Oleh:
Abstrak
Ruas jalan Gunung Sarik Kuranji kota Padang, merupakan salah satu ruas jalan yang ramai dilalui oleh
kendaraan mengangkut tanah galian C ke PT. Semen Padang, Indarung. Saat dilalui diharapkan kondisi
jalan dalam keadaan baik, sehingga kendaraan yang lewat bisa membawa tanah galian dalam jumlah yang
banyak dan lebih efektif, namun jika jalan mengalami kerusakan tentu akan berakibat kendaraan yang
melewati jalan tersebut akan cepat rusak dan biaya operasional kendaraan akan lebih besar.Kondisi
sekarang adalah terjadinya kerusakan yang cukup parah, sehingga kendaraan tidak aman dan nyaman saat
melewati jalan tersebut. Dari hasil survey yang telah dilakukan dengan menggunakan format survey dari
Departemen Pemukiman dan Prasarana Wilayah terdapat kerusakan yang bervariasi mulai dari rusak
ringan, sedang sampai berat. Kerusakan yang terjadi diperparah dengan kelebihan muatan dari kendaraan
yang melalui ruas jalan ini, selain kondisi curah hujan dan volume lalu lintas yang tinggi.Dari hasil
penelitian, kondisi jalan ruas Gunung Sarik Kuranji kota Padang STA 0+000 – 1+000 saat ini, kondisi baik
hanya sebesar 4,76 %, kondisi ringan 23,81 %, kondisi sedang 19,05 %, dan kondisi rusak berat 52,38 %.
Untuk itu perlu dilakukan penanganan segera, terhadap kerusakan yang terjadi dengan memperbaiki jalan
tersebut demi kelancaran lalu lintas kendaraan yang melewati jalan ini.
Abstract
Mountain road section of Padang, a prayer one thing crowded roads passable by vehicles transporting
dugouts C ke PT. Semen Padang, Indarung. When traversed expected road condition hearts good condition,
so the vehicle that pass can bring dugouts hearts period the many and more effective, but if the damage
would result vehicles that pass through these roads will be fast damaged and operating costs vehicle will be
more big. Conditions now is enough severe damage, so the vehicle is not safe and comfortable as it passes
through the streets. From the survey findings has done with survey using format from the Ministry of
Settlement and Regional Infrastructure there is damage that varies from damaged light, medium up weight.
Damage happens compounded with overloaded from vehicles the singer through roads, in addition to
rainfall condition and its high volume traffic. From the results of research findings, road condition Gunung
Sarik Kuranji segment of Padang STA 0 + 000-1 + 000 before currently, good condition only by 4.76%,
23.81% light condition, condition was 19.05%, and condition heavy damage 52,38%. To review the handling
of should be done soon, against damage happens with fixing these roads for smooth traffic vehicles the road
passes singer.
DOI 10.21063/JM.2016.V18.2.26-34
© 2016 ITP Press. All right reserved. 26
Vol.18 No.2 Agustus 2016 Jurnal Momentum ISSN : 1693-752X
27
Vol.18 No.2 Agustus 2016 Jurnal Momentum ISSN : 1693-752X
28
Vol.18 No.2 Agustus 2016 Jurnal Momentum ISSN : 1693-752X
Tabel 2. Hubungan Nilai RCI dengan Kondisi dari korelasi nilai Bump Integreter (BI)
Jalan berdasarkan survey mengunakan alat
No Jenis Permukaan
Kondisi ditinjau Nilai NAASRA, Sedangkan nilai RCI diperoleh
secara visual RCI
1 Jalan tanah dengan Tidak bisa 0 -2
berdasarkan survey secara visual. Namun
drainase yang jelek, dan dilalui sebagai pedoman dalam menentukan kondisi
semua tipe permukaan dan bentuk penanganan jalan
2 Semua tipe perkerasan Rusak berat 2-3
yang tidak diperhatikan banyak lubang
sejak lama (4 – 5 tahun dan seluruh Tabel 3. Hubungan Nilai IRI dan RCI
atau lebih) daerah terhadap Kondisi Jalan
perkerasan
3 Pemiliharaan lama, Rusak 3-4 RCI
Latasbum lama, batu bergelombang, IRI
kerikil banyak lubang
11 - < 16
4 Pemeliharaan setelah Agak rusak, 4-5 (m/km) < 11 % 16 - < 23 % > 23 %
%
pemakaian 2 tahun, kadang-kadang
Latasbum lama ada lubang,
<4 Baik, Sedang, Rusak Rusak
permukaan
Pemeli Pemeliha Ringan, Berat,
tidak rata
haraan raan Pemelihara Peningkata
5 Pemeliharaan baru, Cukup tidak ada 5-6
Rutin Rutin an Berkala n/Kontruksi
Latasbum baru Lasbutag atau sedikit
setelah pemakaian 2 sekali lubang,
tahun permukaan 4–8 Sedang Sedang, Rusak Rusak
jalan agak tidak , Pemeliha Ringan, Berat,
rata Pemeli raan Peningkata
6 Lapis tipis lama dari baik 6-7 haraan Berkala Pemelihara n/Kontruksi
hotmix, Latasbum baru, Berkal an Berkala
Lasbutag baru a
7 Hotmix setelah 2 tahun, Sangat baik, 7-8
Hotmix tipis diatas umumnya rata 8 – 12 Rusak Sedang, Rusak Rusak
permukaan Ringan Pemeliha Ringan, Berat,
8 Hotmix baru (Lataston, Sangat rata dan 8 - 10 , raan Peningkata
Laston) peningkatan teratur Pemeli Berkala Pemelihara n/Kontruksi
dengan mengunakan haraan an Berkala
lebih dari 1 lapi Berkal
a
Sumber : Petunjuk Pelaksanaan Subbidang Jalan Dinas
Pekerjaan Umum (2010) > 12 Rusak Sedang, Rusak Rusak
Berat,P Peningka Ringan, Berat,
eningk tan/ Peningkata
3.3. Hubungan Nilai IRI terhadap RCI atan / Kontruks Peningkata/ n/Kontruksi
Kontru i Kontruksi
ksi
Secara analitis nilai IRI dengan RCI
tidak berhubungan karena nilai IRI didapat Sumber : Petunjuk Pelaksanaan Subbidang Jalan Dinas
Pekerjaan Umum (2010)
29
Vol.18 No.2 Agustus 2016 Jurnal Momentum ISSN : 1693-752X
30
Vol.18 No.2 Agustus 2016 Jurnal Momentum ISSN : 1693-752X
31
Vol.18 No.2 Agustus 2016 Jurnal Momentum ISSN : 1693-752X
STA 1+000 banyak terdapat kerusakan Pengaruh dari lalulintas yang cukup
struktur badan jalan berupa, lobang-lobang padat karena banyaknya jumlah
cukup parah dengan ukuran dan kedalaman kendaraan
yang bervariasi. Bentuk lubang yang terjadi Pengaruh dari beban kendaraan yang
beraneka rangam seperti bentuk bulat, overload
persegi, memanjang dan terjadi pada Passing pemadatan perkerasan jalan
beberapa lokasi di antaranya adalah pada yang tidak merata
STA 0+000, STA 0+050, STA 0+150, STA
0+200, STA 0+250, STA 0+300, STA c. Analisis kerusakan perkerasan jalan yang
0+350, STA 0+400, STA 0+450, STA berlubang (photoles)
0+500, STA 0+500, STA 0+700, STA
0+750, STA 0+950, STA 1+000. Jumlah Kerusakan perkerasan jalan yang
lobang dari STA 0+000 – 1+000 berjumlah berlubang dapat disebabkan oleh beberapa
44 buah dengan luas 486 m² dan luas jalan 5 faktor antara lain :
m² x 50 m² = 250 m²
Makin meningkatnya perkembangan
3.4. Analisis Kerusakan Permukaan Jalan kerusakan jalan lain seperti retak,
amblas, beralur yang tidak segera
a. Analisis kerusakan retak kulit buaya ditangani.
