Prévia do material em texto
Effect of Weigh-in-Motion System Measurement Errors on Load-Pavement Impact Estimation Jorge A. Prozzi1 and Feng Hong2 Abstract: Weigh-in-motion �WIM� technology has found increasing application in the highway and transportation areas for traffic data collection for the purpose of highway capacity analysis, aiding enforcement and, most recently, pavement design. The measurement accuracy of a WIM scale is critical. There are numerous existing studies involving the measurement accuracy per se; however, the implications and effect of the accuracy in the context of pavement design have been rarely examined. To address this issue, two particular aspects are considered in this study. First, with traffic data obtained from WIM stations in Texas, axle load spectra are developed from both statistical data fit and load-pavement impact perspectives. Axle load spectra are shown to be captured best by mixed-lognormal distributions. Second, based on the first aspect, two scenarios are investigated by incorporating two types of errors into the load measurements: �1� A random error component due to the WIM scale intrinsic properties; and �2� a systematic error component due to the improper calibration of WIM system. The relationship between jointly varying measurement error levels and load-pavement impact estimation errors is established. It is demonstrated �and quantified� that both types of errors contribute to load-pavement impact estimation inaccuracy. The random error leads to overestimation of load-pavement impact. The results also show that WIM system calibration is of more importance because load-pavement impact estimation is more sensitive and significantly related to systematic error than random error. In addition, the estimated load-pavement impact is found to be more sensitive to overcalibration �positive bias� than undercalibration �negative bias� condition. In summary, the findings in this study provide an effective and efficient approach to evaluate WIM measurement errors in conjunction with load-pavement impact in pavement design and rehabilitation. The findings in this study could be further applied for the determination of adequate pavement design reliability standards. Results can also be useful to highway agencies in assessing and selecting WIM equipment. DOI: 10.1061/�ASCE�0733-947X�2007�133:1�1� CE Database subject headings: Pavements; Measurement; Weight; Loads; Errors. Introduction Soil and material properties, environmental conditions, and traffic loads are the three major inputs controlling the design of new and rehabilitated pavement structures. Typically, material properties are determined through laboratory and field tests. Environmental conditions, including temperature and moisture, are available through established networks of weather stations and databases. Traffic loads play a critical role in the deterioration of pavement performance. In the current American Association of State High- way and Transportation Officials �AASHTO� pavement design guide the effect of traffic on pavement performance is accounted for in terms of the accumulated traffic load, measured in equiva- lent number of single axle loads �ESALs� �AASHTO 1993�. Fur- ther, in the proposed National Cooperative Highway Research Program �NCHRP� Mechanistic-Empirical Design Guide, traffic load will be accounted for by means of axle load spectra �NCHRP 2005�. In this regard, accurate traffic load data collection is of crucial importance for pavement design and performance analysis. Two approaches can be adopted to collect axle load data. One utilizes static scales and the other uses weigh-in-motion �WIM� equipment. Both approaches have their advantages and disadvan- tages. For static weighing, the most appealing characteristic is its high level of accuracy. However, the disadvantages include a se- ries of inconveniences and inefficiencies. For instance, the in- spected vehicle is required to stop while being weighed so that time loss is a burden for both inspectors and drivers. In addition, more personnel are required to conduct a static weighing opera- tion than those necessary to monitor the WIM system. As a result, static vehicle weighing can only be used to collect limited samples of axle load data. These disadvantages, however, can be overcome by applying WIM technology. Compared with static scale, state-of-the-art WIM technology has gained popularity due to, in large part, its ability to collect continuous traffic data without human interven- tion. Theoretically speaking, as soon as it is set up, a WIM scale is able to continuously collect and record vehicle information including date and time of passage, lane and direction of travel, vehicle class, speed, wheel and axle weight, and axle spacing. At a given location of interest, it is capable of collecting the popu- lation data instead of a small sample. Nevertheless, aspects such as WIM system instability due to sensor technology, environmen- tal effects, pavement conditions and other factors give rise to the 1Assistant Professor, Dept. of Civil, Architectural and Environmental Engineering, The Univ. of Texas, ECJ 6.112, Austin, TX 78712. 