Abstract
Modern industrial systems (e.g., aerospace, manufacturing, energy) increasingly rely on condition-monitoring sensor data to detect early signs of failure and optimize maintenance. Traditional models struggle with the high-dimensional, non-linear, and time-dependent nature of such data, limiting predictive accuracy. Prognostic Health Management (PHM) leverages sensor data to assess system health and predict failures, but modeling degradation in run-to-failure scenarios remains challenging due to evolving conditions and complex dependencies.
This dissertation introduces a deep state-space modeling (DSSM) framework that integrates deep learning, variational inference, and recurrent neural networks (RNNs) to model latent system dynamics. Combining RNNs with variational autoencoders (VAEs), the framework captures long-term temporal dependencies and performs robust probabilistic inference. A hybrid latent state structure—continuous and discrete—allows it to represent diverse degradation behaviors, from gradual wear to sudden faults, enhancing interpretability and robustness.
Extensive evaluation on simulated and real-world datasets, including a wind turbine case study, demonstrates the model’s ability to track latent degradation and accurately estimate remaining life (RL) or time of event (TE). To adapt to dynamic environments, the framework also integrates active learning for selective model fine-tuning, reducing the need for retraining with new data. This approach advances predictive maintenance and reliability assessment in complex, sensor-driven systems.