Zur Kurzanzeige

dc.contributor.advisorDiwanji, Prajakta
dc.contributor.approverWache, Holger
dc.contributor.authorSchweizer, Daniel
dc.date.accessioned2017-10-04T05:33:13Z
dc.date.available2017-10-04T05:33:13Z
dc.date.issued2014-01-31
dc.identifier.urihttp://hdl.handle.net/11654/25535
dc.identifier.urihttp://dx.doi.org/10.26041/fhnw-1177
dc.description.abstractThis paper discusses how usage patterns and preferences of smart home inhabitants can be learned efficiently. Such patterns as a baseline of what constitutes normal behavior of inhabitants allows future smart homes to autonomously achieve positive effects like comfort increases, energy savings or improved safety for elderly residents in assisted living homes. The approach for learning usage patterns chosen by this research project, which was carried out as a Master Thesis at FHNW, uses frequent sequential pattern mining algorithms to mine the event data available in smart homes. While other authors have already published possible solutions or at least approaches to the problem, the information presented herein is unique because it is based solely on reallife smart home event data and not data collected in a laboratory trial and/or enriched by additional sensors. Furthermore the project does not only propose one solution but compares the performance of different algorithms regarding completeness/correctness of the results, run times as well as memory consumption and elaborates on the shortcomings of the different solutions. To be able to solve the challenge of learning usage patterns, this project followed the research onion framework by Saunders, et al. (2009) and the design science research paradigm by Vaishnavi & Kuechler (2004): after a research design and a literature review was done, the available secondary data was analyzed in depth before different solutions (including a brute-force algorithm specifically designed for this project as well as adaptations of the three established frequent sequential pattern mining algorithms PrefixSpan, BIDE+ and GapBIDE) were designed, implemented as prototypes in Java and benchmarked against each other. The main findings of the benchmarking done with the prototypes and of the project as such were:  With all four algorithms a reasonable amount of frequent sequential patterns can be found with an input parameter set of pattern length = 2-5 events, minimum support = 0.01 – 0.001, overlapping patterns, wildcarding deactivated.  The traditional frequent sequential pattern mining algorithms like PrefixSpan, BIDE+ or GapBIDE need pre- and post-processing to be able to mine smart home event data. Additionally, if different minimum and maximum lengths of patterns shall be mined, those algorithms need to be run multiple times to report the correct support count.  The run times vary greatly for the different algorithms, BIDE+ being the slowest of the four algorithms. Both GapBIDE and especially PrefixSpan run significantly faster, however, they are outperformed by the brute-force algorithm WSDD developed for this project.  Wildcarding could not fulfill the potential attributed to it at the beginning of the project because no significantly higher support counts can be found with wildcarding being activated.  While WSDD as the fastest algorithms can be recommended without reservations regarding run times, all four benchmarked algorithms showed bad results regarding memory consumption for certain combinations of input parameters. This paper therefore contains six different propositions for lowering the memory consumption, should memory consumption be a concern.  While the aspect of periodic sequential pattern mining was investigated as part of this research project and a manual analysis of the available data showed that periodic patterns exist in smart home events, no satisfactory mining results could be achieved and it is therefore suggested to look into this aspect in a follow-up research project (e.g. by adapting a state-of-the-art periodic (sequential) pattern mining algorithm to the specifics of smart home event data). Overall it can be said that the thesis could prove part of its statement as true: it is possible to learn, in a run-time efficient way, frequent usage patterns and preferences of smart home inhabitants only from event data available within a smart home. With different use cases it could be shown in a theoretical way that the patterns can be used to achieve the aforementioned positive effects like saving electricity or increasing the comfort of smart home inhabitants.
dc.language.isoen
dc.accessRightsAnonymous
dc.subject.ddc004 - Computer Wissenschaften, Internetde
dc.titleLearning frequent and periodic usage patterns in smart homes
dc.type11 - Studentische Arbeit
dc.audienceScience
fhnw.StudentsWorkTypeMaster
fhnw.publicationStateUnpublished
fhnw.ReviewTypeNo peer review
fhnw.InventedHereYes
fhnw.PublishedSwitzerlandYes
fhnw.IsStudentsWorkyes


Dateien zu dieser Ressource

Thumbnail

Der Eintrag erscheint in:

Zur Kurzanzeige