f--:, ,-, t 8. Performing Orgoni uti on Report No.

Transcription

1 T e:hnl:al Report Doumentation Page 1. Report No. 2. Government Ae uion No. 3. Ruiplent'a Catalo; No. FHWA/TX-~ , Tille t1nd SublitJe 5. Report Date The Use of Lasers for Pavement Crak Detetion i-;---:::--:--:---=--d_e e_m_b-:=-e-:"r 1 _9_ 8 _ 8 _--1 6, PerfC?rmlng Orgoni&ollon Code f--:, ,-, t 8. Performing Orgoni uti on Report No. 7. Aurhorf a) Lynda Donnell Payne, Roger s. Walker 9, Performing Organization Nome and Addrua The University of Texas at Arlington Arlington, Texas ~ ~ ~ 12. Sponaorlng Agen:y Nama and Addrua Texas State Department of Highways and Publi Transportation, D-10 Researh P.O. Box 5051, Austin, Texas Supplamantarv Natu 10. Work Unit ~o. (TRAIS) II. Conrro:t o<. GrGn.t NQ. Researh Study# Type of Report ond Period Covered Interim 14. Sponaoring Ageny Code Study done in ooperation with US Dept. of Transportation, Federal Highway Administration 16. Abatrat This researh was initiated to investigate the apability of using lasers for rak detetion in pavements. If suh a apability ould be developed, it would be used to aid in obtaining and evaluating pavement distress and raking information for the State's P.E.S. proedures, used for maintaining and evaluating pavements. The researh effort has involved three stages. The first two stages were to determine the rak detetion apabilities of the laser probes, used on the Surfae Dynamis Profilometer (DDP). The SDP is owned by the State and used for road profile measurements. After experiments indiated that these probes ould be used for suh detetion, a system was developed to further study this apability and to determine how it ould be used to implement an automated high speed rak identifiation system. The third stage is the implementation of suh a system so its usefulness for P.E.S. data olletion ativities an be determined. This researh report desribes the first two phases of the researh effort. 17, Kay Word a Surfae Dynamis Profilometer (SDP) Lasers, Pavement Distress Measurements, Pavement Crak Identifiation and Reording. 18. Oi atribution Stalamant No restritions. This doument is available to the publi through the National Tehnial Information Servie, Springfield, Virginia Seurity Clouil. (at thia report! 20, Seurity Clauif. (olthla poa l 21. No, of Pagn 22. Prie Unlassified Unlassified 78 Form OOT F ' <8-m Reprodution of ornpletad pave authorized

2 THE USE OF LASERS FOR PAVEMENT CRACK DETECTION by Lynda Donnell Payne Roger S. Walker The University of Texas at Arlington Researh Report Crak Identifiation Using Lasers Researh Projet onduted for Texas State Department of Highways and Publi Transportation in ooperation with the U.S. Department of Transportation Federal Highway Administration Deember 1988

3 The ontents of this report reflet the views of the authors, who are responsible for the fats and auray of the data presented herein. The ontents do not neessarily reflet the offiial views or poliies of the Federal Highway Administration. This report does not onstitute a standard, speifiation, or regulation. There was no invention or disovery oneived or first atually redued in the ourse of or under this ontrat, inluding any art, method, proess, mahine, manufature, design or omposition of matter, or any new and useful improvement thereof, or any variety of plant whih is or may be patentable under the patent laws of the United States of Ameria or any foreign ountry. ii

4 PREFACE This projet report presents interim results from Projet The Projet was initiated to determine the feasibility of using lasers for developing an automated pavement rak detetion and identifiation system. This report provides results of the first two phases of the researh effort. Speial reognition is due Mr. Robert Harris of D-18, for his support in initiating the projet and his many ontributions to this researh efforts. Lynda Donnell Payne Roger S. Walker Deember i

5 ABSTRACT This researh was initiated to investigate the apability of using lasers for rak detetion in pavements. If suh a apability ould be developed it would be used to aid in obtaining and evaluating pavement distress and raking information for the State's P.E.S. proedures, used for maintaining and evaluating pavements. The researh effort has involved three stages. The first two stages were to determine the rak detetion apabilities of the laser probes, used on the Surfae Dynamis Profilometer (SOP). The SOP is owned by the State and used for road profile measurements. After experiments indiated that these probes ould be used for suh detetion, a system was developed to further study this apability and to determine how it ould be used to implement an automated high speed rak identifiation system. The third stage is the implementation of suh a system so it's usefulness for P.E.S. data olletion ativities an be determined. This researh report desribes the first two phases of the researh effort. KEY WORDS: Surfae Dynamis Profilometer (SOP), Lasers, Pavement Distress Measurements, Pavement Crak Identifiation and Reording. iv

6 SUMMARY This projet was initiated to determine the feasibility of using the laser probes on the Surfae Dynamis Profilometer (SOP) owned by the the State Department of Highways and Publi Transportation (SDHPT), for rak detetion and identifiation. If found feasible a system was then to be developed for use on the ARAN measurement vehile, also owned by the State so it ould be used to aid in pavement distress measurements. The SOP was seleted for the initial testing and evaluation as it had existing on-board laser equipment. The initial investigations proved that the lasers on the SOP ould be used for rak detetion. Based on this result, the study has proeeded in obtaining the neessary equipment and developing algorithms and software for implementing an automated rak measuring system whih hopefully ould be used to aid in PES. This report disusses the first two phases of this projet, determining the feasibility of rak detetion using the laser, and obtaining and testing equipment so suh a system ould be implemented. During these first two phases the apabilities and limitations have been identified. To date, it appears that a system an be developed with a limited apability for rak identifiation and reporting whih ould be useful for PES data olletion ativities. The third phase of development and implementation of the automated rak identifiation system is urrently in progress. v

7 IMPLEMENTATION STATEMENT An automated and objetive proedure for rak measurements and reording would provide a signifiant to the State during P.E.S. data olletion proedures. ould be used in many other areas where statistial information regarding pavement raking is desired. savings It vi

8 TABLE OF CONTENTS PREFACE ABSTRACT SUMMARY IMPLEMENTATION STATEMENT LIST OF FIGURES iii iv v vi ix CHAPTER I. II. INTRODUCTION 1.1 Projet and Report Sope 1. 2 Bakground 1.3 Projet Phases 1.4 Distress Types 1.5 Projet Requirements FEASIBILITY Sampling and Update Rates. Resolution, Noise, and Texture. ould Craks Be Deteted? III. 2.4 The Real-time Issue Laser Problems and Limitations CRACK IDENTIFICATION HARDWARE Optoator DAQ Board COMPAQ Portable III. 24 vii

9 IV. TIME AND FREQUENCY ANALYSIS TECHNIQUES Time Series Stohasti Proess Ergodiity and Stationarity Statistial Estimates Power Spetrum Estimation Linear Parametri Modeling 33 v. ANALYSIS OF PAVEMENT CRACKING DATA Introdution Variane Method for Real-Time Crak Counting Spetral Analysis Results Running Mean/Slope Threshold Method Autoorrelation Differene Method AR Proess Modeling Results 46 VI. CONCLUSIONS AND FURTHER RESEARCH. 49 APPENDIX A. DAQ BOARD SCHEMATICS.. 52 B. RUNNING MEAN/SLOPE THRESHOLD LISTING. 59 REFERENCES viii