2Graduate Research Assistant, Dept. of Civil, Architectural and Environmental Engineering, The Univ. of Texas, ECJ 6.510, Austin, TX 78712. Note. Discussion open until June 1, 2007. Separate discussions must be submitted for individual papers. To extend the closing date by one month, a written request must be filed with the ASCE Managing Editor. The manuscript for this paper was submitted for review and possible publication on January 30, 2006; approved on July 26, 2006. This paper is part of the Journal of Transportation Engineering, Vol. 133, No. 1, January 1, 2007. ©ASCE, ISSN 0733-947X/2007/1-1–10/$25.00. JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2007 / 1 concern over its measurement accuracy. Therefore, WIM system reliability for collecting accurate data relies heavily on its peri- odic calibration and maintenance. Due to the importance of proper calibration, research has aimed at developing WIM calibration guidelines or quality assur- ance programs. For instance, Izadmehr and Lee �1987a� sug- gested the type of trucks and minimum number of vehicles required for on-site calibration with Texas WIM systems. The relationship between the tolerance for 95% probability of confor- mity for axle group or gross-vehicle weight �GVW� and three vehicle-speed levels was provided in another study conducted by Izadmehr and Lee �1987b�. Davies and Sommerville �1987� pre- sented a self-calibration approach and an accuracy “funnel” for WIM performance appraisal. Based on existing research, a stan- dard specification was published by the American Society for Testing and Materials �ASTM 2002� to assist highway agents with WIM user requirements and test methods. Apart from the issue of WIM measurement accuracy, the effect of measurement inaccuracy on pavement performance is of inter- est from the pavement design perspective. This is especially criti- cal because the relationship between measurement accuracy and its effect on pavement performance is complex and nonlinear. If a WIM scale is not properly calibrated and underestimates axle load, the expected pavement life would be overestimated resulting in unexpected premature failures. Overestimated loads by the WIM system would lead to the design of thicker and more costly pavement structures than may be actually required. For instance, the report by the Federal Highway Administration �FFwA� focus- ing on long term pavement performance test site WIM data sug- gested that 1% undercalibration �negative bias� of a WIM scale could produce approximately 3% less ESALs estimation, whereas 1% overcalibration �positive bias� could lead to 4.5% more ESALs estimation �FHwA 1998�. The goal of this study is to investigate how the WIM inaccu- racy modeled by measurement errors will affect the assessment of load-pavement impact. Two basic scenarios in terms of two types of WIMmeasurement errors are addressed: random error and sys- tematic error. In the first scenario, the effect of WIM inaccuracy on load-pavement impact estimation will be studied based exclu- sively on the random error. In the second scenario, both the systematic and random errors are assessed simultaneously. Rec- ommendations are presented for the relationship between varying measurement error levels and load-pavement impact estimation error. Load-pavement impact estimation error is the most impor- tant aspect for pavement design reliability. The second section of this paper provides a background of WIM systems concerning measurement errors and specifications. The third section describes the traffic data set used in this study with focus on axle load distributions. The fourth section develops the methodology to address the effect of measurement errors on load-pavement impact estimation. The last section summarizes the main findings. Background WIM technology has been extensively used for traffic monitoring since the early 1950s. A number of WIM system types are avail- able in terms of the specific sensor technology applied. The fol- lowing are some of the frequently used WIM scales with their individual specific sensors �Middleton and Crawford 2001�: • Bending plates, • Capacitive weight mats, • Capacitive strips, • Hydraulic load cells, • Piezoelectric cable, • Quartz-piezoelectric, and • Strain gauge load cells. The advantages and disadvantages pertaining to cost, mainte- nance, and accuracy vary significantly among these sensors. Among the alternatives, bending plates, piezoelectric and load cell WIM scales are the most typical and widely used in the United States. WIM Measurement Errors Generally speaking, measurement errors can stem from the fol- lowing sources: measurement system, inspector, the inspected ob- jects, or data processing �Humplick 1992�. The focus in this study is specifically on the first source. Three major factors should be taken into consideration regarding the accuracy of a WIM scale: �1� roadway factors, among which pavement smoothness and lon- gitudinal and transverse profiles play the central role; �2� vehicu- lar factors, which involve speed, acceleration, tire condition, load, and body type; and �3� environmental factors, including wind, water, and temperature �Lee 1998�. In other words, WIM mea- surement errors result from the joint effect of all these relevant factors. A series of studies have been carried out to quantify WIM measurement error. In those types of studies, test vehicles are used to obtain their static �or reference� and in-motion weights. A minimum of three repeated weighings are required to obtain the static weight, which is calculated