2015 Seventh International Conference on Knowledge and Systems Engineering DF-AMS: Proposed solutions for multi-sensor data fusion in wireless sensor networks Duong Viet Huy Nguyen Dinh Viet Science, Technology and Environ ment Depart ment Ministry of Culture, Sports and Tourism Hanoi, Vietnam huy.duongviet@gmail.co m University of Engineering and Technology Vietnam National Un iversity, Hanoi Hanoi, Vietnam vietnd@vnu.edu.vn Abstract— When using multiple sensor nodes in wireless sensor networks (WSNs) for monitoring (measuring) parameters of the target and sending the result to the base station (BS ), data redundancy is an inevitable problem The measured data often contains the same information, and sending redundant data to BS causes the waste of energy of sensor nodes and the risk of congestion Multi-sensor data fusion in WSNs is a technology of gathering and processing data applied from node to BS It improves the performance of surveillance systems by allowing the obtained sensed information from multiple sensor nodes aggregated to one unified format data packet to send to BS to make decision In this paper, we propose a solution namely DF-AMS (Data Fusion - Average Median Sampling) for sampling and fusing data from the sensor nodes in the cluster to optimize the energy conservation of sensor nodes in clusters and cluster head node to the actual situation because it is difficult to quantify all events from sensing signal and this data is transmitted through a long chain of intermediate nodes to the destination (BS) Therefore, DF in WSNs is studied to increase the accuracy of the conclusions deduced from uncertainty sensing data while saving the energy consumption of sensor as well as the whole network The model uses multiple sensor nodes (each node can monitor many parameters) to track target then send the result to the destination station (BS) Th is causes the wasted energy of sensor nodes and increases the risk of network congestion due to transmission of redundant data (Fig 1) DF on the intermediate nodes between the first sensor node (directly tracks the target) to the BS helps save energy by limit ing the movement of packets on the network (but still preserve the integrity of monitoring data) There are many studies in DF such as Kalman filtering [2], Dempster-Shafer [3], fu zzy logic [4], neural network [5], information entropy [6] ect Most of DF studies have applied to WSNs with clustering, each cluster has a sensor node selected as the cluster head - CH, it processes data from all member nodes in this cluster and sends data fusion result to BS Keywords- wireless sensor networks; data fusion; data aggregation; median sampling; DF-AMS I INT RODUCTION (HEADING 1) Data fusion (or data aggregation) has been a key problem in the monitoring system using sensor nodes increasing in size (nu mber of sensor nodes, the scope of monitoring) and complexity (nu mber of parameters being monitored, fineness of measurement) Typically, sensor nodes are powered by batteries with limited capacity thus energy saving of node and network are always priorit ised in WSNs research There are many advantages in using multip le sensor nodes to track a target such as [1]: tolerance of error (because the results of remaining non-error sensors could be used); better stability and reduction in perceived noise data (as many sensor nodes will increase the dimension of measurement space); multiple independent measurements fro m multip le sensors improve spatial resolution measurements and provide a better source of input data; mu ltiple sensors ensure the continuity of available data, and improve the ability of detection of missed-track events However, DF from mu ltiple sensors also poses many challenges [1] Current technology cannot create sensors which are able to record accurately absolute change of events The data processing in the transmission step from target to BS cannot fix the previous processing step Results of data fusion from mult iple sensor for sensing the same event/object may be worse than data of one single good sensor There is no algorith m that can satisfy all criteria It is very difficult to quantify the overall quality of data compared 978-1-4673-8013-3/15 $31.00 © 2015 IEEE DOI 10.1109/KSE.2015.28 Data Fusion l l l CH BS Base Station l l S2 … Sensor node Sn l S1 Sensing data Target Fig Data fusion model with n sensor node, each sensor node measure l parameters (l≥2) of the target In this paper, we propose a solution namely DF-AMS which includes two phases: sampling sensor nodes (in clusters) by the properties of the sensor nodes; CH performs data fusion of parameters sensed by selected sensor nodes with the application of median method and maximu m method The paper consists of two main parts: the first part presents scientific basis and proposed solution of DF-AMS; the