Data-driven engine state monitoring

Data-driven engine state monitoring: Applying multivariate statistics and machine learning to fault detection and isolation.

Drakoulas, Michail (TU Delft Mechanical, Maritime and Materials Engineering)

Delft University of Technology

2019-08-19

The technological advances of the last decades eventually have led to systems with growing complexity. As a consequence, the task of supervising their condition has become increasingly challenging. Especially for the case of turbocharged diesel engines, this task is even more difficult due to the interaction of the turbocharger with the engine. An effective monitoring system that promptly detects abnormal behavior and provides information regarding the state of the engine is essential to prevent failures and/or system shutdowns. With this paper, the author is proposing a data-driven approach to achieve this objective and then evaluates its performance with the help of a simulation model of a four-stroke diesel engine.
The proposed strategy focuses on evaluating the state of the engine based on a limited number of signals collected during its operation. It applies multivariate methods to analyze the multidimensional collected measurements. At the core of this approach, there is a method known as Independent Component Analysis (ICA), which allows expressing the data in a lower-dimensional space that explains most of the observed variation. This sub-space defined by the dimensionality reduction model makes possible to extract unique features from new deviating observations, thus facilitating the detection and isolation of a problem. Fault detection is achieved with multivariate statistical techniques based on ICA, while fault isolation with the help of a classifier that evaluates the extracted features and attempts to recognize a previously encountered fault pattern.
The application of the proposed monitoring strategy on the simulation model of the case study engine indicated a very robust performance regarding the detection and isolation of the modeled faults. This can largely be attributed to the effectiveness of ICA at exposing abnormal behavior of the system. To the best of the author's knowledge, ICA has not been used in similar applications, thus this work opens the way to further research such data analysis techniques for the purpose of supervising the operation of real engines.

http://resolver.tudelft.nl/uuid:53f1eab4-e01b-4fd6-aca3-536b5260bdbc

2024-08-27

Student theses

master thesis

© 2019 Michail Drakoulas