as the arithmetic mean of the three measurements. In addition, a number of in-motion weights are produced by the underlying WIM scale. The detailed process of determining the measurement errors is referred to ASTM E 1318-02 �2002� and Bergan et al. �1997�. Mathemati- cally, the measurement error of a WIM system can be expressed in terms of percentage difference �relative error� �Davies and Sommerville 1987; Izadmehr and Lee 1987b; Bergan et al. 1997� as ��%� = WIM weight − static weight static weight � 100 �1� where WIM weight�weight recorded by WIM on one pass of a given axle load; static weight�axle load weighed by a static scale. The measurement error, �, is comprised of two independent components according to the nature of the error occurrences: ran- dom and systematic. The random error is described as the statis- tical fluctuations of measurement �in either direction� from the truth, and they are intrinsic to the measurement due to the inabil- ity of the device to determine the truth precisely. On the other hand, systematic errors persistently generate the inaccuracies along one direction, which may be due to factors such as faulty design or inadequate calibration. Provided the WIM system is properly calibrated and installed in a sound road structure and subjected to normal traffic and environmental conditions, only the random errors occur. Typi- cally, random errors exhibit a normal distribution with zero mean �Davies and Sommerville 1987; Izadmehr and Lee 1987b; Bergan et al. 1997�. The standard deviation of the underlying normal distribution ���� is a measure indicating the WIM accuracy or precision �Bergan et al. 1997�. In this paper, �� is defined and used as the WIM accuracy indicator. Fig. 1 illustrates the distri- 2 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2007 butions of random error for weighing GVW by three typical types of WIM equipment: single load cell ���=1.5% �, bending strain ���=5% �, and piezo ���=10% �. The systematic error arises from the calibration bias. The cali- bration bias may be due to the initial improper calibration or the WIM system falling out of calibration during service. A depiction of WIM systematic error is presented in Fig. 2. It is shown that the shift of the random error distribution �in this case ��=5% is fixed� leads to the systematic error. In the example of −10% bias, the WIM system is undercalibrated, whereas +10% bias is an example of overcalibration. In the case of ideal calibration, only random error occurs. In addition, it is implied that when calibra- tion bias occurs, both random and systematic errors exist. WIM Specifications Due to the existence of varying WIM accuracy levels, a review of WIM specifications with focus on accuracy is necessary. ASTM has issued the most recent specifications on highway WIM sys- tems in ASTM E1318-02, where four types of WIM systems— Type I, Type II, Type III, and Type IV—are presented to meet the user’s need based on different applications �ASTM 2002�. Types I and II are used in the data collection procedure, while Types III and IV are for the purpose of enforcement. Type IV is still in the conceptual development process. Past research suggested that the accuracy of WIM is not adequate enough to be employed as a legal basis to write a weight-violation ticket, although recent de- velopments may change this situation �Belfield and Souny-Slitine 1999�. Currently, concerning the ASTM specification for WIM functional performance, the tolerance for 95% probability of con- formity in terms of single axle load are ±20, ±30, and ±15% for Types I, II, and III WIM systems, respectively. The corresponding values in terms of the axle group load are ±15, ±20, and ±10%; and ±10, ±15, and ±6% for the GVW. However, the accuracy that these tolerances can provide for pavement design remains in question, which is addressed in this paper together with some recommendations. Data Set Traffic data in this study were collected from the various WIM systems deployed in Texas. Currently, there are 20 WIM stations in Texas. This research is based on traffic load from the available data set from Texas WIM systems. The case study focus on the data from a randomly drawn WIM Station D516, located on In- terstate Highway 35 �I-35� in San Antonio. The traffic sample from this particular WIM station covers the data collection period from January 1998 to March 2002. The raw data were collected and stored in binary form on a daily basis, and converted into ASCII format after extraction by PAT software REPORTER5 �PAT 1997�. These data were stored in a project database for further analysis. Each passing vehicle is represented by a record containing the vehicle’s detailed traffic information. Overall traffic volume can be obtained by querying the amount of records. Traffic classification information can be found directly from the corresponding database field. Axle load distribution can be obtained from the load counts for the indi- vidual load ranges of each axle set on each traffic class. Traffic Classification Each, vehicle is categorized into one of 13 classes according to the axle number and configuration, and Class 4 to Class 13 are defined as trucks �Traffic Monitoring Guide, FHwA 2001�. For this study, the PAT classificationsystem is preferred in order to avoid conversion into FHwA’s 13-class scheme, thus minimizing