Keenan, Mike (2022) Bridge load assessment using machine learning and weigh-in-motion data. [USQ Project]
|
Text (Project)
Keenan_M_Nguyen_Redacted.pdf Download (2MB) | Preview |
Abstract
Pattern recognition using machine learning algorithms is a mature discipline, but it is still in the research phase in the civil engineering field (Farrar & Worden, 2012). Adding to this body of research contributes to the improvement of our asset management decision-making.
KiwiRail (The New Zealand railway network manager) has an existing process of evaluating axle overloads captured by weigh-in-motion sites that could benefit from further research. This project aims to evaluate how a machine learning model may help.
Overloading is very relevant internationally and has been nominated as one of the top five historical causes of bridge collapse (Zhang et al, 2022). Weigh-in-motion systems capture axle weights and axle spacings but further analysis, such imparted bending moment on any given span length, is a post-processing function. In cases when immediate actions are required following an axle overload then an immediate structural analysis considering also the adjacent axles would be of benefit.
A machine learning model was developed in MATLAB by a process of supervised learning via a simple analytical model constructed in Excel. With enough training data the intent was to obtain an accurate machine learning model such that it could assess a given set of axle loads and spacings and determine if a bending moment limit had been breached or not. A common 6m span length was chosen as the focus area and variables for model input were carefully considered.
As a project outcome a highly accurate machine learning model was established once the training data volume got to approximately 5,000 sets. To get to this stage many variations of training inputs were used and volume of training data was incrementally increased to monitor the effect on accuracy.
A potential future development of this work is to expand the focus area to other span lengths to observe accuracy when more axles are incorporated into the variable set. The analytical model developed for this project was limited to the 6m span length and the required assumptions made it up to 3% non-conservative in outputs. An improved analytical training model is required before the focus area can be expanded.
In conclusion, although moving load analysis lends itself to traditional formula and analytical processing applications this study has shown that a machine learning model may potentially become a viable alternative.
|
Statistics for this ePrint Item |
| Item Type: | USQ Project |
|---|---|
| Item Status: | Live Archive |
| Faculty/School / Institute/Centre: | Current – Faculty of Health, Engineering and Sciences - School of Engineering (1 Jan 2022 -) |
| Supervisors: | Nguyen, Andy |
| Qualification: | Bachelor of Engineering (Honours) (Civil) |
| Date Deposited: | 20 Jun 2023 02:39 |
| Last Modified: | 20 Jun 2023 02:39 |
| Uncontrolled Keywords: | Weigh-in-Motion; Rail Bridge; Machine Learning |
| URI: | https://sear.unisq.edu.au/id/eprint/51896 |
Actions (login required)
 |
Archive Repository Staff Only |
