A Bayesian Approach for Maintenance Condition Assessment

Abstract

1. INTRODUCTION Many transportation agencies use visual inspection to assess and monitor the condition of roadways, and to ensure the quality and consistency of maintenance activities (Adams & Winkelman, 2016; Heymsfield & Kuss, 2013; Markow, 2012). Condition estimates are frequently obtained from visual inspection (Phares, Washer, Rolander, Graybeal, & Moore, 2004). Human visual inspection relies heavily on the ability to follow precise guidelines. For example, a shoulder may be considered eroded when the ruts are deeper than two inches (Nestler, Phetteplace, & Bush, 2017). However, even with precise guidelines, human visual inspections are error-prone and the error may introduce bias that distort results (Demsetz & Cabrera, 1999; Drury & Watson, 2002). Human error is in addition to random error that arises when a transportation agency uses a sample of its facilities to estimate the condition of the full inventory. The quality of the statistical estimate largely depends on the accuracy of the visual inspection, but in practice, the accuracy of the visual inspection is rarely questioned (Dirksen et al., 2013). The authors present a two-phase sampling strategy to account for the human error. Two-phase sampling means to survey some of the samples twice. The first-phase corresponds to the human visual inspection of all the samples. The second-phase serves to assess the quality of the first-phase observations. For that purpose, a subset of the first-phase samples undergoes a confirmation survey performed by different humans who are assumed to be completely accurate. The results of the analysis are expressed in the confusion matrix language. The confusion matrix outcomes describe the behavior of inspectors: the tendency to omit deficient highway features (false negatives); the tendency to classify non-deficient features as deficient (false positives); and the tendencies to make correct classifications (true positives and true negatives). Two parameters measure inspectors’ performance: Probability of Detection (POD) and the False Alarm Rate (FAR). POD and FAR are plotted on Receiver Operating Characteristic (ROC) graphs. A graphical representation of these parameters can help highway agencies identify specific areas of emphasis for inspectors’ training (Fawcett, 2004). For example, if inspectors tend to underestimate or overestimate the condition of pavement markings, future training may emphasize the procedure to assess pavement markings. The confusion matrix outcomes also allow highway agencies to improve the condition estimates (Coste et al., 2007; Levy & Lemeshow, 2008). Using Bayesian techniques, the probability of making correct judgments and the probability of making mistakes are used to adjust the condition estimates. For example, if inspectors tend to underestimate the condition of pavement markings, the Bayesian technique can quantify that tendency and adjust the estimate accordingly. To the best of the researchers’ knowledge, this is the only study that proposes to use quality assurance information to improve the condition estimates.