Unsupervised extraction of patterns and trends within highway systems condition attributes data. (October 2019)

• Record Type: Journal Article
• Authors: Titus-Glover, Leslie
• Is Part Of: Advanced engineering informatics. Volume 42(2019)
• Publication Date: 2019-10
• Subjects: Highway -- Composite health -- Future outlook -- Data abstraction -- Cluster analysis -- Normalization -- Entropy -- Time series; Computer-aided engineering -- Periodicals; Engineering -- Data processing -- Periodicals; 620.00285
• DOI: 10.1016/j.aei.2019.100990
• Languages: English
• ISSNs: 1474-0346

Abstract

Highlights: Methodology to analyze highway asset condition attributes data is presented. Data abstraction and entropy type algorithms are used to compute a health index, CHI. CHI versus time curve salient features are extracted for health pattern recognition. Accuracy of established patterns is tested using internal and external metrics. Executive-level reporting of health matched for over 75 percent of instances. Abstract: Highway agencies combine expert opinions and basic regression modeling techniques to process vast amounts of time series condition attributes data to define highway network health. The health rating exhibit high variability and lack adequate detail for executive-level maintenance planning and resource allocation. This paper presents a new methodology for data abstraction, analysis, and clustering for pattern recognition of highway network health. The methodology describes mathematical and statistical data abstraction algorithms for data preprocessing (smoothening (unweighted moving average), scaling (normalization), and weights derivation (entropy) to compute a composite health index (CHI)), and salient features extraction. Data analysis involved cluster analysis to identify patterns in asset current health and future outlook. The outcome is a characterization of highway network health for executive-level decision making. The algorithms included in this methodology have been successfully applied in the fields of biology, finance, econometrics, bioinformatics, marketing, and social science for pattern recognition. The accuracy of the new methodology is illustrated with an experiment using 463 in-service pavement assets and internal/external metrics (including the degree to which methodology performance classification outcomes conform to national expert opinion). The results from the experiment confirm an accurate and computationally inexpensive methodology, which provides outcomes that compare to real-world pavement condition rating metrics.