error sources. The PAT WIM system used in Texas produces a 15-class-based classification scheme, with truck classes ranging from Class 4 to Class 15. In addition, Class 15 includes the ve- hicles not identified by any of the previous traffic classes and records due to system errors. Fig. 3 presents the traffic classifica- tion scheme employed in the current study. Axle Load Distribution Based on axle spacing and configurations, truck axles are divided into five groups: steering axle, single axle with dual wheels �re- ferred to hereafter as single axle�, tandem axle, tridem axle, and quadruple �quads� axle. Tridem axles exist on Class 7 and Class 11 trucks. Quads exist on Class 11 trucks only. With the indi- vidual axle load magnitude obtained from each WIM station, it is easy to obtain the number of loads �frequency� falling in each bin �discrete load weight interval� for each axle type on a given truck class. The counts for each axle type are normalized in terms of percentages �normalized frequencies� to obtain the discrete axle load distribution. Axle load distribution is also referred to as axle load spectrum. After examining all axle load distributions for all truck classes across the 20 WIM stations, it was found that axle load spectra feature multimodal patterns, with the number of pro- nounced peaks ranging from one to two. The axle load spectra for single and tandem axles of Class 10 �18-wheel truck, Class 9 in FHwA’s 13-class scheme�, is illustrated in Figs. 4 and 5. The Fig. 1. Effect of WIM system accuracy �random error� on measurement error distribution Fig. 2. Effect of WIM system biases �systematic error� on measurement error distribution JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2007 / 3 histograms represent the axle load spectra directly from observa- tion counts. The lines in Figs. 4 and 5 represent the fitted distri- bution functions, which are now discussed. The characteristics of multimodal distribution of axle load spectra lead to the adoption of mixed lognormal distribution to fit the data. The motivations to apply mixed-lognormal distribution are as follows �Prozzi and Hong 2005�: 1. The nature of axle load being positive is captured by the fact that a lognormal random variable is nonnegative. Hence, the sum of the linear combination of lognormal distributions as- signed with positive weights is guaranteed to be nonnegative; 2. The individual peaks of a load distribution can easily be captured by lognormal distributions through meaningful pa- rameters; 3. More important, as shown in the following discussion, the load-pavement impact can be described succinctly and effec- tively through the moment statistics of lognormal distribution. Mixed Lognormal Distribution Function Assume a random variable X has a lognormal distribution X � ln��,��, X � 0 �2� The probability density function �pdf� is f�x;�,�� = 1 �2�x� exp�− 1 2 � ln�x� − � � �2� �3� where � and ��parameters for lognormal distribution. Fig. 3. PAT vehicle classification scheme used in Texas Fig. 4. A representative of single axle load spectrum Fig. 5. A representative of tandem axle load spectrum 4 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2007 Thus, the pdf of the mixed lognormal distribution representing a multimodal load spectrum is given by f�x;�k,�k,Wk� = � k=1 K Wk �2�x�k exp�− 1 2 � ln�x� − �k �k �2� �4� where �k and �k�parameters for each lognormal distribution; k� kth piece of lognormal distribution, referred to as mode or peak; Wk�weight of the kth mode, and �k=1 K Wk=1. Parameter Estimation The first step is to determine the number of modes, K, of lognor- mal distributions to fit the individual axle load spectrum. For instance, with regard to load spectrum with pronounced bimodal patterns, two lognormal distributions are the minimum require- ment for capturing the peaks. If the error by fit function is sufficiently small �e.g., R2�99%� with two mixed lognormal dis- tributions, two distributions are adopted. In other cases, it is found that although the two mixed lognormal distributions can capture the peaks, the central segment �between the two peaks� is not well fitted and may result in significant errors. In such cases, a third lognormal distribution is added as the transition between the two pronounced peaks. The reason for imposing a high fit precision requirement is that load-pavement impact is sensitive to the fit error. For parameter estimation, the nonlinear least square �NLLS� technique is applied to fit the discrete distribution. Con- straints are imposed on the weights with Wk�0 and �k=1 K Wk=1. For example, for a fitting function incorporating three lognormal components, its eight parameters can be obtained simultaneously via NLLS. As representatives, the estimated parameters and data fit sta- tistics for truck tandem classes with the four largest sample sizes are shown in Table 1. The four truck classes are 10, 6, 4, and 9 in descending order of sample size. The parameters for the load spectrum based on tandem load data for all truck classes are pre- sented in the last row of Table 1. The cells without data mean the third peak is not necessary. It is shown that the underlying func- tions fit the data well, with all R2 larger than 99%. As examples, the fitted results are illustrated by the solid lines in Figs. 4 and 5. To date, studies on fitting load distribution