second part presents analysis of efficiency and simulation evaluation of the proposed solution using NS-2 network simulator II sensor nodes operate longer, the energy difference is larger and the median value is smaller than average In this study the node with the highest remaining energy is chosen to transmit data Therefore, we can choose flexibly the median value if it is greater than the average value, or conversely Several studies have applied solutions Sum, Max, Average, Median to multi-sensor data fusion, for example in [10, 11, 12] Data fusion in the cluster is the role of CH The sensor nodes in the cluster send data to the CH alternately in round, frames and time slots in accordance with CDMA and TDMA access In one round, total number of data packets that CH received fro m the nodes in the cluster are usually not much different Therefore, DF solution with median value and mean value can have the similar results With the current measurement by sensor technology, it is difficult to rely on one sensor node which has the largest measured value (Max) Therefore, we use the combined values of median, average, max to support each other to increase reliability and improve energy efficiency DF-A MS SOLUT ION A Concepts 1) Properties conditions The Rough set theory [7] has been studied and proved to be applicable to multi-sensor data fusion in wireless sensor networks [8] The paper will apply the concept of conditional attributes to model "knowledge" data based on the semantics of sensor nodes at the time of DF, including: the semantics of sensor nodes (such as distance, remaining energy, ect.) and semantics of data sense (such as accuracy, number of packets to transmit, ect.) Suppose that a sensor network has n nodes, each sensor node is characterized by the combination of m condition attributes, then the semantic data table has n rows and m colu mns CH bases on this condition attributes to select sensor node, then transmit or fuse data of this particular sensor node to the BS 2) Data attributes, sensor node attributes Data attributes and sensor node attributes are conditional attributes which are used on the semantic aspects, meaning of sensor nodes at the time of DF including physical characteristics and properties of the sens ed data of the target The number of data attributes, sensor node attributes can be changed and quantified by different values depending on the method and the level of required accuracy in modelling sensor network by data attributes Data and sensor node attributes can be used to model sensor networks at a certain time as an information system reflecting (assessment, describing) the current state of the network to serve different purposes In the multi-sensor data fusion problem presented in this paper, data attributes and sensor node attributes are quantified by the values for the most accurate description of sensor nodes and sensing data at the time just before implementing of data fusion Data attributes and sensor nodes attributes are used as inputs in selecting the sensor nodes (output) based on attribute information 3) The average value, the median value For wireless sensor networks, sensor nodes are often scattered randomly which results in different distance values and the relative distances between sensor nodes are often not equal During network operation, the sensor nodes process and transmit electromagnetic waves leading to the loss of energy of nodes When a node transmits data to the destination node by radio waves, it has to amplify radio transmitting power by a function of the square of the distance [9], and the energy consumption of node exponentially proportional to the distance between them At a point, remaining energy of sensor nodes in the network usually have different values If the energy is arranged increase (or decrease), the separation between two random node is not uniform When the network starts operation, the energy of the nodes is full and the variation is small, the median value is often greater than the average value Conversely, when B DF-AMS 1) Sampling A WSN can be divided into several clusters, in this paper, the term cluster is understood as the sub-WSN which cannot be further subdivided and data fusion will be applied to this cluster We use one level data fusion model, means that the CH has only one intermediary ro le node between the sensor in clusters and BS (Fig.1) The sample was selected on both data transmission stages: First, in n sensors of cluster, select q sensor node (qdn) based on the remaining energy of sensor node CH, then require q sensor node to send data to CH; Second, at CH, each q sensor is a sensing data set (each set with k elements, each element is one parameter measurement target) For each parameter, CH chooses q/2 (if q is even) or (q+1)/2 (if q is odd) sensor fuse the data into data set about