Uncertainty-aware fault diagnosis for safety-related industrial systems
Kafunah, Jefkine
Kafunah, Jefkine
Loading...
Identifiers
http://hdl.handle.net/10379/18147
https://doi.org/10.13025/17578
https://doi.org/10.13025/17578
Repository DOI
Publication Date
2024-04-15
Type
Thesis
Downloads
Citation
Abstract
Industry 4.0 (I4.0) has enabled dynamic modern-day industrial environ ments through rapid automation and improved access to real-time data from complex industrial operations. The I4.0 suite of digital technologies, including the Internet of Things (IoT), data analytics, and predictive mod eling, enable intelligent industrial manufacturing systems through data driven decision-making. Further, the systematic integration of physical and virtual worlds through the cyber-physical system (CPS), a core concept of I4.0, enables the construction of expansive factories with high flexibility, adaptability, and even self-awareness. These factories are physically inter connected large-scale industrial plants requiring a higher level of process and quality management strategies to improve overall production safety and efficiency. Recently, data-driven fault diagnosis (FD) models trained on large-scale industrial process datasets using deep learning (DL) techniques have demonstrated the ability to deliver actionable insights required for intelligent process management. However, despite their potentially superior process monitoring capa bilities, DL models have limitations, including excessive data dependency, interpretability challenges, sensitivity to hyperparameters, lack of trans parency, susceptibility to adversarial attacks, and issues with imbalanced data. Furthermore, exposure to gradual changes under different operating conditions in the industrial environment significantly impacts the perfor mance of DL-based FD models. Therefore, it is crucial to design approaches that enhance the reliability of DL-based FD in dynamic industrial environ ments to guarantee the safety and efficiency of industrial systems. This thesis aims to develop techniques leveraging uncertainty estimation for data-informed decision-making under dynamic and uncertain industrial environments, highlighting the potential to enhance the safety and reliabil ity of DL-based FD applications. The proposed approaches enable the gen eration of trustworthy and interpretable fault predictions for data-driven decision-making to facilitate the deployment of DL-based FD models in safety-related industrial environments. First, this thesis proposes an uncertainty-aware ensemble combination method for an ensemble of DL-based FD models to help monitor the sta bility of industrial processes and product quality. The approach generates a continuous multivariate probability distribution as the combined model output, replacing deterministic point estimation techniques that are ineffec tive in capturing the underlying model predictive uncertainties. Next, this thesis proposes a data-driven method for generating syn thetic out-of-distribution (OOD) data based on deep generative networks. This approach leverages an in-distribution (ID) data-supporting manifold of large-scale industrial process data and a combination of strategic man ifold sampling techniques to create realistic OOD data. Generating syn thetic OOD data to augment ID data enhances the DL-based FD model ca pacity for estimating the prediction uncertainty by incorporating insights from OOD data, thereby delivering safe and reliable DL-based FD models for real-world industrial process monitoring. Finally, this thesis proposes a novel approach to enhance the training of DL-based FD systems on imbalanced datasets. The method applies logit weight vectors to the penultimate layer of a deep neural network (DNN), introducing relevant perturbations to influence the network output strate gically. In particular, the approach implements a training regime that fa cilitates the switching between logit vectors to help the classifier focus on samples from the minority classes while effectively generalizing the entire dataset. To demonstrate the effectiveness of the proposed approaches, this thesis explores the problem of monitoring the stability of industrial processes and product quality using case studies on the Steel plate faults and APS failure at Scania trucks datasets.
Funder
Publisher
NUI Galway