Heterogeneous WSN is the outcome of recent advances in sensor hardware design, where the sensor may have multiple capabilities in terms of computing, power supplies, communicating and sensing [17,18]. Hence, a heterogeneous network currently is able to detect multiple attributes of the environment such as humidity, temperature, level of a particular gas in the atmosphere, etc. For example, in an air pollution measuring application, the measurement data may be the fusion of several parameters such as temperature, level of carbon monoxide, level of smoke, etc. In such an application, it is reasonable to have storage and search mechanism for multi-dimensional queries. For example, scientists analyzing the growth of marine microorganisms, in a sea environment surveillance application, might be interested in the multiplex events that occurred within certain temperature and light conditions, e.g., “find all events that have temperatures above 22 Celsius degrees and light levels above 15 Fluxes”.

Range query is another challenge for DCS schemes. For example, a user may be interested in a range rather than a specific point value. For example, air pollution may occur if the level of carbon monoxide is in the range of 30 L/mol to 90 L/mol. A possible query is to find all of the sensing points where the level of carbon monoxide falls into the possible air pollution range. With range queries, users can drill down to improve their search efficiency for the events of interest. The example query presented above illustrates this and may be used by an environmental scientist carrying out a study in a particular forest to identify if carbon monoxide levels cause air pollution and perhaps also to map the results across the region under study with other parameters to identify causal links or to draw a conclusion [13].

Due to sensor hardware imprecision and environmental parameter variations, the similarity search problem in WSN has received considerable research attention. In certain applications or circumstances in addition to an exact match, it is necessary to search within a specified similarity range. For example, a multi-dimensional query on the attributes temperature, carbon monoxide, forest name, location, and smoke level with values 100°| 150 L/mol| Melaleuca| North| 130 L/mol may also wish to identify a similar set of values 90°| 150 L/mol| Melaleuca| South| 130 L/mol. A similarity search may also be useful in many applications such as identifying similar ocean current flows or wildlife activities during habitat monitoring [19]. A traditional approach for a similarity search may be inefficient for the highly energy-constrained sensor network. A possible approach is to search for similar data to the query without collecting data from all of the sensors.

Based on the data stored in sensors, DCS networks can facilitate data aggregation in a fully distributed way. By using data aggregation, traffic generated by producer nodes can be reduced before answering consumer queries. Monitoring building integrity during earthquakes by engineers, habitat monitoring by biologists, monitoring temperature and power usage in data centers by a cluster computer administrator are examples of sensor applications depending on the ability to extract summary (aggregate) data rather than raw data from the network.

In some deployments, sensors may not be uniformly distributed, which means that some sections or zones may be densely populated while others are not. In mobile WSN, sensors may move from one place to another creating additional non-uniformity. In the current state of the art, most DCS schemes are proposed with uniformly distributed sensors. This assumption leads to data losses in overburdened sensors.

Multi-replication reduces the data loss that happens due to node failure or node mobility from one rendezvous zone to another. Furthermore, multi-replication may facilitate data fusion and aggregation. In the literature, most of the DCS schemes fail to meet the key issue of recovering lost data during data replication.

In WSN certain popular events may generate higher data traffic volumes than other events and hence may overburden certain nodes, sections or zones and lead to data loss while others may be left with a light storage load or remain empty. Therefore, it is important to balance the storage load among the sensors to prevent imposing too high a storage load on some nodes.

Geographic Hash Table (GHT) [1] was the first DCS scheme proposed in 2002 by Ratnasamy, et al. The motivation for this research was to make effective use of the vast amount of data gathered by large-scale sensor networks using scalable, self-organizing and energy-efficient data dissemination algorithms. Ratnasamy, et al. use a Distributed Hash Table (DHT) [20] in order to hash keys, for example event name or type, into geographic co-ordinates and store this event in sensor nodes in a geographic location closer to the co-ordinates. Greedy Perimeter Stateless Routing (GPSR) [21] was used to store and/or retrieve data from a sensor node. GHT uses the function put (k, d), where k is a hash key used as the destination geographical location and d is the data, to forward a data packet to the location k using GPSR. The closest sensor node to the geographical location k is chosen as the home node, where data is stored for this event type. Similarly, when a consumer wants to consume/query data of an event type (for this particular case e.g., humidity), GHT again maps the event type to the hash key k and uses get(k) to forward the query to the corresponding spatial/geographical location k. The home node then replies by providing stored data for that event type. In Figure 1, sensors are represented by ‘’. A producer node senses a value and forwards it to home node denoted by ‘’. In turn, another consumer node uses the same hash function and retrieves the stored data from the home node. When storing data, represented by ‘’, the producer node sends data to the target node and in the retrieval process, denoted by ‘’, the query is first forwarded to the home node and replies are then sent back to consumer.

