| Return to Main
Page Evaluation of Rutting in Nova Scotia's Special "B" Asphalt Concrete Overlays JOINT
CSHRP/NOVA SCOTIA BAYESIAN APPLICATION 1.0 INTRODUCTION Wheel path rutting is becoming one of
the most significant manifestations of pavement distress
on Nova Scotia's arterial 100 Series Highway system. The
down sizing of the rail industry within Nova Scotia and
across Canada has resulted in an increase of truck
traffic and truck loading on principle arterial highways.
Heavily loaded trucks, such as tractor trailers comprise
between 20% to 40% of the annual traffic stream and have
caused the extent and severity of rutting to increase
dramatically over the past several years. This has caused
great concern for highway engineers as rutting is a
safety concern in so far as it can cause vehicular
hydroplaning and other associated steering problems. Four basic types of rutting may occur
on a roadway structure (1):
In this study the term
"rutting" refers to plastic deformation only
(item number 2 above). To address plastic deformation, in 1989
the Nova Scotia Department of Transportation and
Communications (NSDOT&C) introduced a High Friction
Asphalt Concrete (AC) mix, commonly referred as Special
"B". Although this asphalt mix design has been
utilized since 1989, the Province of Nova Scotia has not
evaluated the performance of the Special "B"
mix in terms of rutting. The only annual data collected
on Special "B" mixes has been two pavement
monitoring sites initiated by the Canadian Strategic
Highway Research Program (CSHRP) in 1989. These sites
are monitored yearly for several forms of condition
distress, including rutting. This rutting is assumed to
be a result of plastic deformation. Prediction models have become a
commonly accepted method to evaluate and predict pavement
performance. Pavement engineers tend to agree that
predictive modelling is a critical element in determining
the costeffectiveness of rehabilitation alternatives.
Pavement performance models, statistical, subjective, or
otherwise are the basis for maintaining an adequate
highway infrastructure. Decisions on whether to overlay,
cold plane or surface treat are based upon a prediction
of how each alternative will perform in different
situations. From 1990 to 1994, a data analysis
project was undertaken by CSHRP to develop pavement
performance models using the CLTPP (Canadian Long Term
Pavement Performance) project data collected on test
sites across Canada. Bayesian statistical methods were
utilized to develop predictive models for rutting,
cracking, roughness, and a distress index. Bayesian
statistical regression methods provided a means for
developing performance models with the "young"
CLTPP database. The Bayesian methodology combines
information gained through interviewing knowledgable
individuals with available field data within the context
of regression modelling. Using expert knowledge, the
Bayesian methodology allows the analyst to extend the
inference space beyond that provided by the data,
particularly with respect to the age variable. Bayesian
methodology provides an opportunity to evaluate the
performance of the High Friction Special "B" AC
mix in terms of rutting. By implementing Bayesian
methodologies, the existing young performance data
currently available from the CLTPP test sections,
combined with subjective data from knowledgeable
individuals (expert judgement), is sufficient to support
the development of a working predictive model of pavement
performance. 1.1 Project Objectives The primary objectives of this project
were: 1. To utilize the Bayesian statistical
approach utilized by CSHRP to develop and analyse a
predictive model for (plastic) rutting performance for
Nova Scotia's High Friction Special "B"
asphaltic concrete. 2. To exploit a learning opportunity,
in terms of technology transfer, in the use of Bayesian
statistical analysis for pavement performance modelling 1.2 Bayesian Methodology The Bayesian statistical approach
combines prior knowledge and past experience, known as
the "prior" with observed field performance
data to develop a "posterior" model (2). Hence,
as illustrated by the flow chart in Figure 1, a limited
amount of performance data combined with expert opinion
can provide a model for immediate use in predicting
pavement performance. To develop the CSHRP (CLTPP)
pavement performance models, the Fortran coded program
"The Bayesian Regression Analysis Package"
(BRAP version 2.0) was upgraded (3). The new program was
called "BSTAT" (4). Subsequently, another
software program known as "XLBAYES" using
EXCEL™ as an interface was developed by Vemax
Management Inc. and Decision Focus Incorporated, under
contract with CSHRP. The objective was to provide an
optimized and accelerated version of the linear
regression feature in "BSTAT". The methodology developed for the
CSHRP project and the Bayesian statistical analysis
software "XLBAYES" were utilized for this
project. The following Bayesian analysis template was
used to develop the model for this project (4):
2.0 PROJECT TEAM To enhance the technology transfer from the CSHRP Bayesian project, a project team comprised of a consultant from industry and representatives from NSDOT&C and TUNS (Technical University of Nova Scotia) was formed. The team members for this project were: Project Coordinator: Kent Speiran, P.Eng. (NSDOT&C) Lead Analysts: Dr. Nouman All, P.Eng. (Technical University of Nova Scotia (TUNS)) and Tony Ramia, P.Eng. (TUNS/Maritime Testing (1985) Ltd) Data Managers: Romeo Poirier, C.E.T. and Gerard Lee, C.E.T. (NSDOT&C) Expert Judgement Interviewer:
