TELK OMNIKA V ol. 16, No . 5, October 2018, pp . 2179 2190 ISSN: 1693-6930 2179 Vibration-Based Dama g ed Road Classification Using Ar tificial Neural Netw ork Y ud y Purnama* 1 and Fer gy anto E. Guna wan 2 1 Computer Science Depar tment, School of Computer Science 2 Industr ial Engineer ing Depar tment 2 BINUS Gr aduate Prog r am - Master of Industr ial Engineer ing Bina Nusantar a Univ ersity , J akar ta, Indonesia 11480 *Corresponding A uthor , email: 1 ypur nama@bin us .edu, 2 fguna w an@bin us .edu Abstract It is necessar y to de v elop an automate d method to detect damaged road because man ually moni- tor ing the road condition is not pr actical. Man y pre vious studies had demonstr ated that the vibr ation-based technique has potential to detect damages on roads . This research e xplores the potential use of Ar tificial Neur al Netw or k (ANN) f or detecting road anomalies based on v ehicle accelerometer data. The v ehicle is equipped with a smar t-phone that has a 3D accelerometer and geo-location sensors . Then, the v ehicle is used to scan road netw or k ha ving se v er al road anomalies , such as , potholes , speedb ump , and e xpansion joints . An ANN model consisting of three la y ers is de v eloped to classify the r oad anomalies . The first la y er is the input la y er containing six neurons . The n umbers of neurons in the hidden la y er is v ar ied betw een one and ten neurons , and its optimal n umber is sought n umer ically . The prediction accur acy of 84.9% is obtained b y using three neurons in conjunction with the maxim um acceler ation data in x , y , and z -axis . The accur acy increases slightly to 86.5%, 85.2%, and 85.9% when the dominant frequencies in x , y , and z -axis , respectiv ely , are tak en into account beside the pre vious data. K e yw or ds: Damaged Road, Vibr ation-based, Accelerometer , Smar t-phone , Ar tificial Neur al Netw or k Cop yright c 2018 Univer sitas Ahmad Dahlan. All rights reser ved. 1. Intr oduction According to La w of the Repub lic of Indonesia, No . 22, Y ear 2009 about Road T r affic and T r anspor tation ( Undang-undang Repub lik Indonesia Nomor 22 T ahun 2009 T entang Lalu Lintas dan Angkutan J alan ) Ar ticle 24 Section (1), road administr ator or go v er nment shall immediately and should repair an y damaged road that could lead to tr affic accidents [1]. Fur ther more , Section 2 of the la w states that in the case that the damaged road cannot b e repaired, the road admin- istr ator is ob liged to put sign(s) on the d amaged roads to pre v ent tr affic accidents . In the case that tr affic accidents occurred because the road administr ator does not immediately repair the damaged road, the y can be impr isoned or fined. The dur ation of impr isonment v ar ies betw een six months up to fiv e y ears . The amount of fine v ar ies betw een 12 millions up to 120 millions r upiah, depending on the victim condition. Road administr ator needs a method to detect damaged road. It is necessar y to de v elop an automated method to detect d amaged road man ually because monitor ing the road condition is not pr actical. Se v er al research eff or ts to w ards automating damaged road detection ha v e been under tak en. There w ere 3D pa v ement reconstr uction methods [2, 3] and laser imaging method [4]. Those methods required special de vices , making them less economical and difficult to im- plement in real situation. There w ere also vibr ation-based approach using acceler ation sensor a v ailab le on smar t-phones [5, 6]. The pre vious studies had demonstr ated that vibr ation technique has potential to detect damaged road [7, 8]. Ho w e v er , those studies w ere only ab le to diff erentiate damaged roads from un-damaged one , without abilit y to distinguish the types of the road anomaly . In this research, w e intend to de v elop a road anomaly classification. Data are collected b y using a 3D accelerometer Receiv ed October 14, 2017; Re vised Ma y 20, 2018; Accepted J une 13, 2018 ,  accredited  First  Grade  by  Kemenristekdikti,  Decree  No:  21/E/KPT/2018 DOI:  10.12928/telkomnika.v16.i5.7574 Evaluation Warning : The document was created with Spire.PDF for Python.
