Kolloquium - Details zum Vortrag
Sie werden über Vorträge rechtzeitig per E-Mail informiert werden, wenn Sie den Newsletter des kommunikationstechnischen Kolloquiums abonnieren.
Alle Interessierten sind herzlich eingeladen, eine Anmeldung ist nicht erforderlich.
Master-Vortrag: Verbesserung der Fahrzeuganomalienerkennung durch Nutzung von Feature Selection
Max Haupt
Mittwoch, 21. Januar 2026
14:00 Uhr
IKS 4G | zoom
Automotive test benches typically perform the same drive cycle hundreds of times to ensure the reliability of the device under test. Manual anomaly inspection of the resulting data is a time intensive process and costly to maintain. Automated approaches are therefore implemented to help engineers identify anomalies. Due to runtime and memory constraints, it is usually not feasible to subject all sensors available to a unified anomaly detection. Hence, a shortcoming of these approaches is the need for domain knowledge to select the sensors (features) which should be monitored by the automated system. Implementing automated approaches for feature selection is therefore an increasingly relevant topic.
This thesis is specifically aimed at improving the Temporal Variational Autoencoder (TeVAE) model, already used for automated anomaly detection, by introducing feature selection. This thesis makes two contributions. First, the training process of TeVAE is improved by clipping the norm of the gradients and also by introducing a linear learning rate scheduler, resulting in a training time decrease of up to 90% and a significant anomaly detection performance increase. Secondly, a novel feature selection approach based on the Selective Deep Autoencoder (SDAE) architecture for TeVAE is proposed. This includes the use of the Kullback-Leibler Divergence for comparing latent variables of variational autoencoders, which parametrize probability distributions, as well as an additional training process yielding the final model. The proposed adaptation of the SDAE architecture achieves similar anomaly detection performance to state of the art feature selection approaches while determining the number of selected features in an unsupervised fashion.
Additionally, extensive experiments using the adapted SDAE on a simulated 16 feature automotive dataset are conducted. It is shown that the number of features can be reduced significantly at the cost of a minor decrease in anomaly detection performance. In the case explored in this thesis, 50% of the features could be removed without decreasing the anomaly detection performance by more than 20%.
