Thesis (MEng)--Stellenbosch University, 2015. / ENGLISH ABSTRACT: Most industries are being forced to operate at lower costs while delivering more outputs and ensuring a safe working environment. An opportunity to achieve this for asset intensive industries lies within the complex and integrated field of Physical Asset Management (PAM). This study is specifically concerned with the maintenance subset of PAM, more specifically, the proactive maintenance strategy. A field known as prognostics emerges when combining two maintenance tactics, namely predictive and preventative maintenance. Prognostics uses historical failure data from preventative maintenance and variable readings used in predictive maintenance to estimate asset reliability. Reliability is estimated using statistical models commonly known as reliability models or survival models. Variable readings used must describe or portray the health of the assets considered and are called covariates. A problem that exists in the maintenance subset of PAM is concerned with the data needed for the survival models. The historical failure data is difficult to come by or non-existent in industry and the covariate data is often noisy and inaccurate. This poses a problem when wanting to make important maintenance decisions because the prognostics survival models require both the historical failure data and the covariate data. The covariate data is generally acquired by applying Condition Monitoring (CM) to assets, monitoring characteristics reflecting the asset’s health. Prognostics can aid with maintenance decisions because once the equipment reliability has been estimated, it is possible to predict the time that an asset can still operate at its prescribed level of performance. This time of operation, which the asset can still operate, is more commonly known as its residual life (RL). To overcome this problem, six of the most popular survival models found in literature, namely the Accelerated Failure Time Model (AFTM), Additive Hazards Model (AHM), Proportional Covariate Model (PCM), Proportional Hazards Model (PHM), Proportional Odds Model (POM) and the Prentice, Williams and Peterson (PWP), are considered and populated with historical failure data and the covariate data elicited from people. The people whom the data is obtained from are considered as experts in the field this study is conducted in. Also, the data is subjective because each expert has their own opinions and judgement concerning the assets in this study. The purpose of this study is, thus, to investigate whether subjective data can be used to populate survival models, therefore, allowing RL predictions of the assets considered. A guideline consisting of five steps that aid with what system variables to consider as covariates, which people can be selected as experts and selecting the most appropriate survival model, is created and presented. Following the guideline, a case study is conducted on power transformers at an organization in South Africa. Results from the case study reveal that the PCM is the most appropriate survival model reviewed. Using the PCM, RL predictions are made after the models are populated with subjective data and objective industry standard data. The results indicate that the subjective data yielded the same general trends but less conservative estimates when compared to industry standard data. Subjective data can, therefore, be used to populate survival models but this is inherently risky because of the less conservative results noted from this study. This study is based on a single case study, it does prove that it is possible to use the subjective data as an alternative to objective data. It is possible, however, that this characteristic does not apply for other asset types. / AFRIKAANSE OPSOMMING: Die meerderheid nywerhede word onder geweldige druk geplaas om laer bedryfskostes te handhaaf en ter selfde tyd word dit van hulle verwag om hul uitsette te vermeerder en ´n veilige werksomgewing te bied. Bate intensiewenywerhede het ´n geleentheid om hierdie druk te verlig deur gebruik te maak van ´n komplekse en geïntegreerde veld bekend as Fisiese Batebestuur (FB). Hierdie studie is gefokus op die instandhouding onderafdeling van FB, spesifiek die proaktiewe instandhoudingsstrategie. Twee proaktiewe instandhoudingstaktieke, naamlik voorspellende en voorkomende instandhoudingtaktieke, word saamgesmelt en vorm ´n veld bekend as prognostiek. Prognostiek gebruik historiese falingdata van voorkomende instandhouding en veranderlike aflesings vanaf toestandmoniteering toeristing gebruik in voorspellende instandhouding om bate batroubarheid te bereken. Hierdie betroubaarheid word bereken deur gebruik te maak van statistiese modelle bekend as oorlewingsmodelle. Een van die probleme wat voorkom in die instandhouding onderafdeling van FB het te doen met die beskikbaarheid van die data wat benodig word vir die oorlewingsmodelle. Historiese falingdata is selde beskikbaar of bestaan glad nie en die toestandsmoniteering data is dikwels onakuraat. Prognostiek word gebruik om belangrikke instandhoudingsbesluite te motiveer, dus is die beskikbaarheid en betroubaarheid van die nodige data van belange. Om hierdie struikelblok te oorkom bestudeer hierdie studie die gebruik van subjektiewe data bekom vanaf deskundiges in prognostieke oorlewingsmodelle. Die doel van hierdie studie is dus om vas te stel of subjektiewe data gebruik kan word in prognostieke oorlewingsmodelle. Ses oorlewingsmodelle wat gereeld voorkom in literatuur word nagesien in hierdie studie, die modelle sluit in die “Accelerated Failure Time Model” (AFTM), “ Additive Hazards Model” (AHM), “Proportional Covariate Model” (PCM) , “Proportional Hazards Model” (PHM), “Proportional Odds Model” (POM) en die “Prentice Williams and Peterson” (PWP) model. Hierdie modelle word aangevul deur die subjektiewe data wat onttrek is van deskundiges in ´n sekere gebied, vir hierdie studie is die gebied krag transformators. Met gebruik van hierdie modelle kan die betroubaarheid van die betrokke toerusting bereken word. Sodra die betroubaarheid bereken is kan die oorblywende lewe van die toerusting voorspel word. Die oorblywendelewe is die tyd wat ´n stuk toerusting nog moontlik kan werk sonder om te faal. Dit is belangrik omdat nodige instandhoudingsbesluite geneem moet word. Hierdie studie stel ´n metode voor vir die uitvoer van die navorsing en soortgelyke studies. Die metode dui vyf stappe aan wat voorstel watter veranderlikes om te gebruik as kovariate in die oorlewingsmodelle, watter mense as deskundiges gekies kan word, en hoe om die mees toepasslikke oorlewingsmodelle te kies. Nadat hierdie metode voorgestel is word dit toegepas op krag transformators in ´n gevallestudie wat plaasgevind het in Suid Afrika. Vir die gevallestudie is die PCM die meestoepaslikke oorlewingsmodel. Die oorblywende lewe voorspellings wat die metode opgelewer het is met die voorspellings gebaseer op die industriestandaard data vergelyk. Die resultate dui aan dat deskundiges minder konserwatiewe beramings lewer. Dus kan die subjektiewe data gebruik word in oorlewingsmodelle maar die beramings is minder konserwatief en daarom van natuur meer riskant. Hierdie studie se gevolgtrekkings is gebaseer op ´n enkele gevallestudie. Dit is dus moontlik dat die subjektiewe data dalk nie as ´n alternatief gebruik kan word met ander tipes toerusting nie.

Machine Learning is improving at being able to analyze data and find patterns in it, but does machine learning have the capabilities to predict something subjective like a movie’s rating using exclusively objective data such as actors, directors, genres, and their runtime? Previous research has shown the profit and performance of actors on certain genres are somewhat predictable. Other studies have had reasonable results using subjective data such as how many likes the actors and directors have on Facebook or what people say about the movie on Twitter and YouTube. This study presents several machine learning algorithms using data provided by IMDb in order to predict the ratings also provided by IMDb and which features of a movie have the biggest impact on its performance. This study found that almost all conducted algorithms are on average 0.7 stars away from the real rating which might seem quite accurate, but at the same time, 85% of movies have ratings between 5 and 8, which means the importance of the data used seems less relevant.