(aligator crack) Terdapatnya permukaan aspal yang
Kerusakan perkerasan jalan berupa sudah aus
retak kulit buaya (aligator crack) dapat Bahan/material aspal yang tidak sesuai
disebabkan antara lain oleh : dengan mutu
Lapis pondasi yang kurang baik Mutu aggregat base yang tidak sesuai
Kualitas perkerasan jalan yang kurang dengan standar
baik, sehingga menyebabkan Beban kendaraan yang melebihi
perkerasan lemah atau lapis aspal yang standar ketentuan (overload)
rapuh Pemakaian aspal yang kurang
Kemungkinan terdapatnya campuran
Lapisan aspal yang kurang
aspal yang kurang baik
Butiran halus terlalu banyak
Pelapukan pada aspal
Terdapatnya genangan air karena tidak
Penggunaan aspal yang kurang atau
berfungsinya salura samping
tidak sesuai dengan ukuran
(drainase)
Pemadatan yang tidak merata dan tidak Umur jalan sudah melebihi umur
sesuai standar (spesifikasi)
rencana.
Tingginya muka air tanah pada
perkerasan jalan 4. Pembahasan
Lapisan bawah kurang stabil
Alternatif Solusi Permanen
b. Analisis kerusakan berupa penurunan Alternatif solusi permanen yang dapat
(depression) dilakukan untuk berbagai jenis kerusakan
Penurunan (depression) perkerasan yang terdapat pada jalan ruas Gunung Sarik
jalan dapat disebakan oleh beberapa faktor Kuranji kota padang STA 0+000 – STA
antra lain : 1+000 adalah sebagai berikut :
1. Dengan cara membongkar kembali
Tanah dasar tidak kuat mendukung lapis pondasi bawah dan lapis pondasi
beban lalulintas karena daya atas, kemudian dipadatkan dengan cara
dukungnya rendah yang disebabkan per layer dan ditutup kembali dengan
banyaknya tanah humus dan aspal. Selanjutnya baru dilaksanakan
pelapukan. pekerjaan lapis overlay secara
Lapisan pondasi turun yang keseluruhan dengan menambah
kemungkinan disebabkan oleh ketebalan lapis permukaan aspal
tingginya muka air tanah (desain ulang tebal perkerasan).
32
Vol.18 No.2 Agustus 2016 Jurnal Momentum ISSN : 1693-752X
33
Vol.18 No.2 Agustus 2016 Jurnal Momentum ISSN : 1693-752X
34
See discussions, stats, and author profiles for this publication at: https://www.researchgate.net/publication/335107258
CITATIONS READS
0 392
3 authors:
Rizky Rinaldiansyah
UNTIRTA
1 PUBLICATION 0 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Muhammad Fakhruriza Pradana on 11 August 2019.
oleh:
Muhammad Fakhruriza Pradana
Teknik Sipil, Universitas Sultan Ageng Tirtayasa
Email: mfakhruriza@untirta.ac.id
Rizky Rinaldiansyah
Teknik Sipil, Universitas Sultan Ageng Tirtayasa
Email: rizkialdi15@gmail.com
Abstrak : Jalan raya exit Gerbang Tol Cilegon Timur merupakan sarana infrastruktur jalan yang menuju
ke simpang empat/simpang bersinyal Perumahan Cilegon Indah dan terminal Seruni Cilegon. Dengan
pertumbuhan lalu lintas di Kota Cilegon yang semakin padat akan menambah volume kendaraan yang
melewati jalan akses ini, mengakibatkan jalan mengalami kerusakan akibat beban kendaraan yang
meningkat setiap tahunnya. Pada penelitian ini dilakukan perbandingan perhitungan perkerasan kaku
menggunakan metode Bina Marga 2003 dan AASHTO 1993 pada ruas jalan akses tol cilegon timur STA
0+000 – 1+600. Pada lokasi penelitian jalan terbagi atas 2 segmen. Dari hasil perhitungan pada segmen I
metode Bina Marga di dapat tebal perkerasan adalah 20 cm dengan bahan pondasi bawah bahan pengikat
sebesar 10 cm, sedangkan dengan metode AASHTO 1993 didapatkan ketebalan 24 cm menggunakan
bahan pengikat setebal 10 cm. Pada segmen II dengan metode Bina Marga di dapat tebal perkerasan
adalah 18 cm dengan menggunakan bahan pengikat sebesar 10 cm dan metode AASHTO 1993 didapat
ketebalan perkerasan sebesar 23 cm dengan menggunakan bahan pengikat setebal 10 cm.