10 LIST OF FIGURES 1.1 Rutting Pathing Failure Alligator raking Blok raking 1.6 Transverse raking. 1.7 Longitudinal raking 2.1 Laser probe and laser alibration board. 2.2 Laser orientations Calibration board results Laser probe and probe proessing unit. Pulsed, modulated infrared light from GaAs lasers Triangulation priniple.. Laser measurement range. CPU sub-rak with power supply and reeiver-averaging boards Data aquisition (DAQ) boards. 3.7 Crak system in the profilometer 5.1 Power spetral density plots of different raking types Power spetral density plots of different raking severity ix

11 5.3 Running mean/slope threshold tehnique applied to moderate alligator raking data Running meanjslope threshold tehnique applied to severe alligator raking data Raw data Filtered data with r(o)-r(4) value plotted every 16 data points Atual r(o) and r(4) values for the data in Figure

12 CHAPTER I INTRODUCTION 1.1 Projet and Report Sope This projet was initiated to determine the feasibility of using the laser probes on the Surfae Dynamis Profilometer (SDP) owned by the the State Department of Highways and Publi Transportation (SDHPT), for rak detetion and identifiation. If found feasible a system was then to be developed for use on the ARAN measurement vehile, also owned by the state. The SDP was seleted for the initial testing and evaluation as it had existing on-board laser equipment. As will be disussed initial evaluations proved that the lasers on the SDP ould be used for rak detetion. Based on this result, the study has proeeded in obtaining the neessary equipment and developing algorithms and software for implementing an automated rak measuring system for PES. This report disusses the first two phases of this projet, determining the feasibility of rak detetion using the laser, and obtaining and testing equipment so suh a system ould be implemented. The third phase of development and implementation of the automated rak identifiation system is urrently in progress and will be reported on in a later report. This introdutory hapter will first provide a bakground and general understanding of the rak detetion and identifiation problem. Further, it explains some neessary terms and desribes the projet requirements. Chapter two, then addresses the feasibility issue. It desribes the work done to determine if pavement raking ould be deteted with the lasers. The third hapter desribes the hardware designed and built for initial evaluation of a rak detetion system. Chapter four defines and explains the statistial and signal proessing theory used in the rak identifiation algorithms. Chapter five desribes the different rak identifiation algorithms employed. This hapter also desribes results of the data analysis on the test setions used in the study. 1

13 2 Chapter six desribes additional researh, muh of whih is being onduted in the third phase. 1.2 Bakground The evaluation of pavement surfae onditions of the nation's highways is of major interest to transportation engineers. The State has been using suh information in onjuntion with other data in an established proedure for determining the ondition of the State's highway system. This information is essential in determining whih roads should be worked on, and how muh money is needed to omplete the work. The State urrently evaluates pavement surfae onditions by onsidering both road roughness and pavement distress. A measure of road roughness is readily obtained with existing instruments. Pavement distress information is more diffiult to obtain as it requires visual evaluation. Currently, SDHPT personnel attend the annual Pavement Evaluation System (PES) Rater Training Shool and then disperse to their respetive distrits throughout the state to rate pavement surfaes by "walking" the roads. Obviously, this proess is very tedious and time-onsuming. Also, sine so many people are involved in the evaluation, the ratings are often not repeatable. An automated measurement system is needed to simplify the proess and to obtain more onsistent measurements. This researh represents the first attempt by the SDHPT to automate the proess. The researh was made possible when laser probes were purhased for use on the Surfae Dynamis Profilometer (SOP). The SOP is used by the Department to obtain road profile measurements. Two lasers, one in eah wheel path, are used to measure distanes from the bottom of a survey vehile to the road surfae. These distane measurements, along with vertial aeleration measurements from two aelerometers, are used to obtain the road profile by removing the effets of the vehile suspension system [1,2,3]. The laser system disussed in this study is built by Seletive Eletroni Co. (Selom) of Sweden. The devie is alled an optoator. The system's basi omponents are the laser probes and probe proessing units, whih are mounted under the van, and the CPU sub-rak ontaining the power supply and reeiver-averaging boards, whih are installed inside the van. The optoator measures distanes to a surfae using laser probes. Eah probe emits a small infrared light beam that strikes the surfae to be measured. The refleted light is foused onto a position-sensitive detetor in the laser probe allowing aurate distane measurements [5]. Further

14 explanation of the optoator and measuring priniple wi!l be provided in Chapter III. Sine the lasers were available, highway department engineers wanted to know if these lasers ould help identify pavement raking. The intent of this study was first to determine the feasibility of using the existing lasers of the SDP to identify pavement raking. Then, if feasible, the work would be extended to design and implement a system whih would identify the speified raking patterns. Few operational systems for rak identifiation using aser probes have been reported in the literature. Most studies for suh systems have used video data [6,7]. The researh herein does not u~e an elaborate video amera system, only existing lasers. 1.3 Projet Phases As noted above, this study onsisted of three phases and this report is onerned with the first two phases. First, the feasibility of using the existing laser probes on the SDP had to be investigated. This involved determining whether or not the resolution of the laser probes was suffiient to detet raking patterns. Also, the measurement update ra~e had to be onsidered to determine if the laser ould supply the neessary sampling rate for rak identifiation at highway speeds. Another item of interest was the real-time issue. That is, how muh, if any, of the proessing and analysis ould be performed in real-time with the van moving at highway speeds? If real-time omputation was not feasible, what proedures ould be developed to ollet data for later proessing? Phase two was to begin one it was determined that rak3 ould be deteted using the laser probes. This phase would involve designing, testing and implementing both the hardware and software for a system whih ould be used for rak detetion and identifiation. Although it has been stated that phase one was first investigated followed by phase two, this was not exatly the ase. Obviously, some of the issues in phase one ould only be addressed if there existed hardware and software to obtain the raking data. In atuality, the phases overlapped and some of the hardware and software developed will be hanged later based on results obtained. By the same argument the suess of suh systems an only be determined by atual implementation.

15 1.4 Distress Types This researh is only onerned with distress types in asphalt surfaed pavements sine this type of road surfae repres~nts the largest perentage of the highway system in Texas. 4 The distress types whih are urrently reorded by PES raters on asphalt pavements are rutting, pathing, failures, alligator raking, blok raking, transverse raking, and longitudinal raking [8). Eah type will be desribed here for ompleteness; however, not all types are onsidered in this researh. A rut is a surfae depression in the wheel paths. Rutting stems from a permanent deformation in any of the pavement layers or subgrade. It is usually aused by onsolidation or lateral movement of the materials due to traffi loads. Refer to Figure 1.1. Pathes, shown in Figure 1.2, are repairs made to pavement distress. The presene of pathing indiates prior maintenane ativity, and is thus used as a general measure of maintenane ost. A failure is a loalized setion of pavement where the surfae has been severely eroded, badly raked, or depressed. Failures are important beause they identify speifi strutural defiienies whih may pose safety hazards. see Figure 1.3. Alligator raking is a series of interonneting raks aused by fatigue failure of the asphalt surfae under repeated traffi loading. The raking initiates at the bottom of the asphalt surfae where tensile stress and strain is highest under a wheel load. The raks propagate to the surfae initially as one or more longitudinal parallel raks. After repeated traffi loading the raks onnet, forming polygon-shaped, sharp-angled piees that develop a pattern resembling hiken wire or the skin of an alligator. The piees are usually less than 1 foot on the longest side. Alligator raking ours only in areas that are subjeted to repeated traffi loading. Refer to Figure 1.4. Blok raking divides the asphalt surfae into approximately retangular piees. The bloks range in size from approximately 1 foot square to 100 feet square. See Figure 1.5. Craking into larger bloks are generally rated as longitudinal and transverse raking. Blok raking is aused mainly by shrinkage of the asphalt onrete and daily temperature yling. It is not load assoiated, although load an inrease the severity of individual raks. This type of

