Automated 3DGPR Analysis for Concrete Pavement Evaluation
Status: In development
The objective of this project is to identify missing, misplaced, or misaligned dowel and tie bars, voids under joints, and other deficiencies in concrete pavement by automating and improving upon the accuracy and repeatability of the analysis of 3DGPR data. The scope will include collection of 3DGPR data on concrete pavement sections at various field sites and on laboratory test slabs, and using the collected data to develop analysis routines that can be used by DOT personnel to evaluate the conditions of interest. 3DGPR is a relatively new technology for subsurface condition evaluation. It has been implemented by state DOTs through the SHRP2 R06D IAP program and more recently as part of TPF‐5(385) and is currently the primary technology being considered in TPF‐5(504) led by MnDOT. 3DGPR differs from conventional GPR in that it provides detailed information across the width of a lane, enabling it to detect tie bars, dowel bars, and other spatial features that might be missed by conventional GPR systems. Qualitative review of 3DGPR data can reveal important features in small sections of pavement, but qualitative review requires special expertise and is not practical for larger pavement sections. The purpose of this project is to automate the analysis of the 3DGPR data in a way that produces the relevant useful information needed by the owner agency for making decisions. This collaborative project combines the expertise of Infrasense in 3DGPR data acquisition and processing with the expertise of the Marquette University in GPR automated data analysis and machine learning.
Task 1: Obtain 3DGPR data representing concrete pavement conditions
The objective of this task is to obtain 3DGPR data on concrete pavement representing the various conditions to be detected and equipment types and collection settings. This data will be used to develop and train the automated processing analysis to be developed in Task 3. Infrasense has considerable experience collecting and analyzing both ground‐coupled and air‐coupled 3DGPR data. Air‐launched has the advantage of being able to collect data at normal driving speed, and thus can cover large road sections without the need for traffic control. Ground‐coupled systems, operating closer to the pavement, have the advantage of providing better penetration and more detail, but travel speeds are limited, and data collection generally requires MOT. Infrasense will make both types of data collected on jointed concrete pavements in Florida and Georgia available for this task. In addition, MnDOT has agreed, if this project is awarded, to provide data previously collected with both ground‐coupled and air‐coupled systems at PCC pavement sites in Minnesota. MnDOT has also agreed to collect and provide additional data on concrete pavement sections at MnROAD and at additional sites in the state within the limits of their schedule. In addition, WisDOT has agreed that, if available, air‐coupled 3DGPR data planned for collection on concrete pavements in the spring of 2023 will be provided to the project team for evaluation. If determined to be necessary, additional data will be collected on well documented sites to meet the specific needs of this project. In addition, to enable a systematic study of accuracy of the automated data analysis method, small‐scale concrete slabs containing dowel and tie bars with various alignment positions and deficiencies will be constructed and tested in the Marquette University lab using GPR equipment that will simulate 3DGPR.
Dr. Ken Maser, with 35 years of experience with collecting and analyzing GPR data, and 12 years of experience specifically with 3DGPR data, will be the lead investigator on this task.
- Deliverable: 3DGPR data sets associated with different types of concrete pavement constructions, with documentation of construction types and pavement conditions.
Task 2: Processing and preliminary evaluation of 3DGPR data
The objective of this task is to identify and quantify the specific features of the 3DGPR data to be detected with the automated analysis software. To accomplish this, the 3DGPR data sets acquired in Task 1 will be processed sing Kontur’s1 “Examiner” software to convert the data from time to frequency domain with settings that optimize the data quality. Once this data processing takes place the data will be reviewed to assess and quantify the features that are revealed, including the presence, spacing, depth and orientation of dowel and tie bars, the reflections at joints that are indicative of moisture infiltration and voiding, and the features that may be associated with delamination and spalling at the concrete surface.
Consideration will be given to the effects of data resolution (scans per foot) and air vs. ground‐coupled data collection systems. These quantified features will be accessed in the development and training of the automated analysis software as described under Task 3.
As part of this task, the project team will meet with DOT stakeholders to assess and specify the types of conditions to be detected and the desired degree of detection precision. For example, for dowel bar depth and misalignment, this task will seek to specify the depth and angular resolution (+/-x inches, +/-∅degrees). The quantification and specification of the features to be detected will serve as input to the automated analysis, as described in Task 3.
Dr. Ken Maser will be the lead investigator on this task.
- Deliverable:A report describing the features to be detected in the 3DGPR data, their relationship to the conditions of interest, and the quantification and resolution of the specific features.
Task 3: Development of the automated analysis
The objective of this task is to develop software to automate the analysis of the 3DGPR data to detect and evaluate key features in the concrete pavement. The approach will be to adapt a detection algorithm and software successfully developed and tested by Dr. Hayat for detection of stripping under the SHRP2 R06D IAP. The features to be detected in this work (e.g., dowels bars, voids, etc.) will appear as distinctive signatures in the reflected radar traces, and deviations from expected patterns can be associated with defects (e.g., misalignment). The input to the algorithm is a series of GPR scans over a long stretch of pavement, and its output is a set of flagged scans or scan regions representing scan‐to‐scan dissimilarities. The earlier algorithm (developed in SHRP2 R06D IAP) was implemented in MATLAB with an easy‐to‐use graphical user interface. Its efficacy has been validated in two ways. First, the outcomes of the algorithm, when applied to conventional GPR data, agreed with the visual observations of the data by an expert using standard visualization software. Second, the results of the algorithm agreed with the results obtained from coring. Examples of the validation of the algorithm are shown below in Figures 3 and 4 using conventional GPR data collected by the NMDOT on highway NM 264. In the proposed project, this previously developed platform for automatic GPR processing will be adapted to 3DGPR data and to the new application of detecting dowel and tie bar misalignment and other structural defects in concrete pavement.
