Improving the reliability of a Bridge Management System (BMS) using an ANNbased Backward Prediction Model (BPM)

Jaeho Lee, 2 Kamalarasa Sanmugarasa, 3Michael Blumenstein and 1Yew-Chaye Loo

1

Griffith School of Engineering, Gold Coast Campus, Griffith University, QLD 4222, Australia
2
Parsons Brinckerhoff Australia P/L, GPO Box 2907, Brisbane, QLD 4001, Australia
3
School of Information and Communication Technology, Gold Coast Campus, Griffith University,
QLD 4222, Australia

Abstract
The slow adoption of Bridge Management Systems (BMSs) and its impractical future
prediction of the condition rating of bridges are attributed to the inconsistency
between BMS inputs and bridge agencies existing data for a BMS in terms of
compatibility and the enormous number of bridge datasets that include historical
structural information. Among these, historical bridge element condition ratings are
some of the key pieces of information required for bridge asset prioritisation but in
most cases only limited data is available.
This study addresses the abovementioned difficulties faced by bridge management
agencies by using limited historical bridge inspection records to model time series
element level data. This paper presents an Artificial Neural Network (ANN) based
prediction model, called the Backward Prediction Model (BPM), for generating
historical bridge condition ratings using limited bridge inspection records. The BPM
employs historical non-bridge datasets such as traffic volumes, populations and
climates, to establish correlations with existing bridge condition ratings from very
limited bridge inspection records. The resulting model predicts the missing historical
condition ratings of individual bridge elements. The outcome of this study can
contribute to reducing the uncertainty in predicting future bridge condition ratings and
so improve the reliability of various BMS analysis outcomes.
Keywords: Bridge Condition Ratings, Bridge Management System (BMS), Artificial
Neural Network (ANN), Backward Prediction Model (BPM)

1.0
Introduction
The efficient use of public funds for the well-being of bridge networks requires an
effective bridge asset management technology. It is particularly important to optimise
future bridge maintenance, repair and rehabilitation (MR&R) activities with the given
funding and to request suitable future funding based on reliable Bridge Management
System (BMS) outcomes. A BMS, as a computer-based decision support system
(DSS), is used to determine the best possible strategy that ensures an adequate level of
safety for bridges at the lowest possible life-cycle cost [1]. Many bridge agencies
worldwide have begun the transition to BMS-based bridge asset management. A BMS,
based on the results of a deterioration model, provides various important future
estimations for the planning of MR&R activities. The success of a BMS is highly
1

dependent on the accurate estimation of future condition ratings [2]. The condition
ratings are used directly and indirectly as input data for many significant functions in
the commercial BMS package [3]. Fig. 1 presents the uses of bridge condition ratings
and the relationship with many analytical BMS modules in project and network level
analysis. Ideally, a BMS should identify current and future bridge deficiencies and
estimate the backlog of funding requirements. Typical BMS software mainly
functions to [3,4]: (1) forecast future bridge deficiencies; (2) identify a list of
improvement options to correct such deficiencies; and (3) estimate the costs and
benefits of implementing each improvement option.
Fig.1

Re-illustration of relationships between historical bridge inspection data

and BMS outputs [4]
(Note that the relationships with other input sources and BMS outputs are
omitted)

Numerous bridge condition rating and deterioration models have been developed to
reliably determine the bridge life cycle for the remaining years of use and major
MR&R needs. However, the estimations of future structural condition ratings from
BMSs are still not practical for use in developing reliable long-term maintenance
strategies.
From the perspective of bridge agencies, there are a number of shortcomings related
to the use of BMS software. Inter alia these are: (1) Commercial BMS software has
been used for less than 15 years and even those bridge agencies which implemented
BMSs from an early stage, would have only approximately 6 to 7 biennial inspection
records at their disposal. (2) Bridge condition ratings normally do not change much
over short time periods. (3) Approximately 60% of BMS analytical processes rely
heavily on periodic bridge inspection results [5]. These factors lead to the inaccuracy
in predicting the future structural performance of bridges.
The main difficulty faced by current deterioration modelling techniques is the lack of
usable data related to the bridge elements historical behaviour. Based heavily on a
few sets of recent structural condition ratings, current modelling techniques cannot be
expected to produce reliable outcomes. This in turn leads to an unreliable prediction
of future bridge condition ratings. Not withstanding the above, the methodology
presented in this paper is an ANN-based Backward Prediction Model (BPM), which
reliably generates unavailable historical bridge condition ratings. It aims to improve
the accuracy of future structural condition ratings.
1.1 Time-Series Predictions for Insufficient Datasets
Time-series predictions are important resources for making decisions in many
application domains [6]. Various prediction techniques have been applied to
commercial BMS analysis modules. The most frequently used techniques are
Regression, Markov models, Bayesian methods, Fuzzy techniques, Genetic
Algorithms, Case Based Reasoning and Artificial Neural Network (ANN) models.
Specifically, Markov decision processes (MDP) have been used in major state-of-theart BMS software as part of their deterioration modules. To obtain reliable predictions
from conventional techniques, the size of missing patterns from an entire dataset must

be 5% or less [7]. It should be noted that irregularly sampled datasets cannot be used
with conventional prediction methods [8].
This research aims to utilise limited inspection records over a short period to predict
large datasets spanning over a much longer time period. As mentioned in Section 1,
the short history of the BMS adoption and its lack of usable inspection records causes
unreliable long-term bridge performance predictions. Recognising the historical
patterns for aging bridges can be a problem when using commonly available time
series prediction methods.
For any computational prediction methods, the amount of available datasets is
required to be much larger than the target prediction datasets to obtain reliable
prediction results. An ANN-based model also requires a large number of training
datasets to successfully estimate their correlation. It should be noted that existing
ANN-based data-mining techniques have been applied in medical, economics,
engineering and IT fields. While capable of carrying out similar activities as the BPM,
data-mining has had success only in cases where very small proportions of the
datasets are unavailable - much smaller than are required to be generated for an
effective BMS implementation. To rise above the fundamental limitations in timeseries predictions, the proposed neural network model adopts an alternative type of
time-series dataset, which is used to overcome the lack of trends in the existing small
number of bridge element condition ratings.
To address the research problem indicated, the BPM is described in this paper
employing non-bridge historical data to support the lack of trends in the existing
bridge condition ratings to generate the unavailable years of historical condition
ratings and so establishing some comprehensive datasets.
1.2 Outline of the Proposed Model
The research problem regarding the lack of historical bridge data may be solved by
the use of an ANN-based Backward Prediction Model (BPM). A pilot study along
these lines only considered bridge condition ratings amongst the BMS historical data
required. The BPM predicts entire or selected periods of historical bridge condition
ratings to generate unavailable years of datasets. It aims to improve the prediction
accuracy of future bridge condition ratings using a deterioration module.
The bridge condition ratings in existing small numbers of condition ratings do not
change much during a short period of time. As such, it is also difficult to detect
condition rating changes using an ANN-based condition rating prediction model.
However, existing bridge condition ratings can be strengthened by non-bridge factors
including local climates, number of vehicles and population growth in the area
surrounding the bridge. The non-bridge factors are employed to help establish the
correlations between the non-bridge factors and the lack of historical data patterns in
the existing but inadequate bridge condition rating datasets.
Fig.2 schematically describes the mechanism of the BPM. It illustrates the main
function of the ANN technique in establishing the correlation between the existing
condition rating datasets (from year m to year m+n) and the corresponding years
non-bridge factors. The non-bridge factors directly and indirectly affect the variation
of the bridge conditions thereby the deterioration rate. The relationships established

using neural networks are then applied to the non-bridge factors (for year 0 to year m)
to generate the missing bridge condition ratings (for the same year 0 to year m). Thus,
the non-bridge factors in conjunction with the ANN technique can produce the
historical trends that inform the current condition ratings.
Fig.2