have focused on data fit exclusively, such as the work by Mohammadi and Shah �1992�, and Timm et al. �2005�. However, the performance fit error in terms of load-pavement impact estimation is of more interest from the perspective of pavement design and rehabilita- tion. As shown in the next section, load-pavement impact can be estimated through the moment statistic of axle load distribution. In this regard, the error is obtained as the relative difference be- tween the fourth moment from fitted lognormal distribution and that from the discrete load distribution �estimation minus obser- vation�. The corresponding results are presented in the last col- umn in Table 1. It is found that the errors are acceptable for pavement design purposes. For instance, the error for all truck tandems by fitted function is 2.19%. In summary, axle load spectra can be described effectively by mixed lognormal distributions. Not only is good data fit obtained but, most important, it is also determined that fitted functions and their parameters have physical meaning. Methodology Impact of Axle Load on Pavement It was established through the analysis of the American Associa- tion of State Highway Officials �AASHO� Road Test results that the impact of each individual axle load on flexible pavement in terms of serviceability loss can be estimated according to the fourth power law �AASHTO l993; Huang 2003�. The fourth power law implies that pavement damage caused by passing ve- hicles increases exponentially with the increase of their axle load. This relationship was captured by the load equivalence factor �LEF� as follows: LEF = � xr Ls �m �5� where xr�weight of axle load in the rth bin; Ls�load weight on a standard axle with the same number of axles as xr, usually 18 kip for the single axle and it is dependent on pavement structure for the tandem axle, approximately 34 kip; the subscript s represents “standard,” and m�power denoting the relative damage to the pavement of a given load xr, typically around 4. As a result, the load-pavement impact based on a given axle load spectrum of truck Class j can be obtained by summing the contributions from all the loads xr’s in the distribution, denoted as load spectra factor �LSF�, LSFj �under the condition of power m=4� LSFj = � r=1 R � xr Ls �4qr,j �6� where R�total number of load bins and qr,j�normalized fre- quency of load in the rth bin of a given load spectrum of truck Class j. It can be seen that LSFj is the fourth sample moment statistic �DeGrootand Schervish 2002� divided by Ls 4. As the continuous distribution function of each axle load spectrum is available, it is more convenient and equally valid to address the axle load- pavement impact by employing the population moment from the pdf. LSF can be thought of as the number of ESALs of one representative axle in the given spectrum. Table 1. Data Fit Parameters for Tandem Axles for Trucks Class Mixed lognormal distribution parameters R2 Performance fit errorW1 W2 W3 �1 �2 �3 �1 �2 �3 4 0.429 0.571 —a 3.192 3.386 — 0.160 0.083 — 0.997 3.12% 6 0.147 0.565 0.288 2.231 2.462 3.404 0.090 0.347 0.173 0.992 1.90% 9 0.269 0.731 — 2.486 2.841 — 0.164 0.350 — 0.998 −4.33% 10 0.424 0.292 0.285 2.733 3.264 3.488 0.325 0.189 0.065 0.995 3.11% ALL 0.433 0.296 0.270 2.714 3.265 3.487 0.335 0.189 0.065 0.995 2.19% aParameters not available because only bimodal distribution is applied such as Classes 4 and 9. JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2007 / 5 The fourth moment of load spectrum function f�x�, M4, is defined as M4 = E�X4� =� x4f�x�dx �7� Integrating Eq. �7� and the axle load spectra functions, as shown in Eq. �4�, i.e., summing the contribution from all of the axle loads according to their distribution, Eq. �6� is equivalent to LSFj =� � xLs� 4 f j�x�dx or LSFj = � x4f j�x�dx Ls 4 = E�X4� Ls 4 = M4 C �8� where x�axle load weight �kip�; f j�x��axle load spectrum func- tion of one axle type on truck Class j; Ls 4�constant, C. In summary, on the basis of the axle load spectrum function, f�x�, as given in Eq. �4�, axle load-pavement impact can be equivalently estimated by applying Eq. �8� in terms of the mo- ment statistic. Thus, what remains to determine the underlying estimation is the fourth moment of load spectrum function f�x�. Moment for the Lognormal Distribution As is shown in Eq. �8�, the fourth moment of axle load spectrum function is the statistic governing the estimation of load-pavement impact. The fourth moment for a random variable X with lognor- mal distribution can be derived as M4 = E�X4� =� x4 1�2�x� exp�− 12� ln�x� − �� � 2�dx = exp�4� + 8�2� �9� where � and � are as in Eq. �3� or �4�. Thus, with load spectrum fitted by K mixed lognormal distri- butions, the load-pavement impact of an axle type on truck class j, LSFj is obtained as LSFj = � k Wk exp�4�k,j + 8�k,j 2 �/Ls 4 �10� where Wk, �k,j, and �k,j, are as in Eq. �4�, for truck class j. It should be noted, however, that while the fourth moment captures the load-pavement impact in terms of serviceability loss or crack- ing damage, other distress types such as rutting may not be cap- tured by this specific moment �Carpenter 1992; Archilla and Madanat 2000�. It has been argued in several studies that the power may range from around 1 to 4, and not necessary integer �e.g., Christison 1986; Pont et al. 2002�. In those cases, lognormal distribution