Kata kunci : Perkerasan Kaku, Metode Bina Marga 2003, Metode AASHTO 1993, Tebal Perkerasan
Abstract : The Road of East Cilegon exit Toll Gate is the infrastructure that leading to the four-lane
intersection / signalized intersections to Cilegon Indah Housing and to Seruni terminal in Cilegon City. With
the growth of traffic in Cilegon that increasingly crowded will increase the volume of vehicles passing
through these access roads, resulting in road damage due to vehicle loads are increasing every year. This
study focusing on comparison calculation for rigid pavement using Bina Marga 2003 and AASHTO 1993 on
the Road of East Cilegon exit Toll Gate from STA 0 + 000 to STA 1 + 600. The object of this study is divided
into two segments. From calculations in segment I, the Bina Marga method obtained pavement thickness of
20 cm with a foundation using a binding material of 10 cm thick, whereas with the AASHTO method 1993
obtained a thickness of 24 cm using a foundation using a binder 10 cm thick. In segment II with the Bina
Marga method, the pavement thickness was 18 cm using a 10 cm binder and the AASHTO method in 1993
obtained a 23 cm pavement thickness using a 10 cm thick binder.
Keywords: Rigid Pavement, Methode of Bina Marga 2003, Methode of AASHTO 1993, Pavement Thickness
57 | K o n s t r u k s i a
Jurnal Konstruksia | Volume 10 Nomer 2 | Juli 2019
58 | K o n s t r u k s i a
PERBANDINGAN PERENCANAAN PERKERASAN KAKU PADA AKSES JALAN TOL (Fakhruriza – Rindu – Rizky)
59 | K o n s t r u k s i a
Jurnal Konstruksia | Volume 10 Nomer 2 | Juli 2019
60 | K o n s t r u k s i a
PERBANDINGAN PERENCANAAN PERKERASAN KAKU PADA AKSES JALAN TOL (Fakhruriza – Rindu – Rizky)
a. Peta Lokasi
Peta lokasi menunjukan lokasi
penelitian yang akan dilakukan pada
ruas jalan akses tol cilegon timur STA
0+000 – 1+600 Cilegon Banten.
b. Pertumbuhan Lalu lintas
Data jumlah kendaraan di setiap
daerahnya baik itu kabupaten,
provinsi dan kota madya berada pada
Badan Pusat Statistik (BPS). Data yang
dibutuhkan yaitu peningkatan jumlah
kendaraan.
2. Data Primer
Data primer merupakan data yang diperoleh
secara langsung tanpa perantara. Data
primer pada penelitian ini didapat dengan
cara melakukan pengamatan pada lokasi
Gambar 1. Bagan Alir Penelitian
yang akan ditinjau. Data primer pada
penelitian ini yaitu lalu lintas harian rata-
Hasil Penelitian dan Pembahasan
rata dan dan data CBR tanah.
a. Data CBR tanah Pada bagian ini penulis akan menganalisis
CBR tanah didapat dari survey dengan dan membahas perhitungan tebal lapis
jarak antar titik pengambilan sampel perkerasan dari data yang ada dengan
tanah adalah 200 m, yang dilakukan mengikuti langkah perencanaan yang sudah
peneliti dengan menggunakan alat ada baik menggunakan metode Bina Marga
bantu DCP (Dynamic Cone 2003 maupun metode AASHTO 1993.
Penetrometer) yang kemudian Terdapat 2 segmen perhitungan dalam jalan
dianalisis oleh tenaga ahli di ini.
laboratorium FT. UNTIRTA.
b. Lalu Lintas Harian Rata-rata Perhitungan dengan Metode Bina Marga
Untuk mengetahui data lalu lintas 2003 (Segmen I)
harian rata-rata (LHR) di dapat dari
Kondisi lapis keras jalan yang di rencanakan
hasil survey pada ruas jalan akses tol
adalah sebagai berikut:
cilegon timur STA 0+000 – 1+600
1) Panjang : 1,6 km
Cilegon Banten.
2) Kelas Jalan : Kelas II
3) Klasifikasi jalan : Arteri
Tahapan Penelitian
4) Klasifikasi medan : Datar
Pada sebuah penelitian tentunya 5) Jumlah Jalur : 2 Arah
dibutuhkan tahapan/alur yang jelas, terarah 6) Jumlah lajur : 2 Buah
dan terencana dengan baik agar dicapai hasil 7) Kelandaian jalan : < 3%
penelitian yang berjalan sesuai dengan yang 8) Nilai CBR : 21 %
diharapkan. Adapun tahapan-tahapan dalam 9) Umur Rencana : 20 tahun
penelitian ini dapat dilihat pada began alir 10) Pertumbuhan lalu lintas (i) : 10 %
berikut ini. 11) Kuat tarik lentur (fcf) : 4,0 Mpa
61 | K o n s t r u k s i a
Jurnal Konstruksia | Volume 10 Nomer 2 | Juli 2019
12) Kuat Tekan Beton (fc’): 320 kg/cm²: maka didapat nilai CBR efektif = 48%
32 Mpa
13) Bahu Jalan : ada (Beton)
14) Ruji (Dowel) : Ya
15) Koefisien distribusi arah: 0,475
16) Jenis perkerasan kaku: BBDT dengan
Ruji
62 | K o n s t r u k s i a
PERBANDINGAN PERENCANAAN PERKERASAN KAKU PADA AKSES JALAN TOL (Fakhruriza – Rindu – Rizky)
Tabel 2. Tegangan Ekivalen dan Faktor b. Lalu Lintas Harian Rata-rata (LHR). Data
Erosi lalu lintas dilihat pada tabel 1.