2180 ISSN: 1693-6930 in Android smar t-phone . The accelerometer records the v ehicle vibr ation. Our system can detect f e w types of road anomaly . If one is f ound, then our ar tificial neur al netw or k system will classify its type whether pothole or speed-b ump . 2. Resear c h Method 2.1. Rele v ant W orks The siz e of road netw or k that increases massiv ely demands an automatic road monitor ing system. Ho w e v er , the system is hard to de v elop consider ing the comple xity of the road conditions . F or tunately , the roadw a ys and mobile phone netw or ks ha v e g ro wn sim ultaneously in emerging economies . Mukherjee and Majhi [9] demonstr ated the capability of using smar t-phone that has accelerometers and position sensors . This capability can be useful f or autonomous monitor ing roads . The ability of the smar t-phone in recording acceler ations reliab ly is demonstr ated. Guna w an et al. [8] perf or med similar e xper iment that utiliz ed a smar t-phone which w as enr iched with a 3D accelerometer sensor and geo-location sensor . The smar t-phone installed in a v ehicle . Their study f ound that whene v er a v ehicle crosses a pothole , it will vibr ate significantly in z and x directions . Data collected from the pothole case w ere statistically de viated from the nor mal road and the b ump road cases which sho wing potential f or classification pur pose . 2.2. Motion Sensor s on Andr oid Phone Motion sensors on Android Phone retur n a m ulti-dimensional arr a y of data measured b y each sensor at an instance of time [10]. The accelerometer sensor measure acceler ations in three directions , namely , x , y , and z -axis . The or ientations of those axis on a smar t-phone is depicted in Figure 1. The results of the or ientations on the accelerometer responses are the f ollo wing. If the de vice is pushed on the left side (de vice mo v es to th e r ight), the x acceler ation v alue will be positiv e . If the de vice is pushed on the bottom side (de vice mo v es a w a y from user), the y acceler- ation v alue will be positiv e . If the de vice is pushed to w ard the sky with an acceler ation of A m/s 2 , then the z acceler ation v alue equal to A + 9 : 81 m/s 2 , which corresponds to the acceler ation of the de vice ( A m/s 2 ) min us the acceler ation of g r a vity ( 9 : 81 m/s 2 ). The de vice on stationar y condition will ha v e an acceler ation v alue of +9 : 81 m/s 2 , which corresponds to the acceler ation of the de vice ( 0 m/s 2 ) min us the acceler ation of g r a vity ( 9 : 81 m/s 2 ). 2.3. Vibration-based Method Most road an omalies can be char acter iz ed as high-energy e v ents in the acceler ation data, y et not all e v ents are road anomalies . Another thing such as road fixtures (r ailroad crossings and e xpansion joint) can gener ate significant acceler atio n impulse . P assengers slamming th e door or dr iv er br aking suddenly can also produce high energy e v ents . Er iksson et al. [5] and Guna w an et al. [8] used v ehicle acceler ation data as the main Figure 1. The or ientation of the three ax es on Android smar t-phone [10]. TELK OMNIKA V ol. 16, No . 5, October 2018 : 2179 2190 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 2181 source . Smar t-phone which is enr iched with a 3D accelerometer sensor and geo-location sensor is installed into the v ehicle . Figure 2 sho ws the pothole detection flo wchar t used. The detection method w ould be as f ollo w: 1. V ehicle v elocity will be e v aluated. If it is too lo w , this stream of data will be ignored, and ne xt ne w stream of data will be e v alu ated. This process will be repeat until the stream data satisfied the requirement 2. Apply high-pass filter to remo v e acceler ation, br aking, or tur n e v ents 3. z direction acceler ation ( a z ) will be e v aluated against a threshold ( t z ). This stream data w ould be fur ther processed if maxim um of a z ( a max z ) e xceeds ( t z ); Otherwise ne w data stream will be e v aluated (bac k to step 1). 4. Calculate the largest v alue of x direction acceler ation data ( a x ) within the time inter v al cen- tered at the time of a max x occurr ing. The time inter v al ma y v ar y (32, 64, o r 128). This e xtreme v alue will be chec k ed against a threshold ( t x ). Similar to pre vious step , if a max x < t x , this stream data will be ignored and ne w one will be tested (bac k to step 1). 5. Last step is to reject an y data if t max z < t s :v , where t s is threshold and v is the v ehicle tr a v eling v elocity . 