Felix Waldhauser

Abstract

The design of drivetrain components must accommodate a broad range of application scenarios. Due to considerable variability in component lifetimes, conservative design approaches cannot guarantee maximum protection against failures. Sensor technology in combination with advanced reliability methods enable monitoring and failure prediction during operation. Current research projects at the University of Stuttgart are pushing the limits of detectable damage further to increase the predictability and cost-effectiveness of maintenance measures.

Abstract

Efficient and robust fault detection in PHM typically depends on high-quality run-to-failure data, yet such labeled historical data is often scarce. While unsupervised algorithms trained on data assumed to represent normal operation (such as early-stage operation data) can yield good performance, they may be limited by contamination or incomplete coverage of normal states, related to unseen operating scenarios and conditions. This study investigates the effects of incomplete coverage of normal scenarios on fault detection performance. To enhance robustness against such incompleteness without making assumptions on the dataset’s composition, employing weakly supervised anomaly detection by incorporating few labeled fault samples is proposed. Evaluations on degradation testing data show that providing an autoencoder with a few fault labels significantly improves detection performance compared to a fully unsupervised autoencoder. This improvement is shown to be robust under various conditions, including changes in the number and quality of labeled fault samples and variations in the training data's composition. The proposed approach realistically simulates incomplete coverage through selective dataset composition and builds on weak supervision to enhance fault detection, making it particularly suitable when only field data—with no information on its composition—and few labeled fault samples are available.

Abstract

Absence of failure data is a common challenge for data-driven predictive maintenance, particularly in the context of new or highly reliable systems. This is especially problematic for system level failure prediction of analog electronics since failure characteristics depend on the actual system layout and thus might change with system upgrades. To address this challenge, this work pursues a novel sim\-u\-la\-tion-as\-sist\-ed failure analysis methodology enabling automated and comprehensive evaluation of system level failure effects and failure detectability. While results obtained from simulations are suitable for comparative studies, they are confined to the simulation environment. To overcome this limitation, failure simulations are combined with generative models to generate realistic representations of missing failure data. Preliminary results demonstrate the capability of conditional generative adversarial networks (cGANs) to generate operational data of healthy systems, which accurately reflects correlations present in the source dataset. The proposed approach, using simulations as an additional source for generative models, not only targets the scarcity of failure data for highly reliable electronic systems but also ensures the adaptability of predictive maintenance algorithms to accommodate future system modifications and upgrades.