16 5 Figure 1.1 Rutting Figure 1.2 Pathing

17 6 Figure 1.3 Failure Figure 1. 4 Alligator raking

18 7 distress differs from alligator raking in that alligator raks form smaller, many-sided piees with sharp angles. Also unlike blok raks, alligator raks are aused by repeated traffi loadings. Transverse raking, seen in Figure 1.6, onsists of raks or breaks whih travel at right angles to the pavement enterline. Transverse raks are usually aused by differential movement beneath the pavement surfae. They may also be aused by surfae shrinkage due to extreme temperature variations. Although transverse raks may our at any spaing, they will be only onsidered suh for this researh if they our at distanes greater than 10 feet apart. More losely spaed raks are ounted as either alligator or blok. PES data and SDHPT experiene suggests that this assumption will ause only a minor error in statewide PES setions. LongitL:inal raks are parallel to the pavement's enterline r laydown diretion. They may be aused by a poorly onstruted paving lane joint, shrinkage of the surfae due to low temperatures or hardening of the asphalt, or a problem with the subgrade. Refer to Figure 1.7 ( Note the figure also has blok raking). ThJsJ:esearh effort onsidered only three of the seven.ej_stt:ess. types desribed above. Speifially, alligator, - blok, and transverse raking were to be onsidered. Some of the other distress types, partiularly failures and longitu~inal raking, ould ause the raking pattern to be mislassified due to the nature of the sensors used and the method of observation. This should beome lear from later disussions. 1.5 Projet Requirements As previously desribed, this study involved using the existing lasers to identify raking patterns. One laser was to be mounted in eah wheel path, and one in the middle. Obviously, little, or no information aross the lane ould be reorded to help in the identifiation. The laser data was to be reorded and analyzed in real-time at highway speeds if possible. The type, severity and perent area of raking was to be determined from the laser data obtained. Type refers to one of the three types previously mentioned (alligator, blok, or transverse). Severity is determined by the width of the rak. Slight raks are less than 1/8 inh, moderate are 1/8 to 1/4 inh and severe are greater than 1/4 inh wide. Also, the perent of the setion with eah type of rak was to be noted. In the ase of transverse raks, a ount of the

19 -- 8 ',. '-:. +v.,;.f'.j.'_...,:_ ~,~'--.. ;1t'~: r ::, i r Figure 1.5 Blok raking Figure 1.6 Transverse raking

20 9 Figure 1.7 Longitudinal Craking.

21 10 number of raks deteted in a setion length was to be reported. Finally, if the omplete data analysis and reporting ould not be performed in real-time, then at least a reasonable (1 mile) length of data should be reorded in realtime. It ould later be downloaded and further analysis and reporting performed. 10

22 CHAPTER II FEASIBILITY 2.1 Sampling and Update Rates The first question to be addressed in phase one was whether or not the lasers ould provide measurements at a suffiiently fast rate. That is, did the laser update rate meet or exeed the neessary sampling rate? Sine the smallest raks to be deteted were in the 1/8 inh wide range, it was reasonable that a 1/16 inh sampling rate would be required. The update rate of the Selom laser system is fixed with jumpers on the reeiver-averaging board in the CPU sub-rak. This is disussed in Chapter III. However, the maximum update rate (no averaging) is 32,000 samples per seond [4,5]. The neessary sampling rate for 1/16 inh sampling varies from 2816 samples per seond at 10 miles per hour to samples per seond at 50 miles per hour. A omparison of the update rate to the maximum required sampling rate shows that the Selom lasers are able to supply measurements at the neessary speed. Also, sine the update rate is more than twie the required sampling rate it is suggested that the reeiver-averaging boards be jumpered for two point averaging. This will provide a 16K update rate, still exeeding the sampling rate required, and at the same time reduing the noise in the measurements. 2.2 Resolution, Noise, and Texture The laser measurement range, as explained in Chapter III, is inhes. The analog signal from the laser probes varies from 0 to 10 volts. A 12-bit A/D onverter in the probe proessing unit (PPU) onverts the analog signal into a 12-bit digital representation, providing a 2.44 mv or inh resolution. Noise is a major onsideration in determining measurement auray and the ability to detet raking. That is, how muh variability in measurement readings would be expeted if the laser was refleting off a surfae at a onstant distane? 11

23 12 To determine this the range and variane of two data sets was onsidered. In the first, the lasers were benh mounted in the lab and data was olleted with the laser beam refleting off a flat stationary objet. Results from this proedure showed a range of to mv _from the mean with a standard deviation of 7.9 mv. A seond set of data was olleted in the profilometer with the motor running and the van at rest. Here the range was to mv from the mean and a standard deviation of 16.8 mv was observed. These observations were needed to provide insight into reasonable threshold values used in several of the rak detetion algorithms. The texture of a road surfae is another item whih adds variability. In fat, it should be understood that road surfaes of very ourse texture probably do not allow reasonable rak detetion by the methods desribed in this study. 2.3 Could Craks Be Deteted? Phase one of this projet involved determining whether or not the Selom lasers on the profilometer ould detet raks in a road surfae. Two approahes were taken to answer this question. First, short setions of pavement with the desired raking were loated. The setions were marked as to start, end, and the desired path for the driver to take. Laser data was then obtained from the setions with the driver being very areful to follow the marked path. This data was plotted and ompared with slides taken of the marked setion. Results of this omparison were very enouraging. Most of the moderate and severe raks seen in the slides ould easily be reognized in the plots. The first proedure.of driving over a marked setion gave a good idea but it was never known exatly where the laser beam fell. That is, a rak perpendiular to the enterline may be 1/4 inh wide at one point while 1/2 inh over it might be 1/16 of an inh wide. For this reason, that proedure did not give muh insight into how well the lasers would be able to provide severity information. Therefore, a surfae with raks of known width and depth was needed for testing. To provide this known surfae the laser alibration board was built. The laser alibration board, though simple in onept and onstrution, provided valuable information. This board was simply a irular piee of blak plywood suspended from a variable speed motor. Craks of different widths and depths were ut into the board surfae. The board was ut with a