2.4. Data Collection Road anomaly can be defined as abnor mality of the road condition from what it supposed. There are se v er al kinds of road anomaly such as damaged road (pot hole e xistence), speed b ump , r ailroad crossing, or e xpansion joint. This research will f ocus on pro viding a method to detect road anomaly in real time and fur ther classify the types of the road anomaly . There are se v er al things to be prepared bef ore data collection can be perf or med: smar t-phone , v ehi c l e , accelerometer data and the road anomaly it selv es . Smar t-phone and v ehicle: In this research, tw o smar t-phone de vices will be used: De vice A and B . De vice A will be pla ced on the card dashboard, while the de vice B will be placed in the middle of the car floor close to the bac k passenger Accelerometer data: Third par ty softw are that will be used to record the v ehicle acceler ation data. This application will record the acceler ation data and sa v e it in a .csv file . Road anomaly: There are f our kinds of road condition that will be recorded: nor mal, road with pothole , road with speed b ump , and road with e xpansion joint. Figure 2. P othole Detection Flo wchar t [5]. Vibr ation-Based Damaged Road Classification Using Ar tificial ... (Y udy Pur nama) Evaluation Warning : The document was created with Spire.PDF for Python.
2182 ISSN: 1693-6930 2.5. Feature Extraction Ra w accelerometer data ma y not be directly used. These anomalies data are mix ed with noise data, such as passengers slamming the door or dr iv er br aking suddenly that can also produce high energy e v ents . There are se v er al steps bef ore f eatures can be e xtr acted from the r a w data, the y are: 1. Zero Shift: The pur pose of this process to shift each acceler ation data ( x , y , and z ) v alues in data to z ero . All acceler ation data are subtr acted b y theirs median. 2. Sa vitzky-Gola y Filter : The pur pose of this step is to remo v e noise from this acceler ation data. The polynomial order used in this filter is one with fr ame siz e of 41. 3. Deter mine z acceler ation peak point: The moment v ehicle wheel hit the damaged road, the z acceler ation will reach its peak. This point will becomes the median v alue of cutting windo w of data. Number 32 chosen as the siz e of the windo w to co v er more point in time span, because there is a possibility that the peak windo w can be missed. Theref ore data used are 65 points span betw een ( z max 32 ) and ( z max + 32 ). 4. Hamming Windo w a nd F ast F our ier T r ansf or m: F our ier T r ansf or m is implicitly applied to an infinitely repeating signal. Sometimes the star t and end of the finite sample signal do not match, hence mak e it looks lik e a discontin uity in the signal. Applying Hamming Windo w mak es sure that the ends match up while k eeping e v er ything reasonab ly smooth. Sixty-fiv e points that has been acquired bef ore will be applied with Hamming Windo w . 2.6. ANN Model f or Classification ANN is used as classification method because its capability to lear n from e xamples and capture the functional relationships among the hard descr iption of data. The netw or k will be a Mul- tila y er bac k-propagation netw or k. This netw or k will use Sigmoid as its activ ation function.Netw or k par ameter such as p ercentage of tr aining data and n umber of hidden la y ers will be chang ed and tested se v er al times to achie v e the optimal result. Figure 3 sho ws the ANN model used in this study . After pre-processing, there are fiv e input nodes: maxim um x acceler ation data ( a max x ), maxim um z acceler ation data ( a max z ), dominant frequency of x acceler ation ( f dom x ), dominant frequency of y acceler ation ( f dom y ), and dominant frequency of z acceler ation ( f dom z ). The output node w ould be chosen from f our a v ailab le classes of the road condition: nor mal, speed-b ump , pothole , and e xpansion joint. T ab le 1 sho ws the par ameter of neur al netw or k used in this study . If there are 500 data, and ANN set to 10% tr aining set siz e and 50% v alidation set siz e , data composition will be: 50 testing data (r andomly chosen), 225 testing data, and 225 v alidation data. Figure 3. A neur al netw or k model with fiv e neurons in the input la y er and three neurons in the hidden la y er TELK OMNIKA V ol. 16, No . 5, October 2018 : 2179 2190 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 2183 T ab le 1. Neur al Netw or k P ar ameters Used in This Study . P arameter V alue Activ ation function Sigmoid Lear ning r ate 0.3 Momentum 0.2 T r aining time 10000 Number of neuron in the hidden la y ers 2–9 T r aining set siz e 10–90% V alidatian set siz e 50% There are tw o separ ated e xper iments . First e xper iment is to deter mine the reliab le sam- ple siz e: This process deter mines minim um por tion of tr aining data needed to achie v e desired result. T r aining data por tion will be increased g r adually with increment of 10% until 90% por tion of tr aining data. Ev er y m ultiple of 10%, the data set will be classified a hundred times . The optimal tr aining data por tion will be used in the second process . The second process is deter mining the optim um n umber of Neurons: Using pre viously obtained optimal par ameter , n umber of neurons in the hidden la y er will be changed from 2 up to 9. Each v ar iation will be r un f or 100 times classification process . 