Outline of proposed Backward Prediction Model (BPM)

The BPM has been tested using two different types of bridge condition rating datasets
- the National Bridge Inventory (NBI) and BMS condition rating inputs - for the same
bridge provided by the Maryland Department of Transport (DoT), USA. The errors
cope within 10% for the historical bridge condition ratings that are generated. The
magnitude of the prediction error depends on the scale of each Condition State (CS)
(e.g. 20% magnitude of each CS in 5 graded CSs) in element level inspection methods
for a BMS. In order to generate acceptable historical condition rating datasets for
BMS inputs, the proposed model required 4-5 sets of bridge inspection results. The
details of this will be further discussed in Section 3.3.
1.3 Element Level Bridge Inspections
The most widely used inspection method for a BMS operation is the element-level
type. This evolved from the National Bridge Inventory (NBI). NBI information is
submitted annually to Federal Highway Administration (FHWA) by state highway
agencies in the U.S. NBI has been used for more than three and a half decades to
determine the needs of rehabilitation and replacement, considered nation-wide.
However, it was found to be insufficient to establish MR&R. The element level
inspection method has the following advantages which have been reported by FHWA:
(1) more precise bridge condition assessment; (2) more quantitative condition data of
each element beyond deck, superstructure and substructure per bridge; and (3)
sustaining the element level inspection method based BMS [10-12]. Thus, the
proposed study employs commonly used bridge condition rating information based on
an element level inspection method for updating a dataset of the BMS. In addition,
NBI information was used to measure directly the performance of the proposed BPM.
The obtained bridge condition datasets require calibration to fit into the proposed
BPM model, due to the typical ANN input environment. The acceptable numerical
scale for ANN modelling is from -1 to 1 (or 0 to 1). Fig.3 illustrates the scale of NBI
and element level-based condition rating information for this particular study.
Fig.3

Scale of the condition ratings used in the BPM model

The Condition Index (CI) in NBI is scaled between 0 to 9 for the NBI#58 (deck), #59
(superstructure), and #60 (substructure), and every calibration step is titled as a
different Condition State (CS) to express the bridge components condition ratings.
The CI for an element level inspection method consists of 4 or 5 different CSs
(depending on the bridge authorities adoption and customisation) to quantify
condition states of bridge elements.

2.0 Neural Network Modelling

Artificial Neural Networks (ANNs) have attracted world-wide attention over the last
two decades, and are one of the supreme tools for solving the stated research problem,

because they are simple and effective for examining data and developing suitable
models. It is an emerging technique and a promising tool in process identification and
control, owing to its ability to model processes reliability. For the proposed BPM
model, ANNs have a great aptitude in terms of determining missing values and
patterns in historical datasets. They also have a first-rate ability to derive meaning
from complicated or imprecise datasets. They can be used to extract patterns or detect
trends in data, which is useful for seizing unknown areas or periods of data patterns.
Many different ANN models have been developed to achieve various predictions such
as: (1) learning to predict events based on observation of patterns in historical data;
(2) learning to classify unseen data into predefined classes based on observations in
the characteristics of the data; and (3) learning to cluster the training data into natural
groups based on similarity of characteristics [13].
The Back-propagation feed-forward ANN is a universal function approximator that
typically yields better results than traditional approximation methods in practical
applications. Technically, it normally has two different stages, i.e. the training and
testing stages, to induce values as predictions. The training stage is the learning
process to detect patterns of interest in the dataset, and then additional patterns,
unseen by the network previously, are applied as inputs in the testing stage to produce
suitable outputs.
2.1 Composition of the Proposed ANN model
Fig.4 illustrates the proposed single-layer feed-forward back-propagation neural
network model. The model consists of an input layer, hidden layer(s) and an output
layer, whereby neurons exist in the hidden and output layers connected by weights. A
neuron in the hidden layer obtains data from the input layer, which is processed by the
calculation of a weighted sum and subsequently passed to another neuron in the
output layer through a weighted connection.
Fig.4

Structure of ANN-based BPM

In the proposed model, the number of neurons in the hidden layer can be determined
by the number of existing inspection and non-bridge factors in the input and output
layer, respectively. The specifications for the inputs, outputs and functions of the
proposed BPM are detailed in Table 1. The input layer has 21 variables including 4
factors for the number of vehicles, 2 factors for the population growth and 15 factors
for climate. This information is used to train the ANN to determine the correlation
with currently available bridge condition rating data in the output layer. The sigmoid
function as a typical neuronal non-linear transfer function is used in the proposed
model due to its non-linear properties. If a linear transfer function were used in the
proposed model, each of the neuronal inputs would become multiplied by their
identical proportion during training. It may cause the entire system to generate
inappropriate outputs. Hence, the sigmoid transfer function aids to isolate specific
input pathways [14, 15].
Table 1

The components of the proposed neural network model

As mentioned in Section 1, due to the limitation of bridge element condition rating

availability for a BMS's historical requirements, the total eligible data for this study is
only a small amount. It is not a sufficient amount for training to construct reliable

weights using the available dataset, thus additional variables, non-bridge factors, are
required, which have a relationship with the training outputs (condition rating data), to
support the training output's lack of trend. The selected non-bridge factors are suited
for use as variables for bridge condition rating information because bridges are always
exposed to the natural environment and traffic conditions. All variables used as nonbridge factors have been available for many years as publicly-accessible data.
2.2 Sample datasets for BPM
As mentioned in Section 2.1, non-bridge factors can be used to add trends into
existing bridge condition ratings, thus the sensitivity of these factors affects the
reliability of backward predictions.
Based on the approximate bridge locations provided from the Maryland DoT,
historical vehicle registrations, census population, and climate data have been
obtained from the Federal Highway Administration, U.S. Census Bureau, and the U.S.
Department of commerce National Oceanic & Atmospheric Administration,
respectively.
The historical vehicle registrations are plotted in Fig.5 (a) in terms of 4 different
variables: passenger vehicle, bus, truck and total number of vehicles. The historical
population changes are plotted in Fig.5 (b) in terms of 2 different variables: bridge
location - city and state. The historical temperature and precipitation changes are
plotted in Fig.5 (c) & (d), respectively.
Fig.5

Raw data of non-bridge factors for BPM inputs

The details of raw datasets for NBI and BMS inputs for the different tests are detailed
in Fig.6 and Table 2, respectively. The initial assumption of the BPM is that the
components and elements of the bridge have excellent condition ratings when it was
built.
Fig.6

Raw data of NBI for BPM outputs (Bridge #0301xxxx1)

Table 2 Raw data of actual condition ratings (Element #234 on Bridge

#0301xxxx1)

3.0 Validation of the BPM

The entire timeframe of the bridge data used in the BPM is from the year 1966 to
2004. Amongst these, on 5 occasions, inspection results were used as ANN-based
BPM training inputs and outputs (from 1996 in 2-year increments to 2004). The
assumed condition rating at year zero (1966) of the bridge has also been used. The
remaining years (from 1968 to1994 with 2-year increments) of historical condition
ratings can be generated by the proposed BPM.
Generated historical condition ratings are compared with existing information to
assess the reliability of results. As shown in Fig.7, the timeframe of Tests #1 and #2
for the proposed BPM using NBI information is described in Fig. 7 which shows the
timeframe of the inputs (Fig.7 (c) and (f)) and their results (Fig.7 (d) and (g)). The

two different results are compared with the existing NBI (Fig.7 (e) and (h)) to
measure the BPM performance using the raw datasets.
Fig.7

BPM timeframe (Test#1 and 2 using NBI for performance measurements)

NBI data, which are used to generate the historical condition ratings between 1968
and 1994, are employed in the first test. Fig.8 shows the results of the generated
historical condition ratings for decks, superstructures, and substructures. 78.6% of the
generated data can be directly compared with the actual NBI data to measure the
prediction errors. There are two different modes (test #1 and 2) are conducted to
validate this model and are detailed in Sections 3.1 and 3.2.
Fig.8

