Abstract
In modern industrial operations, the reliability and efficiency of machinery is crucial forminimizing operating costs and thus maintaining competitiveness. Predictive maintenance has become a key strategy in this regard, enabling industry to foresee equipment
failures before they occur and thus reduce unplanned downtime, saving resources in the
long run. However, the challenge lies in accurately predicting the timing and nature of
these failures, especially for non-linear dynamic systems that are inherently complex
and for which there are often no simple analytical solutions.
In this thesis, the application of the Sparse Identification of Nonlinear Dynamics (SINDy)
concept is investigated in the context of predictive maintenance. SINDy is a data-driven
methodology for identifying the equations of nonlinear dynamic systems from time series
data. By using sparse techniques, SINDy is able to extract the most important dynamics from a pool of potential candidate functions, resulting in a simplified but accurate
mathematical model of the analyzed system.
The main objective of this thesis is to investigate how SINDy can be used to apply predictive maintenance to industrial machinery. Traditional predictive maintenance methods often rely on empirical models or machine learning algorithms that may not fully
capture the underlying physics of the systems, leading to potential inaccuracies in failure prediction. In contrast, SINDy offers a more physics-based approach that is able to
uncover the underlying dynamics of machine behavior. This capability is particularly
valuable for modeling wear processes as well as for detecting deviations in system parameters that could indicate impending failures.
The methodology used in this thesis involves the application of SINDy to a simulated
data set of a non-linear oscillator exhibiting complex dynamics with a simple system
equation. In addition, an error is applied to the system equation to simulate the wear
behavior of a system.
The results of this study show that SINDy not only effectively captures the nonlinear
dynamics of industrial systems, but also offers significant advantages in terms of model
interpretability and computational efficiency. The models generated by SINDy are typically sparse, i.e. they ideally contain the minimum number of terms to describe the
nonlinear dynamical system, which facilitates their interpretation and analysis. Moreover, these models are computationally efficient to evaluate, which makes them well
suited for real-time monitoring and prediction tasks.
In summary, this work represents an experimental approach to the use of SINDy in
predictive maintenance and demonstrates potential to change the way industry approaches the maintenance of complex machinery. The results suggest that SINDy can
increase the accuracy of failure predictions, with the benefit of providing a deeper understanding of the underlying dynamics of industrial systems. As the industry adopts
more sophisticated predictive maintenance techniques, SINDy proves to be a powerful tool to overcome the gap between data-driven modeling and practical maintenance
applications.
| Date of Award | 2024 |
|---|---|
| Original language | English (American) |
| Supervisor | Hans-Georg Brachtendorf (Supervisor) & Martin Steiger (Supervisor) |