24 desired irumferene so it ould easily simulate a road surfae passing under the laser probes at speeds from 1 to 30 miles per hour by varying the rotational speed. Three different sets of raks were ut into the board. Craks within eah set were the same depth. That is, one set of raks was 1/8 inh deep, one set was 1/4 inh, and the third set was 3/8 inh in depth. Five raks of varying width were ut in eah set. They were 1 inh, 1/2 inh, 1/4 inh, 1/8 inh, and 1/16 inh. Figure 2.1 shows the benh mounted laser probe, PPU, and the laser alibration board. One important observation whih ame to light while working with the alibration board was that orientation signifiantly affeted measurements. As will be disussed in Chapter III, the laser beam whih strikes the measured surfae is longer in one diretion than the other. It was found that the ability to detet slight raking was signifiantly improved by having the laser beam fall aross a rak perpendiular to the enterline instead of into the rak. That is, orientation 2 in Figure 2.2 gave muh better results. Also, orientation 1 gave invalid data readings on the bak side of pratially every rak. Invalid data is typially aused by an insuffiient amount of laser light falling on the detetor. Orientation 2 showed no invalid data. This observation an be explained by the fat that the entire beam fell into the rak in orientation 1 and the path of the refleted light bak to the detetor was obstruted by the rak wall as the beam neared the bak side of the rak. Figure 2.3 provides a plot of laser measurements obtained from the alibration board at 15 miles per hour using orientation 2. It an be seen that the 1.inh down to the 1/8 inh raks are easily reognized. However, the 1/16 inh rak is not as easily deteted. In fat, its true depth is not refleted in the plot. The reason is that the distane value represents the average distane measurement of all the area overed by the laser spot. Sine the beam does not ompletely fall into the rak, the true depth of slight raking annot be aurately measured. This will ause a problem beause slight raking an easily be lost in the variability seen in noise and texture. 2.4 The Real-time Issue The ability to detet and provide detailed analysis of pavement raking at highway speeds up to 50 miles per hour annot be performed by the hardware built in this initial study. Real-time analysis at speeds of 50 miles per hour with 1/16 inh sampling requires a proessing time less than 71 miroseonds.

25 14 Figure 2.1 Laser probe and laser alibration board -

26 15 Orientation # 1 Orientation # 2 Diretion of travel Dire1ion of travel Crak Crak - I Laser Spot Laser Spot Figure 2.2 Laser orientations

27 DATA SAWPLES Figure 2.3 Calibration board results Two revolutions of the alibration board are represented in the plot above. Note 3 sets of raks with 5 raks eah are inluded in eah revolution. Details of depth and width are desribed on page 17.

28 The system desribed in this study has the ability to give an approximate rak ount in real-time or to ollet a setion of data in real-time whih will later be downloaded, analyzed, and reported off-line. The real-time rak ount feature is based on a variane alulation of one inh inrements of data. These alulations an be performed in approximately 40 miroseonds. It should be emphasized that this is only an estimate of raking and is very sensitive t variane threshold values supplied by the operator. Chapter VI will address the real-time issue again in a disussion of upgrades and further researh. 2.5 Laser Problems and Limitations Initial work in determining the sensitivity of the lasers to pavement raking used the Selom lasers installed in the profilometer. Based on results obtained from the alibration board experiments, a deision was made to obtain new lasers whih had a redued spot size. The laser probes with the larger spot size ould not detet 1/16 inh raking and even did a poor job of deteting 1/8 inh raking. As expeted, the new lasers did a muh better job of deteting less severe raks. Unfortunately, with the new laser system ame many problems and delays. The new lasers showed an abnormally high sensitivity to sunlight. In fat, results were so bad that the laser probes and probe proessing units had to be sent bak for modifiation. Following the modifiations the probes were again benh tested both in the lab and outside in sunlight. Results obtained indoors or in a shaded area were aeptable; however, one again, when exposed to sunlight an abnormally high perentage of invalid data measurements were obtained. Selom tehniians were again onsulted. This time Selom suggested hanging the F-stop in the detetor's lens system. To determine the best F-stop to use, data was olleted from the laser alibration board in diret sunlight. Changing the F-stop from its preset 1.4 position to 4.0 seemed to eliminate the invalid data problem. The lasers were then field tested with mixed results. Suffiient data was olleted to ontinue the study. Meanwhile, the laser probes and probe proessing units were one again shipped bak to Selom for further modifiation and alibration. It should be noted that Selom engineers have sine suggested not to hange the F-stop more than two settings. They now reommend a setting of 2.8. ''7 1.

29 CHAPTER III CRACK IDENTIFICATION SYSTEM HARDWARE The three basi hardware omponents of the initial onfiguration for the rak identifiation system are the optoator, the DAQ board and the COMPAQ Portable III personal omputer. The optoator obtains a distane measurement using non-ontat lasers. The data aquisition board aquires the data from the optoator at a speified sampling rate, temporarily stores the data in onboard RAM and performs some preliminary proessing of the data as well as data redution. Finally, the COMPAQ aepts a redued data set and stores it for final proessing and analysis. 3.1 Optoator The optoator is an optoeletroni measurement system whih measures the distane to an objet with high speed and pre1s1on. Most importantly, the measurement is made without ontating the measured surfae. The basi omponents of the optoator are the non-ontat laser probes, the probe proessing units (PPU), and the CPU sub-rak whih ontains the power supply and the reeiver-averaging boards whih reeive and proess data from the gauge probes. A laser probe and probe proessing unit are shown in Figure 3.1. The gauge probe ontains a pulsed, modulated (32KHz) and intensity-ontrolled laser diode, a position sensitive photodetetor and an appropriate lens system. The laser diode is a lass III b gallium-arsenide (GaAs) laser whih entails the risk of eye damage if the beam hits the eye diretly [4). The GaAs laser in the gauge probe gives off pulsed, modulated invisible infrared light as shown in Figure 3.2. Eah pulse in the 16 pulse burst is 350 ns. The bursts our at a frequeny of 32 KHz whih aounts for the 32 KHz data rate of the serial data passed to the reeiver-averaging board. The light from the laser beam passes through a lens whih fouses the light in the enter of the measurement range. The spot size whih strikes the ground surfae is approximately 1/4 inh by 1/16 inh. 18

30 19 Figure 3.1 Laser probe and probe proessing unit..

31 The optoator measures the distane to an objet by use of the triangulation priniple, as illustrated in Figure 3.3. From a light soure, L, a onentrated light beam is direted onto the surfae of the measured objet, 01. The light beam will strike the surfae at point A and the sattered light refletion is foused through a lens to a point A' on a position sensitive detetor. If the distane of the measured objet is hanged by X, the laser beam will hit point B on surfae 02 and be foused at point B' on the detetor. Sine the relative position of the light soure, the lens and the detetor are fixed, the relation between X and X' is known and distane measurements an be obtained. The maximum measurement range, 01-02, as well as the standoff distane must be onsidered when mounting the laser probes. Selom's gauge probe type 2008 requires a standoff distane of 355mm (13.98 inhes) and has a measurement range of 256mm (10.08 inhes) [5]. Therefore, to obtain orret measurements, the laser probes should be mounted suh that the distane from the bottom of the probe to the ground surfae (middle of the measurement range) is approximately 14 inhes. When orretly mounted, distanes plus or minus 128mm (5.04 inhes) from the alibrated ground level an be aurately measured. Refer to Figure 3.4. Measured surfaes whih do not fall within the measurement range will result in invalid readings. The PPU proesses the analog signal from the laser probe. It applies bandpass and anti aliasing filters to the signal. The PPU onverts the analog signal into a serial digital form whih an be transmitted over long distanes to the reeiveraveraging boards loated in the CPU sub-rak. The serial digital output inludes the 12 bit value from the analog to digital onverter as well as 3 invalid data bits. The probe proessing unit determines invalid data if the refleted laser beam is not orretly deteted by the position sensitive detetor in the probe. For example, if the measured surfae is out of the measurement range, the invalid data bits would reflet this and the data ould be proessed aordingly. Another funtion of the PPU is to ontrol the intensity of the laser light emitted by the GaAs laser diode in the probe. This is done through a feedbak mehanism. The reeiver-averaging boards are loated in the CPU subrak as shown in Figure 3.5. There is one board for eah laser probe. Eah board reeives serial data from the gauge probe at a rate of 32 KHz and is apable of reduing the data rate by forming the average of a number of measurements. The data rate, also referred to as updating frequeny, is set by jumpers on the board. The update frequeny ranges from a maximum of 32 KHz (no averaging) down to 62.5 Hz in powers of 20

32 usee 16 pulaes 16 pulaea 16 pulaea Figure 3.2 Pulsed, modulated infrared light from GaAs lasers

33 22 x L A I X Figure 3.3 Triangulation priniple B.

