Recently, there is a strong demand for the condition assessment of a large population of highway infrastructure with various advanced non-destructive evaluation (NDE) technologies and advanced processing techniques, which allow automated identification of hidden defects and damages. To date, the interpretation of NDE is generally carried out in a point-by-point manner, which treats each data point in isolation and ignores its spatial context. Also, NDE techniques have been employed to capture snapshots of condition, and thus all NDE data analysis and interpretation techniques focus exclusively on data collected at a single time-period. Over the last decade, several researchers have begun to deploy NDE over multiple time periods with the goal of not just capturing a snapshot of condition, but also tracking the deterioration rate and how the spatial patterns of deterioration evolve with time. This new approach to data collection essentially creates a time series of NDE data; however, current data analysis and interpretation approaches are only capable of dealing with NDE data as a snapshot.
Artificial Intelligence (AI) algorithms (e.g., Machine Learning (ML), Deep Learning (DL)) also have great potential to automate the NDE data analysis process. ML is generally defined as a set of algorithms that are capable of parsing and learning from data through specific user-defined features, and then applying what was learned to make informed decisions. For ML to work in practice, it is necessary for an analyst to provide feedback as to whether the resulting decisions are accurate or not. In contrast, DL, which is a subset of ML, does not require this feedback to improve decisions and can operate more autonomously, since the features are also obtained by the DL models. These benefits, while impressive, do come with some additional requirements, namely, access to much larger data sets. In either manifestation, the goal of these algorithms is to automatically process big data, identify patterns and relationships implicit within the data, and leverage these patterns and relationships to provide improved predictions. As such, both ML and DL appear highly compatible with NDE data analysis, which also aims to discern patterns and relationships from massive datasets. These methods can be used for multi NDE data processing to automate infrastructure condition assessment. The general procedure involves collecting and preprocessing data; labeling/parsing data; defining appropriate learning model(s); training the model, and testing or evaluating the model on the new data (outside the training samples).
The objective of this study is to establish how AI (inclusive of both ML and DL algorithms) can improve the reliability and quality of NDE data analysis and signal processing for the condition assessment of highway infrastructure. Of particular interest will be the identification of the most suitable learning models and optimization methods to process NDE data, and to develop a framework to guide the implementation of such techniques to support reliable data interpretation. This work could result in future applications for asset management (improving predictive modeling for planning maintenance and reconstruction).
Data from past field tests (inclusive of LTBP data collection and SHRP2 activities) will be available for this study.
Artificial intelligence; Machine learning; Deep learning; Nondestructive evaluation; data fusion; data analysis; highway infrastructure; bridge; Asset management; LTBP