Original title:
Multivariate Anomaly Detection System with System-Dynamics-Reflective Neural Architectures
Authors:
HRUBÝ, Filip Document type: Master’s theses
Year:
2024
Language:
eng Abstract:
The aim of this thesis is to propose and explore machine learning based anomaly detection in dynamical systems for industrial applications where training data is limited and subject to noise and disturbances. The focus is on neural network architectures that reflect the system dynamics, i.e. where analogies can be drawn to the state space representation of dynamic systems in a discrete-time domain. A comparison of such architectures with conventional neural networks such as multilayer perceptron or LSTM networks is shown. First, the background methods are defined, followed by an examination of the different neural architectures and learning algorithms and their comparison. Then, the dynamical system for generating data is described and data sets with anomalies are generated. The core of the contribution is the proposed method for anomaly detection via a heat map of the neural weight behaviour over time and its investigation and comparison for different networks and learning algorithms. Finally, the results of the algorithms are compared based on the analysis and discussion, and the feasibility of the approach is evaluated.
Keywords:
anomaly detection; dynamical systems; explainable AI; gradient descent; heatmap; higher-order neural units; in-parameter-linear nonlinear neural architectures; Levenberg-Marquart; physics informed machine learning; shallow neural networks Citation: HRUBÝ, Filip. Multivariate Anomaly Detection System with System-Dynamics-Reflective Neural Architectures . České Budějovice, 2024. diplomová práce (Mgr.). JIHOČESKÁ UNIVERZITA V ČESKÝCH BUDĚJOVICÍCH. Přírodovědecká fakulta
Institution: University of South Bohemia in České Budějovice
(web)
Document availability information: Fulltext is available in the Digital Repository of University of South Bohemia. Original record: http://www.jcu.cz/vskp/75389