BPM results for Bridge#0312xxx1

(Note that the number of prediction results in each year is 66 which is the combined
number of learning rates (lr:0.0-0.5) and momentum coefficients (mc:0.0-1.0) in the
neural network configuration)

3.1 Backward Comparisons

As mentioned, most of the generated data from Test #1 can be directly compared with
the existing historical NBI datasets. The comparisons between the results for each
bridge component and its NBI records are plotted in Fig.9. Most artificially-generated
historical condition ratings are obtained within a prediction error scale of less than
10%. However, year 1982 in deck, year 1984 in superstructure, and years 1982,
1984 and 1986 in substructure exhibit larger errors than the maximum allowance
(10%). The main reason for generating imprecise outcomes is that as mentioned
earlier in Section 2.1, the proposed model was developed based on the correlations of
condition ratings and their corresponding non-bridge factors in the ANN training
stage. Nevertheless, this is adequate for historical condition ratings, because they are
ranked within the same CS (60% CS2 < 80%).
In the case where the ANN training datasets using existing condition ratings, do not
have a relevant correlation with the non-bridge factors, the BPM cannot provide
reliable historical condition ratings for a specific year. For example, depreciations of
condition ratings caused by sudden physical damages to a bridge are not influenced
by the non-bridge factors used in the BPM yielding unreliable predictions.
The results of the backward predictions are validated by comparing them with
existing historical condition ratings (Test #1). However, the actual element-level
condition ratings for BMS inputs are only available in a small number of datasets and
are not applicable to the backward comparison method used in this section for BPM
validation. Therefore, the forward comparison (Test #2) is used in Section 3.2 to
validate the BPM.
Fig.9

Performance measurements: existing NBI vs BPM results (Bridge #

0312xxx1)

3.2 Forward Comparisons

The other validation method for the BPM is conducted in this section, called the
forward comparison (Test #2). The BPM of the ANN training inputs in Test #2

utilises the results of Test #1 between 1968 and 1994, as illustrated in Fig. 7 (f). The
BPM produces the future condition ratings between 1996 and 2004. The results are
also compared with existing condition ratings (year 1996-2004) and summarised in
Table 3. Each year of condition ratings predicted using the ANN-based BPM provides
satisfactory results within the error allowances. Therefore, the forward comparison
method constructed in this section can be used to validate the BPM results using the
actual BMS condition rating inputs.
Table 3

Summary of prediction performance for test 2

3.3 Minimum Inputs for the BPM

To verify the minimum BPM input requirements and the response with the number of
inspection records in the BPM, additional tests are also conducted for the various
years of condition ratings. The BPM assumes that the bridge agency only retained up
to 5 sets of historical condition ratings.
Fig.10 shows the forward comparisons when the BPM is used for various years of
condition ratings. The figure demonstrates the average prediction errors, which
gradually decrease when the number of inspections is increased. The effective range
of inputs for the number of inspections is from 2 sets of records, which partially meet
the maximum error allowance. It is demonstrated that the proposed BPM can provide
satisfactory results when more than 4 sets of inspection records are used as its inputs.
Fig.10

Average errors for different numbers of training inputs (bridge#

0312xxx1)

4.0 BPM for BMS Condition Ratings

The BPM is validated using two different methods as described in Section 3. The
BPM is still required to deal with actual BMS inputs to demonstrate the contribution
of the BPM on the research problem.
The BPM test for BMS actual condition rating inputs is conducted by using one
element in the superstructure on the same bridge (Bridge# 3210xxx1) in Section 3.
Thus, bridge specifications and its lifecycle are identical. The bridge sample datasets
obtained (Element #234: Reinforced Concrete Pier Cap) are the actual BMS condition
rating inputs and have been collected to periodically update the BMS database. In
general, the element-level inspection results contain more detailed condition states
than the NBI (component level of condition ratings). In the element-level inspection
method, the condition ratings of Element #234 are quantified using five different
condition states.
Fig.11 describes the BPM timeframe for the obtained bridge Element #234. It shows
the time in the number of years for: (a) the entire bridge life cycle; (b) available
condition ratings; (c) BPM inputs; (d) generated historical condition ratings; (e) inputs
for validation; (f) forward-prediction results; and (g) result comparisons of forwardpredictions with the existing condition rating datasets. Only 5 historical condition
rating datasets (from 1996 to 2004 with 2-year increments) are available to be used as
BPM input values as detailed earlier in Table 2.

Fig.11

BPM timeframe of Element #234 in Bridge #3210xxx1 (Test using BMS

inputs for performance measurements)

The average quantity of each CS on this element#234 between 1996 and 2004 is 80%,
16.2%, and 3.8% of the total elements in CS1, CS2, and CS3, respectively. The BPM
generates historical condition ratings from 1968 to 1994 in three different proportions
of the condition state as shown in Fig.12. Fig.12 (c) shows that 3.8% of total elements
have historically deteriorated more than the others. In other words, maintenance
activities (repair or replacement) on these numbers of elements have been done
historically.
Fig.12

Backward-prediction results for Element #234 for Bridge #0312xxx1

(Note that the number of prediction results in each year is 66 which is the combined
number of learning rates (lr:0.0-0.5) and momentum coefficients (mc:0.0-1.0) in the
neural network configuration)

In addition, the format of the final results has to be modified to conform to the type of
element level inspection. The BPM outcomes cannot be used directly as BMS inputs.
Hence, the BPM results are required to go through a simple post-calibration process.
Table 4 shows an example of BPM results in 1982. The element quantity can be
determined by using Equation 1, from that, the yearly element quantity is shown in
Table 5. These graph and table represent the BPM results using 5 existing condition
rating datasets (1996-2004), which are employed to generate historical condition
ratings from 1968 to 1994 with 2-year increments. The average proportion of each CS
is also shown.
Table 4

Conversion from BPM results to BMS input format (Year 1982 of Element
#234 for Bridge #0312xxx1)

Table 5

Results of BPM for BMS inputs

Total quantity
of subjected =
element

Proportion of
Total quantity of
Average
BPM results subjected element proportion of
per CS
per CS
CS

[1]

4.1 Performance Measurements

To validate the BPM outcomes (1968-1994), the forward comparison method
(Test#2) is used. This validation method already tested in earlier BPM modelling has
been described by using NBI datasets in Section 3.2. This is because condition ratings
between 1967 and 1994 are not available for direct comparison.
For the forward comparison (Test #2), the back-prediction results (1968-1994) are
used as input datasets in this test to generate the condition ratings for the present years
(1996-2004). The BPM generated condition ratings are then directly compared with
existing condition rating datasets. The BPM results are shown in Fig. 13 in four
different proportions of element quantity.
Fig.13

Performance measurements of Element #234 for Bridge #0312xxx1

9

(Note that the number of prediction results in each year is 66 which is the
combined number of learning rates (lr:0.0-0.5) and momentum coefficients
(mc:0.0-1.0) in the neural network configuration)
The identical calibration processes using Equation 1 are performed to conform to the
type of element-level inspection format as shown in Table 4. Table 6 shows the final
results from the BPM as well as its prediction errors. The yearly average prediction
errors are less than 10%, which is acceptable. Therefore, the generated historical
condition ratings (1966-1994) by the BPM can be used as historical condition ratings.
Table 6 Prediction errors of the BPM using forward comparisons
5.0 Case Study
This section provides brief descriptions and results of the case study. This case study
was conducted to present an additional validation for the BPM.
The obtained bridge sample datasets and its corresponding years of non-bridge factors
including climates and population growth for these case studies are provided by Road
and Traffic Authority New South Wales (RTA NSW), the Australian Government
Bureau of Meteorology and the Australian Bureau of Statistics (ABS).
RTA NSW is one of the earlier BMS adopters among bridge authorities in Australia
and has utilised the PONTIS BMS software, which is based on an element-level
inspection method to collect bridge condition ratings. The main reason for adopting
their bridge sample data was that they hold the largest quantity of bridge element
condition rating datasets for BMS software in Australia.
RTA NSW provided 10 different bridge sample data with 15 different types of bridge
elements. Most given bridges were built during the 1960s and 1970s, with an
approximate average life cycle of 40 years.
The number of bridges element inspection datasets obtained and used for the BPM is
mostly between 4 to 6 of the inspection records for the 10 to 12 years of historical
bridge condition ratings. The most typical bridge element types selected for the case
studies are defined in the RTA bridge inspection procedure manual [16] as detailed in
Table 7. The total number of the most typical bridge elements used in the case studies
is nine from seven bridges.
Table 7