the target and send this result to the BS Suppose CH has selected q node in n node of cluster satisfying the conditions resulting data fusion with q sensors which have similarly good (or better) results with n sensors We use the remaining energy properties of sensor nodes to select suitable sensor node immed iately after the cluster is established CH can determine the energy of nodes and is arranged in ascending ES = {ES , ES , ES k ,ES n-1 , ESn }, ES =Min, ES n =Max Let Eavg is average remain ing energy, EMed is median value of n sensors in the cluster, Đ n ã E Avg ă Ư ES i / n â ES (n+1)/2 (1) i f n is odd EMed = (2) (ES n/2 + ES (n+1)/2)/2 if n is even Let SDFI = {S i | d i d n, ES i ≥ EAvg or ESi ≥ EMed } a collected sensor which satisfied (1) or (2) and is selected first 3) DF-SM algorithm a) Convention by CH, k is the number of nodes which is not selected by CH (k d n), q is the number of nodes (cardinal number) of SDFI , (3) q = |SDFI| = n-k 2) Data Fusion With SDF.ii SDFI, d ii d q, VSDF.ii is the set of l measurements parameter values of sensor S ii where VSDF.ii = {VSii.M1 , VSii.M2 ,…, VSii.Ml } and l≥2 Then, table data received (at CH) fro m q sensor quantitative measurements of l parameters of the target is a matrix size (q x l) in Table Sign n Si ESi EAvg EMed ESelect SDFI q SDF.ii l Mj TABLE – MEASURED DAT A AT DATAFUSION Sensor node SDF.1 S DF.2 … SDF.ii … S DF.q M1 VSDF.1.M1 VSDF.2.M1 … VSDF.ii.M1 … VSDF.q.M1 Measurement parameters M2 … VSDF.1.M2 … VSDF.2.M2 … … … VSDF.ii.M2 … … … VSDF.q.M2 … Ml VSDF.1.Ml VSDF.2.Ml … VSDF.ii.Ml … VSDF.q.Ml J µ SDFII Med.Sq.M1 Avg(Ml, µ) Vmax.qMl VDF.Mj.sent In this paper, we assume that the higher the amount of data (packets / data) which sensor nodes produce the more accurate and more knowledge of the target properties Suppose, for each parameter, the sensor can measure the number of d ifferent levels This is consistent with the fact that although sensor manufacturers strive to measure the value change to reflect the best target to ensure the smooth and asymptote to the variation of the target to be monitored However, it is very difficult to manufacture this sensor b) Data processing model The algorith m consists of two main processing phases as shown on Fig Phase 1, immed iately after the cluster is established, based on the remain ing energy of sensor node, CH selects sensor node with more energy than average (or med ian) to be in SDFI, only the selected nodes are allo wed to continue sending data to the CH in next steps Phase 2, just before the end of the cycle, CH selects Ml measurement results of the q sensor nodes and chooses the corresponding med ians Ml The selection of node based on med ian values of Ml has SDFII (with µ sensor node) With each parameter Ml , CH will fuse the data by average values and maximu m value, then send this result to the BS TABLE – LEVEL, MEASURE VALUES OF THE PARAMETERS Le vel X1 X2 XJ Signification Number of sensor nodes of the cluster Sensor ith of cluster (1 d i d n) Remaining energy of Si Average energy of the n sensor node Median value at median position (in n) Energy landmark for reference/select Set of sensor with minimum energy at ESelect Cardinal number of SDFI , q = |SDFI| Sensor iith of SDFI (1 d ii d q) Number of measurements parameter Parameter jth (1 d j d l) Number of measurement level of Ml Median position in q sensor (1 d µ d q) Set of sensor from µ, |SDFII|=µ Ml value at µ Ml average value of µ sensor node Maximum value of Ml in q sensor node Mj last results that CH sent to BS Value of the parameter measurement (each parameter may have a different level) M1 M2 Ml VX 1.M1 VX 1.M2 VX 1.Ml VX 2.M1 VX 2.M2 VX 2.Ml VXJ.M1 VXJ.Ml VXJ.M2 Phase Suppose, for each parameter (M1 , M2 ,…, Ml ), sensor can measure J different level, in Table A value VSii.Ml in Table is getting a VXJ.Ml in Table CH Fro m Table 1, sorted ascending SDF.q measured values of Ml respectively Let Med.S q.M1 = VSDF.µ.Ml be Ml median value at median position µ of q sensor, where µ=(q/2)+1 (if q is even) or µ=(q+1)/2 (if q is odd) SDFII is set of sensor node satisfying conditions VSDF.ii.Ml ≥ Med.Sq.M1 (µ d ii d q) SDFII is a matrix size (µ x l) (4) Let Avg(Ml , µ) be a Ml average value of µ sensor node Let Vmax.q Ml = Max (VSDF.q.Ml ) be Ml maximu m value of Ml parameter measurement of q sensor node (5) Then, CH sends sensing data to BS (including l values) about the target at time of DF as follows: VDF.Mj sent = (Avg(Ml , µ) + Vmax.q Ml )/2 (6) Ml Median Ml CH sorted by energy ( n nodes) … SDF I (q nodes) M2 Median M2 M1 Median M1 EAvg / EMed SDF II ( µ) … M1 M2 … Ml … AVG MAX M1 sent Phase AVG MAX M2 sent … AVG … Base Station (BS) Fig Data processing model DF-AMS MAX Ml sent c) 10 11 12 13 14 15 16 17 18 19 20 TABLE THE MAIN PARAMETERS Algorithm Set n = num_cluster_nodes; SDF I = Ø For {set i 1} {$i