| Return to Main
Page Design of a Long Term Pavement Monitoring System for the Canadian Strategic Highway Research Program
Section 4 RECOMMENDED
CONFIGURATION OF AN INTEGRATED PAVEMENT MONITORING,
PERFORMANCE PREDICTION, AND PAVEMENT MANAGEMENT SYSTEM 4.1 PAVEMENT MONITORING,
PERFORMANCE PREDICTION, AND PAVEMENT MANAGEMENT In Section 2 of this report, we
presented and illustrated a comprehensive design for a
statistical analysis approach that could accept the
information assembled under the CLTPP program and
calculate its implications for pavement performance
prediction. Such implications were determined
probabilistically as well as deterministically. In
Section 3, we then indicated how the estimated pavement
performance parameters b could be input to a pavement
management system of rather general design. This section
integrates the previous two discussions, presenting a
design for an integrated long term pavement monitoring,
pavement performance prediction, and pavement management
system. We will present not only the integrated design
and its rationale, but we will also present a gameplan
for implementing it. One of the key objectives of this
report is to analyze and interrelate the various
techniques of pavement monitoring, statistical analysis
of pavement data, and pavement management. We have
surveyed the various techniques available and categorized
them in a way we find useful for designing and proposing
the most workable system. In this section, we will
recommend what we consider the best and most workable
pavement monitoring, performance prediction, and pavement
management system that can meet the true needs of
pavement decision makers and highway constituents in
Canada and the United States. We will use Figure 41 to represent
the critical relationships between pavement monitoring,
pavement performance prediction, and pavement management.
Beginning at the left hand side of the figure, the box
entitled CLTPP represents the acquisition of long term
pavement performance data under the anticipated CLTPP
program. (It represents the LTPP program as well.) The
CLTPP box also implicitly represents insertion of such
data into a statistical data management system for use in
boxes further to the right in the diagram. The CLTPP
box represents the activities of site selection,
selection of variables to monitor, installation and use
of monitoring equipment, accumulation of pavement data at
the selected sites, and delivery of the data to the
statistical data management system. Following accumulation of pavement data under the CLTPP program, there are two possible ways of manipulating that data. The first, represented by the topmost portion of the "Performance Prediction" box in Figure 41, is the regression approach. The "plain vanilla" regression approach was described in substantial detail in Section 2. As outlined there, the regression approach proposes to accept the CLTPP data and perform standard regression analysis on standard mathematical forms postulated to represent pavement deterioration using standard commercial statistical packages.] As discussed in Figure 41, the regression approach completely ignores any prior knowledge or data and will be inherently beset by small sample size, autocorrelation, heteroskedasticity, multicollinearity, and other difficult problems. As indicated in the figure, the
regression approach gives a "point estimate" of
the deterministic pavement deterioration relationship. In
particular, the regression approach gives a single
numerical estimate of the parameters b of the function
f(x,b). Using that single numerical estimate of the
parameters b, the function f(x,b) provides a single
numerical estimate of pavement performance y as a
function of contributory variables x. The single
numerical estimate, designated a point estimate, is
handed rightward as shown in the figure. Because is it a
single deterministic estimate, the result of regression
analysis can only support deterministic pavement
management systems. It cannot support probabilistic
pavement management systems. Returning to the "Performance
Prediction" box in the figure, the lower portion of
the box represents the Bayesian approach. Notice that the
Bayesian approach accepts as input precisely the
same CLTPP data as does the regression approach.
Section 2.3 showed that precisely the same statistical
data base was used by the Bayesian as the regression
approach. The regression and Bayesian examples in that
section both used the identical data in Tables 1 and 2.
However, Figure 41 also indicates that the Bayesian
approach uses any and all preexisting information and
data in addition to the CLTPP data. It uses whatever
prior information or knowledge is available. The Bayesian
approach allows prior information to be correctly
factored into pavement performance predictions and
balanced against the emerging CLTPP data as it comes
in. This feature makes the Bayesian approach uniquely
able to deliver benefits from the very first day and
deliver increasing benefits the more long term monitoring
that occurs. As indicated in Figure 41, the
Bayesian approach produces a complete probability
distribution over pavement performance as a function of
contributory environmental and maintenance variables.
