Chen, Zheyuan
2022.
Data mining for fault diagnosis in steel making process under industry 4.0.
PhD Thesis,
Cardiff University.
Item availability restricted. |
Preview |
PDF (PhD thesis)
- Accepted Post-Print Version
Download (3MB) | Preview |
PDF (Cardiff University Electronic Publication Form)
- Supplemental Material
Restricted to Repository staff only Download (129kB) |
Abstract
The concept of Industry 4.0 (I4.0) refers to the intelligent networking of machines and processes in the industry, which is enabled by cyber-physical systems (CPS) - a technology that utilises embedded networked systems to achieve intelligent control. CPS enable full traceability of production processes as well as comprehensive data assignments in real-time. Through real-time communication and coordination between "manufacturing things", production systems, in the form of Cyber-Physical Production Systems (CPPS), can make intelligent decisions. Meanwhile, with the advent of I4.0, it is possible to collect heterogeneous manufacturing data across various facets for fault diagnosis by using the industrial internet of things (IIoT) techniques. Under this data-rich environment, the ability to diagnose and predict production failures provides manufacturing companies with a strategic advantage by reducing the number of unplanned production outages. This advantage is particularly desired for steel-making industries. As a consecutive and compact manufacturing process, process downtime is a major concern for steel-making companies since most of the operations should be conducted within a certain temperature range. In addition, steel-making consists of complex processes that involve physical, chemical, and mechanical elements, emphasising the necessity for data-driven approaches to handle high-dimensionality problems. For a modern steel-making plant, various measurement devices are deployed throughout this manufacturing process with the advancement of I4.0 technologies, which facilitate data acquisition and storage. However, even though data-driven approaches are showing merits and being widely applied in the manufacturing context, how to build a deep learning model for fault prediction in the steel-making process considering multiple contributing facets and its temporal characteristic has not been investigated. Additionally, apart from the multitudinous data, it is also worthwhile to study how to represent and utilise the vast and scattered distributed domain knowledge along the steel-making process for fault modelling. Moreover, state-of-the-art does not iv Abstract address how such accumulated domain knowledge and its semantics can be harnessed to facilitate the fusion of multi-sourced data in steel manufacturing. In this case, the purpose of this thesis is to pave the way for fault diagnosis in steel-making processes using data mining under I4.0. This research is structured according to four themes. Firstly, different from the conventional data-driven research that only focuses on modelling based on numerical production data, a framework for data mining for fault diagnosis in steel-making based on multi-sourced data and knowledge is proposed. There are five layers designed in this framework, which are multi-sourced data and knowledge acquisition, data and knowledge processing, KG construction and graphical data transformation, KG-aided modelling for fault diagnosis and decision support for steel manufacturing. Secondly, another of the purposes of this thesis is to propose a predictive, data-driven approach to model severe faults in the steel-making process, where the faults are usually with multi-faceted causes. Specifically, strip breakage in cold rolling is selected as the modelling target since it is a typical production failure with serious consequences and multitudinous factors contributing to it. In actual steel-making practice, if such a failure can be modelled on a micro-level with an adequately predicted window, a planned stop action can be taken in advance instead of a passive fast stop which will often result in severe damage to equipment. In this case, a multifaceted modelling approach with a sliding window strategy is proposed. First, historical multivariate time-series data of a cold rolling process were extracted in a run-to-failure manner, and a sliding window strategy was adopted for data annotation. Second, breakage-centric features were identified from physics-based approaches, empirical knowledge and data-driven features. Finally, these features were used as inputs for strip breakage modelling using a Recurrent Neural Network (RNN). Experimental results have demonstrated the merits of the proposed approach. Thirdly, among the heterogeneous data surrounding multi-faceted concepts in steelmaking, a significant amount of data consists of rich semantic information, such as technical documents and production logs generated through the process. Also, there Abstract v exists vast domain knowledge regarding the production failures in steel-making, which has a long history. In this context, proper semantic technologies are desired for the utilisation of semantic data and domain knowledge in steel-making. In recent studies, a Knowledge Graph (KG) displays a powerful expressive ability and a high degree of modelling flexibility, making it a promising semantic network. However, building a reliable KG is usually time-consuming and labour-intensive, and it is common that KG needs to be refined or completed before using in industrial scenarios. In this case, a fault-centric KG construction approach is proposed based on a hierarchy structure refinement and relation completion. Firstly, ontology design based on hierarchy structure refinement is conducted to improve reliability. Then, the missing relations between each couple of entities were inferred based on existing knowledge in KG, with the aim of increasing the number of edges that complete and refine KG. Lastly, KG is constructed by importing data into the ontology. An illustrative case study on strip breakage is conducted for validation. Finally, multi-faceted modelling is often conducted based on multi-sourced data covering indispensable aspects, and information fusion is typically applied to cope with the high dimensionality and data heterogeneity. Besides the ability for knowledge management and sharing, KG can aggregate the relationships of features from multiple aspects by semantic associations, which can be exploited to facilitate the information fusion for multi-faceted modelling with the consideration of intra-facets relationships. In this case, process data is transformed into a stack of temporal graphs under the faultcentric KG backbone. Then, a Graph Convolutional Networks (GCN) model is applied to extract temporal and attribute correlation features from the graphs, with a Temporal Convolution Network (TCN) to conduct conceptual modelling using these features. Experimental results derived using the proposed approach, and GCN-TCN reveal the impacts of the proposed KG-aided fusion approach. This thesis aims to research data mining in steel-making processes based on multisourced data and scattered distributed domain knowledge, which provides a feasibility study for achieving Industry 4.0 in steel-making, specifically in support of improving quality and reducing costs due to production failures.
Item Type: | Thesis (PhD) |
---|---|
Date Type: | Completion |
Status: | Unpublished |
Schools: | Engineering |
Subjects: | T Technology > TA Engineering (General). Civil engineering (General) |
Date of First Compliant Deposit: | 20 March 2023 |
Last Modified: | 20 Mar 2023 11:19 |
URI: | https://orca.cardiff.ac.uk/id/eprint/157811 |
Actions (repository staff only)
Edit Item |