TY - THES

T1 - Classification of Multivariate Time Series Data using Machine Learning and System Redundancy Analysis

AU - Lang, Elliot

N1 - no embargo

PY - 2024

Y1 - 2024

N2 - This thesis investigates the level of redundancy within batches of sensor data; the explicit goal of which being to evaluate the system's reaction to the inactivity of one or more sensor units. Moreover, this thesis explores the applicability and viability of different machine learning algorithms for classifying the state of the analysed machine based on batches of acceleration data. The redundancy within the data is measured by quantification of the inherent dimensional coverage; based on the result, the data dimensionality can be reduced. Possible permutations of the loss of one or two of the six available sensor units are analysed in the context of the remaining dimensional coverage. This gives an indication of the certainty with which analyses based on the reduced data can be evaluated. Furthermore, several possible supervised machine learning algorithms, capable of multi-class classification, are identified and their applicability to the labelled data batches is assessed. An array of promising classification methods were identified and applied to each batch of data. The successful classification is quantified by measuring each method's predictive accuracy, the number of correctly identified machine states, and the training time each method requires including the optimization of its hyperparameters. The results of both of these processes show a robustness to the loss of sensors within the system, independent of spatial location, as well as some promising machine learning classification algorithms, capable of identifying the machine's state.

AB - This thesis investigates the level of redundancy within batches of sensor data; the explicit goal of which being to evaluate the system's reaction to the inactivity of one or more sensor units. Moreover, this thesis explores the applicability and viability of different machine learning algorithms for classifying the state of the analysed machine based on batches of acceleration data. The redundancy within the data is measured by quantification of the inherent dimensional coverage; based on the result, the data dimensionality can be reduced. Possible permutations of the loss of one or two of the six available sensor units are analysed in the context of the remaining dimensional coverage. This gives an indication of the certainty with which analyses based on the reduced data can be evaluated. Furthermore, several possible supervised machine learning algorithms, capable of multi-class classification, are identified and their applicability to the labelled data batches is assessed. An array of promising classification methods were identified and applied to each batch of data. The successful classification is quantified by measuring each method's predictive accuracy, the number of correctly identified machine states, and the training time each method requires including the optimization of its hyperparameters. The results of both of these processes show a robustness to the loss of sensors within the system, independent of spatial location, as well as some promising machine learning classification algorithms, capable of identifying the machine's state.

KW - Machine Learning

KW - Classification

KW - Redundancy

KW - Decision Trees

KW - Neural Networks

KW - Discriminant Analysis

KW - k-Nearest Neighbours

KW - Ensemble Classification

KW - Naive Bayes

KW - Machinelles Lernen

KW - Systemredundanz

KW - Decision Trees

KW - Neural Networks

KW - k-Nearest Neighbours

KW - Discriminant Analysis

KW - Naive Bayes

KW - Ensemble Classification

KW - Classification

U2 - 10.34901/mul.pub.2024.065

DO - 10.34901/mul.pub.2024.065

M3 - Master's Thesis

ER -