Files
Mouchet_37541200_2017.pdf
Open access - Adobe PDF
- 4.45 MB
Details
- Supervisors
- Faculty
- Degree label
- Abstract
- In many industrial processes, faults are susceptible to occur and can sometimes have dramatic and/or expensive consequences due to failures and unplanned downtimes. Therein, operators have always been interested in detecting fault occurrences as early as possible. Though at beginning, fault detection methods relied mostly on the operator’s expertise and knowledge, growing interest has been paid to automatic data-based fault detection tools. In that regard, data analysis and machine learning has been recently used to develop fault detection methods. The advantage of those is that they require few information on the process except a large set of data. This dissertation has two main goals. The first one is to deliver an overview of the fault detection field so that we can easier wander around that field. Therefore, the three first chapters will be dedicated to the presentation of a few methods that turns out to give nice results in some articles. We will also mention several commercial software and their approach to fault detection. The second goal of this work is to implement fault detection methods on data from wind turbines. Therefore, we will make extensive use of neural networks. We will compare results of simple methods with those of more complicated ones. We will see that more powerful (model-based) methods lead to non-negligible time gain in fault detection. We will also tackle the problem of dealing with slight changes in the process (e.g. due to aging/component replacement) by progressively updating the training data set and relearning the model for the process. Through this work, we hope to provide an insight on fault detection methods and to demonstrate the possibilities of some of them. Further work may be dedicated to the evaluation of the monetary value of those methods (how we can capitalize on the time gained in fault detection, how much is that gained time worth).