| description abstract | Predicting the deterioration of high mast light poles (HMLPs) can support capital planning and asset management for highway agencies, such as inspection and maintenance prioritization decisions. This paper aims to develop a data-driven framework for predicting the performance of in-service HMLPs along highways by considering multiple factors including structural and environmental factors. The proposed framework consists of a pipeline of machine learning techniques including data cleaning, feature engineering and data integration, model development, and feature importance identification. Two data-driven models (XGBoost and logistic regression) are developed with the implementation of data oversampling to address imbalanced data issues. The importance level of selected factors is identified to provide insight into the underlying influential mechanism of various factors on the deterioration of HMLP elements because of both fatigue and corrosion effects. The proposed framework is implemented for one element of HMLP as an illustrative example using structural data from manual inspection reports, climatological data (e.g., wind and snow events) from the National Oceanic and Atmospheric Administration (NOAA), and geographical locations from the Geographical Information Systems (GIS) database. The results show that XGBoost with oversampling outperforms other methods. With the implementation of oversampling, the balanced accuracy of the XGBoost model increases from 0.69 to 0.74. The average precision of the best-fit model is 0.88, representing a high level of precision in predicting positive instances correctly across all classes. Feature importance analysis indicates that cumulative time of natural wind with speed from 5 to 15 m/s (days) is the most significant factor affecting the deterioration of HMLPs (797, 1,316, and 1,716 days on average in the data set for condition states 1, 2, 3+4 respectively), followed by the cumulative time of natural wind with speed over 15 m/s (days) and distance to the nearest coastline, suggesting the significance of considering these features from multisource data. The proposed data-driven approach can predict the deterioration of each in-service HMLP by considering the large variety and complex combination of influential factors, and is expected to constitute an important basis for inspection and maintenance decision making. | |