34 23 LJ STANDOFF (13.98 IN.) RESOLUTION 2.44 mv MEASUREMENT RANGE (10.08 IN.) (0-10 VOLTS) GROUND LEVEL Figure 3.4 Laser measurement range -

35 two. Output from the reeiver-averaging boards i~ the mea~ured di3~ane value represented as 12 bit parallel da+a plus a data invalid bit and a data ready flag. This 12 bit p3rallel data value is input to the data aquisition board (DAQ) whih interfaes to the COMPAQ's PC bus DAQ Board The data aquisition board initially used to determine the measuring harateristis and apabilities for the projet is a speially designed board whih uses the Motorola 6800u proessor and plugs into one of the system expansion slots in the COMPAQ Portable III expansion module (See Figure 3.6). Its funtion is to reeive the laser data from the optoator and perform some preliminary proessing of the rak data before passing it on to the COMPAQ Portable III for final rak identifiation and setion analysis. The DAQ board is atually made up of two boards. Shematis for the boards are inluded in Appendix A. The main board ontains the M68000 miroproessor, stati RAM, EPROM, serial and parallel I/0 and is apable of running independently of the other. The seond board is an auxiliary memory board whih only ontains buffers and an additional 512K of stati RAM. This board is used when large amounts of data needs to be stored in real-time. The main DAQ board features inlude an 8 MHz Motorola miroproessor, 64K stati RAM, 64K EPROM; two Motorola parallel interfae and timer hips, an Intel 8251 USART and the IBM PC interfae. The 8 MHz M68000 provides 500 nanoseond bus yles. The stati RAM and EPROMs have 100 and 200 nanoseond aess time, respetively. This allows memory reads and writes with no wait states. THe M68230 PI/T hips are programmed in the 16- bit port mode to provide the parallel interfae for two lasers. The timers on the M68230 provide interrupt signals at the required sampling rate. The Intel 8251 USART gives an RS- 232 ompatible serial interfae running at 9600 BAUD. The serial interfae is used for most of the ommuniations between the DAQ and the COMPAQ. The IBM PC interfae provides an 8-bit parallel interfae for downloading large amounts of laser data to the COMPAQ. 3.3 COMPAQ Portable III The COMPAQ Portable III is the user's interfae to the entire system. From the COMPAQ's keyboard the user an run diagnosti heks on the system, ollet a speified amount of

36 25 Figure 3.5 CPU sub-rak with power supply and reeiveraveraging boards

37 26 data, download rak data to the COMPAQ for storage and subsequent proessing, or enter a real-time rak ounting mode. The programs whih provide detailed rak identifiation and setion analysis reside on the COMPAQ. When the user runs a setion of road to be analyzed, the data is olleted on the DAQ boards and then downloaded to the COMPAQ for off-line analysis. The real-time rak ount mode provides a rough estimate of the number of raks seen as the van moves at highway speeds. This estimate is performed by the DAQ board using a variane measure. In this mode the COMPAQ is used to issue the ommand to the system and to display the rak ount. Figure 3.7 shows the system as it is urrently running in the profilometer.

38 27 Figure 3.6 Data aquisition (DAQ) boards.. 'J. -~ '.L Figure 3.7 Crak system in the profilometer

39 CHAPTER IV TIME AND FREQUENCY ANALYSIS TECHNIQUES This hapter provides some of the basi onepts in the theory of time series analysis needed in the proessing of rak data. Most important among these are the onept of a stohasti proess, a stationary proess, the autoovariane funtion of a stationary proess, the frequeny ontent of a time series, and linear parametri models. Several lassial texts are inluded in the bibliography and may be referened for a more detailed treatment of the subjet [11,12,13). It should be noted that all equations given in this hapter assume real-valued time series. Sine omplex-valued time series are not onsidered, the omplex onjugation operator needed for the stritest definition of autoorrelation and autoovariane has been omitted. 4.1 Time Series A signal whih is ontinuous in time is a ontinuous time series. A disrete time series is simply a sequene of measurements or observations taken at speifi instants of time. Often a disrete time series is a sampling of a ontinuous time series. Typially th~ observations are taken at equispaed time inrements and denoted x(n). A ontinuous time series may be obtained by measurements taken from a physial instrument. Suh a series is bandlimited and ontains no frequenies higher than the maximum frequeny response of the measuring instrument. To analyze a ontinuous time series in disrete form the sampling interval must be determined suh that all information present in the original signal is maintained. This sampling rate must equal or exeed twie the highest frequeny present in the signal and is generally referred to as the Nyquist rate [14]. A signal from whih the series was obtained ould be deterministi or stohasti in nature. If it is possible to predit future values of the series exatly, the signal is deterministi. If future values an only be approximated based on statistial harateristis of past observations, the signal is a statistial or stohasti time series. 28

40 Stohasti Proess The possible values of the time series at a given time t are assumed to be desribed by a random variable X(t) and its assoiated probability distribution. An observed value x(t) at time t represents one of the infinite number of possible values of the random variable X(t). The probability distribution funtion F(x(t)) defined by F(x(t)) = Prob(X(t) < x(t)) is the probability that random variable X(t) has a value less than or equal to x(t). The behavior of the time series at all sampling times is desribed by an ordered set of random variables {X(t) ). The statistial properties of the time series are desribed by assoiating a probability distribution funtion with eah random variable in the set. The ordered set of random variables {X(t)} and the assoiated probability distribution funtions is alled a stohasti proess. An observed time series x(t) is only one of an infinite number of possible realizations of the stohasti proess. Th~ olletion of all sequenes that ould result as realizations of the stohasti proess is alled an ensemble of sample sequenes. The expetation of a random variable X(t) at time t, denoted by E{X}, is given by E{X} f J x p(x) dx = x Here x is the observation at time t and p(x) is the probability density funtion of X(t). This implies that the mean, x, is based on values x taken from all possible ensembles of the random variable at time t. The expeted value of the squared magnitude of random variable X is E{jXI 2 } = J lxl 2 p(x) dx -~ is the mean squared value of X. The variane of a random variable is the mean squared deviation of the random variable from its mean,

41 30 var{x} r = J lx- E{X}I 2 p(x) dx -m An indiation of the statistial relationship of one random variable Xl at time tl to another X2 at time t2 is given by the autoorrelation r{x1x2} = E{X1X2} This represents the engineering definition for autoorrelation, as first suggested by Weiner. The autoorrelation of a stohasti proess with the mean removed is the autoovariane, given by {XlX2} = E{(Xl- E{Xl}) (X2- E{X2}) = r{xlx2} - xl x2 If the random proess has zero mean for time tl and t2 then {X1X2} = r{x1x2}. Also, if the random variables Xl and X2 are mutually independent or unorrelated then {X1X2} = 0. This implies that there is no relationship between the two random variables and knowing values for Xl does not help in prediting a value of X Ergodiity and Stationarity The definitions of mean, variane, autoorrelation, and autoovariane desribed above are based on statistial ensemble averaging. That is, they were based on observations at a partiular time t. In pratie one does not have the luxury of an ensemble of waveforms from whih to evaluate these statistial desriptors. Typially these statistial