3. Result and Anal ysis 3.1. T ypical Acceleration Data This section sho ws ho w each road anomaly aff ects the accelerometer data. Figure 4 sho ws acceler ation data when a v ehicle crosses a nor mal road. The best indicator is that z acceler ation tends to sta y at g r a vity acceler ation which is +9 : 81 m/s 2 . Using this inf or mation can be concluded that v ehicle crosses nor mal road will ha v e its z acceler ation relativ ely sta ys at +9 : 81 m/s 2 . An y r ise or f all from this v alue is the indicator of road anomaly . Figure 4. T ypical acceler ation data when the test v ehicle crosses a road without road anomalies . Vibr ation-Based Damaged Road Classification Using Ar tificial ... (Y udy Pur nama) Evaluation Warning : The document was created with Spire.PDF for Python.
2184 ISSN: 1693-6930 Figure 5 sho ws acceler ation data when a v ehicle crosses a nor mal road then hits a pot- hole . Region in betw een the sixth and eighth seconds is when the v ehicle hits the pothole . Notice that star ting from the nor mal v alue of g r a vity acceler ation, the z acceler ation f alls to 5– 7 m/s 2 . It is when the front wheel hits the base of the pothole . After that the z acceler ation star ts to r ise significantly to 12–13 m/s 2 . It is when the front wheel e xits the pothole . The ne xt drop is caused b y the rear wheel hitting the pothole base . Identical to the pre vious one , this one is also f ollo w ed b y another r ise when the rear wheel e xits the pothole Figure 6 sho ws acceler ation data when a v ehicle crosses a nor mal road then hit a speed b ump . Region in betw een the ele v enth and thir teenth second is the time when the v ehicle hits the speed b ump . When the front wheel hits the speed b ump , it giv es significant increase to z acceler ation from g r a vity acceler ation v alue to about 13-14 m/s 2 . After that z acceler ation star ts to f all off because the front wheel has passed through the speed b ump . Figure 7 sho ws acceler ation data when a v ehicle crosses a nor mal road then passing an e xpansion joint. Region in betw een the fifth and sixth second is the data recorded when the v ehicle crosses the e xpansion joint. When the wheel hits the e xpansion joint, it drops the z acceler ation to about 7-8 m/s 2 . Then the z acceler ation r ises significantly to about 15 m/s 2 . 3.2. Determining the Reliab le Sample Siz e This study e v aluates a v ar iation of the tr aining data siz e to the accur acy of the ANN prediction. The approach is of the f ollo wing. Firstly , the tr aining siz e is fix ed at 10% of the total sample siz e . The remain data are equally divided f or the v alidation and testing stages . F or these fix ed siz es , the data are resampled f or a hundred times using a Monte Car lo sim ulation. This procedure is repeated f or the tr aining siz e of 20%, 30%, ..., 80% and 90%. The eff ects of the data siz es on the accur a cy are sho wn in Figure 8. The ANN model tr ained using 10% data is only about 15% accur ate or about 85% misclassify the cases . The accur acy increases almost steadily with the increasing of the tr aining data siz e until the data siz e reaches 50%. After the siz e , the accur acy still slightly v ar ies with the data siz e . The highest accur acy is obtained f or 80% tr aining data siz e . Figure 5. T ypical acceler ation data when the test v ehicle crosses a pothole . TELK OMNIKA V ol. 16, No . 5, October 2018 : 2179 2190 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 2185 Figure 6. T ypical acceler ation data when the test v ehicle crosses a speed b ump . Figure 7. T ypical acceler ation data when the test v ehicle crosses a speed b ump . 3.3. Determining the Optim um Number of Neur ons Increasing the n umber of neurons increases the capability of the model to fit more com- ple x relationship . Ho w e v er , this comple xity ma y happen due to o v er fitting. A good ANN netw or k model should be gener al and not o v erfit to a specific case . A minim um n umber of neurons is usu- Vibr ation-Based Damaged Road Classification Using Ar tificial ... (Y udy Pur nama) Evaluation Warning : The document was created with Spire.PDF for Python.
