comparisons are made between the unknown data and the stored model tuples. Consequently, the classification success rate improves with an increasing TTL. In Fig.10, corresponding to TTL=4, the classification success rates for the SD, SDO, FDD, BIBD and the blind flooding schemes are 47%, 47%, 53%, 70%, and 100%, respectively, for N=2. However, when the diversity is increased to N=3, the classification success rate for each dispatch scheme increases to more than 70%. For example, that of the BIBD scheme increases to 93%. In Fig. 11, corresponding to TTL=5, the classification success rate for each dispatching scheme is greater than 80% for N=3. Significantly, the classification success rate for the BIBD schemes attains a value of 100%. In other words, the unknown data can be consistently classified by the WSN when using the BIBD dispatching scheme and the diversity-driven selective forwarding method with settings of TTL=5 and N=3.
Figs. 12 and 13 show the variation of the number of traversed nodes crossed by the unknown data when specifying TTL=4 and TTL=5 under Scenario 1, respectively. In the case of N=1, the unknown data is broadcasted to all the local neighboring nodes and is forwarded only along the single path having the highest diversity. As a result, the number of traversed nodes is approximately equal for each dispatching scheme irrespective of the TTL value. However, when TTL or N increases, the unknown data is forwarded by a greater number of nodes, and
thus the number of nodes traversed by the data also increases, irrespective of the tuple dispatching scheme. Nonetheless, it can be seen that the BIBD scheme results in fewer traversed nodes than the other dispatching schemes as the path diversity and time-to-live parameter settings are increased. Moreover, the classification success rate of the proposed forwarding scheme associated with the BIBD dispatching scheme is the same as the success rate of the blind flooding scheme, but the number of traversed nodes is far fewer than that of the blind flooding scheme.
Figs. 14 and 15 show the variation of the classification success rate and the variation of the number of traversed nodes crossed by the unknown data as a function of the path diversity for TTL=5 under Scenario 2. In the top-1 diversity case (i.e. N=1), the unknown data is broadcasted to all the local neighboring nodes and is then forwarded by the neighboring node with the highest diversity. As shown, the corresponding classification success rate is less than 10% and the number of traversed nodes is approximately equal for each of the dispatching schemes. When the value of N is increased, the unknown data is forwarded to a greater number of sensor nodes, and thus the classification success rate increases and the number of nodes traversed also increases. Nonetheless, it can be seen that the BIBD scheme results in better classification success rate and fewer traversed nodes than the other dispatching schemes as the path diversity parameter settings are increased. Furthermore, since the model set size and the deployment area in Scenario 2 are much larger than those in Scenario 1, the classification
TTL=4 30 80 130 180 230 280 1 2 3 top-N diversity T ra ve rs ed n od es SD SDO FDD BIBD Flooding
Fig. 12. Number of nodes traversed by unknown data for various diversity paths and TTL=4 in Scenario 1
TTL=5 30 130 230 330 430 1 2 3 top-N diversity T ra ve rs ed n od es SD SDO FDD BIBD Flooding
Fig. 15. Number of nodes traversed by unknown data for various diversity paths and TTL=5 in Scenario 2
TTL=5 30 130 230 330 430 1 2 3 top-N diversity T ra ve rs ed n od es SD SDO FDD BIBD Flooding
Fig. 13. Number of nodes traversed by unknown data for various diversity paths and TTL=5 in Scenario 1
TTL=5 4% 14% 24% 34% 44% 54% 64% 74% 84% 94% 1 2 3 Top-N diversity C la ss if ic at io n su cc es s ra te SD SDO FDD BIBD Flooding
Fig. 14. Classification success rate for various diversity paths and TTL=5 in Scenario 2
success rate under Scenario 2 on the same path diversity setting is worse than that under Scenario 1 and the number of traversed nodes under Scenario 2 is larger than that under Scenario 1.
In WSNs designed to perform a monitoring and tracking function, a 100% classification success rate can be obtained by using a blind flooding scheme to forward the unknown data to all the nodes in the network. However, this approach incurs significant communication and processing overheads; particularly in large-scale networks. The simulation results presented above have shown that this problem can be resolved by using an appropriate model dispatching scheme to distribute the model set amongst the individual sensor nodes and then applying a selective diversity-driven forwarding method to propagate the unknown information through the network. Of the tuple dispatching schemes considered in the simulations (i.e. SD, SDO, FFD and BIBD), it has been shown that the BIBD scheme yields an improved classification success rate and a lower number of traversed nodes as the TTL and path diversity parameters are increased. The improved performance of the BIBD dispatching scheme is to be expected since the BIBD scheme has improved uniformity characteristics.
7. DISCUSSION AND FUTURE WORK
This Section will describe several issues that were not addressed in the paper. First, the mentioned above experimental results only compare the performance of the proposed diversity-driven selective forwarding approach against that of a blind flooding approach in terms of the classification success rate and the energy costs. In our future work, it is valuable to benchmark the performance of the proposed diversity-driven selective forwarding approach against that of a blind flooding approach in terms of network overhead and retrieval latency.