Typical bridge elements tested

The selected bridge elements have 3 to 5 different condition states (CSs) for the
evaluation of their bridge element condition ratings. As illustrated in Fig.14, all tested
actual condition states are calibrated to suit the BPM input environment.
Fig.14

Scales of bridge condition states (CSs) for the BPM

Similar types of non-bridge factors (as outlined in Section 3 & 4) are employed for
the BPMs training data, such as historical population growth and climates
surrounding the bridge area. Two historical population variables (city and state
population growth) and 5 different climate variables (maximum temperatures > 40 C,
maximum temperatures > 35 C, minimum temperature < 0 C, highest maximum

10

temperature, lowest minimum temperature) were used as non-bridge factors in this

case study.
Among the selected bridge elements in the case study, this section demonstrates one
of 3-CS's bridge elements (the element code BELA: Elastomeric Bearing Pad). The
available input datasets for the BPM consists of 5 inspections as detailed in Table 8
and have 1-2 year regular inspection intervals. The timeframe for the case study of
element code BELA is described in Fig.15 showing the time in number of years for:
(a) the entire bridge life cycle; (b) available condition ratings; (c) BPM inputs; (d)
generated historical condition ratings; (e) inputs for validation; (f) forward-prediction
results; and (g) result comparisons of forward-predictions with the existing condition
rating datasets.
Table 8

Element condition ratings (3-CSs) obtained for the case study

Fig.15

BPM timeframe for Element code BELA (Bridge# 5xx0)

5.1 BPM Modelling and Results

The average quantity of each CS on this element between 1994-2005 is concentrated
on CS1 showing 100% of the total elements. The BPM generates historical condition
ratings from 1974 to 1992, with 2-year increments using 100% proportions of CS1.
The BPM generated 10 historical condition ratings for the past 20 years from 1974 to
1992 of Element code BELA. The yearly average predictions are shown in Fig.16.
Most past conditions are positioned in CS1 and CS2. However, there are some major
condition improvements between 1980-1982 and 1988-1991, which are observed
from the backward prediction results.
Fig.16

Backward prediction results for Element BELA on Bridge# 5xx0

The outcomes from neural network modelling cannot be used directly as BMS
database inputs. The format of BPM results has to be calibrated to conform to the
element level inspection format. Thus, generated past condition ratings require postcalibrations using Equation 1 as given in Section 4.0. As a result, the yearly numbers
of elements from 1974 to 1992 per condition state can be determined. These results
are presented in Table 9 in accordance with the BMS software input format.
Table 9

BPM results as BMS inputs (Element BELA for Bridge #5xx0)

5.2 Evaluation of BPM prediction

Another neural network model is required for BPM verification. The validation of all
elements obtained is conducted using the second validation method (generated
historical condition rating datasets are used as input datasets for this neural network
training) presented in Section 4. This is because this particular type of inspection
method was not previously employed, hence, direct comparisons between generated
and existing datasets are not possible. However, as shown in Section 4, the forward
comparison method is validated to measure the BPM prediction accuracy.
The ANN training datasets used the BPM results collected from 1973 to 1992
including the assumed element condition ratings at the year of bridge construction as

11

shown in Section 5.1. To compare these condition ratings with the existing datasets,
results from the neural network model are converted to follow the format of the
element-level inspection method using Equation 1.
As a result, in CS1, accurate results using the neural network model for this element
are obtained as shown in Fig.17 (a). Similarly, Fig.17 (b) and (c) shows respectively
the acceptable prediction errors in CS2 and CS3 of up to about 2.5%. The maximum
error in 3-CS's element condition ratings are about 33% (16.5%).
Fig.17

Comparisons on each condition state between forward-manner prediction

results and existing condition ratings from 1994 to 2005 for Bridge #5xx0

However, using the prediction errors in CS1, the actual number of elements is greater
than the predicted number of elements as shown in Table 10. In other words, the
prediction results present a lower risk. Consequently, the generated historical element
condition ratings for Element code BELA are acceptable to be recorded in the BMS
database as historical bridge inspection records.
Table 10 Comparisons between forward-predictions and existing condition ratings
for Bridge #5xx0
5.3 Summary of case studies
The other typical bridge elements are modelled to demonstrate the capability of the
BPM. The test methods, which generate the missing bridge elements' condition
ratings and their validations, are identical to those presented in Section 5. As shown in
Table 11, the overall results from the case studies meet the pre-defined maximum
prediction errors for the nine bridge elements. The maximum average yearly
prediction errors are found to be 10.21%, 9.26% and 4.40% respectively for the three
different condition states (CSs). These BPM results are considered satisfactory.
Table 11 Average condition rating differences of the 9 typical bridge elements

6.0 Discussion
The development of the ANN-based Backward Prediction Model (BPM) is described
in this paper. The BPM uses non-bridge factors as supplementary historical data to
overcome the lack of historical bridge data in terms of its quantity and patterns. The
non-bridge factors including local climates, traffic volume and population growth in
the area surrounding the bridge are employed to help establish the correlations
between the non-bridge factors and the lack of historical data patterns in the existing
condition ratings.
To establish the methodology of backward manner prediction, sample structural
condition rating datasets made available by the Maryland Department of
Transportation (Maryland DoT), and based on five existing condition rating datasets
(or 26% of the total record), ensured that the BPM was able to generate 14 missing
datasets (or 74%) for the intervening years when proper inspection records were
missing. The average ratio of the generated and existing datasets is about 3. The
average prediction errors of the generated bridge condition ratings are between 6.7%7.5% over a period of 20 years.

12

To confirm the BPM methodology established, case studies were performed using 9
typical bridge elements from the Road and Traffic Authority of New South Wales
(RTA NSW). Test methods to generate missing bridge elements condition ratings
and their validation methods were identical to those presented in the pilot study. The
maximum yearly prediction errors of three different condition state scales (CS 3, 4
and 5) are 18.61%, 23.89% and 4.40% respectively. These are satisfactory as
compared to the maximum errors of 33.33%, 25% and 20% for 3CSs, 4CSs and 5CSs,
respectively.
In addition, the methodology described in this paper aims to establish a possible
solution and a very initial approach for determining the unavailable historical
structural condition ratings. Further study may be required to explore the types and
numbers of non-bridge factors for the BPM. As further critical non-bridge factors are
identified and incorporated into the model, the accuracy of correlations with a small
number of structural condition ratings, as well as future condition rating predictions,
will improve.