CBR Eff
tanah
Tegangan Setara
Faktor Erosi
Dengan Ruji/Beton bertulang
c. Pertumbuhan Lalu Lintas. Tingkat
Dasar
STRT STRG STdRG STrRG STRT STRG STdRG STrRG
pertumbuhan lalu lintas di analisa untuk
(%)
35 0.83 1.25 1.05 0.84 1.73 2.33 2.36 2.39 memprediksi arus beban lalu lintas yang
48 0.825 1.225 1.03 0.83 1.72 2.32 2.34 2.36
50 0.82 1.2 1.01 0.82 1.71 2.31 2.32 2.33
akan datang. Nilai didapat dari
Sumber : Hasil Perhitungan, 2016 perhitungan dengan data yang diperoleh
f. Menentukan Faktor Rasio Tegangan dari BPS Banten. Faktor pertumbuhan
(FRT). Faktor Rasio Tegangan (FRT) lalu lintas (R) tiap jenis kendaraan adalah
didapat dengan cara membagi Tegangan sebagai berikut :
Ekivalen dengan, Kuat Tarik Lentur (fcf). 1) Mobil Penumpang :
(1+0,10)20-1
TE TE
FRT STRT = Fcf = 4,25 Mpa= 4,25 Mpa = 0,25
0,825 R= = 57,274
0,10
TE TE 1,225 2) Pick Up :
FRT STRG = Fcf = 4,25 Mpa= 4,25 Mpa = 0,41 (1+0,10)20-1
TE TE 1,03
R= 0,10
= 64,202
FRT STdRG = Fcf = 4,25 Mpa= 4,25 Mpa = 0,34
3) Bus :
TE TE 0,83
FRT STrRG = = = = 0,26 R=
(1+0,010)20-1
= 115,379
Fcf 4,25 Mpa 4,25 Mpa 0,010
g. Menentukan Jumlah Repetisi Ijin Fatik 4) Truk :
dan Repetisi Ijin Erosi. Menghitung (1+0,010)20-1
R= = 64,202
repetisi ijin fatik dan repetisi beban ijin 0,010
a. Data Pendukung
Data Pendukung dalam analisi ini adalah:
1) Klasifikasi jalan : Arteri
2) Status jalan : Urban
3) Umur rencana : 20 tahun Sumber : Hasil Perhitungan, 2016
4) CBR : 21%
5) Mutu Beton:320kg/cm2 (32Mpa)
63 | K o n s t r u k s i a
Jurnal Konstruksia | Volume 10 Nomer 2 | Juli 2019
64 | K o n s t r u k s i a
PERBANDINGAN PERENCANAAN PERKERASAN KAKU PADA AKSES JALAN TOL (Fakhruriza – Rindu – Rizky)
Hal yang sama terjadi pada segmen II Marga (Studi Kasus Jalan Kawasan
dimana metode Bina Marga memiliki Industri Krakatau Cilegon). Jurnal
ketebalan lebih rendah 5 cm Fondasi Volume 2 No 1. Cilegon.
dibandingkan metode AASHTO. Undang-undang Republik Indonesia No. 38
Perbedaan ketebalan tersebut Tahun 2004, Tentang Jalan.
disebabkan karena masing-masing
metode yang digunakan mempunyai
parameter yang berbeda dalam
mengelompokan tingkat kepadatan
kendaraan.
Daftar Pustaka
65 | K o n s t r u k s i a
Jurnal Konstruksia | Volume 10 Nomer 2 | Juli 2019
66 | K o n s t r u k s i a
ScienceDirect
International Journal of Pavement Research and Technology 11 (2018) 582–593
www.elsevier.com/locate/IJPRT
Received 7 July 2017; received in revised form 19 November 2017; accepted 18 December 2017
Available online 29 December 2017
Abstract
The determination of traffic volume in asphalt pavement design has not considered the effects of seasonal variability on vehicle axle
loads in China. By analyzing the effects of seasonal variation on the axle loads, formulas for the conversion of modified axle load to
traffic volume have been derived and presented, in accordance with specifications in China. This study includes the following: derivation
of formulas for converting modified axle loads to traffic volume, classification of 3 different seasons, determination of seasonal distribu-
tion and influencing coefficients, and application of modified axle load conversion formulas, among others. By comparison of results
between modified and unmodified formulas applied in a sample asphalt pavement design in Harbin, China, the study demonstrates that
the modified axle load conversion formulas are scientific and rational. The results of asphalt pavement design obtained using the mod-
ified axle load conversion formulas increase the thickness and whole strength of the asphalt pavement structure; in addition, it can more
efficiently satisfy the requirements of an actual vehicle running on the pavement.
Ó 2017 Chinese Society of Pavement Engineering. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Keywords: Traffic volume; Vehicle axle load; Axle load conversion formula; Seasonal influence; Modified formula
https://doi.org/10.1016/j.ijprt.2017.12.007
1996-6814/Ó 2017 Chinese Society of Pavement Engineering. Production and hosting by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
H. Zhang et al. / International Journal of Pavement Research and Technology 11 (2018) 582–593 583
resistance performance was determined [7]. A three- In this study, the axle load conversion in traffic volume
dimensional finite element model was utilized to simulate in accordance with the specifications in China is modified
the AASHO road test pavement sections and to develop by annual seasonal factors. The project aims to explore a
load equivalency factor (LEF) formulae for various pave- reasonable method for calculating traffic volume. Research
ment sections under various axle loadings [8]. The fatigue findings and methods are presented in Fig. 2.
damage caused by multiple-axle configurations is a key By analysis of different seasonal effects on traffic axle
issue for pavement design, and a method for the computa- load or pavement structure, the modified axle load conver-
tion of the fatigue life of bituminous mixtures under sion formulas to traffic volume can be derived and pre-
multiple-axle configurations was presented [9]. All afore- sented in accordance with the current principles of
mentioned research has been focused on axle load spec- asphalt pavement design in China. Using the aforemen-
trum and its model, but the modification of axle load tioned modified formulas and by predicting and calculating
affected by seasonal factors has not been involved to any the deflection of pavement and the resilient modulus of
significant extent. subgrade and pavement materials in different areas and
Fig. 1 shows how seasonal variability affects pavement seasons, the seasonal influencing coefficients in the modi-
performance under the same traffic loads. fied formulas can be determined. With the application of
In accordance with the specifications in China, aspects modified formulas in asphalt pavement design, the scien-
which contribute toward seasonal variability in traffic vol- tific rationality of the modified formulas is further
ume mainly include the following: demonstrated.
(1) Requirements with respect to subgrade consistency in 2. Modified methods of axle load to traffic volume formula
asphalt pavement design.
(2) Strength indexes of asphalt and asphalt mixtures, 2.1. Classifications of different seasonal effects
which are determined based on only one temperature.