Second, the simulation section doesn’t clearly discuss how the value of TTL should be chosen, yet. In fact, to determine a fixed TTL value is not easy even if the sensor networks use the blind flooding scheme to search the unknown data. The TTL is varied from the density of the sensor deployment and the needed classification success rate. From the experimental results, the value can be summarized as follows. In the simulation scenario1, 1450 sensors are deployed in the 3000*3000 m2 area. If we want to achieve the classification success rate of 100%, the value of TTL should be chosen equal or larger than 5. Moreover, in the simulation scenario 2 (4858 sensors deployed in the 10000*10000 m2 area), the value of TTL should be chosen lager than or equal to 6 to achieve a higher classification success rate. Finally, sensors in all of WSN applications will be lost once their energy resources are consumed. When a number of the sensor nodes run out of batteries, the classification decision of the unknown data will be infected either in the proposed search scheme or the blind flooding scheme. To overcome the effect, a timeout mechanism can be implemented to the node identifier discovery protocol as described in Section 5. The live sensors would use the discovery protocol to update their diversity path information over a period of time. If we consider redeploying sensor nodes in the surveillance region when a number of sensors run of batteries, the redeployed sensors can trigger an event to execute the identifier discovery protocol. The sensors
in the network will get new two-hop neighboring node identifiers to update their diversity path information. The study falls outside the scope of the present study and is considerable for our future work.
8. CONCLUSION
This study has addressed the problem of improving the performance of WSN monitoring applications designed to answer queries for replicated data. In general, WSN sensors have only limited memory and storage capabilities, and thus it is impossible to store all of the model tuples monitored by the network on a single sensor. Furthermore, once the energy resources of a sensor are fully consumed, any information held at this sensor is irretrievably lost unless it is duplicated on other nodes within the network. Accordingly, this study has proposed four different tuple dispatching schemes for distributing the model tuples of interest amongst the sensor nodes of the network during the initialization phase, namely Sequential Dispatching (SD), Sequential Dispatching with Overlap (SDO), Fixed Distance Dispatching (FDD), and Balanced Incomplete Block Dispatching (BIBD). Furthermore, to improve the efficiency of the search process performed in the event that a sensed object cannot be classified at the local node, an efficient diversity-driven selective flooding scheme has been proposed in which the known correlation between the sensor identifiers and the indexes of the model tuples stored at each node is used to forward the unknown data along only those paths with a higher diversity. In general, the simulation results have shown that the proposed forwarding scheme yields a high classification performance irrespective of the tuple dispatching scheme applied. Of the four dispatching schemes, the BIBD scheme yields the highest classification success rate and the lowest power consumption due to the improved uniformity characteristics of the model tuple distribution within the network.
ACKNOWLEDGMENT
This study was supported in part by the National Science Council of Taiwan under Grant No. NSC 98-2219-E-006-010. We are also very grateful to thank the anonymous reviewers for their comments and suggestions.
REFERENCES
[1] Hebbert John Ruser, “Combinatorial Mathematics”, Mathematical Association of America 1963.
[2] ALOKE DEY, “Theory Block Design”, Halsted Press 1986.
[3] Douglas B. West, “Introduction to Graph Theory,” Second edition, Prentice Hall 2001.
[4] Yogan K. Dalal and Robert M. Metcalfe. "Reverse Path Forwarding of Broadcast Packets". Comm. ACM, pp. 1040-1048, vol. 12, December 1978.
[5] C. Shen, C. Srisathapornphat, and C. Jaikaeo, “Sensor Information Networking Architecture and Applications,” IEEE Pers. Comm., pp. 52–59, Aug. 2001.
[6] Akyildiz, I.F., Weilian Su, Sankarasubramaniam, Y., Cayirci, E.,”A survey on sensor networks”, IEEE Communications Magazine, Aug 2002.
[7] C. Intanagonwiwat, R. Govindan, D. Estrin, and F. Silva, “Directed diffusion for wireless sensor networking”, IEEE/ACM Trans. Netw., pp. 2-16, vol. 11, 2003.
[8] S. Ratnasamy, B. Karp, S.Shenker, D.Estrin, R.Govindan, L.Yin, and F. Yu, “Model-centric storage in sensornets with GHT, a geographic hash table”, Mobile Networks and Applications, pp. 427-442, vol. 8, 2003. [9] Feng Zhao, Leonidas J. Guibas, “Wireless Sensor Networks: An
Information Processing Approach”, First edition, ELSEVIER 2004. [10] J. Gehrke and S. Madden, “Query Processing in Sensor Networks”, IEEE
Pervasive Computing, pp. 46-55, vol. 3, 2004.