This probability distribution is the most complete
representation possible of what is known at any point in
time about pavement deterioration. It contains not only
the mean or expected rate of pavement deterioration but
also uncertainty about the mean. A trivial special case of the
probabilistic representation can be obtained by using
only the mean of the probability distribution over the
pavement deterioration parameter, i.e., "throwing
away" the rest of the probability distribution and
using only the mean value. By doing this, we are in
effect extracting a point estimate from the Bayesian
probabilistic approach, a point estimate that can be
input to deterministic pavement management systems in
precisely the same way that the point estimate from
regression analysis can be input. The only difference
between using the mean of the Bayesian probabilistic
estimate and the regression result, a difference we
consider critically important, is that the mean from the
Bayesian approach contains the correct balance between
prior information and CLTPP data while the regression
approach is unable to consider any prior
information at all. It considers only CLTPP data with
caveats about insufficient statistical significance. In
summary, the inputs to deterministic pavement management
systems are similar in form but quite different in
substance between the Bayesian and regression performance
prediction approaches. Nonetheless, it is important to
emphasize that both the regression and Bayesian
approaches are capable of "feeding"
deterministic pavement management systems. Returning to the Bayesian statistics
box in Figure 41, notice that the probabilistic output
can be fed directly to a semiMarkovian probabilistic
pavement management system. In particular, the full
probability distribution over pavement deterioration
derived from the Bayesian statistical approach summarized
in Section 2.3 can be delivered wholescale and
automatically to semiMarkovian pavement management
systems where they can be processed directly. The results
of the semiMarkovian pavement management systems
systematically hedge against environmental, traffic, or
other uncertainties while the results of deterministic
pavement management systems contain no concept of hedging
or risk mitigation through diversification. It is
important to emphasize that the semiMarkovian pavement
management approach reduces to the deterministic approach
as a trivial special case. Stepping back and viewing Figure 41
in a gestalt sense, the current proposal before the SHRP
in the United States begins at the upper left, proceeds
along the top arrow to the top of the Pavement Prediction
box (regression analysis), and proceeds thereafter along
the top arrow to the top of the Pavement Management box
(deterministic pavement management system). The current
SHRPLTPP program follows the topmost path, and only the
topmost path, through the diagram. As we have stated
repeatedly, this approach has the shortcomings that · it cannot overcome small sample size
limitations and caveats that are inevitable in the
CLTPP program. · it cannot consider any information
other than LTPP information, i.e., it cannot balance
prior information with LTPP information. It will
therefore not appeal to engineers in the field who have
built their careers around conventional wisdom and
established practice. · it cannot deliver definitive results
until many years after program initiation. The body of
statistics emanating from the LTPP program in the United
States will not be sufficiently large to support decision
making until test sections have been monitored through
most of one life cycle. · it cannot provide definitive input
to probabilistic pavement management systems including
semiMarkovian and Markovian systems. Probabilistic
systems are far superior, both computationally and
conceptually, to deterministic systems. These shortcomings can be easily
overcome using the approach we recommend, which will be
explained in the next section. 4.2 RECOMMENDED GAMEPLAN FOR
CANADIAN SHRP LONG TERM PAVEMENT PERFORMANCE PROGRAM We believe the technical discussions in
this report support the recommendation presented in this
section. The recommendation is rather obvious given the
structure of Figure 41. Rather than proceeding along
the topmost path in the figure, we recommend that CSHRP
initiate the capability to proceed along the bottommost
path in the figure. In particular, we recommend that
CSHRP implement the bottom path in Figure 41 in a
systematic, incremental, three phase program such as that
described here. This three phase program does not
represent a major departure from the LTPP program in the
United States but borrows from it rather liberally.
Furthermore, because the Bayesian statistical approach
can be rendered identical with the regression approach by
using the diffuse prior assumption, our proposal will not
necessarily deliver radically different results from the
United States program. On the contrary, the United States
program will be achieved as a trivial special case of the
CLTPP program. The CLTPP program will be much richer
in its ability to consider prior information as well as
CLTPP data in designing and maintaining Canada's
highways, and it will be quicker to generate palpable
benefits and attract and retain enthusiastic
participants. In short, the proposal should be viewed as
a prudent, targeted improvement over the United States
SHRPLTPP program, extending it in several critical
areas. The recommended three phases and the
requisite tasks comprising each phase will be detailed
below. We have also provided estimates of the time it
would take to accomplish each task and the level of
effort in manmonths required. These estimates are
subject to revision pending more detailed discussions
with CSHRP. PHASE I: ASSEMBLING AND
PROCESSING CLTPP INFORMATION USING BAYESIAN STATISTICAL
METHODS Task I.1. Paper Design of Bayesian Module. The first task is to outline carefully on paper a design for the overall Bayesian statistical methodology to be employed. The paper design should contain a
detailed description of the data anticipated, the data
management package that will be used to store the data,
the Bayesian statistical modules that will be used to
retrieve the data and calculate results, the way those
results will be delivered downstream to pavement
managers, the way those results will be and the way
pavement managers should use those results, properly
balancing prior information with CLTPP results, and
careful consideration of design flexibility to
accommodate inevitable future changes. CLTPP should require a comprehensive
summary design of the entire system as indicated in
Figure 41, complete with descriptions of each box and
the information that will be handed between boxes.