Eric Theriault, P.Eng. (Jacques Whitford &
Associates) 2.1 Model Selection A critical step in the modelling
process is to explicitly define the model. Hence, for
this project the following statement was developed: "To develop a model which will
predict the degree of rutting due to plastic deformation
in the High Friction Special "B" type AC
overlays over AC pavement on Nova Scotia's arterial
highways". It should be recognized that the model
developed for this project was limited to the province of
Nova Scotia. Hence, it was assumed that the environmental
conditions would be constant and these effects could be
omitted from the modelling exercise. 2.2 Dependent Variable (y) Statistical regression models have dependent and independent variables and take the following form:
The dependent variable for this model
was selected to be rutting (plastic deformation) measured
in millimetres(mm). A rut, as defined by SHRP (Strategic
Highway Research Program), is a longitudinal surface
depression in the wheel path (5). The dependent variable
is an average measurement of rutting in the outer and
inner wheel paths. For this project, pavement rutting is
measured per the SHRP LTPP methodology. In this
procedure, the maximum rut depth is measured with a 1.2
meter straight edge and recorded in millimetres at 15
meter intervals for each wheel path. This technique has
been adopted by CSHRP in their analysis of CLTPP
data. However, in the CSHRP procedure, a transverse
profile is obtained using the Dipstick profiler, which
then is analysed to provide rutting measurements as
defined by the SHRP straight edge technique. 2.3 Independent Variables (xi's) Many variables contribute to the
instability rutting of asphaltic concrete. A preliminary
literature review identified 18 variables. However to
achieve a workable model, all 18 variables could not be
utilized. Reducing the number of variables to a
manageable set (less than 7) was accomplished by first
defining the following eight variables as design
parameters: asphalt penetration, asphalt liquid
viscosity, penetrationviscosity number (PVN), Marshall
airvoid content, voids in mineral aggregate (VMA),
Marshall stability, Marshall flow, and voids filled in
asphalt (VFA). These design parameters, although affect
rutting, would be considered to be within design
specifications for the modelling process. The remaining 10 variables were then
presented to our expert panel for final selection. A
document was developed to solicit expert opinion, based
on experience and judgement, as to the importance of 10
variables affecting rutting. The variables included:
Asphalt Content, Field Compaction, Insitu Air Voids,
Blend Sand Content, % Passing 5000 um sieve, % Passing 80
um Sieve, % Fractured Faces, Bitumen Filler Ratio,
Thickness of AC Overlay and Traffic/Year (ESALS
Equivalent Single Axle Loads). Each variable was defined
in terms of its relationship to rutting. The experts were
asked to rank the variables in order of importance in
terms of rutting predictability and to weight the
relevance of each variable on a scale of 1 to 10 (10
being a dominant affect). Two key factors influenced the final selection of the independent variables. First, practical consideration demanded that the maximum number of variables fall within the range of 5 to 7, a limitation on the expert encoding process. Second, the data at the CLTPP test sections had to support the selected variables. Having considered these factors, as well as solicited expert opinion, the following independent variables were selected for use in the model: 2.4 Model Type and Functional Form In the modelling process, the model can
be derived from mechanistic principles, empirical
evidence, or a combination of both. However, since an
empirical model best exploits the power of Bayesian
methods (6), the model for this project was based on
pragmatic evidence. Therefore, the model type was chosen
to be empirical, derived from field evidence at Nova
Scotia's CLTPP test sections and expert opinion. Purely empirical methods offer no
guidance in initially selecting the functional form of
the relationship between the dependent variable Yi
(rutting) and the selected independent variables (xi's) A
literature search was conducted to assist in postulating
the functional form; however no models stated in terms of
instability rutting were found. Therefore, the CSHRP
Bayesian project took the approach of moving from the
simple to the complex by adopting an additive linear
functional form for the first iteration and then moving
to a linearlog model in the second iteration (4). The functional form of the additivelinear model (first iteration) was as follows: To improve the linear relationship model and to account for more rutting in early years and less in the later years, the CLTPP project investigated a transformed model which accounted for the cumulative effect of traffic (kESALs) on the pavement. This was accomplished by adopting a linearlog functional form (second iteration) as described below: 3.0 CSHRP TEST SECTIONS In 1989, the Canadian Strategic Highway
Research Program (CSHRP) initiated two pavement test
sites in Nova Scotia as part of the CLTPP project: 1. Highway 102, north bound lane at km 92.2 near Hilden. 2. Highway 103, west bound lane, 3.15
km west of Hebb's Cross. Each of these test sites has three test
sections. Across Canada there are a total of 24 CLTPP
test sites, each of which has between 24 test sections.