Similarity Data Storage (SDS) [19] proposes an efficient spatial-temporal similarity search scheme for both static and dynamic WSN. Efficient data aggregation and query, similarity search for multi-attribute data and spatial temporal search are identified as three major challenges faced by DCS schemes. SDS is an important approach that utilizes spatial-temporal and similarity search functionalities and aims to reduce overhead, energy consumption and search latency. In SDS, a deployed large-scale WSN field is considered as a rectangular field. The entire field is divided into small rectangular zones. Each zone has a dedicated sensor node named as zone head. It is assumed that each node in the network is configured with three information elements: (1) Number of zones horizontally n_x and vertically n_y., (2) the zone ID assignments scheme—IDs are assigned within the zone sequentially from left to right, and (3) the zone ID and geographical location. A node in the zone with ID_i can calculate its Euclidean distance from another zone ID_j using , where δx_i,j = (ID_j − ID_i)%n_x and δy_i,j = (ID_j − ID_i)/n_x. A head node in a zone is responsible for communication with other zones. All other nodes inside a zone are connected with the head node.

The Similarity Search Algorithm (SSA) [22] was proposed by Chung, et al. based on the Hilbert Curve over a DCS structure. SSA is successful in searching similar data without collecting data from all of the network sensors. Being motivated from the Hilbert Curve concept, the authors divide the network recursively into 4^l square quadrants where l denotes the number of levels. The center (indexing node) of each square quadrant (cell) is denoted by I. It is important to select the indexing nodes to avoid performance degradation due to too many indexing nodes and on the other hand a lack of enough storage space due to too few indexing nodes. If the total memory space for storing data is A and the memory size of each sensor is z, then the number of indexing nodes n can be defined as: . So, the number of levels l is defined by: l = log₄. The entire data range of an event is referred to by R where R_L and R_U, respectively, denote the lower bound and upper bound. R is divided into n equal sub-ranges each being equal to r i.e., n.r = R. So, the data sub-range for which I_ID is responsible is defined as . Figure 2 illustrates level 1 and level 2 assuming that the data range R of an event is (0, 1). Detected events are mapped to a particular segment if the event falls in the range of that cell. Two parameters are used to record the minimum and maximum values of each segment. Initially these two parameters are set to 0. When data is inserted into an index node, the values are updated accordingly. For example, if a sensor detects an event with a value of 0.2 then the data is sent to I₀ as it belongs to the range [0, 0.25] and both parameters of this cell will be updated to. Being a mini-repository of an entire distributed database, each sensor has knowledge of its local data and hence lacks global knowledge of the entire sensor database. The distributed nature of the data throughout the sensor network is one of the major challenges when processing similarity search queries. To overcome this challenge, the adjacent index nodes in SSA along the Hilbert curve have data of similar values and thereby avoid the need to collect data from all of the sensors in the network. This data mapping based on the concept of the Hilbert curve is simple and easy to implement. However, when storing multidimensional attributes it is not clear whether SSA maintains a separate Hilbert curve or not and in this case how it responds to the multidimensional queries.

In Dynamic Load Balancing (DLB) [23], Liao, et al. mention unbalanced distribution of data among sensors as one of the major constraints for most of the DCS techniques. To address this issue, Liao, et al. propose a grid-based DLB approach that relies on two schemes: (1) A cover-up scheme to deal with the problem of a storage node whose memory space is depleted and (2) multi-threshold levels to achieve load balancing in each grid and all nodes get load balanced. DLB divides the whole network into a grid with cells of the same size in such a way that all the nodes inside a cell are within one hop distance. Each grid is numbered with positive coordinates (X, Y) called grid IDs. A sensor node can calculate its grid ID (X, Y) using the following equation:

and    (1)

Each node has a virtual grid ID and virtual co-ordinates that are initially equal to the actual grid ID and co-ordinates. Initially, each node broadcasts a message within its grid by limited broadcast to exchange the information to build a ‘Grid_Node’ table. A producer node uses the hash function on the event type to map the event type into a grid and transform the event type into a grid ID using the above equation. The center of the grid is called a grid point. The node, after detecting an event, sends a Put packet to the grid ID and uses GPSR to forward this packet to the node closest to the grid point.

