Neufeld, DenizDenizNeufeld2024-06-062024-06-062024https://fis.uni-bamberg.de/handle/uniba/95413Dissertation, Otto-Friedrich-Universität Bamberg, 2023Hydraulic systems are important in the functioning of everyday life, as well as critical infrastructure by driving various systems such as automotive, aircrafts, and construction machines. To assure their functionality according to specification over their complete lifetime, reliability tests are conducted in test benches to check a representative number of samples for long amounts of time. During the tests, physical system inputs and outputs such as pressures, electric currents and voltages are recorded for analysis. The goal of this thesis is to research anomaly detection methods based on the recorded data to support root cause analysis and reduce testing time. Since the first decision makers in case of an error are the responsible engineers and technicians, it is important to focus on methods that are both robust and easily understandable for non-experts of data science. Several challenges stem from the given problem. With the start of a test bench, there is often no prior measurement data available for model training, and samples are not repeated after a successful test run. Furthermore, there are various kinds of signals with different properties, from digital bus system signals to fluid flow or electrical currents. If a complete system is assessed, different failure modes can occur based on different sub-components. This leads to many possible failure constellations and therefore a large number of relevant features. There are multiple fields of research which are related to this problem, ranging from the general field of time series analysis to more specific condition monitoring research in the domain of engineering. This means there exist similar problem types in state of the art, but also limited closely related work. The following work not only focuses on research on the anomaly detection on hydraulic test benches per se, but also investigates data visualization, anomaly detection in physical systems, multivariate time series anomaly detection, and vibration analysis. The original contribution to knowledge of this thesis are advances in multiple aspects of anomaly detection with a focus on hydraulic test benches. First a collection of methods for the visualization of multivariate, repeating time series is provided. It supports engineers in viewing data and detecting anomalies visually by aligning the signals along the time and the amplitude axis. From this, this thesis examines a model-based method for a digital twin of the tested system using recurrent neural networks. Challenges with this approach are shown and their root cause described in depth. Additionally, data-based methods are examined: One, an unsupervised, statistical method is developed for anomaly detection on multivariate, periodical data in the time domain, as well as its robustness and limits of this method towards concept drift are investigated. It is further shown how the results of the method can be visualized for root-cause analysis. Second, a supervised approach is followed on vibrational data using convolutional neural networks (\acrshort{cnn}s). For this, preprocessing in the frequency-domain its influence on model performance is researched. Due to the black box nature of \acrshort{cnn}s, an explainable artificial intelligence method is developed to make the relevant features of the data in the frequency domain interpretable for engineers and system experts. The XAI method is verified in quantitatively using an accessible data set specifically designed for this task. Based on the shown research, this thesis presents possibilities for future work, specifically with a focus on the enormous amounts of data collected. During a test bench run, millions of time series can be collected, but the visualization and anomaly detection methods developed are not yet perfected towards this fact. Due to the size of the data, dealing with the data and the results of the anomaly detection methods must become more efficient as well, be it for displaying, labeling or for judging the output of the algorithms. While the methods shown can be used on subsets of the data, an important aspect of future work is the clustering of data to reduce the amount of data to support inspection and labeling. Additionally, the higher-level visualizations with a drill-down functionality need to be developed as well to guide users towards relevant information.engAnomaly detectionHydraulic systemsCondition Monitoring004doctoralthesisurn:nbn:de:bvb:473-irb-954138