Subsequent tasks should not be initiated until CSHRP
approves the paper design. This task can be accomplished over a
three (3) month period with approximately five (5)
manmonths effort. Task I.2. Selection of Tentative
Commercial Statistical Package. Following approval of
the paper design, the next task will be to review the
existing commercial statistical packages to determine
which will be the most suitable for the CLTPP Bayesian
statistical program. CLTPP should insist that the
survey be rather broad, encompassing issues such as
mainframe/PC compatibility, user friendliness, ease of
storage of raw CLTPP data, easy ability to add
applications modules (e.g., Bayesian statistical
procedure), graphics and report writing, universality of
use, and ability to interface directly with other
statistical packages. It is relatively unimportant as to
what regression procedures are embedded, for the Bayesian
procedure will supersede them. The best possible outcome of this task
will be to identify a package that can interface easily
with whatever package the SHRPLTPP program selects (if
it is not the identical package) and can best meet the
efficiency and flexibility needs of the CLTPP program
design in Figure 41 and Task I.1. In DFI's experience,
it is rather easy to write interfaces between different
commercial packages. As a result, it is a better strategy
to select the packages most consistent with CLTPP's
needs on a standalone basis and thereafter design and
implement whatever interface software CLTPP needs to
interface with SHRP. Such a strategy also allows CLTPP
to adapt to inevitable changes in the SHRPLTPP
configuration, changes over which CLTPP has no control,
and yet at the same time meet its own unique needs. We should reiterate that the objective
of this task is not necessarily to achieve lockstep
compatibility with the SHRPLTPP software design.
Rather, the objective of this task is to maximize the
probability of successfully meeting CLTPP's objectives
in behalf of Canadian participants. The three paramount
criteria should be · Suitability of the package to house
the Bayesian statistical module designed in Task I.1. · Ease and efficiency of use. · Ability to accommodate field use as
well as headquarters use, which probably necessitates
personal computer implementation. DFI has a great deal of experience
reviewing and implementing custom software applications
within commercial statistical and data management
packages. DFI is a retail supplier of Oracle and an
active user of SAS. This task should require 12 months to
complete with 23 manmonths of effort. Task I.3. Detailed Paper Design of
Bayesian Statistical Module. Following approval of a
statistical package, the next task is to provide a
detailed and comprehensive paper design of the Bayesian
statistical module as outlined in Section 2.3. The paper
design should begin with the design of the prior
probability distribution, likelihood function, and
posterior probability distribution based on the linear
model in Section 2.3. The first extension of the simple model
in Section 2 will be to extend to the multivariate case
with pavement deterioration functions that are linear in
the coefficients b. Such extension is straightforward,
requiring an extension of the single variable
probabilistic approach in Section 2 to the multivariate
case. In extending to the multivariate case,
consideration must be give to several key issues:
The paper design should be sufficiently
specific so as to serve as a programming guide for the
basic nonautocorrelated, homoskedastistic Bayesian
statistical case. In particular, the design should be
sufficiently detailed that a contractor other than the
design contractor could use the documentation to
implement the Bayesian statistical module. By securing a
design at this level of detail, CLTPP can ensure
comprehensiveness and costeffectiveness of subsequent
implementation. The deliverable on this task should be
a detailed technical report outlining the Bayesian
statistical equations to be programmed, the logic flow
chart, necessary inputs, and model outputs. This task will require 12 months
calendar time and 24 manmonths professional time. Task I.4. Implementation and Testing of Bayesian Statistical Module. Following approval by CLTPP of the Bayesian statistical design in Task I.3, implementation will commence. The programming language will be selected so as to be maximally compatible with
the statistical package selected in Task I.2 and yet at
the same time to meet the speed and other requirements of
the application at hand. Implementation is anticipated to
be in FORTRAN, PASCAL, or C. DFI writes code using rigid
internal documentation and comment standards. Similar
coding standards should be specified and adhered to by
whatever coding contractor is selected. Initial testing of the Bayesian
statistical module should occur in "test bed"
mode. By "test bed," we mean that all inputs
should reside in a flat ASCII file and all outputs should
be written to terminal or ASCII file output. Such testing
minimizes the possibility that another software system
such as a data management system is failing to deliver
outputs properly to the module being tested or that
program or memory control in the competing software
system is interfering with the module being tested. An important requirement of the test
program will be to mandate that CLTPP will be able to
specify several data bases to be used to test the
Bayesian statistical module. CLTPP should endeavor to
provide data bases that are realistic and difficult. For
example, CLTPP should provide a difficult data base
with exacerbated small sample size problems, a
systematically autocorrelated data base, and a very large
data base to ensure that the Bayesian statistical module
is running properly with such data bases. The contractor should be asked to
guarantee the correctness of the Bayesian statistical
module by entering into a maintenance contract with
CLTPP for a period of five years at a nominal budget.