demonstrates its full advantage over other distribu- tions since its moment statistics do exist and can be easily ob- tained as in Eq. �9�. However, other competing distributions are inappropriate in this regard. For example, if normal distribution is adopted as load spectrum function, under the condition of power equal to 3.5, it can be shown that its moment does not exist. The procedure to calculate the LSF under different power values via moment approach is identical to that presented earlier. Load-Pavement Impact Estimation under Measurement Errors Two scenarios are evaluated to study the effect of WIM measure- ment errors on load-pavement impact estimation. First, load- pavement impact estimation is derived under the condition of ideal calibration �with zero calibration bias and involving random error only� of a WIM scale. The axle load spectra with mixed lognormal distributions aforementioned are used as the reference �without measurement errors�. A comparison is made between the estimated load-pavement impact with random measurement error �normal distribution, �� not equal to zero� and that from the ref- erence. The second scenario investigates the load-pavement im- pact estimation under the presence of bias resulting from lack of WIM calibration. In such a case, not only systematic error but also random error is involved as the latter is unavoidable. Scenario I As mentioned previously, under general conditions, the WIM scale measurement errors exhibit a normal distribution. Assuming that for, a given axle load �of certain axle type and specific truck class�, X=x, the observation by a WIM scale with random errors is a random variable, denoted as, X�, then �1 = X� − X X � N�0,�� 2� �11� where �1�axle load relative error under Scenario I; the subscript 1 represents Scenario I; and ���indicator of WIM accuracy. Hence, the variable X� conditional on axle load x also has a normal distribution X� X = x � N�x,x2�� 2� �12� The estimated load-pavement impact by observed load X� con- ditional on X=x, denoted as LSFX� X=x, is LSFX� X=x = � �x��4gX� X�x� x�dx� Ls 4 = E�X� X = x�4 Ls 4 �13� where gX� X�x� x��pdf of load observation X� conditional on X=x, see Eq. �12�; E�X� X=x�4�the fourth moment of x� condi- tional on x; and Ls, see Eq. �10�. It is shown in Eq. �13� that the moments for the normal distri- bution are required for the solution, which can be derived on the basis of moment-generating function �DeGroot and Schervish 2002� of random variable X� �t� = EX�„exp�tx��… =� exp�tx�� 1�2�� exp�− �x� − � 2 2�2 �dx� �14� where and ��parameters of normal distribution of variable X�, as in Eq. �12�. As a result, the fourth population moment becomes Mnormal 4 = E�X�4� = 3�4 + 6�2 2 + 4 �15� where the superscript 4 in Mnormal 4 represents the “fourth power,” the remainders of the superscript represent the power value nu- merically, and “normal” in the subscript indicates the moment corresponds to the normal distribution. The estimated conditional axle load-pavement impact factor LSFX� X=x can be determined by substituting Eq. �15� in Eq. �13� 6 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2007 LSFX� X=x = �3�� 4 + 6�� 2 + 1�x4 Ls 4 �16� As established previously, the variable of axle load X for each individual axle type follows a mixed lognormal distribution �see Eq. �4��. Therefore, the estimated axle load-pavement impact �with random measure errors� based on load spectrum for a given axle type can be obtained by integrating the contribution from its overall axle loads, denoted as LSFj E LSFj E =� � x�4gX� X�x� x�fXj �x�dxdx�/Ls4 = �3�� 4 + 6�� 2 + 1�� k Wk exp�4�k,j + 8�k,j 2 �/Ls 4 �17� where superscript “E” represents “estimated;” gX� X�x� x� is same as in Eq. �13�; fX j �x� is the axle load �without measurement errors� distribution function of certain axle type on a truck class j, see Eq. �4�, and the parameters in the summation, see Eq. �10�. By comparing Eqs. �17� and �10�, it is indicated that under random measurement error, an additive portion of �3�� 4+6�� 2��kWk exp�4�k,j +8�k,j 2 � /Ls 4 is introduced into the load- pavement impact estimation. Additionally, the always-positive value in the additive term implies that the random measurement error results in overestimation of the load-pavement impact. The extent of overestimation depends on the magnitude of the WIM accuracy indicator, ��. Scenario II When a WIM scale is not properly calibrated �biased�, the axle load measurement is subject to systematic error and the measured weight differs from the actual value. In such cases, both system- atic error and random error should be considered. Assuming that, for a given actual axle load �of certain axle type on a truck class�, X=x, the observation by a WIM scale with both errors is also a random variable, denoted as, X�, then Fig. 6. Load-pavement impact estimation error versus WIM accuracy indicator Fig. 7. Three-dimensional �3D� representation of the load-pavement estimation error under both WIM random and systematic errors JOURNAL OF TRANSPORTATION ENGINEERING © ASCE /JANUARY 2007 / 7 �2 = X� − X X � N��,�� 2� �18� where �2 is axle load relative error under Scenario II; subscript 2 represents Scenario II; � is the calibration bias, which is 0 if ideally calibrated; and �� is as in Eq. �11�. The variable X� conditional on axle load X=x has a normal distribution under biased WIM