(3) Effects of seasonal variability on pavement deflection In accordance with the influence of seasonal variability
in the refurbishment of an old asphalt pavement. on annual traffic axle loads, the year is divided into 3 differ-
ent seasons, as shown in Table 1. The influence on the axle
Analysis of the 3 aforementioned aspects indicates that load for each season is the same.
these aspects have not entirely considered the effects of sea- The determination of 3 different seasons depends on the
sonal variability on traffic volume. In fact, the design different areas in China because of the large natural vari-
parameters (deflection, resilient modulus, etc.) of subgrade ability of the climate; thus, the months constituting the 3
and pavement materials can be influenced by temperature seasons vary across different areas in China. The months
and humidity of different seasons. The resilient modulus for each of the 3 seasons in the study are divided based
of the subgrade soils can be significantly influenced by on example of an expressway in the Harbin area of China.
in situ moisture content, and the resilient modulus of the On the basis of the 3 aforementioned seasons, according
pavement materials can be influenced by temperature. to the seasonal distribution coefficients of traffic volume
Thus, the pavement structure can be damaged given the (K1, K2, K3), the original traffic volume n is calculated as
same traffic axle loads because of seasonal variability, or n ¼ n1 þ n2 þ n3 , or the design traffic volume N is calcu-
the volume of traffic axle loads can increase for the same lated as N ¼ N 1 þ N 2 þ N 3 . The formulas for converting
pavement structure. axle loads to traffic volume are shown in Formula (1)
[10–12].
Vehicle load
Pavement structure
Determination of K1 K2 K3
…
N
Therefore,
l
Xn Yn
n m N 0 ¼ K n1 N 1 þ K n2 N 2 þ K n3 N 3 ; ð3Þ
Training Therefore,
N 0 b
N0 P
¼ :
End N P
Initial Training
Table 2
Input data of coefficient and temperature and precipitation in different areas.
Month 1 2 3 4 5 6 7 8 9 10 11 12
Coefficient Suqian 0.40 0.60 1.04 1.11 1.14 1.04 1.10 1.18 1.23 1.05 1.04 1.07
Nanjing 0.82 0.81 0.99 1.05 0.97 0.96 0.99 1.01 1.18 1.08 1.09 1.05
Hangzhou 1.02 0.98 0.96 0.92 0.97 1.07 1.04 1.03 1.05 1.00 1.01 0.99
Xian 0.93 0.94 0.97 1.04 1.01 0.97 1.13 1.04 1.02 1.10 0.96 0.89
Max./°C Suqian 5 8 14 21 26 31 32 31 27 22 14 8
Nanjing 7 9 14 20 26 29 32 32 27 22 16 10
Hangzhou 8 10 14 21 25 29 33 33 28 23 17 11
Xian 5 8 14 21 26 32 32 31 25 19 12 6
Min ./°C Suqian -3 -2 3 10 15 20 24 23 18 11 4 -2
Nanjing -1 1 5 11 16 21 25 24 19 13 7 1
Hangzhou 1 3 6 12 17 21 25 25 21 15 9 3
Xian -4 -1 4 9 14 19 22 21 16 10 3 -3
Precipitation /mm Suqian 17 21 33 48 64 107 248 141 82 44 27 15
Nanjing 39 49 76 86 96 169 192 129 82 54 52 29
Hangzhou 72 85 125 127 157 211 147 148 150 78 61 51
Xian 7 10 26 45 58 58 101 77 95 62 26 7
586 H. Zhang et al. / International Journal of Pavement Research and Technology 11 (2018) 582–593
X m b Pm2 b Pm3 b
0 Pi K p2 P i
; N 03 ¼
K p3 P i
N ¼ C 1 C 2 K n ni i¼1 C 1 C 2 n2i Ps i¼1 C 1 C 2 n3i Ps
, and
P
N 01 ; N 02 ; N 03
i¼1 s
b b represents, separately, the traffic volume in 3
X m1
Pi Xm2
Pi seasons (Table 1).
0 0
N1 ¼ C 1 C 2 K n1 n1i ; N2 ¼ C 1 C 2 K n2 n2i ;
i¼1
P s i¼1
P s In summary, comparison of the 2 modified formulas
X m3 b yields the same results even though Kn and Kp are differ-
0 Pi
N3 ¼ C 1 C 2 K n3 n3i ently defined. The values of Kn and Kp can be determined
i¼1
P s based on BP neural network technology in subsequent
N 0 ¼ N 01 þ N 02 þ N 03 research.
ð5Þ
3. Determination of seasonal distribution and influencing
where coefficients of traffic volume
N 01 ; N 02 ; N 03 represent the modified design traffic volume
Pm b 3.1. Determination of seasonal distribution coefficients of
in 3 seasons, times/season. N 0 ¼ i¼1 C 1 C 2 K n ni PP si is the traffic volume
modified conversion formula of axle loads based on the
seasonal influencing coefficients of traffic volume (Kn1, 3.1.1. Principle and prediction method of BP neural network
Kn2, Kn3). N 0 is divided into 3 parts, namely, N 01 ¼ technology
Pm1 b Pm2 b Backpropagation (BP) neural network technology is
Pi
i¼1 C 1 C 2 K n1 n1i P s ; N 02 ¼ Pi
i¼1 C 1 C 2 K n2 n2i P s ; N 03 ¼
composed of large neural elements through their mutual
Pm3 b
connection. The general structure includes the input layer,
Pi
i¼1 C 1 C 2 K n3 n3i P s , and N 01 ; N 02 ; N 03 , separately, represent
implied layer, and output layer. It includes two parts, i.e.,
the traffic volumes in 3 seasons (Table 1). the input and output data system and the predicting sys-
tem. It can express the unknown function by the training
2.1.2. Modified formula based on the seasonal influencing network by using input and output data, and the network
coefficients of the axle load predicting system can be used as the output. Its structure is
The seasonal influencing coefficient of the axle load (Kp) shown in Fig. 3. The neural elements of the input layer,
is defined as hidden layer, and output layer are n, l, m; that is, the net-
P0 work structure is n-l-m. The calculation procedure for BP
P 0 ¼ K p P or K p ¼ : ð6Þ neural network technology is shown in Fig. 4 [13,14].
P
Where
Therefore,
1 X1, X2 . . . Xn are the input values of the neural network.
K p ¼ ðK n Þb ð7Þ Y1, Y2 . . . Ym are the output values (predicted value);
The formula for converting axle load to traffic volume, and
based on Kp, is modified as Formula (8): xij, xjk are the right values.