42 estimates are obtained from a single waveform x(n) by substituting time averages for ensemble averages. Here x(n) represents a disrete time series. For a stohasti proess to be aurately desribed by time averages instead of ensemble averages the proess must be ergodi. Ergodiity requires a ertain amount of stationarity; that is, the statistis must be independent of the time origin seleted. A random proess is wide sense stationary if its mean is onstant for all time indies and its autoorrelation depends only on the time index differene m, where m=n2-n1. The variable m denotes the time lag, that is, the number of time inrements between time n2 and time n1. All results reported in this study assume the data is wide sense stationary or at :east loally stationary suh that time averages an be substituted for ensemble averages. 4.4 Statistial Estimates If a stohasti proess is ergodi then 31 E{X1} = E{X2} = E{X3} = = E{XN} and the mean, x, an be estimated by N X = 1/N E x(n) n=1 The autoorrelation and autoovariane funtions no longer depend on the time index of the random variable, only the time index differene. The time index differene is referred to as the lag and denoted by m. The autoorrelation, r, and the autoovariane,, then beome and r(m) = E{x(n+m) x(n)} (m) = E{(x(n+m) - = r(m) - x 2 x)(x(n) - x)} Assuming ergodiity, the autoorrelation and autoovariane an be estimated by

43 32 N-m r(m) = 1/(N-m) E x(n) x(n+m) n=1 and N-m (m) = 1/(N-m) Z (x(n) - x) (x(n+m) - x) n=1 4.5 Power Spetrum Estimation Spetral analysis is any signal proessing method that haraterizes the frequeny ontent of a measured signal. In spetral analysis one is typially interested in obtaining a spetral plot whih represents the distribution of signal strength at eah frequeny. Peaks in the spetral plot show whih frequenies are predominant in the signal. Most power spetrum estimation is aomplished by either the autoorrelation or the diret method [15]. The latter method has beome the most popular beause of the fast Fourier transform (FFT) algorithm developed in 1965 [16]. The FFT is a fast, effiient algorithm for omputing the disrete Fourier transform (OFT) of a time series. The OFT determines a sampled periodogram in whih the values of the periodogram for only a disrete number of equally spaed frequenies is omputed rather than evaluating over the ontinuous range of frequenies. The method for alulating the power spetra in this study was first proposed by Welh [17]. This method segments the data, applies a window to eah segment, determines the periodogram of eah windowed segment, and then alulates the average periodogram, whih is alled the modified periodogram. With this method the data segments may be overlapped. This method of periodogram averaging redues the variane of the spetral estimate. The essential features of this method are desribed below. The available time series x(n), 0 < n < N-1, is divided into K overlapping segments of length L. The segments overlap by L/2 samples. The total number of segments then beomes K = (N - L/2)/(L/2) where any frational portion of K is trunated. segment then beomes The ith data

44 33 xi(n) = x(il/2 + n) w(n) where 0 < n < L-1, 0 < i < K-1 of length L.- Typially, eithe~ window is used. and w(n) is a window funtion a retangular or Hamming The DFTs of eah of the K data segments are then omputed using the FFT algorithm by M-1 xi (k) = E n=o xi(n) exp(-jkn(2w/m)) where 0 ~ k ~ M-1 and 0 < i < K-1. must be > L. M is the DFT length and The modified periodograms, Si(k), are then averaged to produe the spetrum estimate K-1 S(2~k/M) = 1/KU E Si(k) i=o for 0 < k ~ M-1, 0 ~ i < K-1 and and Si(k) = IXi(k) 12 L-1 U =! w 2 (n) n=o 4.6 Linear Parametri Modeling Many disrete time stohasti proesses an be approximated by a linear regression model. In this model, the input driving white noise series w(n) and the observed output time series x(n) are related by the linear differene equation

45 34 x(n) = b 0 w(n) + b 1 w(n-1) bqw(n-q) - a 1 x(n-1) apx(n-p) This may be rewritten in the form x(n) = p E aix(n-i) + i=1 q E b w(n-i). 1 1=0 This general regression model is alled an autoregressivemoving average (ARMA) model. If all ai = 0, then x(n) = q ~ b w(n-i). 1 1=0 and the proess is known as a moving average model of order q and represented MA(q). If all bi x(n) = o, i > o, then p = - E a x(n-i) + b 0 w(n). 1 1=1 and the proess is known as an autoregressive model of order p; that is, AR(p). Any one of the three parametri models desribed above may be expressed in terms of the other two models. An ARMA or MA model of a finite number of parameters may be desribed by an AR proess, generally of infinite order. Similarly, an ARMA or AR proess an be expressed as a MA model of infinite order. This observation is important beause it suggests that any of the three models may be seleted and a reasonable model obtained if a suffiiently large order is used. Of the three models, the AR model has mathematial harateristis whih have allowed the development of a number of effiient algorithms. Speifially, AR models have linear solutions; whereas, solving for ARMA or MA parameters involves nonlinear equations.

46 35 Estimates of the AR parameters ai an be obtained as solutions to the p+1 linear equations given by r(o) r{-1) r{-p) 1 r(1) r{o) r(-p+1) 0 = r(p) r(p-1) r{o) 0 These linear equations are ommonly referred to as the Yule Walker equations. The autoorr;lation matrix is both Toeplitz and Hermitian beause r(-k) = r (k), where* represents omplex onjugation. These properties allow more effiient solution than the standard Gaussian elimination. The method for solution of the Yule-Walker equations that takes advantage of these properties was developed by Levinson and is ommonly referred to as the Levinson-Durbin algorithm [31,32].

47 CHAPTER V 5.1 Introdution ANALYSIS OF PAVEMENT CRACKING DATA The methods first investigated to identify pavement raking are omputationally intensive and annot be performed in real-time with the hardware developed in this study. Eah of these methods involve first filtering the data and then applying various statistial tehniques to identify raking. Data is filtered to remove the low frequeny ontent of the signal. Low frequeny omponents inlude suh things as wheel boune, vehile suspension effets, bumps and hills in the setion. The two methods whih onsistently gave best results were the running meanjslope threshold tehnique and the autoorrelation differene method. These are disussed in detail in Setions 5.4 and 5.5, respetively. Another tehnique onsidered was modeling the data as an AR proess and then examining the AR oeffiients. This method would allow rak identifiation and lassifiation if eah raking type would give distintly different AR parameter values and the same type raking would give similar oeffiients. The AR modeling results are disussed in Setion 5.6. As stated, the methods mentioned above give detailed analysis and annot be performed in real-time with existing hardware. It was desired to develop a tehnique, even a rough estimate, whih ould perform in real-time with the hardware desribed herein. A tehnique, using a variane measure, has been implemented whih provides a rak ount in real-time. This is disussed in the following setion. 5.2 Variane Method for Real-Time Crak Counting Although detailed rak identifiation and lassifiation annot be obtained using the DAQ board and COMPAQ at highway speeds, an estimate of the number of raks seen is possible using a simple variane alulation. This method simply alulates the variane every 16 data points (1 inh) and ompares that statisti to a threshold level provided by the operator. If the variane for that inh of data surpasses the 36