7.0 Conclusions
It is generally recognised that for most bridges there are big time gaps between the
dates of construction, adoption and implementation of relevant Bridge Management
Systems (BMSs). As a result, suitable bridge inspection records, and of course the
bridge element condition rating datasets, are missing or unavailable for the
intervening years. The missing data, which is required as input to the relevant analysis
modules of any BMS, is a major factor contributing to the unreliable BMS outcomes
currently experienced by most bridge agencies.
The main focus of the proposed research is to rectify this untenable situation faced by
these agencies in most countries. To this end, the so-called Back Prediction Model
(BPM) is developed as part of this research effort. Capable of generating the missing
data, the BPM incorporates ANN techniques and operates on the limited existing
inspection records and historical non-bridge factors such as local climates, number of
vehicles and population growth in the area surrounding the bridge. A total of nine
case studies conducted in this research confirms the reliability of the BPM to help
produce accurate BMS outcomes.
In conclusion, the main goal of this research is to develop an appropriate methodology
for an effective bridge condition rating model to produce the historical data of the
relevant bridge elements. The BPM is believed to be a useful methodology for an
effective implementation of BMS packages and deserves wider application. However
further insight into the appropriate non-bridge factors is required as recommended
below:
- Parametric studies should be conducted to determine the optimum numbers and
types of non-bridge factors for different locations and types of bridges. This should
lead to more effective operations of the BPM by excluding the unnecessary nonbridge factors to yield a superior correlation between condition ratings and the wellchosen non-bridge factors.

13

- Development of a Long-term Bridge Performance (LTBP) model should be carried

out using a comprehensive structural condition rating database. This should lead to
more reliable future structural condition rating predictions. This will also enhance
the reliable outcomes of many other analytical BMS modules.

Acknowledgements
The resources used for the present study were provided by Maryland State
Department of Transportation, U.S. The writers wish to thank Mr. Earle Freedman
and Matt Zulkowski from Maryland Department of Transportation for tracking our
information requests. Authors also would like to thank Mr. Perumynar Siva of the
New South Wales Road and Traffic Authority, who has provided the important
historical bridge datasets which enabled the case studies to be successfully completed.
In particular, Professor Waheed Uddin from The University of Mississippi for
insightful comments and suggestions.

Abbreviations
ANN
Artificial Neural Network
BMS
Bridge Management System
BPM
Backward Prediction Model
CI
Condition Index
CS
Condition State
DB
Database
DoT
Department of Transportation
IMS
Infrastructure Management System
MR&R
Maintenance, Repair & Rehabilitation
NBI
National Bridge Inventory

References
[1] D.M. Frangopol, E.S. Gharaibeh, J.S. Kong and M. Miyake, Optimal NetworkLevel Bridge Maintenance Planning Based on Minimum Expected Cost, In the
Proce. of the Transportation Research Record, Florida, (2000) 26-33.
[2] S. Madanat, Optimal infrastructure management decisions under uncertainty,
Transportation Research Part. C., 1 (1993) 77-88.
[3] G. Hearn, Segmental Inspection for Improved Condition Reporting in BMS, In
the Proce. of the Eighth Int. Bridge Management Conference, Denver, Colorado,
(1999) B-3/1-8.
[4] B. Godart and P.R. Vassie, Review of existing BMS and definition of inputs for
the proposed BMS, Deliverable D4: BRIME Report (1999 a).
[5] G. Hearn, R.L. Purvis, P.D. Thompson, W.H. Bushman, K.K. McGhee and W.T.
McKeel, Bridge Maintenance and Management: A look to the future, In the
Proce. of the TRB 81st Annual Meeting: A3C06:Structures Maintenance and
Management (2000) 1-7.
[6] A. Weigend and N. Gershenfeld, Time Series Prediction: Forecasting the
Futureand Understanding the Past, Addison-Wesley, Reading, MA, (1994) 5966.
[7] B.G. Tabachinick and L.S. Fidell, Using Multivariate Statistics, Allyn and
Bacon (2001).

14

[8]
[9]

[10]
[11]
[12]
[13]
[14]
[15]
[16]

T. Karna, F. Rossi and A. Lendasse, LS-SVM functional network for time series
prediction, In the Proce. of the of XIVth European Symposium on Artificial Neural
Networks (ESANN 2006), Bruges. Belgium, (2006) 473-478.
R.G. Mishalani and M.R. McCord, Infrastructure Condition Assessment,
Deterioration Modeling and Maintenance Decision Making: Methodological
Advances and Practical Considerations, Journal of Infrastructure Systems, 12(3)
(2006) 145-146.
E.P. Small, T. Philbin, M. Fraher and G.P. Romack, Current Status of Bridge
Management System Implementation in the United States, Eighth International
Bridge Management Conference, Denver, Colorado, (1999) A-1/1-16.
J.H. Milligan, R.J. Nielsen and E.R. Schmeckpeper, Implementing PONTIS As
a Bridge Management Tool in Idaho, N04-04, National Institute for Advanced
Transportation Technology, University of Idaho, (2004) 2-5.
Federal Highway Administration, Element Level Bridge Inspection (Bridge
Management and Inspection Technologies), FHWA-RC-BAL-04-0015,
Washington DC (2006).
A.K. Smith, Introduction to neural networks and data mining for business
applications, Melbourne: Emerald, Vic. Eruditions Publishing (1999).
J. Anderson, An Introduction to Neural Networks, The MIT Press, Cambridge,
MA. (1995).
M. Nelson and W. Illinworth, A Practical Guide to Neural Nets, AddisonWesley Publishing Company, Reading, MA, (1990) 104.
Road and Traffic Authority of New South Wales, RTA Bridge Inspection
Procedure, RTA NSW (1999).

Legend of Figures
Fig.1

Re-illustration of relationships between historical bridge inspection data

and BMS outputs [4]
(Note that the relationships with other input sources and BMS outputs are
omitted)

Fig.2

Outline of proposed Backward Prediction Model (BPM)

Fig.3

Scale of the condition ratings used in the BPM model

Fig.3 (a) NBI
Fig.3 (b) BMS element condition ratings

Fig.4

Structure of ANN-based BPM

Fig.5

Raw data of non-bridge factors for BPM inputs

Fig.5 (a) Historical vehicle change
Fig.5 (b) Historical population change
Fig.5 (c) Historical Temperature change

15

(MMXT: Mean Maximum; MMNT: Mean Minimum; MNTM:

Mean; EMXT: Highest; EMNP: Lowest; DT90: Max. Number of day
90F; DX32: Max. Number of day 32F; DT32: Min. Number of
day 32F; DT00: Min. Number of day 0F)
Fig.5 (d) Historical precipitation change
(TPCP: Total Precipitation; TSNW: Snow and Sleet Total Fall;
MXSD: Snow and Sleet Max. Depth; DP01: Number of Days 0.1;
DP05: Number of Days 0.5; DP10: Number of Days 1.0)
Fig.6

Raw data of NBI for BPM outputs (Bridge #0301xxxx1)

Fig.7

BPM timeframe (Test#1 and 2 using NBI for performance measurements)

Fig.8

BPM results for Bridge#0312xxx1

Fig.8 (a) Deck
Fig.8 (b) Superstructure
Fig.8 (c) Substructure

Fig.9

Performance measurements: existing NBI vs BPM results (Bridge #

0312xxx1)
Fig.9 (a) Deck
Fig.9 (b) Superstructure
Fig.9 (c) Substructure

Fig.10

Average errors for different numbers of training inputs (bridge#

0312xxx1)

Fig.11

BPM timeframe of Element #234 in Bridge #3210xxx1 (Test using BMS

inputs for performance measurements)

Fig.12

Backward-prediction results for Element #234 for Bridge #0312xxx1

(Note that the number of prediction results in each year is 66 which is the
combined number of learning rates (lr:0.0-0.5) and momentum coefficients
(mc:0.0-1.0) in the neural network configuration)
Fig.12 (a) About 80% of the total quantity
Fig.12 (b) About 16.2% of the total quantity
Fig.12 (c) About 3.8% of the total quantity

Fig.13

Performance measurements of Element #234 for Bridge #0312xxx1

(Note that the number of prediction results in each year is 66 which is the
combined number of learning rates (lr:0.0-0.5) and momentum coefficients
(mc:0.0-1.0) in the neural network configuration)
Fig.13 (a) About 84.20% of the total quantity
Fig.13 (b) About 14.16% of the total quantity
Fig.13 (c) About 1.62% of the total quantity
Fig.13 (d) About 0.02% of the total quantity