calibration condition X� X = x � N��1 + ��x,x2�� 2� �19� Hence, the axle load-pavement impact by a given axle type on certain truck class under the biased calibration condition is esti- mated using a similar approach as presented in Scenario I, de- noted as LSFj E�b� LSFj E�b� =� � x�4gX� X�x� x�fXj �x�dxdx�/Ls4 = �3��4 + 6��2�1 + ��2 + �1 + ��4�� k Wkexp�4�k,j + 8�k,j 2 � Ls4 �20� where the superscript “E�b�” represents “estimated under calibra- tion bias;” gX� X�x� x� is conditional distribution of observation X, see Eq. �19�; and the other variables, see Eq. �17�. The result in Eq. �20� suggests that, similar to Scenario I, the estimated load-pavement impact under the biased calibration situ- ation is also comprised of two components: �1� a multiplicative term �3�� 4+6�� 2�1+��2+ �1+��4�, including both systematic and random errors, and �2� the load-pavement impact obtained with no measurement errors occurring �as before�. Provided the coef- ficient term is greater than one, it implies an overestimation of load-pavement impact; whereas if the term is less than one, the load-pavement impact is underestimated. Furthermore, the extent of over- or underestimation is determined by both the magnitude of WIM accuracy indicator, ��, and calibration bias, �. To better understand how WIM measurement error affects load-pavement impact estimation, a numerical illustration is pre- sented based on the findings just discussed. With respect to WIM measurement under only random error, a range of the WIM accu- racy indicator, ��, from 0 to 20% is evaluated to address how these random errors affect the estimation of pavement perfor- mance. Fig. 6 shows the relationship between varying WIM ac- curacy indicator �� �representing random errors� and relative er- rors of load-pavement impact estimation. It can be seen that the load-pavement impact estimation error is lower than the WIM accuracy indicator, ��, for �� below 15%. Above this, the load- pavement impact estimation errors increase. When calibration bias occurs, both random and systematic er- rors should be addressed. Load-pavement impact estimation error shows a significant variation. Figs. 7 �in three dimensions �3D�� and 8 �in two dimensions �2D�� illustrate the effect of a series of combinations of WIM accuracy indicator �� and calibration bias � on load-pavement impact estimation. It is shown that both over- estimation and underestimation occur. It is also indicated that load-pavement impact estimation error is more sensitive to the calibration bias than to WIM accuracy indicator ��, since the estimation error covers a larger range along systematic errors than that of random errors. Next, the sensitivity of load-pavement impact estimation to calibration bias is examined. A typical WIM scale with accuracy indicator ��=8% is used to this purpose as shown in Fig. 9. It was found that 10% overcalibration resulted in 51% overestimation of load-pavement impact, which is even more significant than previ- ously reported in FHWA-RD-98-04 �1998�; whereas the 10% un- dercalibration produced results similar to those in the 1998 report, approximately 31% underestimation of load-pavement impact. Conclusions This research study analyzes the effects of WIM measurement errors on load-pavement impact estimation. Two types of errors, random error �represented by the accuracy indicator ��� and sys- tematic error �represented by the calibration bias �� were identi- Fig. 8. Two-dimensional �2D� representation of the load-pavement estimation error under both WIM random and systematic errors 8 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2007 fied and evaluated for WIM systems. Axle load spectra for differ- ent truck classes were established to address the traffic load characteristics effectively. A methodology based on moment sta- tistics was developed to represent load-pavement impact under different WIM measurement error conditions. Traffic data from WIM systems in Texas were used to demonstrate the varying load-pavement impact evaluation biases. To realize the objective of this study, axle load spectra func- tions were established at first with considerations, not only fo- cused on statistical data fit but also on load-pavement impact implications. These functions are mixed-lognormal distributions. It was found that axle load-pavement impact given a load spec- trum could be expressed in terms of the moment statistic of log- normal distributions. WIM random measurement error was first introduced in Sce- nario I to analyze its effect on load-pavement impact estimation. Two moments jointly addressing the random error variable �normal distribution� and axle load variable �mixed lognormal distribution� were integrated to represent the estimation of load- pavement impact. It was found that WIM measurement random error always leads to overestimation of load-pavement impact. In Scenario II, both WIM measurement random error and sys- tematic error were included �representing biased calibration of a WIM scale�. Calibration bias was accounted for by shifting the normal distribution of the random error. It was found that com- pared with the case of random error only, WIM systematic error contributed more significantly to the error of load-pavement im- pact estimation. For a WIM scale with a typical accuracy indica- tor, 10% overcalibration results in as much as 