Machine learning methods for corrosion and stress corrosion cracking risk analysis of engineered systems

Download files
Access & Terms of Use
open access
Copyright: Jiang, Peng
Altmetric
Abstract
Demand for risk analysis on corrosion and stress corrosion cracking (SCC) related failure events is increasing globally. Analyzing such failure events is complex since multiple factors, such as physical, biochemical and mechanical factors, may take place. It requires cross-disciplinary knowledge and long-term data collection from sites and lab tests. Therefore, corrosion and SCC risk analysis often needs collaboration among experts of different areas. However, it is difficult to develop rules or knowledge for guiding risk analysis because failure events occur on different engineered systems in different environments, leave alone the fact that the valuable data is barely shared among industries. This thesis introduces a new method for the automatic risk analysis of corrosion and SCC related problems. This method uses machine learning to handle tasks in different scenarios, including environment-oriented and system-oriented risk analysis. The former studies one or alike systems in different environments while the latter studies different systems in the same environment. Ensemble methods and support vector machine (SVM) are built to handle environment-oriented risk analysis problems by predicting failure of bolting systems in underground coal mines. Ensemble methods produces the relative importances of features and suggests optimization on data collection strategies for water types. SVM reconstructs location indicators through kernel methods and gives reliable predictions for occurrence of failure. SVM and deep learning are employed to handle system-oriented risk analysis by predicting inhibition efficiencies of corrosion inhibitors. A standard process including feature engineering and cross validation is proposed to assess the generalization performance of models. Overfitting is observed due to mismatched training and test sets. Transfer learning is utilized to overcome overfitting by transferring knowledge between different domains, leading to improvement in generalization performance. All of the machine learning methods that are employed in this thesis are packaged in a Python library.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Author(s)
Jiang, Peng
Supervisor(s)
Crosky, Alan
Creator(s)
Editor(s)
Translator(s)
Curator(s)
Designer(s)
Arranger(s)
Composer(s)
Recordist(s)
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
2018
Resource Type
Thesis
Degree Type
PhD Doctorate
UNSW Faculty
Files
download public version.pdf 5.58 MB Adobe Portable Document Format
Related dataset(s)