X m b
KpP i 3.1.2. Predicting results
N0 ¼ C 1 C 2 ni
i¼1
Ps
X m1
b Xm2 b (1) Input data
K p1 P i K p2 P i
N 01 ¼ C 1 C 2 n1i ; N 02 ¼ C 1 C 2 n2i ; On the basis of the factors influencing monthly distribu-
i¼1
Ps i¼1
Ps tion coefficients, the max and min temperature and precip-
X m3 b
K p3 P i itation in different areas are used as input data, as shown in
N 03 ¼ C 1 C 2 n3i Table 2.
i¼1
Ps
N 0 ¼ N 01 þ N 02 þ N 03 (2) Predicted results
ð8Þ
The corresponding training and predicted results
where through data training are shown in Fig. 5.
Kp1, Kp2, Kp3 are the seasonal influencing coefficients of The predicted results of the monthly distribution coeffi-
axle load in 3 seasons. cients of traffic volume in Harbin, China are shown in
P b
N 0 ¼ mi¼1 C 1 C 2 ni Pp s i is the modified conversion for-
K P Table 3. The predicted results in Table 3 are compared
and verified with the local WIM (weigh in motion) data
mula of axle loads on the basis of the seasonal influencing
at the expressway in Harbin (shown in Fig. 6). The com-
coefficients of axle load (Kp1, Kp2, Kp3). N 0 is divided
Pm1 b parative results show that the predicted results are
into 3 parts, namely, N 01 ¼ i¼1 C 1 C 2 n1i p1P s i ; N 02 ¼
K P
reasonable.
H. Zhang et al. / International Journal of Pavement Research and Technology 11 (2018) 582–593 587
Distribution coefficient
Distribution coefficient 1.1 1.05
1.0
1.00
0.9
0.95
0.8
0.90
0.7
0.6 0.85
0.5 0.80
0.4 0.75
0 5 10 15 20 25 30 35 40 0 2 4 6 8 10 12
Sample Month
1.6
Monthly distribution coefficient
1.4
1.2
0.8
0.6
0.4
0.2
0
1 2 3 4 5 6 7 8 9 10 11 12
Month
Predicting data WIM data 1 WIM data 2 WIM data 3
WIM data 4 WIM data 5 WIM data 6 WIM data 7
Fig. 6. Comparison between predicted results and local WIM data in Harbin.
588 H. Zhang et al. / International Journal of Pavement Research and Technology 11 (2018) 582–593
Table 4
Classification results for 3 different seasons (Harbin, China).
Favorable seasons Normal seasons Unfavorable seasons
Month 1 + 2(1st half) 2(2nd half) 7(Rainy season)+ 4(Spring thaw season)
+11(2nd half)+12 +3 + 5+6 + 9+10 + 11(1st half) 8(Rainy season)
Ave. temperature/°C 16.55 14.6 23 7.1
Max. temperature /°C 10.95 21 27.9 13.7
Min. temperature /°C 21.95 8.3 18.3 0.4
Precipitation /mm 4.6 49 142.7 18.4
Ave. sunshine/h 5.8 8.7 8.6 8.3
Table 5
Determination of seasonal distribution coefficients of traffic volume (Harbin, China).
Favorable seasons (K1) Normal seasons (K2) Unfavorable seasons (K3)
Coefficients (K) 0.2083 0.5196 0.2721
Table 6
Subgrade resilient modulus in 3 different seasons, based on BP neural networks (Harbin, China).
Favorable seasons Normal seasons Unfavorable seasons
Rainy season Spring thaw season
Resilient modulus/MPa 147.3 37.9 27.0 20.7
Table 7
Temperature conversion results of asphalt pavement in 3 different seasons (Harbin, China).
3 different seasons Max. temperature Min. temperature
Favorable seasons 0.7 20.8
Normal seasons 40.5 1.7
Unfavorable seasons Rainy season 49.5 9.6
Spring thaw season 31.3 4.4
Table 8
Resilient moduli of asphalt layers in 3 different seasons, as determined based on BP neural networks (Harbin, China).
Favorable seasons Normal seasons Unfavorable seasons
Rainy season Spring thaw season
Resilient modulus/MPa 11279.2 2040.8 1417.4 2954.0
Table 9
Seasonal influencing coefficients in traffic axle load under different ratios (Harbin, China).
Subgrade: asphalt layers Favorable seasons Normal seasons Unfavorable seasons
(%) (Kn1) (Kn2) (Kn3)
<50:50 Influencing coefficients of traffic volume 0.0019 1 5.6572
=50:50 (Kn) 0.0014 1 3.5946
>50:50 0.0009 1 2.1665
Subgrade: asphalt layers Favorable seasons Normal seasons Unfavorable seasons
(%) (Kp1) (Kp2) (Kp3)
<50:50 Influencing coefficients of axle load (Kp) 0.2381 1 1.4894
=50:50 0.2190 1 1.3420
>50:50 0.2000 1 1.1945
Note: Subgrade: asphalt layers(%) – The percentage of resilient modulus between subgrade and asphalt layers.
Table 10
Test data of asphalt pavement deflection in different seasons (Harbin, China).
Months Resilient modulus of subgrade/MPa Resilient modulus of asphalt layers/MPa Pavement deflection/0.01 mm
Favorable seasons 12 132.2 7139.0 2.39
1 151.2 7636.0 2.33
Normal seasons 5 40.0 1991.0 17.2
6 37.9 2020.8 17.4
6 41.6 2045.4 16.8
9 42.9 1593.0 18.1
9 35.5 1530.5 19.6
Unfavorable seasons 4 27.0 1417.4 22.4
7 20.7 2954.0 19.9
(2) Determination of the resilient moduli of asphalt lay- (3) Determination of seasonal influencing coefficients
ers in 3 different seasons
The resilient modulus of subgrade and asphalt layers
In accordance with Formula (10), the temperature con- can change with varying humidity and temperature in dif-
version results for the asphalt pavement in Harbin, China ferent seasons. Therefore, the resilient modulus of subgrade
are obtained, as shown in Table 7. and asphalt layers in different seasons is not a constant.
The seasonal influencing coefficients in axle load can be
T max ¼ T air þ 0:25T air þ 15:8
ð10Þ determined based on the different ratios of the resilient
T min ¼ T air 0:25T air 3:0 modulus of subgrade and asphalt layers.