In total there are 65 test sections across Canada. Each
test section is 150 meters long and the AC overlays are
currently between 24 years old (4). Annual data on these sites has been
collected as part of a data base for the CLTPP project.
Data is collected yearly for all sites and includes (7): 1. Surface distress mapping and photologging. 2. Benkelman beam deflection testing (yearly and continuously in spring for peak deflection). 3. FallingWeight Deflectometer deflection testing (biannually). 4. Longitudinal and transverse profiles using "dipstick" profiler. 5. Traffic data, including general and classification data, as well as weight data collected with permanent and portable WeighInMotion systems. 6. Climate for sitespecific deicing chemical application rate and load restriction periods. 7. Maintenance data on all agencies' regular maintenance practice. 8. Skid resistance measurements. 9. Video logging the pavement and surrounding area. This was conducted once in 1993. The main objectives of the CLTPP project have been (4): 1. To evaluate Canadian practice in the rehabilitation of flexible pavements and to subsequently develop improved methodologies and strategies. 2. To identify ways to increase
pavement life through the development of costeffective
pavement rehabilitation procedure based upon the
systematic observation of in service pavement
performance. After a five year process of data
collection and analyses, there was a need to expand on
these general objectives. The data collected could be
analysed to provide the maintenance and rehabilitation
strategy that is most costeffective for any given
asphalt pavement. Hence, the initiation of the Bayesian
pavement performance model projects. 3.1 Test Site Cross Sections In accordance with the CLTPP project, both test sites in Nova Scotia were overlayed with AC in 1989. The rehabilitation strategy varied for each section of the test site in terms of asphalt concrete type and thickness. Prior to the 1989 rehabilitation AC overlays, the test sites had pavement structure cross sections as presented in Tables 2 and 3, and Figures 2 and 3. The CLTPP project provided NSDOT&C the opportunity to utilize the High Friction Special "B" AC mix within the rehabilitation strategy for existing pavement structure cross sections. The High Friction Special "B" AC mix was used on two test sections on Highway 102 and all three test sections on Highway 103. A summary of the rehabilitation strategies are presented in Table 4. 3.2 Nova Scotia's Special "B" AC Mix Instability rutting has been experienced by all roadway agencies. Most agencies now agree a more rut resistant asphalt concrete mix and better aggregates are the most effective solutions to this problem. In 1989,NSDOT&C took a positive step by incorporating the Special "B" mix into the specifications for 100 series highway projects. Compared to a conventional Type "B" AC mix, the High Friction Special "B" (BHF) mix coarse, composed of large top size aggregates, (see the combined aggregate gradations in Table 5), and incorporating, as it does, more crushed faces than the conventional Type "B" mix. When Special "B" was introduced, a minimum of 80% of the coarse aggregate had to have crushed (fractured) faces. This specification was revised in 1995 and now requires 95% of coarse aggregate to have crushed faces. 3.3 Expert Judgement The process of developing a predictive
model by the Bayesian methodology allows quantification
of past experience in a statistical format similar to
that which would be obtained from field trials. The data
is collected from individuals who are knowledgeable in
the subject matter of the model. This form of data
collection is commonly referred to as "encoding
expert judgement". To develop a statistically
acceptable Bayesian model, the expert judgement must
arise from experience and intimate knowledge of the
subject because their judgement is the foundation for
developing the model. Some of the desired attributes for
selecting the experts in the Bayesian methodology used in
this project were (2): 1. Fifteen or more years of experience with pavement engineering. 2. Substantial experience in observing the field performance of pavements, including occurrence of physical distress. 3. Familiarity with material property definitions and specification requirements. 4. Experience with construction practices in the recent past. 5. Familiarity with pavement design
