Please use this identifier to cite or link to this item:
|Title:||A predictive maintenance framework for data analysis and resource optimization in industrial IoT||Authors:||Ong, Kevin Shen Hoong||Keywords:||Engineering::Computer science and engineering::Computer systems organization::Special-purpose and application-based systems
Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
|Issue Date:||2022||Publisher:||Nanyang Technological University||Source:||Ong, K. S. H. (2022). A predictive maintenance framework for data analysis and resource optimization in industrial IoT. Doctoral thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/163043||Project:||001860-00001||Abstract:||Ubiquitous sensors and networks are critical elements that seamlessly integrate into our daily lives and enable diverse commercial and engineering applications, including the Internet-of-Things (IoT). By extension, the Industrial IoT (IIoT) entails continuously monitoring revenue-generating assets, such as wafer dicing, robotic pick and place, and industrial washing machines, for anomalous patterns and minimizing unplanned breakdowns of critical machines via predictive maintenance (PdM). However, PdM must grow beyond equipment maintenance to achieve prescriptive maintenance, and we identified several research gaps, which we highlight and address in the following paragraphs. In the first study, we consider that unplanned breakdown of critical equipment interrupts production throughput in IIoT, and data-driven PdM becomes increasingly important for companies seeking a competitive business advantage. Manufacturers must manually allocate competent manpower resources in the event of machine failure. Furthermore, human errors have a negative rippling impact on both overall equipment downtime and production schedules. To address these issues, we formulate the complex resource management of humans and machines as a resource optimization problem. We developed a maintenance repair simulator (MRS) game and conducted real-world experiments to support our findings. These results are contrasted against our proposed deep reinforcement learning (DRL) method to evaluate the efficacy of AI-based complex resource management for PdM applications. In addition, DRL improves its accuracy with more training data, self-learns an optimal maintenance strategy, and incorporates human feedback as part of its learning strategy. Notably, this study analyzes one machine for maintenance to limit the research scope of and validate our hypothesis. In the second study, preliminary examination of existing literature reveals that existing IIoT-based PdM frameworks do not consider complex real-time production states, machine health, and maintenance manpower resources. For this reason, we propose a generic PdM optimization framework to help maintenance teams prioritize and resolve maintenance task conflicts. Specifically, the PdM framework jointly optimizes edge-based machine network uptime and manpower allocation in a stochastic IIoT-enabled manufacturing environment utilizing model-free Deep Reinforcement Learning (DRL) methods. Since DRL requires a significant amount of training data, we propose and demonstrate the use of Transfer Learning (TL) method to help DRL learn more efficiently by incorporating expert demonstrations, termed TL with demonstrations (TLD). TLD reduces training wall-time by 58% compared to baseline methods, and we undertake numerous experiments to illustrate the performance, robustness, and scalability of TLD. PdM research focuses on improving RUL prediction accuracy via end-to-end models (i.e., raw sensor to RUL prediction) to increase factory productivity. Since machine remaining useful life (RUL)(i.e., ground truth information) is often unavailable, an RUL prediction model's success is not readily transferable to similar applications. Besides, DL-based RUL models require considerable training data, including rare failure data, and critical monitoring of model performance drift is not actively researched. To address these concerns, we first present an attentive multi-branch feature network (AMBFNet) model to improve RUL prediction performance and benchmark AMBFNet against comparable work on real-world datasets. Secondly, we present an incremental learning-based approach for monitoring field-deployed edge-based RUL model performance drift. Notably, the AMBFNet model surpasses state-of-the-art RUL prediction models in terms of prediction error, and we are the first to report incremental learning results for a popular PdM dataset. In summary, this thesis encloses several critical contributions to the corpus of knowledge in the PdM research topic. Specifically, we propose transforming PdM into a holistic maintenance strategy via AI-based methods, such as DRL. In this regard, DRL is used to learn data-driven decision-making strategies and provide actionable recommendations to augment human decision-makers, thereby termed ``augmented intelligence". Furthermore, we consider hybrid deep learning methods to improve RUL prediction performance and batch-based incremental learning to mitigate model drift. Finally, unifying the proposed contributions creates a generic PdM framework for data analysis in IIoT to jointly manage resources (i.e., humans and machines) and achieve good RUL prediction performance via AI-based methods.||URI:||https://hdl.handle.net/10356/163043||DOI:||10.32657/10356/163043||Schools:||School of Computer Science and Engineering||Organisations:||Robert Bosch (South East Asia) Pte Ltd||Rights:||This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0).||Fulltext Permission:||open||Fulltext Availability:||With Fulltext|
|Appears in Collections:||SCSE Theses|
Updated on Dec 8, 2023
Updated on Dec 8, 2023
Items in DR-NTU are protected by copyright, with all rights reserved, unless otherwise indicated.