48 threshold, the ount is inremented and displayed on the COMPAQ. This alulation takes approximately 40 ~iroseonds on the DAQ board, well within the 71 miroseond requirement for 50 miles per hour. The variane is alulated on 16 raw data values. Sine unfiltered data is used, there exists variane in the data due to the fators previously mentioned whih ontribute to low frequeny ontent of the signal. However, beause. only 16 data points are used in eah alulation these omponents do not ontribute as heavily to the variane value as high frequeny raks and thus filtering an be negleted to save alulation time. The auray of this tehnique is highly dependent upon the operator entering meaningful threshold values. More investigation is needed to determine reasonable threshold limits for various pavement textures. 5.3 Spetral Analysis Results Typially, one of the first things that should be onsidered about any measured signal is its frequeny ontent. As previously disussed, spetral analysis provides this information. Of partiular interest in this study was a determination of whether or not the different raking types displayed harateristi power spetra. Also, it seemed reasonable that raking of the different severity types might show harateristi peaks at different frequenies. The proedures desribed below provide information about the frequeny ontent of pavement raking data. The first question addressed was whether or not eah raking type had its own harateristi spetrum. Here several data segments of 1000 data points in length were identified from the test setions for eah of the desired types. The types onsidered were moderate alligator, moderate blok, and no raking. Moderate transverse raking was not inluded beause by using 1000 data points (5.2 feet) a single rak may or may not have been seen in the data; thus, it would appear as blok or no raking. A typial power spetrum for these three types is shown in Figure 5.1. Three important observations an be made from that figure. First, no raking appears as virtually a straight line. There are no frequenies or range of frequenies whih are predominant. A flat power spetrum indiates white noise: that is, the signal is ompletely random and there is no orrelation in the data. A seond observation is that data with raking shows no notieable peaks at any frequenies but does onsistently show more power at the lower frequenies (greater than 1/4 inh wavelengths). This suggests that data 37

49 with raking is orrelated and statistial measures suh as autoorrelation and autoovariane will be appropriate. Finally, moderate alligator shows more power than moderate blok at the low frequenies. This implies that more raking means more orrelation of the data and larger autoorrelation and autoovariane values should be observed. Figure 5.2 shows typial power spetral results for setions with moderate versus severe alligator raking. Initially it was felt that perhaps different severity (widths) of raking might show peaks at different frequenies. This has not been observed. However, onsistent with previous results, a higher degree of raking again shows more power at wavelengths less than 1/4 inh. Slight alligator raking was not inluded beause it is not believed that the lasers are aurately measuring slight (less than 1/8 inh) raks. In summary, the spetral analysis results indiate that data obtained from road pavements with no raking is unorrelated. Data from pavement with raking is orrelated: in fat, the higher the degree and severity of raking, the more orrelated the data is. 5.4 Running Mean/Slope Threshold Method The basi idea behind this method is that a running mean, representing ground level, is maintained and eah new data value is ompared with this mean to determine if it is a value taken from a rak or not. The term running is used beause the mean must be onstantly updated using the new data points to maintain an aurate representation of ground level. Data points whih are determined to represent a rak or a surfae too muh above ground level, perhaps an extraneous rok or spikes in the data, do not ontribute to the running mean alulation. The running mean is an average alulated from the last N data points whih have been determined to be at ground level. N is user seletable, typially 4 to 8. The simplest way to apply this tehnique is simply to ompare eah new data value to the running mean. If it is below a threshold distane from ground level then identify it as a rak, do not inlude it in the mean, and advane to the next point. If it is less than a threshold distane below the running mean then it is not a rak and the value replaes the "oldest" value used in the mean alulation and a new running mean is determined. Unfortunately this will not provide aurate rak identifiation for raks with gently sloping walls. The problem is that although the values are dereasing they may not exeed the threshold using the tehnique desribed above and so they are inluded in the running mean. This lowers the mean value and makes it even more diffiult for the next point to be identified as part of a rak. The 38

50 eo 70, '0 "" 50 0 lo'l n : O.J FRACTION OF SAWPUNG FREQUENCY a No Croking + Wod. Blok <> Wod. Alll9ator Figure 5.1 Power spetral.density plots of different raking types

51 ,J ,..._.0 't! '-" a Ill :! FRACTION OF SAMPUNG FREQUENCY 0 I.Cod. Alligator + Sev. AJii9<Ji:or Figure 5.2 Power spetral density plots of different raking severity

52 problem is solved by looking ahead up to L lookahead points, assuming all values are onstantly below the mean, for a value exeeding the threshold before updating the running mean. L is a user supplied parameter. If the threshold is exeeded within L points, then eah of the dereasing data points are identified as part of a rak and will not be inluded in the mean. The auray of this method depends on the number of data points used in the mean, the number of data points allowed in the lookahead for threshold violation, and the threshold value itself. After plotting and examining results from various types of raking in the test setions, it is believed that about 85% of the raking an be identified using 4 points for the mean and lookahead value and 35 for a threshold level. 41 This tehnique performs better if the data is first filtered to remove the DC omponent and longer wavelengths. highpass Butterworth filter is typially applied to the raw data. A Figures 5.3 and 5.4 show the results of applying this algorithm to moderate and severe alligator raking, respetively (5.2 feet) filtered data points have been plotted in both figures. Above the filtered data is a plot representing whether or not a rak has been seen. Ground level is plotted at 200 on the Y-axis and raks at 100. The running meanjslope threshold algorithm is inluded in Appendix B. 5.5 Autoorrelation Differene Method The autoorrelation is a statisti whih measures the orrelation of data at different time inrements apart. Assuming ergodiity, the autoorrelation lag m, denoted r(m), tells if data points m time inrements apart over a length of data are related. The autoorrelation value will be approximately zero if the data is unorrelated. As shown by the power spetral analysis results of Setion 5.3, data with raking is orrelated. Data with sharp raks will show large orrelation for a lag or two but the autoorrelation value dereases rapidly as the number of lags inreases. Data with longer wavelength omponents, suh as bumps, show high autoorrelation values for longer lag times. Setion 5.2 disussed a "quik and dirty" way of identifying raks in unfiltered data by alulating the variane, (O), every 16 data points and then omparing that value to a threshold. That method was, at best, an estimate. However, beause the data was not filtered and only a simple variane alulation was needed, it did meet the real-time

53 DATA POINTS (5.2 FT.) Figure 5.3 Running mean/slope threshold tehnique applied to moderate alligator raking data

54 43 ~0, ~ ,_ DATA POINTS (5.2 FT.) Figure 5. 4 Running meanjsl ope threshold tehnique applied to severe alligator raking data

55 requirement. The autoorrelation differene method is an enhanement of the simple variane method. Using this method the data is first filtered with a highpass filter. Filtering removes the DC omponent and muh of the variability aused by hills, tire boune, and vehile suspension effets. This an be seen by omparing the raw data plot in Figure 5.5 with the plot of filtered data in Figure 5.6. With the DC omponent removed, the data now approximates a zero mean proess and the autoorrelation lag 0 is an estimate of the autoovariane lag o whih, by definition, is the variane. The autoorrelation differene method involves determining the spread between r(o) and r(m) alulated for every one inh (16 points) blok of data. This differene is then ompared with a threshold value. As disussed previously, r(o), an estimate of the variane for zero mean data, is large for data with raking. r(o) will also be large if the data varies too far from the zero mean as is the ase on a rough road when the filter is not able to keep the data suffiiently lose to a zero mean. This is illustrated in the last 100 data points plotted in Figure 5.6. r(m) is the autoorrelation for data points in the 16 point blok whih are m time lags apart. r(m), m is typially 4, will derease more rapidly if variane in the data is higher frequeny, that is, sharp raks. Using the property r(o) ~ r(m) and exam1n1ng the four ases for relative values of r(o) and r(m) provides justifiation for this tehnique. 44 CASE I: CASE II: CASE III: r(o) small and r(m) small implies a small differene and no raking. r(o) small and r(m) large is not possible by property r(o) > r(m). r(o) large and r(m) small implies a large differene and raking present. CASE IV: r(o) large and r(m) large implies a small differene and no raking. Figure 5.6 shows filtered data with the r(o)-r(4) value plotted over the sixteenth point of eah blok of data. Also, any differene greater than 1000 is plotted as 1000 so all information ould be plotted on a reasonable sale. As an be seen from the plot, a threshold of 200 identifies all raks exept the one at point A on the plot. Here a shortoming in the algorithm is illustrated. That is, when one 16 point blok ends and another begins in the middle of a small rak it may not be deteted.