The following subtasks will be carried out:
- Perform noise removal from the 3DGPR data by applying temporal filtering to the data.
- Establish correlation between subsurface features to be detected and their signatures in the 3DGPR data using physics‐based modeling of GPR response, deep learning, and signal processing to extract features/signatures from the raw 3DGPR response relevant to the quantification of misalignment and defect features. Simulated 3DGPR will also be generated using gprMAX software to identify signatures of subtle defects and dowel and tie bar misalignment. The features identified in this subtask and Task 2 coupled with the 3DGPR simulations will form a 3DGPR library, which will be used to train our machine learning algorithm described below in subtask (5).
- Develop metrics, based on signal processing, designed to represent features of interest.
- Generate metric traces for each lateral point on the 3DGPR trace to form 3D feature surfaces.
- Exploit spatial features in the lateral and longitudinal directions in each 3D feature surface to identify subsurface defects or misalignment of dowel and tie bars. Develop a PIDCGAN (Physics Informed Deep Convolutional Generative Adversarial Network), a novel machine learning algorithm, to identify defects and dowel and tie bar misalignment. PIDCGAN will use the physics of GPR to regularize the training of the network, which enables effective training with less data. The GAN portion of the signal network consists of a discriminator that detects the type of pavement feature and a generator that produces examples of each class of features and defects. The PIDCGAN is trained using both the features extracted using processing techniques and the raw 3DGPR scans from our 3DGPR library. In addition to identifying a GPR Scan as an indicator of a pavement feature or defect, the PIDCGAN will highlight where on the scan image it is located.
Dr. Majeed Hayat at Marquette University will be the lead investigator on this task, assisted by Drs. Richard Povinelli, Baolin Wan, Qindan Huang, James Richie, and Jaime Hermandez.
- Deliverable: An automated algorithm for the detection of subsurface defects and dowel and tie rod misalignment. Algorithm input is the series of 3DGPR scans generated in Task 1 over a length of concrete pavement, and its output is the set of flagged scans and the feature or defect type and degree identification.
Task 4: Evaluation and validation of the automated analysis
The objective of this task is to test, verify, and update the analysis software developed in Task 3 ‐ The software will be tested against those data sets acquired in Task 1 that have not been used in the development of the software, through the following subtasks:
- As a starting point, the efficacy of the algorithm will be examined by testing against 3DGPR scans obtained from gprMAX simulations. Noise will also be added to the simulations at different SNR levels to test the robustness of the algorithm to noise and the degrees of defects and misalignment.
- The algorithm will be then tested using the 3DGPR data obtained from the pavement data set developed in Task 1.
- Validate the algorithm on the test portion of the 3DGPR library, developed in Task 2, that was not used for development of the algorithm. We will analyze results with a confusion matrix, which indicates how many scans are correctly and incorrectly classified. Additionally, we will analyze the rate of incorrectly identifying a GPR scan as having a feature of defect of interest and the rate of missing features or defects in GPR scans.
Dr. Majeed Hayat will be the lead investigator on this task, assisted by Drs. Ken Maser and Povinelli.
- Deliverable: A report documenting the testing and validation of the automated detection algorithm.
Task 5: Development of a user interface and output specifications description
The objective of this task is to enable users who are not GPR experts to implement and utilize the software developed in Task 3. As part of SHRP2 R06D IAP, Dr. Hayat’s team developed a Graphical User‐Interface (GUI) that facilitated the use of defect‐detection algorithm. The GUI was capable of scanning GPR data and providing graphical output of GPR anomalies automatically. This included the ability to select and interpret raw data extracted from conventional GPR files, and associated metadata, process them, and present detection results to the end user in graphics mode, Excel tabular format, and MATLAB mode. In this NRRA project, we will use our experience in developing such a GUI to build a similar capability for the algorithm developed in this project. Hence, this task will involve development of a user interface to allow users to specify the 3DGPR data set(s) to be analyzed and the type of output to be provided. Specifically, we will develop a GUI that is capable of scanning 3DGPR data and providing graphical output of concrete pavement features and defects such as the presence, spacing, and misalignment of dowels and tie bars. This includes the ability to select and interpret raw data extracted from files (and associated metadata, including scan rate and scan numbers), process them, and present detection results concisely to the end user in graphics mode, Excel tabular format, and, if desired, in Matlab mode; the latter can be inspected and further analyzed by the user in a Matlab environment. Documentation for the user interface will also be provided.
This task will be led by Dr. Majeed Hayat, and he will be assisted by Dr. Ken Maser.
- Deliverable: Software including modularized and optimized code, multiple iterations of working GUI, final GUI, and a user manual for the GUI.
Email the Project Team
Principal Investigator: Ken Maser, Senior Principal, Infrasense, Inc, firstname.lastname@example.org
Technical Liaison: Shongtao Dai, MnDOT, email@example.com
Technical Advisory Panel (TAP): Contact us to join this TAP
- Hoda Azari, FHWA
- Haluk Sinan Coban, WisDOT
- Paul S. Collins, Kontur
- Shongtao Dai, MnDOT (TL)
- Rob Golish, MnDOT
- Ian Rish, Georgia DOT
- Nicholas Schaefer, SSI
- Jake Sumeraj, Illinois Tollway
- Michael Wallace, MnDOT
- Eyoab Zegeye, MnDOT
- Initial Project proposal (PDF), 5/5/2023