2186 ISSN: 1693-6930 ally required to pro vide a gener ic model. T o find this gener ic model, the neur al model accur acy is computed f or a v ar ious n umber of neurons . The results are depicted in Figure 9. F or the tw o- neuron case , the accur acy v ar ies widely from around 57% up to around 91%. Ho w e v er , f or the cases where the n umber of neuro ns is thr ee and nine , the accur acy v ar iation is relativ ely constant from one case to the others . The figure suggests that the most op tim um n umber of neurons is three . 3.4. Determining the Significance Features F eature selection is the process of selecting a subset of rele v ant f eatures f or use in the classifier model constr uction. Sometimes the data collected ma y redundant or irrele v ant. F ea- tures selection ma y help eliminate this possibility b y pre v enting loss of inf or mation. Theoretically , smaller n umber of f eatures can decrease the classifier w or kload, hence decreasing the modelling and tr aining time of the classifier . This also increases the classifier perf or mance b y maintaining its accur acy . T o perf or m f eature selection, the condition of the data in each class m ust firstly be obser v ed. The distr ib ution of f eatures of the classification is sho wn in Figure 10. The dominant frequency of x in nor mal class is 1.52, which is identical in other classes too . Meanwhile , the dominant frequency of y in nor mal and pothole case both has score 1.52, while speedb ump and e xpansion joint ha v e 1.21 and 1.49. Only the dominant frequency of z that has v ar ied score f or each class . Using these f acts , fur ther classification is perf or med b y reducing the n umber of f eatures in v olv ed in th e classifier . T ab le 2 sho ws which f eatures presence in each classifier . Each classifier used 80% tr aining data and three neurons in the hidden la y er . This classifier is resampled f or a Figure 8. The eff ects of the por tion of the tr aining data to the classification accur acy of the road anomalies . TELK OMNIKA V ol. 16, No . 5, October 2018 : 2179 2190 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 2187 hundred times using a Monte Car lo sim ulation. After testing the classifier , the results are depicted in Figure 11. Classifier A that used all the f e atures has accur acy of 85.2%. Classifier B has 46.1% accur acy , which is the w orst amongst another classifier . Classifier B did not include a max x as its f e atures . F rom this result can be predicted that a max x is a significance f eatures . Classifier C accur acy is 83.3%. Its accur acy is slightly lo w er than classifier A. This clas- sifier did not ha v e a max y as its f eature . Meanwhile Classifier D has second lo w est accur acy at 75% Figure 9. The eff ect of the n umber of neurons in the hidden la y er to the classification accur acy of the road anomalies . T ab le 2. The Combination of F eatures Studied in The Research. Case F eatures a max x a max y a max z f dom x f dom y f dom z A X X X X X X B X X X X X C X X X X X D X X X X X E X X X X X F X X X X X G X X X X X H X X X Vibr ation-Based Damaged Road Classification Using Ar tificial ... (Y udy Pur nama) Evaluation Warning : The document was created with Spire.PDF for Python.
2188 ISSN: 1693-6930 Figure 10. The f eatures distr ib ution f or each class in this classification. TELK OMNIKA V ol. 16, No . 5, October 2018 : 2179 2190 Evaluation Warning : The document was created with Spire.PDF for Python.