On the basis of the relationships of the resilient modulus
where
of subgrade and asphalt layers, the seasonal influencing
coefficients in Harbin, China are determined, as shown in
Tmax – maximum temperature of asphalt surface, °C.
Table 9.
Tmin – minimum temperature of asphalt surface, °C.
Tair – air temperature, °C.
Lat – latitude (north latitude, 45 °C). 3.2.2. Determination of seasonal influencing coefficients in
axle load based on the relationships between pavement
The relationship between the resilient modulus of deflection and resilient modulus of asphalt layers
asphalt layers and pavement temperature, as determined The test data of asphalt pavement deflection in different
based on BP neural network technology, is established in seasons in Harbin, China are presented in Table 10.
Formula (11). On the basis of test data in Table 10, the regression rela-
tionship between the resilient modulus of asphalt layers
Et
¼ 3:1928 arctanð0:0779t þ 0:4395Þ þ 4:6090 ð11Þ and pavement deflection is established, as follows:
E20
lg l ¼ 1:9607 lg E þ 5:7465
Generally, the resilient modulus of asphalt layers at 20 °
C ranges between 1600 and 2000 MPa; 1991 MPa is chosen The resilient modulus of asphalt layers under favorable
as the resilient modulus of asphalt layers at 20 °C. The resi- seasons is 11279.2 MPa, and the calculation result of pave-
lient moduli of the asphalt layers in 3 different seasons are ment deflection under favorable seasons is 0.0236 mm. The
shown in Table 8. influence of seasonal variability on the coefficients, based
590 H. Zhang et al. / International Journal of Pavement Research and Technology 11 (2018) 582–593
Table 11
Seasonal influencing coefficients in axle load based on the relationship between pavement deflection and the resilient modulus of asphalt layers (Harbin,
China).
Favorable seasons Normal seasons Unfavorable seasons
Deflection /0.01 mm 2.36 17.4 19.9
Deflection coefficient ac 0.0017 0.0708 0.0961
Kp 0.2087 1 1.2607
Kn 0.0011 1 2.7378
According to specification of asphalt pavement design The daily traffic volume of asphalt pavement is 4463.0
(JTD D50-2006) in China, the theory of elastic layered sys- times/d, and the annual traffic volume of asphalt pavement
tem (M-E method) is used for asphalt pavement design. In is 1629003.8 times/d. Therefore, the modified ESAL based
Table 13
Conversion results of modified ESAL based on the influence of seasonal variability (BZZ-100 kN).
Favorable seasons Normal seasons Unfavorable seasons
Unmodified annual traffic volume N (times/y) 1629003.8
Seasonal distributing coefficients of traffic 0.2083 0.5196 0.2721
volume K
Seasonal traffic volume Ni (times/season) 339369.0 846450.7 443184.1
① <50:50 Kn 0.0019 1 5.6572
N0 644.8 846450.7 2507181.1
=50:50 Kn 0.0014 1 3.5946
N0 475.1 846450.7 1593069.6
>50:50 Kn 0.0009 1 2.1665
N0 305.5 846450.7 960158.3
② Kn 0.0011 1 2.7378
N0 373.3 846450.7 1213369.6
① ②
<50:50 =50:50 >50:50
Modified annual traffic volume N0 (times/y) 3354276.6 2439995.5 1806914.6 2060193.6
Coefficient of lanes 0.5
Based on ld and rR of asphalt layers Modified ESAL (107) (times) 4.275 3.110 2.303 2.626
Unmodified ESAL (107) (times) 2.098
Based on rR of base layers Modified ESAL (107) (times) 5.447 3.962 2.934 3.346
Unmodified ESAL (107) (times) 2.673
Increasing ratio (%) 103.77 48.24 9.77 25.17
Note: ①–Subgrade: asphalt layers (%).
②–Relationships between pavement deflection and resilient modulus of asphalt layers.
H. Zhang et al. / International Journal of Pavement Research and Technology 11 (2018) 582–593 591
Table 14
Comparative results of pavement design deflection, based on modified and unmodified ESAL.
Modified ESAL (107) (times) ① ②
<50:50 =50:50 >50:50
4.275 3.110 2.303 2.626
ld (0.01 mm) 17.9 19.0 20.2 19.7
Unmodified ESAL (107) (times) 2.098
ld (0.01 mm) 20.6
Note: ①, ②–Same with ① and ② in Table 13.
Table 15
Comparative results of pavement allowable tensile stress, based on modified and unmodified ESAL.
rm (MPa)
① ② ③ ④ ⑤
Modified ESAL (107) (times) ⑥ <50:50 4.275 0.39 0.32 0.26
=50:50 3.110 0.42 0.35 0.28
>50:50 2.303 0.45 0.37 0.30
<50:50 5.447 0.24 0.13
=50:50 3.962 0.25 0.13
>50:50 2.934 0.26 0.14
⑦ 2.626 0.44 0.36 0.30
3.346 0.25 0.13
Unmodified ESAL (107) (times) 2.098 0.46 0.38 0.31
2.673 0.26 0.14
Note: ①–Asphalt concrete-13 (AC-13); ②–Asphalt concrete-16 (AC-16); ③–Asphalt concrete-20 (AC-20); ④–Cement stabilized macadam; ⑤–Cement
and lime stabilized sand and gravel soil; ⑥–Subgrade: asphalt layers (%); ⑦–Relationships between pavement deflection and resilient modulus of asphalt
layers.
on the influence of seasonal variability is calculated using The comparative results of ld and rR, based on modified
Formula (5). The conversion results are shown in Table 13. and unmodified ESAL, are presented in Tables 14 and 15.
Table 13 shows the large variations with different ratios
between subgrade and asphalt layers; moreover, the influ- 4.4. Design of pavement structure
ence of unfavorable seasons is more evident than other 2
seasons. Compared with unmodified formulae, the ESALs According to the results of the aforementioned design
calculated by modified formulae considering seasons’ influ- traffic volume, the highway is designed as an expressway
ence have increased about 25%. In fact, the ESALs consid- with an asphalt pavement and 4 lanes; the servicing time
ering seasons’ influence should be bigger than that without is 15 years. The structure and corresponding design param-
consideration of seasons’ influence, therefore, the modified eters of asphalt pavement are shown in Fig. 7 and Table 16.
formulae better reflect reality and it is scientific.