Load Balanced Data-Centric Storage (LB-DCS) [24] is an organic approach that relies on the home perimeter for data replication and thereby overcomes the unbalanced load constraint in DCS-GHT [1]. LB-DCS functions on top of three mechanisms: (i) A density estimation protocol that is used to estimate the network density f, which is included in put and get protocols; (ii) a modified hashing function that includes f in its parameter list; and (iii) a storage protocol enforcing QoS in the selection of the number of replicas for data storage. In [24,25], the authors show that depending on the event type, the number of local replicas should be different. So, when a producer node produces any event it also specifies a value in terms of the parameter q to specify the number of replicas. The put primitive takes q along with two other parameters: datum d and meta-datum k. Depending on the value of q, the home node selects q neighbor nodes using the ball method to replicate that event. This dispersal method is iterative. A home node, H (with co-ordinates X_H, Y_H) considers a ball with a radius r (randomly selected value). The home node sends a request for storage to all sensors within the range of this ball denoted by:

   (2)

In turn, when a sensor receives a storage request it acknowledges the request to H. H calculates the number of acknowledgments (q') received from the nearest sensors. If q' < q, then H sends a storage request to sensors. This time it considers only the sensors in . This process continues until H receives q acknowledgements or exhausts sensors within the perimeter. Apart from this QoS, the authors also include non-uniform hashing that can be used to balance the load even in non-uniform distribution such as a Gaussian distribution of sensor nodes in a network. In such a non-uniform WSN, LB-DCS applies two distributed protocols, referred to as proactive (Broadcast) and reactive (Stripes and Fatstripes), to compute the density approximation f of each zone. The density approximation is used to bias the hash function enabling the distribution of target co-ordinate pairs for storage according to the network distribution. For non-uniform hashing, the Rejection Method [26] has been used in LB-DCS.

In Tug-of-War (ToW) [27], the authors propose a data-centric mechanism where queries and events meet at a point that is selected based on the relative frequencies of events and queries. ToW operates in two modes referred to as write-one-query-all and write-all-query-one. The former allows a sensor to store an event in the nearest mirror image but queries need to be disseminated to all mirror images. Conversely, in the latter operation node events must be stored in all mirror images to facilitate a query node that disseminates the query to the nearest mirror image. ToW takes its motivation from the Structured Replication (SR) mechanism in GHT. In SR, to alleviate a node’s load, a detected event is stored in the nearest mirror image. This mechanism alleviates the storage cost but as a query node has no idea which image node may have the data, the query node needs to query all of the images and this increases the query cost. Like SR-GHT, a home node and a set of 4^r− 1 mirror images are assigned in ToW for each event class c and here r is referred to as the system resolution. The mode a sensor node will operate depends on r. The system resolution is determined based on the relative query frequency and events detected for a particular class of event. To minimize the communication cost, ToW adjusts the rendezvous point on the fly on an optimal basis. With a given system resolution, nodes define the mode of operation while in the selected mode nodes need to find an optimal value of r in order to minimize the communication cost. In write-one-query-all mode, the total communication cost per unit interval is defined by:

   (3)

By using an optimal r, the cost of ToW in this mode per event is defined by:

   (4)

Here, C_e, C_q, f_e, f_q, k and n denote storage cost, query cost, event frequency, query frequency, number of the event detection node and total number of nodes deployed in the network, respectively. With the given value of k, f_e, and f_q,is minimized by the following value of r:

   (5)

Similarly, in write-all-query-one mode, the total communication cost is denoted by and the optimal value of r is defined by:

   (6)

   (7)

Using an optimal r, the cost of ToW in this mode per event is defined by:

   (8)

Thus, in order to determine the mode of operation, the resolution must be non-negative. With the given value of k, f_e, and f_q, if a nonzero optimal system resolution r exists in one mode then in other mode no non-zero optimal r can exist. Hence, finding a non-zero optimal r in a mode is essential to determine the mode of operation.

In Quadratic Adaptive Replication (QAR) [16], the authors recognize multi-replication as one of the most important DCS research areas and propose to enhance ToW by a flexible solution that permits selecting a more adaptive number of replicas. ToW is based on a geometric replication formula that calculates the number of replicas, N_r = 4^d, where d is the network-depth. Thus, the main drawback of ToW is its inflexibility in selecting the number of replicas, which are limited to 1, 4, 16, 64, and so on. In contrast, QAR calculates the number of replicas as N_r = d² allowing the number of replicas to grow in a quadratic fashion. Furthermore, QAR provides a mathematical model that can find the optimal number of rendezvous nodes based on the ratio of consumption and production traffic.

Considering information brokerage as one of the prime concerns, the authors in [28] propose a Double Rulings scheme storing data replica at a curve instead of one or multiple isolated sensors. In this approach, information hints can be left along the trail when data travels from source to a rendezvous node at no extra communication cost. Hence, discovering relevant data also becomes easier for the consumer. The scheme provides greater flexibility to design a network’s retrieval strategy subject to the current network load and energy level. The paper presented a number of possible retrieval mechanism such as GHT retrieval, distance-sensitive retrieval, aggregated retrieval and double rulings retrieval. The flexibility in retrieval curves removes the possibility of having bottlenecks around the rendezvous node since the retrieval curve may not necessarily visit the rendezvous node. A local recovery scheme is achieved by storing data on the boundary when the sensors in a certain region are destroyed. Compared to SR-GHT, the double rulings scheme improves data robustness by imposing much lower communication costs for replication through the organization of replicas along a closed curve. The paper also proposed a new routing technique referred to as greedy routing on a curve, which adds extra complexity when implemented. Furthermore, flat replica in sensor nodes along the replication curve deplete the storage capacity of the network quicker than other schemes.