Fig.14

Scales of bridge condition states (CSs) for the BPM

Fig.15

BPM timeframe for Element code BELA (Bridge# 5xx0)

16

Fig.16

Backward prediction results for Element BELA on Bridge# 5xx0

(Note that the number of prediction results in each year is 66 which is the
combined number of learning rates (lr:0.0-0.5) and momentum coefficients
(mc:0.0-1.0) in the neural network configuration)

Fig.17

Comparisons on each condition state between forward-manner prediction

results and existing condition ratings from 1994 to 2005 for Bridge #5xx0
Fig.17 (a) CS1
Fig.17 (b) CS2
Fig.17 (c) CS3

Legend of Tables
Table 1

The components of the proposed neural network model

Table 2

Raw data of actual condition ratings (Element #234 on Bridge

#0301xxxx1)

Table 3

Summary of prediction performance for test 2

Table 4

Conversion from BPM results to BMS input format (Year 1982 of Element
#234 for Bridge #0312xxx1)

Table 5

Results of BPM for BMS inputs

Table 6

Prediction errors of the BPM using forward comparisons

Table 7

Typical bridge elements tested

Table 8

Element condition ratings (3CSs) obtained for the case study

17

Table 9

BPM results as BMS inputs (Element BELA for Bridge #5xx0)

Table 10 Comparisons between forward-predictions and existing condition ratings

for Bridge #5xx0
Table 11 Average condition rating differences of the 9 typical bridge elements

18

Input

Project-level Output

Network-level Output

Inventory

A1. General Queries

B1. List/count bridge satisfying

Inspection

A2. Inspection condition

history

B2. List/count inspection overdue

Maintenance

A3. Maintenance history

B3. List/count bridges that are

substandard

A4. Traffic history

B4. List/count poor condition

bridges

Load carrying
capacity

A5. Load carrying capacity

history

B5. List/count bridges with traffic

restrictions

Posting history

A6. Posting history

B6. Budget for optimal

maintenance plan

A7. Prediction variation of

load carrying capacity

B7. Number of bridges with

deffered maintenance

Traffic

A8. Prediction variation of

bridge condition

B8. Long term cost of

maintenance

A9. Estimating
maintenance cost

B9. Prioritised maintenance plan

A10. Estimaing traffic

delay cost

B10. Prediction mean load

carrying capacity

A11. Optimal maintenance

plan

B11. Prediction mean bridge

condition for given budget

A12. Test result history

B12. Routing of vehicles

B13. History of different types of
maintenance
B14. History of occurrence of
different types of defect
B15. History of the occurrence of
substandard bridges
B16. History of the performance
of elements/components
B17. Cost rates for different
maintenance options
B18. History of the performance
of different maintenance methods

Fig.1. Re-illustration of relationships between historical bridge inspection data and

BMS outputs (Godart and Vassie, 1999 a)
(Note that the relationships with other input sources and BMS outputs are omitted)

Present
time

Year of
construction

Missing Information

Existing
Information

Historical
bridge
condition
ratings

Generate Historical Information

Correlation
by ANN
Non-Bridge Factors

m+n

Time
[Years]

Fig.2. Outline of proposed Backward Prediction Model (BPM)

Excellent Condition
Very Good Condition
Good Condition
Satisfactory Condition
Fair Condition
Poor Condition
Serious Condition
Critical Condition
Failure Condition
Failed Condition

1.0
0.9
0.8
0.7
0.6
0.5
0.4
3
0.3
2
0.2
1
0.1
0
0.0
Scale for the
NBI
#58, 59 & 60 ANN model
9
8
7
6
5
4

(a) NBI

Excellent CS 1
Condition

1.0
0.9
0.8

CS 2

0.7
0.6

CS 3

0.5
0.4

CS 4
Failure CS 5
Condition

0.3
0.2

0.1
0.0
Element level Scale for the
inspection ANN model

(b) BMS element condition ratings

Fig.3. Scale of the condition ratings used in the BPM model

Output
Layer

Climates

...

...
...

Population
growth

...

...

Number of
vehicles

Hidden
Layer

...

Input
Layer

Non-bridge
factors

Available
condition ratings

Fig.4. Structure of ANN-based BPM

0.60
Passenger vehicle
Bus
T ruck
T otal vehicle

0.40

Population
[10^7]
Number
of population
[x107]

7
Number
]
Number of
of vehicles
Vehicles [x10
[10^7]

0.50

0.30
0.20
0.10
0.00
-

0.40
BALT IMORE
Baltimore
City
CIT Y
Maryland
MARYLAND

0.30
0.20
0.10
0.00

1966

1972

1978 1984 1990

Time [Years]

1996

2002

1966

(a) Historical vehicle change

M M XT
EM XT
DX32

M M NT
EM NP
DT32

M NTM
DT90
DT00

0.11
0.07
0.03
-0.01
1966

1972

1978

1984

1990

1996

1972

1978 1984 1990

Time[Years]
[Years]
Time

1996

2002

(b) Historical population change

2002

T P CP
MXSD
DP 0 5

0.12
Precipitation[inches,
[inches,10^3]
x103]
Precipitation

0.15
Temperature
F, 10^3]
Temperature[[F,
x103]

0.50

0.10

T SNW
DP 01
DP 10

0.08
0.06
0.04
0.02
0.00

-0.05

T ime[Years]
[Years]
Time

(c) Historical climate change

(MMXT: Mean Maximum; MMNT: Mean
Minimum; MNTM: Mean; EMXT: Highest;
EMNP: Lowest; DT90: Max. Number of day
90F; DX32: Max. Number of day 32F;
DT32: Min. Number of day 32F; DT00:
Min. Number of day 0F)

1966

1972

1978 1984 1990

T ime [Years]
[Years]
Time

1996

2002

(d) Historical precipitation change

(TPCP: Total Precipitation; TSNW: Snow and
Sleet Total Fall; MXSD: Snow and Sleet Max.
Depth; DP01: Number of Days 0.1; DP05:
Number of Days 0.5; DP10: Number of Days
1.0)

Fig.5. Raw data of non-bridge factors for BPM inputs

Condition Ratings [10]

Assumed condition ratings

for the proposed BPM

0.8
0.6
0.4
Deck
Superstructure
Substructure

0.2
0
1966

1974

1982
1990
T ime [Years]

1998

Fig.6. Raw data of NBI for BPM outputs (Bridge #0301xxxx1)

(a) Entire Timeframe

66

68

70

72

74

76

78

80

82

84

86

88

90

92

94

96

98

00

02

04

74

76

78

80

82

84

86

88

90

92

94

96

98

00

02

04

96

98

00

02

04

96

98

00

02

04

96

98

00

02

04

Test #2 using NBI

Test #1 using NBI

(b) Available NBI

(c) BPM inputs

66

(d) BPM Results

68

(e) Compare the results

with NBI

70

72

74

76

78

80

82

84

86

88

90

92

94

74

76

78

80

82

84

86

88

90

92

94

74

76

78

80

82

84

86

88

90

92

94

(f) Inputs for validation

(input from test#1 results)
66

68

70

72

(g) Results for the validation

(h) Compare the results
with NBI

Time [Years]

Fig.7. BPM timeframe (Test#1 and 2 using NBI for performance measurements)

1.00

1.00

NBI

NBI

0.80
Condition Ratings

Condition Ratings

0.80
0.60
0.40
Predictions
(Average values )

0.20
0.00

0.60
0.40
Predictions
(Average values )

0.20
0.00

1967 1971 1975 1979 1983 1987 1991 1995

Time [Years]

1967 1971 1975 1979 1983 1987 1991 1995

Time [Years]

(a) Deck

(b) Superstructure
1.00

Condition Ratings

0.80
0.60
0.40

NBI

Predictions
(Average values )

0.20
0.00
1967 1971 1975 1979 1983 1987 1991 1995
Time [Years]