The deliverable on the five year contract will be to
correct all errors identified in the Bayesian statistical
module in the five years following its initial
development. The deliverable on this task will be
the basic multivariate Bayesian statistical module
operating without autocorrelation or heteroskedasticity.
The task will require approximately six (6) calendar
months and approximately 3 manyears to complete. Task I.5. Interface of Bayesian
Statistical Module with Selected Commercial Statistical
Package. Following development and testing of the
Bayesian statistical module in Task I.4, the next task is
to interface the module with the selected statistical
data management package. Such interface should have the
following attributes: · all statistical data should be
stored in the data management package and should be
amenable to all reporting and other features of the
package. · the prior probability distribution
must be stored in the data management package and should
be amenable to all reporting and other features of the
package. · the Bayesian statistical module must
be able to accept inputs directly from the statistical
package. It will not be acceptable to implement
"dirty" interfaces such as writing interim
files for subsequent readin by the Bayesian module. · The Bayesian statistical module must
be able to write results directly back into the
statistical package as well as to terminal or file
output. This task will require three (3)
calendar months to complete and approximately 3
manmonths. Task I.6. Incorporate Modifications to Deal with Data Complexities. Following successful implementation in Task I.5, we will next incorporate capabilities to deal with the most troublesome complexities that reside in the data base:
The first activity will be to extend the basic Bayesian module from Task I.4 to diagnose and deal with autoregressive data. In particular, we will extend the Bayesian module to accept more complex dynamic pavement deterioration functional forms and more complex dynamic error terms. Deliverables will include written documentation of the augmented capability as well as implementation of software within the basic Bayesian module. Dealing with autoregression will increase the predictive power of the data that are gathered on the CLTPP program. The second activity will be to extend
the basic Bayesian module from Task I.5 to diagnose and
deal with multicollinearity in the data. The approach is
rather standard in classical regression analysis but will
have to be extended somewhat to accommodate the Bayesian
approach. Deliverables will include written documentation
of the augmented capability as well as implementation of
software within the basic Bayesian module. The final activity will be to extend
the basic Bayesian module from Task I.4 to diagnose and
deal with heteroskedastic errors in the data. There is no
fully general way to deal with heteroskedasticity, for
there is an infinity of prospective rules that might
govern regression error terms. On the other hand, it will
be quite productive to postulate structural relationships
among sites and time periods that can diagnose critical
heteroskedastic errors in the data and correct the bias
in statistical estimates that might otherwise occur. The
approach here will be to develop simple and general
structural relationships that could lead to
heteroskedastic errors and implement capability to
diagnose and correct them. Deliverables will include
written documentation of the augmented capability as well
as implementation of software within the basic Bayesian
module. This task will require approximately six (6)
calendar months and 18 manmonths to complete. Task I.7. Assembly of Hypothetical Data Base to Test Statistical Package with Bayesian Statistical Module. The interface should be tested with one or more of the test data bases supplied by CLTPP in Task I.4. The test data bases selected must be entered into the statistical data management package, and successful reading, calculating, and writing by the Bayesian statistical module should be demonstrated. The test will be considered correct when the test runs under this task (with full operation of the statistical data base package) are shown to be consistent with the test runs in Task I.4. Furthermore, the proper operation of
the autocorrelation, multicollinearity, and
heteroskedasticity enhancements from Task I.6 should be
tested. This task should require 12 calendar months and
1 man month. Task I.8. Preparation of User
Documentation and Training Program. Phase I will
culminate with the preparation of complete user
documentation in the form of a User's Manual for the
Bayesian module operating within the commercial data
management package. The User's Manual must be suitable
for ongoing, independent, inhouse operation of the
system. An important supplement to the user
documentation is a training program. DFI will implement a
training program complete with presentation materials
(i.e., slides) and case studies suitable for onsite
presentation to support initial use of the system as well
as "tune up" use of the system months or years
after initial installation. The training program will
initially be designed to be delivered by DFI. However, as
it ages and progresses, it is intended that the training
program will become routinized to the point where
inhouse staff can administer it themselves on an
ongoing basis. This task will require 3 calendar months and 6 manmonths to complete. |