In contrast to pure replication, Michele Albano et al propose an alternate approach by incorporating erasure coding in DCS. In [29], the authors use the Redundant Residue Number System (RRNS) to encode data into a set of fragments based on n out of m code that guarantees the survival of the data in the event of the loss of a limited number of fragments (up to m-n). The put(d:k) primitive function first selects a set of sensors N_k to store the data based on the number of fragments to be computed, then it multicasts a storage request of d:k to the sensors in N_k.. Prior to sending the storage request, each sensor p (p belongs to N_k) is assigned with a module denoted by m(p), which is used to compute the fragments. The assigned module is randomly chosen from a library of m = n + r pairwise prime moduli m₁,….,m_{n + r}. Once a sensor p is assigned with a module m(p) for storing the pair d:k, it stores x:k in its memory where x is the fragment given by the residue of d module m(p). Upon receiving a request from a query node for data d:k stored by sensors in N_k, the reconstruction might happen using one of two strategies. In the first strategy, each sensor in N_k retrieves the fragment x:k that corresponds to d:k and sends it to the query node. In the second strategy, one sensor in N_k reconstructs d:k after collecting all of the fragments x:k from N_k and sends this reconstructed data to the query node. The first method ensures reliable data reconstruction at the cost of high energy consumption, on the other hand, the second method trades off energy consumption with reliability requiring only local communication in reconstructing the data, and only one final packet is sent to the sink.

Distributed Index for Features (DIFS) [30] in sensor networks is designed to provide communication load balance across the index keeping the search efficiency of a quad tree [31]. DIFS also uses a geographic hash similar to GHT [1]. DIFS constructs a multiply rooted hierarchical index that differs from traditional binary and quaternary trees. In DIFS, a non-root node can have multiple parents. Nodes are responsible for storing information for a specific range within a particular geographic region. A node covering a small area stores a wider range of values while a node covering a large area stores a smaller range of values. DIFS efficiently supports range queries related to values distributed in a range. Unlike QUAD tree, each child in DIFS has bfact parents where bfact = 2ⁱ, i ≥ 1. The range of values in the histogram of a child is bfact times the range of values maintained by its parents. The range of index values decreases by the factor bfact as it progresses up the index hierarchy. Suppose an event is detected and the value of an attribute (say flux density) is 57 and this attribute value is to be recorded in the index. This attribute value would be recorded in a local leaf node covering a range of 0~255 (considering the range of flux density 0~255) and would also be stored in the parent of the local leaf node that covers a range of 0~63, in a grandparent covering 48~63 and in a great grandparent with a value range of 56~59. Here, a value of bfact = 4 and hence the range of values covered by the grandparent is one fourth of that covered by the parent while the geographic area covered by grandparent is four times that of the parent and so on.

In Figure 3, the width and height of the total covering region of a network is eight units (w = 8 and l = 8). Considering bfact = 2, the dimension of the system-specified minimum covering region for this case is 1 (, where h is the number of levels) and 1 (). Suppose that an event has been found in the vicinity of geographical co-ordinates (7.5, 4.5) with an attribute value of 8. For the time being, it is assumed that nothing is known a priori about the expected distribution of the attribute except that it always falls in the range [0~8]. The hash for the key “attribute:0:8” will return a location somewhere in the bounding box defined by the corner (7, 4) and (8, 5). So, a message containing the string “attribute”, the co-ordinates (7.5, 4.5) and the value 8 will be sent to leaf-level (level 0) index node, say . Since, bfact = 2, the level 0 leaf index node will forward a histogram containing a count for values 7~8 to a level 1 index covering the region (6, 4) to (8, 6). The level 1 node will then forward a histogram containing counts for values 4~8 to a level 2 node, covering the region (4, 4) to (8, 8). This process continues until the value is forwarded to node covering the entire network bounded by (0, 0) and (8, 8).

DIFS provides communication efficiency and fast searching of data by name, value range and location. Furthermore, a communication bottleneck at the root of the search tree is avoided using multiple parents of a child thereby maintaining the search efficiency that can be achieved by drilling down through a search tree using a hierarchical approach.