(c) Substructure
Fig.8. BPM results for Bridge#0312xxx1
(Note that the number of prediction results in each year is 66 which is the combined number of learning
rates (lr:0.0-0.5) and momentum coefficients (mc:0.0-1.0) in the neural network configuration)

1.00

Average
prediction

0.80

NBI

Condition Ratings

0.60
0.40
0.20
0.00

Average
prediction

0.80

NBI

0.60
0.40
0.20

94
19

92
19

86

19
8

19

Time [Years]

(a) Deck

(b) Superstructure
1.00

Average
prediction

0.80
Condition Ratings

84

Time [Years]

19

78

19
8

19

76

19
7

2
19
94

19
9

19
8

19
8

19
8

19
8

19
7

19
7

19
7

0.00

19

Condition Ratings

1.00

NBI

0.60
0.40
0.20

94
19

92
19

88
19

84

82

86
19

19

19

76

78
19

19

19

74

0.00

Time [Years]

(c) Substructure
Fig.9. Performance measurements: existing NBI vs BPM results (Bridge #
0312xxx1)

20

Average errors (%)

16

Superstructure

Deck

Max. error
allowance

12
8
4

Substructure

0
0

2
3
4
No. of input data set(s)

Fig.10 Average errors for different numbers of training inputs (bridge# 0312xxx1)

10

(a) Entire Timeframe

66

68

70

72

74

76

78

80

82

84

86

88

90

92

94

96

98

00

02

04

96

98

00

02

04

96

98

00

02

04

96

98

00

02

04

96

98

00

02

04

(b) Available BMS inputs

(c) BPM inputs

Test using BMS Inputs

66

(d) BPM Results

68

70

72

74

76

78

80

82

84

86

88

90

92

94

68

70

72

74

76

78

80

82

84

86

88

90

92

94

(e) Inputs for validation

(input from test results)
(f) Results for validation
(g) Compare test results
with existing BMS inputs

Time [Years]

Fig.11. BPM timeframe of Element #234 in Bridge #3210xxx1 (Test using BMS
inputs for performance measurements)

Formatted: Left

11

0.8

0.8
Condition Ratings

Condition Ratings

Average condition
ratings for 80% of
total quantity

0.6
0.4
0.2
0
1965

0.6
Average condition
ratings for 16.2%
of total quantity

0.4
0.2

1971

1977
1983
T ime [Years]

1989

1995

(a) About 80% of the total quantity

0
1965

1970

1975

1980 1985
T ime [Years]

1990

1995

(b) About 16.2% of the total quantity

Condition Ratings

0.8
0.6
0.4
Average condition
ratings for 3.8% of
total quantity

0.2
0
1965

1970

1975 1980 1985

T ime [Years]

1990

1995

(c) About 3.8% of the total quantity

Fig.12. Backward-prediction results for Element #234 for Bridge #0312xxx1
(Note that the number of prediction results in each year is 66 which is the combined number of learning
rates (lr:0.0-0.5) and momentum coefficients (mc:0.0-1.0) in the neural network configuration)

12

0.8

0.8

0.6

Condition Ratings

Condition Ratings

Average condition
ratings for 84.20%
of total quantity

0.4
0.2

0.6

0.2
0

0
1995

1997

1999
2001
Time [Years]

2003

1997

1999
2001
Time [Years]

2003

2005

(b) About 14.16% of the total quantity

1
Condition Ratings

Average condition
ratings for 1.62%
of total quantity

0.8

1995

2005

(a) About 84.20% of the total quantity

Condition Ratings

Average condition
ratings for 14.16%
of total quantity

0.4

0.6
0.4
0.2
0

0.8
0.6
0.4
Average condition
ratings for 0.02%
of total quantity

0.2
0

1995

1997

1999
2001
Time [Years]

2003

2005

(c) About 1.62% of the total quantity

1995

1997

1999
2001
Time [Years]

2003

2005

(d) About 0.02% of the total quantity

Fig.13. Performance measurements of Element #234 for Bridge #0312xxx1

(Note that the number of prediction results in each year is 66 which is the combined number of learning
rates (lr:0.0-0.5) and momentum coefficients (mc:0.0-1.0) in the neural network configuration)

13

Excellent
Condition CS 1

1.00

1.0

1.00
0.825

CS 1

0.875
0.75

0.66
CS 2

CS 2

0.33
Failure
Condition

CS 3

3 CS

0.165
0.00
Scale for the
ANN model

CS 2
CS 3

4 CS

0.5
0.4

CS 4

0.25
CS 4

0.7
0.6

0.375

CS 3

0.9
0.8

0.625
0.50

0.495

CS 1

0.125
0.00
Scale for the
ANN model

0.3
0.2

CS 5
5 CS

0.1
0.0
Scale for the
ANN model

Fig.14 Scales of bridge condition states (CSs) for the BPM

14

(a) Entire timeframe

73 74

76

78

80

82

84

86

88

90

92

94

96

98

03

05

94

96

98

03

05

94

96

98

03

05

94

96

98

03

05

94

96

98

03

05

(b) Available BMS inputs

(c) BPM inputs

73

(d) BPM results

74

76

78

80

82

84

86

88

90

92

73 74

76

78

80

82

84

86

88

90

92

(e) Inputs for validation

(input from test results)
(f) Results for validation
(g) Compare test results
with existing BMS inputs

Time [Years]

Fig.15 BPM timeframe for Element code BELA (Bridge# 5xx0)

15

Condition Ratings

0.99

0.66

0.33

Average condition ratings

for 100% of total quantity
during 1994-2005

0.00
1973 1975 1977 1979 1981 1983 1985 1987 1989 1991 1993
Time [Years]

Fig.16 Backward prediction results for Element BELA on Bridge# 5xx0

(Note that the number of prediction results in each year is 66 which is the combined number of learning
rates (lr:0.0-0.5) and momentum coefficients (mc:0.0-1.0) in the neural network configuration)

16

100
Forward prediction results
Existing condition ratings

Number of Elements

80

60

40

20

0
1994

1996

1998
Time [Years]

2003

2005

(a) CS1
20

Number of Elements

16

12

8
Forward prediction results

0
1994

1996

1998

2003

2005

2003

2005

Time [Years]

(b) CS2
10

Number of Elements

4
Forward prediction results

0
1994

1996

1998
Time [Years]

(c) CS3
Fig.17 Comparisons on each condition state between forward-manner prediction
results and existing condition ratings from 1994 to 2005 for Bridge #5xx0

17

Table 1. The components of the proposed neural network model

Training Algorithm

Back-propagation feed-forward

Transfer Function

Log-sigmoid function

Inputs

Number of vehicles (4 factors)

Population growth (2 factors)
Climates (15 factors)

Output

Bridge condition ratings (1 output)

Table 2. Raw data of actual condition ratings (Element #234 on Bridge #0301xxxx1)
Year of
inspection

Total
quantity
(%)

CS1
(%)

CS2
(%)

CS3
(%)

CS4
(%)

CS5
(%)

1996

350
(100)

280
(80)

50
(14)

20
(6)

1998

350
(100)

280
(80)

50
(14)

20
(6)

2000

350
(100)

280
(80)

50
(14)

20
(6)

2002

350
(100)

283
(80)

67
(19)

3
(1)

2004

350
(100)

283
(80)

67
(19)

3
(1)

Table 3. Summary of prediction performance for test 2

Year
1996

1998

2000

2002

2004
Mean
Errors
(%)

Bridge
component
Deck
Superstructure
Substructure
Deck
Superstructure
Substructure
Deck
Superstructure
Substructure
Deck
Superstructure
Substructure
Deck
Superstructure
Substructure
Deck
Superstructure
Substructure

NBI
records
0.600
0.600
0.600
0.600
0.600
0.600
0.600
0.600
0.600
0.600
0.600
0.600
0.600
0.600
0.600

Average
prediction
0.568
0.567
0.569
0.590
0.593
0.591
0.665
0.666
0.669
0.590
0.595
0.591
0.555
0.556
0.557
3.20
3.10
3.20

