Implementing Machine Learning with Highway Datasets

Every year MDOT invests millions of dollars into testing geomaterials, digitizing historic records and capturing inventory and condition data. Massive amounts of tabular data, documentation and imagery, which are relevant for planning and engineering purposes, continue to be accumulated. The engineering characteristics of pavement or other materials can be estimated in the early phase of the project. Additionally, scheduling and construction can be optimized by smart decision-making assistance enabled by such machine learning models. This project enhanced the existing machine learning models with newly available data and developed and tested new machine learning models for datasets of interest including drilling and pavement data, project duration and highway right-of-way (ROW) image datasets. Various machine learning models were developed and trained for the selected highway datasets including drilling data, pavement falling weight deflectometer (FWD) data, scheduling estimates using reinforcement learning, ROW image QA/QC processing and object detection for three types of image objects, pavement core thickness datasets. Tabular data neural network models for drilling data and pavement data were trained and used for dependent variable prediction. A state-of-art object detection model using YOLO (You Only Look Once) v3 algorithm were also trained and tested for detecting traffic barrier end treatments and insect blocking in ROW images. Reinforcement learning models for drilling project schedule estimation were developed and tested using historical data records. Furthermore, random forest model was also trained and tested for one type of drilling data – groundwater depth. These machine learning models can potentially be used to assist with the decision-making process in project planning and construction and some of these models have been integrated with existing working process in MDOT SHA. Improved cost effectiveness of the agency can be achieved by enhancing analysis capabilities and improving decision-making through incorporating machine-learning into planning and engineering work processes.  This report is available from the Maryland State Govenment web site at https://www.roads.maryland.gov/OPR_Research/MD-21-SHAUM5-23_Machine-Learning_Report.pdf