In PathDCS [32] messages are routed to locations using a tree-structure, and the path to a location is identified using landmarks, which are shared reference points. The interesting novelty of this approach is that an event type is mapped to a path instead of mapping the event type to a spatial location. The path is defined by an initial landmark and then followed by a set of procedural directions. The landmarks are also called beacon nodes, which are elected randomly or manually configured. Standard tree construction techniques are used to build trees rooted at each of these beacon nodes. This ensures that all nodes know how to reach the beacons. A path consists of a sequence of b_i pairs (b_i, l_i), where b_i is a beacon and l_i is a length. A beacon node b_i is the beacon whose identifier is closest to the hash function h(k, i). The packet is first routed to beacon b₁, and then it is sent l₂ hops toward beacon b₂ using the tree rooted at b₂, and so on, until it arrives at the previous i − 1 segment. The packet is then sent l_i hops toward the next beacon b_i. In addition, the first segment length l₁ is equal to the distance to the first beacon b₁, whereas segment lengths for i > 1 are given by:

   (9)

The PathDCS algorithm has two parameters: the total number of beacon nodes denoted by B and the number of path segments P. The performance of PathDCS largely depends on these two parameters and by varying the number of B, PathDCS trades off the control traffic overhead due to tree construction versus load on the beacons. Data is also been locally replicated by flooding within k-hops of the destination.

A Hierarchical Voronoi Graph Based Routing (HVGR) [33], the Voronoi graph is used to construct and maintain a virtual hierarchy of sensor nodes. The Voronoi approach is a self-organizing hierarchical graph that is constructed without the help of GPS or other geo-location devices. The network is divided into different levels of regions based on landmarks. A landmark selection algorithm is used to select at most m_i(i = 1, 2, ….) landmarks in an (i − 1)^th level region. Then the network is divided into first level sub-regions based on first level landmarks. Each node in the first level sub-regions broadcasts a landmark packet to the entire network. By receiving this packet every node estimates its distance from all first level sub-region landmarks. Each node selects one landmark as its representative. Then each first level sub-region is again divided into second level sub-regions with second level landmarks. Each sensor node again chooses the closest landmark as its second representative. This process continues until the last level sub-regions are small enough so that each node knows all other nodes in its sub-region as shown in Figure 4.

For landmark selection the authors use the optimized random landmark selection algorithm [34]. According to this algorithm, a random node, say ‘L₁’, starts the landmark selection process becoming the master landmark of first level landmark nodes. It then selects m₁− 1 nodes for m₁ first level networks. Each first level landmark then again acts as master and continues the selection process in its zone to select a second level master node for the second level of the network. A master landmark node stops the landmark selection process once it finds that all sensor nodes in its region are within its communication range. Figure 4 shows an example of the basic routing algorithm that is used to route packets from the source (S) to the destination (D). The first level landmark for the destination is L₁. The packet first moves hop by hop toward L₁ until it reaches the edge of first level region, R₁. Now, nodes in R₁ know the second level landmark and the packet now moves hop by hop toward the second level landmark, L₂. Again once it reaches the node B of region 2 in Figure 4, the packet starts moving toward L₃. Finally, the packet reaches the lowest level sub region whose entry node ‘C’ knows the path of the destination and forwards the packet to the destination. The interesting part of this algorithm is it never overwhelms the landmark, as the packets never go to the landmark, rather the packets move toward the landmark until the packets reach any node within the region.

Multidimenstional Attribute (MDA) [35] proposes an energy-efficient & scalable multi-dimensional range query mechanism (MDA_s) to retrieve and store data. It builds an in-network distributed data structure by mapping multi-dimensional attributes to their corresponding range spaces. The authors consider a few assumptions in their work: (1) The sensors are uniformly and densely deployed; (2) each node can sense multiple events; and (3) each node maintains a neighbor table via periodic beacon message exchanges and knows its own geographic location. A source node detects an event with a set of attribute values, E = {a_1, a₂... a_n}. It is assumed that all attribute values are scalar and normalized to (0~1). An event is assigned with a k bit code where k = 2m. If 0 < a_i < 0.5, the i^th bit code is assigned to 0 otherwise 1. For x_j=a_2j-1 and y_j=a_2j the attribute values of E are assigned by a serial bit code B = {x₁, y₁, x₂, y₂……x_m, y_m}. For example, an event E = {0.8,0.7,0.4,0.3} is assigned by a four bit code B={ x₁, y₁, x₂, y₂}={1,1,0,0}. Assignment of code B can be achieved by generalizing 0 < a_i < 0.5 and 0.5 < a_i < 1 to 0 < 2a_i < 1 and 1 < 2a_i < 2, respectively:

   (10)

   (11)

This code B is mapped to a range space R = [x_low – x_up, y_low − y_up]. Hence, it is necessary to calculate (x_low, y_low) and (x_up, y_up) to find the range space R.