procedures. To develop and encourage the Bayesian
methodology technology transfer, expert consultants were
drawn from NSDOT&C and private industry consultants
for the expert knowledge they could bring to this
project; 1. Expert 1 NSDOT&C 2. Expert 2 NSDOT&C 3. Expert 3 Private Industry Consultant 4. Expert 4 NSDOT&C 5. Expert 5 NSDOT&C 3.4 MODEL INPUTS 3.4.1 NEW DATA The Bayesian methodology allows the
incorporation of objective data, commonly referred to as
new data. The ideal source of new data is field data,
such as those measurements taken in test sections. In
selecting the source of new data, it is necessary to
ensure that data exists for the selected independent
variables. The data should have frequency of collection
and the units for each variable should be compatible with
the model variables. Hence, for the model developed in
this project, the only pavement sites with yearly rutting
measurements on the Special "B" AC mix were the
CLTPP test sections. The five test sections with the
Special "B" mix represented five sets of new
data for the model. This new data was obtained from
NSDOT&C and included 25 observations of rutting. The
data was entered into a tabulated format using the
"EXCEL" software in preparation for using the
"XLBAYES" software. To prepare for the
linearlog model (Model 2) the data is similarly
prepared; however, the logarithm of the (age*traffic)
variable was computed and tabulated manually. The
regression analysis is then conducted on the
"transformed" data. 3.4.2 Data Quality Assurance A strength of the Bayesian methodology
is that a small and young "new data base" can
be used. It is still necessary though, to have good
quality data. To provide quality data for the model
development, quality assurance checks on the data are
necessary to ensure the assembled data is accurate and
avoid compatibility problems. In the original model run, which is not
included in this report, the data obtained were averaged
values for the test sections at the two test sites.
Hence, there was limited variability in the data. This
resulted in problems with running the "XLBAYES"
software. The data would not support the development of a
5 variable model. The data presented in this project is
the result of doing a quality assurance check on
CSHRP's CLTPP test site data. 3.4.3 Prior Data Prior data was collected from the
experts. In the Bayesian methodology, there are many ways
to collect the prior data, including the following (6): 1. Full Orthogonal Matrix 2. Incremental Orthogonal 3. Market Questionnaire (NonOrthogonal) 4. Card Sort (NonOrthogonal) 5. DataDerivation The Full Orthogonal Matrix is
considered the simplest method. It was utilized for this
project and is the only method discussed in this report.
The Full Matrix method is basically a systematic method
of enumerating a series of questions which correspond to
combinations of independent variable settings. For each
setting the expert is requested to respond with their
'prediction' on the dependent variable (Y). To solicit the expert judgement for the Bayesian model, an encoding document was developed which included: 1. A definition of the model 2. A definition of the dependent variable 3. A definition of the independent variables and the inference space for judgement coding. 4. A matrix for encoding judgement. The complete encoding package is
included in Appendix A. The encoded expert judgement
matrices are presented in Appendix B. The data collected using the encoding matrix is tabulated to resemble the format of the new data. Assuming the expert's responses in the matrix are observations, it is possible to generate a model, given the specified functional form, which completely characterizes the expert's judgement. This model is called the "prior". In this project, the prior data had 48 observations of rutting. The "XLBAYES" software has a classical regression feature that can be utilized for calculating the prior. All model runs for this project used the "NPrior" method in the "XLBAYES" software. From the this classical regression analysis, the variance/covariance matrix, the means of the regression coefficients, degrees of freedom, and the residual variance are required as input to the NPrior Bayesian regression analysis. |
