| Περίληψη : | Σκοπός της παρούσης διπλωματικής εργασίας είναι η μελέτη του προβλήματος προσδιορισμού ακραίων τιμών σε ασύρματα δίκτυα αισθητήρων. Συγκεκριμένα στόχος είναι η εξέταση των υπαρχουσών μεθόδων προσδιορισμού ακραίων τιμών στο προαναφερόμενο περιβάλλον, καθώς και η πρόταση νέων αλγορίθμων και μεθόδων σε αυτήν την κατεύθυνση. Οι μέθοδοι που προτείνονται χρησιμοποιούν και εκμεταλλεύονται υπάρχουσες οργανώσεις ασύρματων δικτύων αισθητήρων σε συστάδες και αφορούν την ανάπτυξη σχημάτων προσδιορισμού ακραίων τιμών υιοθετώντας την έννοια του κατακερματισμού. Κάτι τέτοιο αποτελεί καινοτομία καθώς καμία από τις μεθόδους που μελετήθηκαν σε σχετικές ερευνητικές εργασίες δεν προβαίνει στην ανάπτυξη τεχνικών με ανάλογο τρόπο. Η πειραματική αξιολόγηση των προτεινόμενων μεθόδων, αναδεικνύει και επικυρώνει την εφαρμοσιμότητα, την ακρίβεια και την αποδοτικότητα από άποψη κατανάλωσης ισχύος των προτεινόμενων προσεγγίσεων.
Wireless sensor networks are becoming popular in many environmental, industrial, military and other application areas. Sensor nodes or motes constitute inexpensive, tiny devices equipped with limited power supply resources, low processing and constraint memory capabilities that are placed in, often harsh, environments of interest. Their role regards the collection of data in a continuous or on demand fashion so as to serve sampling procedures, develop alert mechanisms, detect abnormal behavior in the studied environment as well as support decision making processes.Despite their restricted capabilities, sensor nodes form powerful networks due to synergy effects. In wireless sensor networks topologies are built ad-hoc, while communication is achieved by taking advantage of the overlaps between mote radio ranges. Administrators and users of the network are able to pose queries and receive corresponding answers examining undergoing phenomena. The obtained measurements that are taking part in a query answer are transmitted throughout the network and reach the query point in a multihop fashion.Because of hardware limitations motes are often prone to failures involving the acquisition of spurious measurements or total malfunction. As a result, network value and utility are substantially reduced. For instance, in case of using a wireless sensor network as an alert mechanism infrastructure, faulty readings lead to erroneous actuation of the corresponding response procedures. On the other hand, excluding unusual measurements from final query answers would prevent the detection of interesting phenomena that might take place within the network setting and which are obligate to be reported.Therefore, a challenge occurs involving the invention of outlier detection techniques that will enable the detection and report of real outlying values which should be further examined while simultaneously separating faulty measurements. Nonetheless, another primary goal regards the network lifetime prolongation. Network lifetime is usually defined as the time period between the beginning of network function and the timepoint when the network becomes non functional (i.e an important percentage of nodes die, so the network is partitioned). As a result, proposed techniques should minimize power consumption caused by processing and communication activities.The above problem definition is the subject of this thesis. An outlier is defined as a value produced by a sensor node and substantially differs from the measurements of its peers. Based on the aforementioned outlier definition, we develop innovative outlier detection techniques that make good use of existing clustered network organization approaches.Clustering is a useful approach to reduce energy dissipation in sensor networks. It is often coupled with data fusion to extend sensor lifetime. Each cluster elects one node as the cluster head. Data collected from sensors are sent to the cluster head first, and then forwarded to a sink or base station. Compared to the source of the query, a cluster head is closer to sensors within the same cluster. Cluster heads can fuse data from sensors to minimize the amount of data to be sent to the source. This can greatly reduce the energy consumption of sensor networks. We take advantage of the underlying network organization by electing certain motes as bucket nodes to which compact data representations of sensor measurements are transmitted.The proposed techniques foster the concept of hashing to group similar sensor value vectors into bucket nodes. To the best of our knowledge this is the first outlier detection mechanism that adopts the idea of hashing to develop a suitable scheme. The ultimate goal of the grouping achieved via hashing is the comparison of corresponding vectors and the determination of those nodes that produce unusual, outlying measurements. A support parameter is used to act as an indicator of nodes lacking an adequate number of similar value vectors. The general architectural framework proposed comprises of the following stages:• Sensor nodes obtain measurements of monitored attributes and temporarily store values in a window of predefined size.• After collecting a number of values equal to the window size, motes transform those data so as to derive compact representations and achieve reduced energy consumption. To do so certain time series analysis and locality sensitive hashing techniques are utilized.• Each node sends the constructed representations of the data to their corresponding clusterhead.• Clusterheads determine the bucket node to which data should be hashed based on a hash key obtained from time series linear models or locality sensitive hashing bitstrings. • Bucket nodes receive data representations and proceed to vector comparison and outlier report. We further provide a thorough investigation of existing similarity measures and the way they could be incorporated in our schemes. Experimental evaluation results are presented to exhibit the applicability, high accuracy and energy efficiency of our methods.
|
|---|
Copyright © 2013 Βιβλιοθήκη Ο.Π.Α. Με την επιφύλαξη παντός δικαιώματος.