In code B, x = x₁, x₂… x_m are used to calculate X of R and y = y₁, y₂… y_m are used to calculate Y of R. In code B, if x₁ = 1 its value is in between 0.5 and 1; if x₂ = 1 its value is in between 0.5 and 0.5 + (0.5)². So, X_low = x₁(0.5) + x₂(0.5)² + ……. + x_m(0.5)^m. Similarly Y_low = y₁(0.5) + y₂(0.5)² +……. + y_m(0.5)^m.

Again, X_up is defined by:

X_up= 1 − (x₁(0.5) +x₂(0.5)² +……. + x_m(0.5)^m)

So,

So,

   (12)

Example:

Hence, the range space is defined by R = (0.5–0.625, 0.875–1). Figure 5 illustrates the MDA_s storage and query process.

In [13,19,22,30,35,36], the range query mechanism is implemented in different format. Chung et al. [22] propose an efficient technique, illustrated in Section 3.3, for a DCS similarity search that includes both a point and range query. A range query is divided into sub-queries and then the sub-queries are forwarded to their corresponding indexing node. For example, the level of carbon monoxide found in the air in a forest may be represented by levels that range from 50 L/mol to 150 L/mol. This range is split equally into four (4¹=4, four index node referred as I_0, I_1, I₂, I₃) sub-ranges [50, 75), [75, 100), [100, 125), [125, 150]. Hence, if a range query is to find data within [60, 80), then the sub-ranges of I₀ and I₁ are located as they cover the given query.

Li, et al. [13], proposed a method called DIM, which includes both a point and a range query in a multidimensional DCS model. In DIM, each sensor is linked as a node in a tree structure where each node represents a range of values. A root node represents the entire range of values and splits into two equal parts for left and right child nodes. This process continues for each non-leaf node until leaf nodes are reached.

DIFS [30], referring to Section 3.10, performs a data fusion based on data conveyance through the network. The routing is designed on top of a quad tree in a manner that balances the communication load across the index, and the range is maintained along the sensor node hierarchy.

In MDA [35], described in Section 3.13, a range query is split into a k-dimension range of multiple sub-queries. After splitting, each attribute sub-query is assigned with a bit code and tuples of k-bit codes, referred to as code B, are produced. The code B are then mapped to range space R and data is stored in nodes located in the mapped R. The state-of-the-art approaches in the field of range query DCS are summarized in Table 1.

In SDS [19], described in Section 3.2, a data item denoted by d consists of keywords denoted by v^d. Thus, the data item d is represented by d = (v₁^d, v₂^d,………, v_m^d). For a similarity search, the following formula is proposed in SDS to calculate the similarity between data items d₁ and d₂:

   (13)

Here, m is the number of attributes, w_i is the weight for each attribute based on their significance as shown in Table 2 and B(i, j) is a Boolean function returning 1 if v_i^d1=v_i^d2 and 0 otherwise.

For example, a data item in a WSN application is characterized by five attribute lists such as object, model, color, direction and division with the weight value shown in Table 2. Consider the following two data items denoted by d₁ and d₂:

Hence,

SDS uses Locality-Sensitive Hashing (LSH) [37] to transform d to a series of hash values. Using LSH, data items that have common keywords will have the same hash value while similar data items will have similar hash values. For example, if the difference between d₁, d_2, and d₃ is d₁ > d₂ > d_3, their hash values conform to h_d1 > h_d2 > h_d3, where h_d is the hash value of d. A data item with a hash value h is mapped to the first zone with ID ≥ h. For example, if the value of h_d1returned by LSH is five, then the value of h_d2 (d₁ and d₂refer to the data item shown above) would approximately be 6.67 or 3.75. Hence, d₁ will be stored in the zone with ID ≥ 5 and d₂ will be stored in the zone with ID ≥ 3.75 or ID ≥ 6.67. Thus, a query denoted by d_q with 80% similarity will search in zones with IDs: .

In SSA [22], referring to Section 3.3, when data is queried, the query is first sent to the specific cell, with the range of this cell in the queried data. The SSA similarity search mechanism contains two phases known as the similarity search query resolving phase and the query probing phase. The query resolving phase determines an indexing node that is most likely to provide an answer for the query. If the answer does not match exactly then the query probing phase is initiated to find the closest possible answer.

In the query resolving phase, the function locate (V_q) is used to find the target node I_T with the indexing node I_ID such that R_L + (I_ID− 1)∙r ≤ V_q < R_L + I_ID∙r, where V_q is the search value given by the query. The query is then forwarded to the Target Node I_T to retrieve data. If an exact match to V_q is found, then the query execution is finished, otherwise the query probing phase comes in action. In the query probing phase there can be the two following possible cases:

In SSA, three functions are proposed referred as backward probing, forward probing and bi-directional probing where backward probing and forward probing are used to deal with Case 1 and bi-directional probing is used for Case 2.