Difference
0.032
0.033
0.031
0.010
0.007
0.009
0.065
0.066
0.069
0.010
0.005
0.009
0.045
0.044
0.043

Table 4. Conversion from BPM results to BMS input format (Year 1982 of Element
#234 for Bridge #0312xxx1)

Condition
State

Proportion
1

Proportion
2

Proportion
3

CS1
CS2
CS3
CS4
CS5
CS1
CS2
CS3
CS4
CS5
CS1
CS2
CS3
CS4
CS5

Prediction
results
98.48%
1.52%
0.00%
0.00%
0.00%
13.64%
86.36%
0.00%
0.00%
0.00%
1.52%
87.88%
10.61%
0.00%
0.00%

Total
elements
350
350
350
350
350
350
350
350
350
350
350
350
350
350
350

Total number of elements

Average
proportion
of element
from 1996
to 2004

80.0%

16.2%

3.8%

Number
of
elements

Number
of
elements
for a BMS

275.76
4.24
0.00
0.00
0.00
7.73
48.97
0.00
0.00
0.00
0.20
11.69
1.41
0.00
0.00

276
4
0
0
0
8
49
0
0
0
0
12
1
0
0

350.00

350

Table 5. Results of BPM for BMS inputs

Year
CS1
CS2
CS3
CS4
CS5
Total
Year
CS1
CS2
CS3
CS4
CS5
Total

1968
(%)
91.10
8.90
0.00
0.00
0.00
100.00
1982
(%)
81.05
18.54
0.40
0.00
0.00
100.00

Total quantity
of subjected =
element

1970
(%)
94.31
5.69
0.00
0.00
0.00
100.00
1984
(%)
78.79
17.47
3.68
0.06
0.00
100.00

1972
(%)
80.75
17.46
1.78
0.00
0.00
100.00
1986
(%)
80.00
17.75
2.25
0.00
0.00
100.00

1988
(%)
88.65
11.35
0.00
0.00
0.00
100.00

1974
(%)
79.52
17.42
3.05
0.00
0.00
100.00
1990
(%)
80.00
16.20
3.63
0.17
0.00
100.00

1976
(%)
80.51
18.92
0.58
0.00
0.00
100.00
1992
(%)
80.00
16.20
3.80
0.00
0.00
100.00

Proportion of
Total quantity of
Average
BPM results subjected element proportion of
per CS
per CS
CS

[Equation 1]

1978
(%)
92.76
7.24
0.00
0.00
0.00
100.00
1994
(%)
80.00
16.49
3.51
0.00
0.00
100.00

1980
(%)
91.35
8.65
0.00
0.00
0.00
100.00
Average
(%)
84.20
14.16
1.62
0.02
0.00
100.00

[1]

Table 6. Prediction errors of the BPM using forward comparisons

Results
1996 Existing data
Error
Average error
Results
1998 Existing data
Error
Average error
Results
2000 Existing data
Error
Average error
Results
2002 Existing data
Error
Average error
Results
2004 Existing data
Error
Average error

CS1

CS2

CS3

CS4

CS5

Total

84.20%
80.00%
4.20%

14.18%
14.29%
0.11%

1.52%
0.00%
1.52%

0.00% 100.00%
0.00% 100.00%
0.00%

84.41%
80.00%
4.41%

13.98%
14.29%
0.31%

0.02%
0.00%
0.02%

0.00% 100.00%
0.00% 100.00%
0.00%

87.65%
80.00%
7.65%

10.76%
14.29%
3.53%

1.40%
0.00%
1.40%

0.00% 100.00%
0.00% 100.00%
0.00%

91.29%
80.00%
11.29%

6.96%
19.14%
12.18%

0.54%
0.00%
0.54%

0.00% 100.00%
0.00% 100.00%
0.00%

81.86%
80.00%
1.86%

16.35%
19.14%
2.79%

0.10%
5.71%
5.62%
2.29%
1.58%
5.71%
4.13%
1.78%
0.20%
5.71%
5.52%
3.62%
1.20%
0.86%
0.34%
4.87%
0.85%
0.86%
0.00%
1.12%

0.93%
0.00%
0.93%

0.00% 100.00%
0.00% 100.00%
0.00%

Table 7 Typical bridge elements tested

Element
Code

Description

No. of
condition states
(CSs)
3

Units

BELA

Elastomeric Bearing Pad

ea

CDSL

Concrete Deck Slab

CPIL

Concrete Pile

CPIR

Concrete Pier (excl. any Headstock or Piles)

CPRG

Concrete Pre-tensioned Girder

JASS

Assembly Joint Seal

LBGI

Steel(L) Beam / Girder (Load Bearing)

MMAS

Brick / Masonry / Reinforced Earth

RMET

Metal Railing

Table 8. Element condition ratings (3CSs) obtained for the case study
Bridge
number

5xx0

Construction
year

1973

CS1

CS2

CS3

14/09/1994

Total
element
quantity
80

80

29/11/1996
3/03/1998
26/03/2003
14/12/2005

80
80
80
80

80
80
80
80

0
0
0
0

0
0
0
0

Inspection
date
(dd/mm/yyyy)

Structure
element type
code

BELA

Condition State (CS)

Table 9. BPM results as BMS inputs (Element BELA for Bridge #5xx0)
1974
(%)

1976
(%)

1978
(%)

1980
(%)

1982
(%)

0.66 <CS1 0.99

96.97

95.45

100.00

96.97

95.45

0.33 < CS2 0.66

1.52

1.52

0.00

3.03

4.55

0.00 CS3 0.33

1.52

3.03

0.00

0.00

0.00

Total

100.00

100.00

100.00

100.00

100.00

Year

1984
(%)

1986
(%)

1988
(%)

1990
(%)

1992
(%)

0.66 <CS1 0.99

98.48

100.00

83.33

100.00

100.00

0.33 < CS2 0.66

1.52

0.00

16.67

0.00

0.00

0.00 CS3 0.33

0.00

0.00

0.00

0.00

0.00

100.00

100.00

100.00

100.00

100.00

Year

Total

Table 10. Comparisons between forward-predictions and existing condition ratings for
Bridge #5xx0

1994

1996

1998

2003

2005

Results
Existing data
Error
Average error
Results
Existing data
Error
Average error
Results
Existing data
Error
Average error
Results
Existing data
Error
Average error
Results
Existing data
Error
Average error

CS1

CS2

96.25
100.00
3.75

2.50
0.00
2.50
2.50%
2.53
0.00
2.53
1.69%
2.53
0.00
2.53
1.69%
2.53
0.00
2.53
1.69%
2.50
0.00
2.50
2.50%

97.47
100.00
2.53
97.47
100.00
2.53
97.47
100.00
2.53
96.25
100.00
3.75

CS3

Total

1.25 100.00%
0.00 100.00%
1.25
0.00 100.00%
0.00 100.00%
0.00
0.00 100.00%
0.00 100.00%
0.00
0.00 100.00%
0.00 100.00%
0.00
1.25 100.00%
0.00 100.00%
1.25

Table 11 Average condition rating differences of the 9 typical bridge elements

Description
Assembly Joint Seal
Brick / Masonry / Reinforced Earth
Elastomeric Bearing Pad
Concrete-Deck Slab
Concrete-Pile
Concrete-Pier
Concrete-Pre-tensioned Girder
Metal Railing
Steel(L)-Beam / Girder

No. of
condition
states (CSs)
3
3
3
4
4
4
4
4
5

Max. error
allowance
(%)
33.33
33.33
33.33
25.00
25.00
25.00
25.00
25.00
20.00

Prediction difference (%)

Min

Max

1.69
3.95
1.14
0.71
11.03
1.63
0.73
5.81
1.14

2.50
18.61
9.52
4.18
12.21
4.47
1.56
23.89
4.40