TY - BOOK

T1 - Statistical process monitoring of networks

AU - Malinovskaya, Anna

PY - 2024

Y1 - 2024

N2 - The digital information revolution offers rich opportunities for scientific progress; however, the amount and variety of data available require new analysis techniques in order to adequately address the growing complexity of processes. These requirements have influenced the development of networks and integrated their application in various disciplines. This thesis addresses the detection of changes in networks, combining network theory and sta-tistical process monitoring to create improved techniques for network monitoring. Considering networks as graph-structured data with either fixed or dynamic nodes and edges or as a model based on artificial intelligence, three directions of network monitoring are identified, namely, random network monitoring where the networks represent random variables, fixed network monitoring where the networks are assumed to be fixed structures, and monitoring of artificial neural networks. The idea of using different modelling techniques and control charts to monitor network-related processes connects contributions in this thesis to the outlined monitoring directions.The first developed approach shows how multivariate control charts can be used to detect changes in dynamic networks of various types generated by the temporal exponential random graph model in an online manner. This monitoring procedure allows for many applications in different disci-plines which are interested in analysing networks of medium size.Next, the monitoring framework to detect anomalies in the network with a given structure but a random process on it by combining the generalised network autoregressive model with nodespecific time series exogenous variables and the cumulative sum control chart based on residualsis presented. This approach can be of particular interest for guaranteeing the safety of the infrastructure but also for foreseeing possible accidents.The third contribution is dedicated to the development of a monitoring procedure for artificial neural network applications that applies a non-parametric multivariate control chart based on ranks and data depths. The core idea is to monitor a low-dimensional representation of input data calledembeddings that are generated by artificial neural networks to detect nonstationarity in a processed data stream. In addition to the development of three monitoring approaches, a fourth contribution, namely the extension from the pure detection of the change point to the identification of its cause is presented. The investigation includes a proposal for an automated inspection procedure, bringing together a control chart for quantile function values and a graph convolutional network.

AB - The digital information revolution offers rich opportunities for scientific progress; however, the amount and variety of data available require new analysis techniques in order to adequately address the growing complexity of processes. These requirements have influenced the development of networks and integrated their application in various disciplines. This thesis addresses the detection of changes in networks, combining network theory and sta-tistical process monitoring to create improved techniques for network monitoring. Considering networks as graph-structured data with either fixed or dynamic nodes and edges or as a model based on artificial intelligence, three directions of network monitoring are identified, namely, random network monitoring where the networks represent random variables, fixed network monitoring where the networks are assumed to be fixed structures, and monitoring of artificial neural networks. The idea of using different modelling techniques and control charts to monitor network-related processes connects contributions in this thesis to the outlined monitoring directions.The first developed approach shows how multivariate control charts can be used to detect changes in dynamic networks of various types generated by the temporal exponential random graph model in an online manner. This monitoring procedure allows for many applications in different disci-plines which are interested in analysing networks of medium size.Next, the monitoring framework to detect anomalies in the network with a given structure but a random process on it by combining the generalised network autoregressive model with nodespecific time series exogenous variables and the cumulative sum control chart based on residualsis presented. This approach can be of particular interest for guaranteeing the safety of the infrastructure but also for foreseeing possible accidents.The third contribution is dedicated to the development of a monitoring procedure for artificial neural network applications that applies a non-parametric multivariate control chart based on ranks and data depths. The core idea is to monitor a low-dimensional representation of input data calledembeddings that are generated by artificial neural networks to detect nonstationarity in a processed data stream. In addition to the development of three monitoring approaches, a fourth contribution, namely the extension from the pure detection of the change point to the identification of its cause is presented. The investigation includes a proposal for an automated inspection procedure, bringing together a control chart for quantile function values and a graph convolutional network.

M3 - Doctoral thesis

T3 - Wissenschaftliche Arbeiten der Fachrichtung Geodäsie und Geoinformatik der Leibniz-Universität Hannover

CY - Hannover

ER -