In [38], the authors propose four application profiles where two are aggregation profiles called ConsAggr and ProdAggr. Data aggregation is proposed using replication nodes proposed in QAR, see Section 3.7. In ConsAggr the profile consumer queries dominate production queries. In this profile, events produced in a given area are aggregated using the replication node for that area. In contrast, with the ProdAggr profile the production traffic dominates consumption traffic. Replicas aggregate events received from producers to achieve an effective overall traffic reduction. In Resilient Data Centric Storage (RDCS) [36], three types of event query called List, Summary and Attribute-based are proposed for the DCS scheme. A query request for all stored data for a particular event type is referred to as a List while a query request for aggregated or summarized data for a particular event type is called a Summary query. On the other hand, an Attribute-based query requests data for all events that match certain constraints based on attribute values. However, it is not mentioned how the monitor node will aggregate data for an event type to give a response to the summary query. TinyDB [9], madwise [39] and TAG [40] are some of the aggregation mechanisms proposed mainly for WSN databases, which can be adapted to DCS networks as well.

Load Balanced Data Centric Storage (LB-DCS) [24], Section 3.5, deals with non-uniformity of the sensor network based on two mechanisms. At first it estimates network distribution and then exploits data dissemination methods based on the estimation. A closest sensor node, called the sentinel, to the watch point, which is the region center, is responsible to identify its neighbors by sending broadcast messages. Every sentinel node sends its own region’s sampling density estimation toward other sentinel nodes using both proactive (Broadcast) and reactive protocols (Stripes and FatStripes).

From GHT on [1], illustrated in Section 3.1, a home node is called a hotspot if too many events with the its key are detected. To address this issue the authors have extended GHT by employing Structured Replication (SR) referred to as SR-GHT. SR-GHT introduces a hierarchical scheme where one node can have 4^d− 1 mirror images of the root home node.

The hierarchy depth is denoted by d. As shown in Figure 6, a decomposition of two (d = 2) is deployed having 4²− 1 (15) mirror images at different levels. A producer node can store the detected event to its closest mirror node reducing the storage cost to from . In the retrieval phase, GHT needs to forward queries to all mirror images. The query is first sent to the root node, which then forwards it to all mirror images in level 1 and then level 1 nodes forward this query to level 2 mirror images. This again increases the retrieval cost to from . Table 3 shows a summary of multi-replication DCS schemes based on their functionalities.

In SDS [19], Section 3.2, the head node keeps a copy of all data to be replicated, which may be unrealistic due to storage space availability, and if any head node fails, then data cannot be recovered from the head node. ToW [27] also replicates the root node into 4^d− 1 mirror images like GHT but the authors used a combing routing protocol, taking advantage of the grid structure (see Figure 6) to create a shorter replication tree for transferring messages among replica nodes. Like GHT and ToW, QAR also proposes replication of the home node but in a more adaptable quadratic evolution N_r = d². QAR also provides an analytical model defining the optimal number of replicas (N_r^*), thus minimizing the overall network traffic. However, Unlike GHT, ToW and QAR replicate producer data to all replicas in their second mode of operation. Nevertheless, the mode of operation is not deterministic since the mode of operation depends on the resolution (see Section 3.6) and hence ToW and QAR might also suffer from the data loss problem during their first mode of operation, i.e., write_one_query_all. SSA [22] creates a mirror of the index node using a mirror Hilbert Curve and mirror mapping function. However, the replication process is not explained in the paper. RDCS [36] proposes at most one mirror node for each zone, which will keep a copy of the zone’s index node. An index node stores all data directed toward its responsible zone and therefore creation of a hotspot surrounding the index node is possible. Nevertheless, the hotspot issue has not been mentioned in the selection process for the mirror node. Dynamic random replication for DCS [41] and double rulings [28], described in Section 3.8, are two important research outcomes related to multi-replication research for DCS. Relative to previous work, random replication is a simpler and more flexible technique that enables an effective reduction of network traffic. The schemes, specifically, consider the case where nodes can determine the current set of N_r replicas associated with a given application by generating N_r random spatial locations with a hash function , where app is the application’s name and epoch is a shared time identifier employed to change replicas over time. The paper demonstrates that by placing a replica node randomly in the network it is possible to outperform ToW, QAR and GHT. Also by changing replication nodes over time it is possible to equalize the energy burdens across the network and hence a 60% improvement in network lifetime may be possible. However, the wireless network in the model is assumed to be static and the application is considered to be spatially homogenous.