51% overestimation of load-pavement impact, while it is 31% underestimated by 10% undercalibration. The results suggest that load-pavement impact estimation error is more sensitive to overcalibration than under- calibration of a WIM scale. More important, the findings suggest that WIM calibration is worthy of more attention for collecting traffic data for pavement design and rehabilitation. The results of this study can be applied to evaluate over- or underdesign of pavement in terms of traffic data provided by WIM. In addition, the results could be used to aid with selection of the appropriate WIM system in terms of accuracy and potential biases from the perspective of pavement design and rehabilitation. Thus, it could also be used to select the WIM equipment accord- ing to the requirement of pavement design reliability. Acknowledgments The writers wish to thank the Texas Dept. of Transportation �TxDOT� for providing the funds and support for this research through TxDOT Project No. 0-4510. Joe Leidy and Rich Rogers from the Construction and Materials Division are particularly ac- knowledged as well as Dr. German Claros from TxDOT Research and Implementation Office. References American Association of State Highway and Transportation Officials �AASHTO�. �1993�. AASHTO guide for design of pavement struc- tures, Washington, D.C. American Society for Testing and Materials �ASTM�. �2002�. “Standard specification for highway weigh-in-motion �WIM� systems with user requirements and test method.” ASTM E 1318-02, Philadelphia. Archilla, A. R., and Madanat, S. �2000�. “Development of a pavement rutting model from experimental data.” J. Transp. Eng., 126�4�, 291– 299. Belfield, K. M., and Souny-Slitine, N. �1999�. “Truck weight limit en- forcement technology applicable to NAFTA traffic along the Texas- Mexico border.” Research Rep. No. 167209-1. Center for Transporta- tion Research, The Univ. of Texas at Austin, Austin, Tex. Bergan, A. T., Berthelot, C. F., and Taylor, B. �1997�. “Effect of weigh- in-motion accuracy on weight enforcement efficiency.” International Road Dynamics, Technical Paper, �http://www.irdinc.com/english/ html/tech-ppr/index.htm� �June 1, 2005�. Carpenter, S. H. �1992�. “Load equivalency factors and rutting rates for the AASHO road test.” TransportationResearch Record. 1354, Trans- portation Research Board, Washington, D.C. Christison, J. T. �1986�. Pavement responses to heavy vehicle test program—Part 2: Load equivalency factors, vehicle weights and di- mensions study, Vol. 9, Roads and Transportation Association of Canada, Ottawa. Davies, P., and Sommerville, F. �1987�. “Calibration and accuracy testing of weigh-in-motion systems.” Transportation Research Record. 1123, Transportation Research Board, Washington, D.C., 122–126. DeGroot, M. H., and Schervish, M. J. �2002�. Probability and statistics, 3rd Ed., Addison-Wesley, Boston. Federal Highway Administration �FHwA�. �1998�. “WIM scale calibra- tion: A vital activity for LTPP sites.” Techbrief No. FHWA-RD-98- 104, U.S. Dept. of Transportation, Washington, D.C. Federal Highway Administration �FHwA�. �2001�. Traffic monitoring guide, U.S. Dept. of Transportation, Washington, D.C. Huang, Y. H. �2003�. Pavement analysis and design, Prentice Hall, Engle- wood Cliffs, N.J. Humplick, F. �1992�. “Highway pavement distress evaluation: Modeling measurement error.” Transp. Res., Part B: Methodol., 26B�2�, 135– 154. Izadmehr, B., and Lee, C. E. �1987a�. “On-site calibration of weigh-in- motion systems.” Transportation Research Record. 1123, Transporta- tion Research Board, Washington, D.C., 136–144. Izadmehr, B., and Lee, C. E. �1987b�. “Accuracy and tolerances of weigh-in-motion systems.” Transportation Research Record. 1123, Transportation Research Board, Washington, D.C., 127–l35. Lee, C. E. �1998�. “Factors that affect the accuracy of WIM systems.” Proc., 3rd National Conference on Weigh-in-Motion, St. Paul, Minn. Middleton, D., and Crawford, J. A. �2001�. “Evaluation of TxDOT’s traffic data collection and load forecasting process.” FHWA/ TX-01/1801-1, Texas Dept. of Transportation, Austin, Tex. Mohammadi, J., and Shah, N. �1992�. “Statistical evaluation of truck overloads.” J. Transp. Eng., 118�5�, 651–665. National Cooperative Highway Research Program. �2005�. “Mechanistic- empirical design of new and rehabilitated pavement structures.” Fig. 9. Sensitivity of performance estimation on calibration bias ���=8% � JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2007 / 9 Research Project No. 1-37A, Washington, D.C., http://www.trb.org/mepdg/� �Aug. 8, 2005�. PAT Traffic Control Corporation, Inc. �1997�. REPORTER, data evalua- tion software (program description and user’s manual), Chambers- burg, PA. Pont, D. J., Steven, B., and Alabaster, D. �2002�. “The effect of mass limit changes on thin-surface pavement performance.” Proc., 7th Int. Symp. on Heavy Vehicle Weights and Dimensions, Delft, The Nether- lands. Prozzi, J. A., and Hong, F. �2005�. “Optimum statistical characterization of weigh-in-motion data based on pavement impact.” Proc., 85th An- nual Meeting of the Transportation Research Board, Washington, D.C. �CD-Rom�. Timm, D., Tisdale, S. M., and Turochy, R. E. �2005�. “Axle load spectra characterization by mixed distribution modeling.” J. Transp. Eng. 131�2�, 83–88. 10 / JOURNAL OF TRANSPORTATION ENGINEERING © ASCE / JANUARY 2007