56 45... will :::l-o...j : ~ ~ 0 ~ 2.2 ';{~ '-/ 2.1 2, , DATA POINTS (5.2 FT.) Figure 5.5 Raw data..

57 Figure 5.6 and 5.7 taken together illustrate eah of the ases desribed above. Figure 5.6 shows the differene r(o) - r(4) while Figure 5.7 shows the atual values of r(o) and r(4)'. For example, Case III was large r(o) and small r(4). The atual values are plotted in Figure 5.7 and then Figure 5.6 an be examined to see the harateristis of the data and the atual differene value. The autoorrelation differene method has been applied to several of the test setions with good results. In fat, raking identified by this method ompares favorably with that identified by the running meanjslope threshold method. One drawbak of this method, however, is that it will not be able to aurately detet rak width. 5.6 AR Proess Modeling Results This tehnique was investigated to determine whether or not the oeffiients obtained by modeling rak data as an AR proess ould suessfully be used to lassify raking types and severity. The assumption was that raking of the same type and severity would show similar oeffiients while the oeffiients would be signifiantly different for a different type and/or severity of raking. First, several setions of test data were modeled to determine the number of oeffiients to use. It was found that only the first three oeffiients ontributed signifiantly; that is, beyond three lags the oeffiients were essentially zero. This was also substantiated by the fat that the variane of the white noise, the error term, ould only be dereased to a ertain level by adding AR terms; beyond that, it really did not improve the model by adding additional terms. Having determined that three terms should be used in the model, different types of raking were then examined. Bloks of data one foot in length were examined. It was found that data with more raking showed higher autoorrelation values and the oeffiients were signifiantly larger than data with no raking. However, the resolution required to provide the detailed information needed simply was not there. This tehnique ould tell if there was a large amount of raking or little to no raking in eah one foot blok, but that was all. Sine the details, suh as approximate number of raks or severity, ould not be asertained, this method was not onsidered further. 46

58 , ~ o.e 0.7 Case III +~ '""'... '0 t: 0 ":I 0 (::. '-" Case IV 0 -o.1 -o DATA POINTS (5.2 FT.) Filtered Dole + r(o)-r( 4) Figure 5.6 Filtered data with r(o)-r(4) value plotted every 16 data points..

59 "1:7 " 2.2 2,.8,.6, Case IV, ;lo Case III VI :I!i1 a ~ , o 0 -Q.2 + -Q.4 -o.6 + -o.e / t;: I a r(o) + r(4) Figure 5.7 Atual r(o) and r(4) values for the data in Figure 5.6

60 CHAPTER VI CONCLUSIONS AND FURTHER RESEARCH The report desribes the first two phases of researh Projet for developing an automated method of obtaining and evaluating pavement distress and raking information for PES. For these initial two phases, the use of two lasers, one in eah wheel path, are used to obtain raking data whih is proessed on a Motorola based data aquisition board and the COMPAQ Portable III. For detailed analysis, the data is filtered to remove the DC omponent and long wavelengths before proessing. The data is analyzed using several different statistial tehniques. Two tehniques in partiular have been shown to be very reliable. These are the running meanjslope threshold and the autoorrelation differene methods. Software still needs to be written whih take the results from these two methods and provide the detailed reporting, that is, the perentage and severity of eah raking type within the setion. Several important onlusions an be made as a result of this initial study. First, alligator and blok pavement raking an be deteted using the Selom lasers n~junted in the wheel paths. However, it is unlikely that transverse raking an be aurately identified. It is believed that additional lasers must be installed to obtain data aross the lane before the system will be able to provide this information. Using only two lasers it is simply too likely that something in 30 to 50 feet of data will appear to be a rak even on smooth pavement. With multiple lasers, transverse raking would be identified only after eah laser aross the lane had deteted a rak within the same foot or two of data. It should perhaps be pointed out that multiple lasers would also allow rutting to be deteted. Reall, rutting is one of the seven distress types urrently reported by PES. Three lasers are being investigated in Phase 3. Another issue whih remains unresolved is whether or not slight (less than 1/8 inh) raking is aurately deteted. The old lasers with 3/8 by 1/8 inh spot size ould not detet them. The new lasers with 1/4 by 1/16 inh spot size have performed reasonably well on the laser alibration board but have not been thoroughly field tested due to the forementioned problems. 49

61 Any user of this system must understand the limitations imposed by trying to detet raking using only two narrow beams of laser light running parallel to the enterline. Obviously, massive amounts of information aross the lane is not available. Due to the nature of the sensors used, raks deteted in failures and longitudinal raking patterns will be mislassified as alligator or blok raking. There is little that an be done to prevent this using lasers as sensing devies. The distress types suh as failures, pathing and longitudinal raking an only be deteted if the entire lane is examined using video ameras as desribed by other researhers [6,7]. Video system provide muh more detail, but this extra detail presents problems in proessing out the unwanted information. A system with a small luster of lasers along and in between eah wheel path would seem to provide the best hoie, however, would likely be to ostly. As pointed out numerous times, the algorithms developed for detailed identifiation and analysis annot be performed in real-time with the hardware developed in this initial study. Prototype boards whih are wirewrapped, suh as the DAQ board built for this projet, are limited to lok speeds less than 10 MHZ beause of noise problems, regardless of the maximum lok frequeny allowed. Therefore, to obtain faster speeds, printed iruit boards must either be built or purhased. Also, to obtain more omputing power a 32-bit miroproessor should be onsidered over the 16-bit It is believed the open arhiteture VMEsystem developed by Motorola should provide needed hardware upgrades for this projet. The VMEsystem allows the user to purhase a basi ardage whih has the VMEbus interonnet standard. The user an then onfigure the system for his speifi needs by purhasing individual VMEmodules whih simply plug into the VMEbus with the widely aepted euroard onnetor. Typial VMEmodules are miroproessor boards, memory boards, various ontroller boards, and I/O boards. The VMEsystem arhiteture allows the user to onfigure a multiproessor system with both loal and shared memory. A multiproessor VMEsystem is urrently being assembled for this projet. For this system VMEmodules with the miroproessor interfae to the PC. Eah of these VMEmodules will be dediated to proessing the data from a single laser. This system should provide the omputing power needed to filter the data and identify raking, at least with the autoorrelation differene method, in real-time. It is still questionable whether or not the running mean/slope threshold method, whih provides severity information, will run in real-time. It may very well be the ase that the data will be filtered and raks identified in 50