In DLB [23], Section 3.4, to balance the load among sensors in a grid, a scheme named the ‘cover-up’ scheme is used. According to this scheme every node of a grid has storage threshold levels. For instance, a sensor node has two threshold levels, e.g., 1^st level = 30 and 2^nd level = 60. A grid point forwards packets to the closest grid node for storage. When the node closest to the grid point reaches the first threshold level it modifies its virtual co-ordinate to (∞, ∞) in order to hide its original location. Hence, the geographic routing protocol forwarding event data to a storage node will ignore the original closest node and find a new node that is the next closest node to the grid point. This process continues until all nodes reach their first threshold level. The last node, farthest from the center, reaching this threshold broadcasts that the first threshold has been reached and the second threshold storage level is established. However, following this mechanism, at some point,all of the nodes within a grid could become saturated. If this occurs, the paper proposes an extended grid that uses adjacent grids to the saturated one to select a new home node.

The authors introduce two types of load balancing schemes in SDS [19]: (1) storage load balancing and (2) routing load balancing. In storage load balancing, the storage load is balanced among different zones. Every zone maintains a threshold, denoted by φ, of the percentage of a zone’s used storage to indicate when a zone is at risk of being overloaded. When a zone’s storage crosses φ, it starts forwarding all storage requests to a lightly loaded neighbor zone. Every zone calculates its storage status and forwards this status to its neighbor zones periodically. In routing load balancing instead of routing packets using one specific route, SDS calculates more than one shortest route between a source or relay and destination. This reduces congestion in any specific route. The calculated shortest path between H_a (X_a, Y_a) and H_b (X_b, Y_b), as shown in Figure 7, is not unique. For simplification let us consider |X_a – X_b| = |Y_a – Y_b| and then the maximum number of inflexion points in the horizontal and vertical routing process is C = |X_a – X_b|∙C. The number of choices in choosing j inflexion points among C is , which is equivalent to the problem of putting C ball into j boxes. There are 2^j possible routes for each choice with j inflexion points, hence the maximum number of possible shortest routes between H_a and H_b is . When two nodes frequently communicate to each other, each query issuer randomly chooses one at each time from the list of all the shortest paths calculated.

In HVGR [33], Section 3.12, for balancing the load among sensor nodes, the name based routing approach is modified. During the landmark selection, the network is divided unevenly. For example, an event E has a probability 1/m₁ of being assigned to the first level landmarks (L₁, L₂,…). If the network is divided unevenly (L₁ > L₂), L₂ is more likely to be overloaded earlier than L₁. Hence the load is balanced by assigning a task to regions in proportion to the region size. An event is stored in a node in L_k’s first level region if:

   (14)

Here, N_i is the number of nodes in landmark L_i’s first level region and .

In LB-DCS [24], for handling load balancing across a dynamic network, the hash function includes a network density estimation f. To estimate sensor density, the WSN is divided into n x n non-overlapping square regions with a side length p. A sensor node closest to the center of a zone is referred to as a watch point and is called a sentinel node. Each sentinel node broadcasts a request to its neighbors to count them. The number of neighbors is used as an estimation of the local density in the region. One proactive (Broadcast) and two reactive protocols (Stripes and FatStripes) are used to deliver the estimate computed by sentinels to other sensor nodes. After collecting estimates a sensor uses and to calculate the density in each region and the final approximation based on the first computation, respectively. Here, w_ij is the density estimated by the sentinel for the i, j region and m is the number of neighbors in this region. In two situations referred to as false zeroes and over reporting, d ' i,j might give a misleading density estimation value. A region with sparse nodes near the watch point or a concentration of nodes near the border may report a zero or very low density representing false zeroes. In another situation, a region with a high concentration of nodes near a watch point and sparse nodes near borders may report a higher density than normal representing over reporting. To overcome these situations the final approximation d_i,j is computed as a weighted mean of the approximation computed in the first step for region ij. Table 4 shows a summary of load balancing DCS schemes based on their functionalities.

The DCS scheme might incur high update traffic due to the lack of an optimally synchronized routing algorithm. Initial research in DCS focused on designing an efficient technique that could be used to map data to the rendezvous node/zone rather than concentrating on developing an optimal packet routing technique. Most of the DCS methods, for data storage and search routing, rely on a locating system (e.g., GPS) that places an energy burden on the WSN. However, in the last quarter of this decade a few DCS techniques were proposed putting focus on enhancement of packet routing. These different routing algorithms can broadly be divided into two categories namely point-to-point routing and tree based hierarchical routing. In point-to-point routing the deployed network field is divided into zones or sectors and data usually propagates from one zone or rendezvous node to another in a multi-hop point-to-point fashion. In contrast, tree based hierarchical routing relies on a tree-construction technique dividing the whole network into a tree-structure and provides a mapping of data transfer paths with minimal assumption about the underlying infrastructure. DIM [13], KDDCS [42] are developed based on K-D trees while HVGR [33] and PathDCS [32] were developed on top of hierarchical Voronoi Graph and path based tree structure respectively. Table 5 summarizes DCS schemes based on these two types of routing algorithm used.