[11] Min Sheng, Jiandong Li, and Yan Shi, “Relative degree adaptive flooding broadcast algorithm for ad hoc networks”, IEEE Transactions on Broadcasting, June 2005.
[12] N. Sadagopan, B. Krishnamachari, and A. Helmy, "Active query forwarding in sensor networks," Elsevier Ad Hoc Networks, 2005. [13] Kwangil Lee; Yao, T.S , “Dynamic Path Adaptation Routing Protocol for
Mobile Ad Hoc Networks”, IEEE Vehicular Technology Conference, 2006.
[14] Taeck-Geun Kwon and Sukho Lee,"Model placement for continuous media in multimedia DBMS", IEEE International Workshop on Multi-Media Model Management Systems, Aug 1995.
[15] C. Jaikaeo, C. Srisathapornphat, and C.-C. Shen, “Querying and tasking in sensor networks,” in Proc. 14th International Symposium on Aerospace/Defense Sensing, Simulation, and Control, pp. 184–197, vol. 4037, 2000.
[16] P. Bonet, J. Gehrke, and P. Seshadri, “Towards sensor model systems”, in Proceeding of the Second International Conference on Mobile Model Management, pp. 551-558, vol. 43, 2001.
[17] S. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong, “TAG: a tiny aggregation service for ad-hoc sensor networks”, SIGOPS Opererating System Review, pp. 131-146, vol. 36, 2002.
[18] Y. Yao and J. Gehrke, “ The cougar approach to in-network query processing in sensor networks”, in SIGMOD Record, 2002.
[19] S. Ratnasamy, B. Karp, L. Yin, F. Yu, D. Estrin, R. Govindan and S. Shenker, "GHT: A Geographic Hash-table for Model-centric Storage in Sensornets", ACM International Workshop on Wireless Sensor Networks and Applications (WSNA), September 2002.
[20] S. Ratnasamy, D. Estrin, R. Govindan ,B. Karp, S. Shenker , L. Yin and F. Yu, “Model-Centric Storage in Sensornets”, SIGCOMM 2002. [21] S. Shenker, S. Ratnasamy, B. Karp, R. Govindan, and D. Estrin,
“Model-centric storage in sensornets”, SIGCOMM Computer Communication Review, pp. 137-142, vol. 33, 2003.
[22] S. Madden, M. J. Franklin, J. M. Hellerstein, and W. Hong,”The design of an acquisitional query processor for sensor networks”, in Proceedings of the 2003 ACM SIGMOD international conference on Management of model, 2003.
[23] Chih-Hung Chao and Jung-Shian Li, “A Novel Bib-Based Parallel Download Scheme” in Proceedings of IEEE Asia-Pacific Conference on Circuits and Systems, vol. 1, pp. 461-464, Dec. 2004.
[24] C.S Reghavendra, Krishna M. Sivalinkgan and Taieb Znati, “Wireless Sensor Networks”, Springer 2006.
Chih-Hung Chao graduated from the Feng Chia University, Taiwan, with BS degree in the Department of Electronic Engineering. He received MS degree from the Department of Electrical Engineering, National Cheng Kung University, Taiwan. He obtained his PhD degree from the Department of Electrical Engineering, Institute of Computer and Communication Engineering, National Cheng Kung University. He is an information manager of the Network Communication Section of the Library Information Center, National University of Kaohsiung, Taiwan. His research interests include wireless and mobile network protocols, peer-to-peer computing, and communications.
I-Hui Liiii received BS degree in Computer Science at Tamkang University, MS degree in Computer Science and Information Engineering at National Chiao Tung University, and PhD degree Department of Computer Science and Engineering at National Chung-Hsing University, Taiwan. She is currently a faculty member in the Department of Information Networking and System Administration of Ling Tung University, Taiwan. Her research interests are in data mining, optimization algorithms, wireless sensor networks, and software engineering.
Chong-Yi Yang studied in the Department of Computer Science at I-Shou University, where he received a B.S. degree in 2005/6; and in the Department of Electrical Engineering, National Cheng Kung University where he received a M.S. degree in 2008/6. He is a project manager in the department of information security unit, Data Communication Business Group, Chunghwa Telecom Company, Taipei City 100, Taiwan.
Jung-Shian Li is a full professor in the department of electrical engineering, National Cheng Kung University, Taiwan. He graduated from the National Taiwan University, Taiwan, with BS and MS degrees in electrical engineering. He obtained his PhD at the Technical University of Berlin, Germany. He teaches communication courses and his research interests include wired and wireless network protocol design, network security, and network management. He is currently involved in funded research projects dealing with network security testbed, common criteria, intrusion prevention system, router active queue management, VoIP security, and P2P architectures. He is the division director of computer and network center, NCKU. He serves on the editorial boards of the International Journal of Communication Systems and guest editor for wireless medium access control in IEEE wireless communication magazine.
