Predictive maintenance for power plants.

The present PhD thesis delves into the topic of predictive maintenance (PdM), which is a maintenance strategy for detecting, predicting, and planning the maintenance needs of machinery based on data obtained from installed sensors. The main aim of the thesis is to explore several aspects of PdM. Three major topics, as described in the included research questions, are addressed through the thesis and the contributed papers. First, the scalability of PdM models is addressed, as it is important to understand how best to implement a PdM system that is maintainable, robust, and requires little fine-tuning. This is explored mainly through Papers A and B included in the thesis. Paper A compares classical, statistically based models from the field of statistical process control with more advanced deep learning models for condition monitoring (CM). Paper B presents a case study of CM using a simple exponentially weighted moving average model for CM of a particular generator fault in a large population of offshore wind turbines. In these two papers, we find that the simplicity of models should be preferred, as it provides several benefits, such as sufficiently good model performance, interpretability, robustness to sparse amounts of data and diverse operating conditions, and lastly, vastly increased likelihood of adoption by maintenance personnel due to their more transparent nature compared to black-box models such as deep learning neural networks. Second, the economic and maintenance benefits of a PdM system at scale are explored along with potential challenges. This is primarily addressed in Papers B and C. Paper C is an overview and comparison of the PdM maintenance strategy with more traditional maintenance approaches. It also delves into relevant literature to explore the potential benefits of PdM. In Paper B, we find that the economic benefits of a large-scale PdM system, in this case for CM of wind turbines, can be highly economically beneficial as large savings are shown to be achieved if faults can be detected with enough lead time to enable optimal planning of maintenance visits. Similarly, we find through Paper C that there are indeed considerable benefits to be had from a well-functioning PdM system despite large initial investments and potential implementation difficulties. The third issue addressed throughout the project is the optimal utilization of available data and exploring how to derive value from available data. This is addressed in Paper D, which presents an existing feature fusion model for extracting improved equipment health indicators from existing data by fusing it in a linear combination. It does this by optimizing a set of metrics that measure any given feature’s suitability as a health indicator and, subsequently, prediction of the remaining useful life of the targeted equipment. Moreover, two novel models are proposed that extend the original model in that they can model non-linear relationships. It is found in the paper that health indicators obtained from one of the non-linear models may provide a slight performance gain in terms of RUL prediction errors compared to the linear model. However, another important message of this paper is that some data may simply be unsuited for the task, such as RUL prediction. This is the case if there is not enough information in the data related to the remaining lifetime of the equipment. Lastly, we also find that cumulative operation hours are a universally applicable predictor of RUL. Lastly, we provide perspectives on the research undertaken during the project and how it relates to business value, as insights on how best to implement a PdM system in a large industrial organization were the main drivers that initiated the PhD project.