No description available.

Assessment of machine health and prediction of future failures are critical for maintenance decisions. Many of the existing methods use unsupervised techniques to construct health indicators by measuring the disparity between the current state and either the healthy or the faulty states of the system. This approach can work well, but if the resulting health indicators are insufficient there is no easy way to steer the algorithm towards better ones. In this thesis a new method for health indicator construction is investigated that aims to solve this issue. It is based on measuring distance after transforming the sensor data into a new space using a feed-forward neural network. The feed-forward neural network is trained using a multi-objective optimization algorithm, NSGA-II, to optimize criteria that are desired in a health indicator. Thereafter the constructed health indicator is passed into a gated recurrent unit for remaining useful life prediction. The approach is compared to benchmarks on the NASA Turbofan Engine Degradation Simulation dataset and in regard to the size of the neural networks, the model performs relatively well, but does not outperform the results reported by a few of the more recent methods. The method is also investigated on a simulated dataset based on elevator weights with two independent failures. The method is able to construct a single health indicator with a desirable shape for both failures, although the latter estimates of time until failure are overestimated for the more rare failure type. On both datasets the health indicator construction method is compared with a baseline without transformation function and does in both cases outperform it in terms of the resulting remaining useful life prediction error using the gated recurrent unit. Overall, the method is shown to be flexible in generating health indicators with different characteristics and because of its properties it is adaptive to different remaining useful life prediction methods. / Estimering av maskinhälsa och prognos av framtida fel är kritiska steg för underhållsbeslut. Många av de befintliga metoderna använder icke-väglett (unsupervised) lärande för att konstruera hälsoindikatorer som beskriver maskinens tillstånd över tid. Detta sker genom att mäta olikheter mellan det nuvarande tillståndet och antingen de friska eller fallerande tillstånden i systemet. Det här tillvägagångssättet kan fungera väl, men om de resulterande hälsoindikatorerna är otillräckliga så finns det inget enkelt sätt att styra algoritmen mot bättre. I det här examensarbetet undersöks en ny metod för konstruktion av hälsoindikatorer som försöker lösa det här problemet. Den är baserad på avståndsmätning efter att ha transformerat indatat till ett nytt vektorrum genom ett feed-forward neuralt nätverk. Nätverket är tränat genom en multi-objektiv optimeringsalgoritm, NSGA-II, för att optimera kriterier som är önskvärda hos en hälsoindikator. Därefter används den konstruerade hälsoindikatorn som indata till en gated recurrent unit (ett neuralt nätverk som hanterar sekventiell data) för att förutspå återstående livslängd hos systemet i fråga. Metoden jämförs med andra metoder på ett dataset från NASA som simulerar degradering hos turbofan-motorer. Med avseende på storleken på de använda neurala nätverken så är resultatet relativt bra, men överträffar inte resultaten rapporterade från några av de senaste metoderna. Metoden testas även på ett simulerat dataset baserat på elevatorer som fraktar säd med två oberoende fel. Metoden lyckas skapa en hälsoindikator som har en önskvärd form för båda felen. Dock så överskattar den senare modellen, som använde hälsoindikatorn, återstående livslängd vid estimering av det mer ovanliga felet. På båda dataseten jämförs metoden för hälsoindikatorkonstruktion med en basmetod utan transformering, d.v.s. avståndet mäts direkt från grund-datat. I båda fallen överträffar den föreslagna metoden basmetoden i termer av förutsägelsefel av återstående livslängd genom gated recurrent unit- nätverket. På det stora hela så visar sig metoden vara flexibel i skapandet av hälsoindikatorer med olika attribut och p.g.a. metodens egenskaper är den adaptiv för olika typer av metoder som förutspår återstående livslängd.

Remaining useful life (RUL) prediction is a problem that researchers in the prognostics and health management (PHM) community have been studying for decades. Both physics-based and data-driven methods have been investigated, and in recent years, deep learning has gained significant attention. When sufficiently large and diverse datasets are available, deep neural networks can achieve state-of-the-art performance in RUL prediction for a variety of systems. However, for end users to trust the results of these models, especially as they are integrated into safety-critical systems, RUL prediction uncertainty must be captured. This work explores an approach for estimating both epistemic and heteroscedastic aleatoric uncertainties that emerge in RUL prediction deep neural networks and demonstrates that quantifying the overall impact of these uncertainties on predictions reveal valuable insight into model performance. Additionally, a study is carried out to observe the effects of RUL truth data augmentation on perceived uncertainties in the model.

Reducing unscheduled maintenance is important for aircraft operators. There are significant costs if flights must be delayed or cancelled, for example, if spares are not available and have to be shipped across the world. This thesis describes three methods of aircraft health condition monitoring and prediction; one for system monitoring, one for forecasting and one combining the two other methods for a complete monitoring and prediction process. Together, the three methods allow organizations to forecast possible failures. The first two use decision trees for decision-making and genetic optimization to improve the performance of the decision trees and to reduce the need for human interaction. Decision trees have several advantages: the generated code is quickly and easily processed, it can be altered by human experts without much work, it is readable by humans, and it requires few resources for learning and evaluation. The readability and the ability to modify the results are especially important; special knowledge can be gained and errors produced by the automated code generation can be removed. A large number of data sets is needed for meaningful predictions. This thesis uses two data sources: first, data from existing aircraft sensors, and second, sound and vibration data from additionally installed sensors. It draws on methods from the field of big data and machine learning to analyse and prepare the data sets for the prediction process.