4.5 Design of pavement thickness
4.3. Calculation of pavement design deflection (ld) and
allowable tensile stress (rR) The thickness of the pavement structure is calculated
using Formula (13). Design thickness is finally determined
The asphalt pavement design deflection (ld) and allow- by the greater value between the thickness in ls ld and
able tensile stress (rR) are calculated using Formula (12). thickness in rs rR.
0:2 2pd
ld ¼ 600ðESALÞ A c As Ab ls ¼ 1000 ac F
rs ð12Þ E1 ð13Þ
rR ¼ rm ¼ prm
Ks
where where
ld – asphalt pavement design deflection, 0.01 mm. ls – asphalt pavement calculating deflection, 0.01 mm.
Ac, As, Ab – coefficients of highway grade, surface types, E1 – resilient modulus of the asphalt surface, MPa.
and base types, respectively. p, d – coefficients of the standard axle load (BZZ-100
rR – asphalt pavement allowable tensile stress, MPa. kN).
rs – split strength of pavement materials, MPa; and ac – coefficient of pavement theoretical deflection.
Ks – strength coefficient of pavement materials. F – comprehensive coefficient of pavement deflection.
592 H. Zhang et al. / International Journal of Pavement Research and Technology 11 (2018) 582–593
Pm b
Asphalt concrete-13 (AC-13)
In unmodified formula N ¼ i¼1 C 1 C 2 ni
Pi
Ps
; there is
Asphalt concrete-16 ((AC-16)
no consideration about the seasonal influence to axle loads,
Asphalt concrete-20 (AC-20) but actually the axle loads on pavement will be influenced
by different seasons. In modified formula
P b
Cement stabilized macadam N 0 ¼ i¼1 C 1 C 2 K n ni PP si ; the traffic volumes ni in three
m
Table 17
Comparative results of pavement thickness design, based on modified and unmodified ESAL.
Modified ESAL (107) (times) Unmodified ESAL (107) (times)
⑥ ⑦
<50:50 =50:50 >50:50
Based on ld and rR of asphalt layers 4.275 3.110 2.303 2.626 2.098
Based onrR of base layers 5.447 3.962 2.934 3.346 2.673
ld (0.01 mm) 17.9 19.0 20.2 19.7 20.6
Thickness of design layer (ls ld) (cm) 29.4 26.0 22.6 24.0 21.5
rm (MPa) ① 0.39 0.42 0.45 0.44 0.46
② 0.32 0.35 0.37 0.36 0.38
③ 0.26 0.28 0.30 0.30 0.31
④ 0.24 0.25 0.26 0.25 0.26
⑤ 0.13 0.13 0.14 0.13 0.14
Thickness of design layer (rs rR) (cm) 19.9 19.8 16.8 16.8 16.8
Thickness of design layer (cm) 30.0 26.0 23.0 24.0 22.0
Increasing thickness (cm) 8 4 1 2
Increasing ratio (%) 36.4 18.2 4.6 9.1
Note: ①, ②, ③, ④, ⑤, ⑥, ⑦– Same with ①, ②, ③, ④, ⑤, ⑥ and ⑦ in Table 15.
H. Zhang et al. / International Journal of Pavement Research and Technology 11 (2018) 582–593 593
(3) On the basis of modified formulas and seasonal distri- [4] Yanqing Zhao, Yiqiu Tan, Changhong Zhou, Determination of axle
bution coefficients of traffic volume, the seasonal load spectra based on percentage of overloaded trucks for mechanis-
tic-empirical pavement design, Road Materials and Pavement Design
influencing coefficients of axle loads in Harbin, China 13 (4) (2012) 850–863.
are determined based on the strong relationship [5] Józef Judycki, Determination of equivalent axle load factors on the
between subgrade and asphalt layers. Determination basis of fatigue criteria for flexible and semi-rigid pavements, Road
of the coefficients renders the modified formulas Mater. Pavement Design 11 (1) (2010) 187–202.
applicable. [6] Sue Ahn, Srivatsav Kandala, J. Uzan, Mohamed El-Basyouny,
Impact of traffic data on the pavement distress predictions using the
(4) Using the example of asphalt pavement design, com- mechanistic empirical pavement design guide, Road Mater. Pavement
bined with the expressway in Harbin, China, the com- Des. 12 (1) (2011) 195–216.
parative design results between the modified and [7] Hongxin Guan, Hao Liu, Qisen Zhang, Xu Yang, Equivalent
unmodified formulas are analyzed. The design results conversion method of asphalt pavement axle-load skid-resistance
further demonstrate that the modified formulas are performance, J. Highway Transport. Res. Dev. 6 (4) (2012) 18–22
(English Edition).
scientific and rational. [8] Chenming Kuo, Shenhsian Lin, An analytical study of axle load
(5) The study indicates that the design results obtained equivalency factors of concrete pavements, J. Chin. Inst. Eng. 24 (2)
using the modified axle load conversion formulas (2001) 119–130.
increase the thickness and general strength of the [9] F. Homsi, D. Bodin, S. Yotte, D. Breysse, J.M. Balay, Fatigue life
asphalt pavement structure. It can more efficiently modelling of asphalt pavements under multiple-axle loadings, Road
Mater. Pavement Des. 13 (4) (2012) 749–768.
satisfy the requirements of an actual vehicle running [10] Ministry of Communication, Specification of asphalt pavement
on the pavement. In addition, improved economic design (JTD D50–2006), China Communication Press, Beijing, China,
and social benefits will be further discussed at next 2006.
research. [11] Zhongda Chen, Wu Jianmin, Mingshuo Feng, Jingliang Dai, Season-
correcting coefficient of axle load conversion for asphalt pavement, J.
Chang’an Univ. 23 (6) (2003) 1–3.
[12] Xiaozhou Han, Wu Liang, Pu Sun, Research on season-correcting
Acknowledgments coefficient in axle load conversion of asphalt pavement in season-
frozen areas, Shanxi Architect. 37 (23) (2011) 149–151.
The authors gratefully appreciate the supports from the [13] Yu Tengjiang, Determination of Seasonal Influencing Coefficients in
province key laboratory of road in Northeast Forestry Axle Load Conversion Formula of Asphalt Pavement (Dissertation),
Northeast Forestry University, Harbin, China, 2015.
University and the foundations for the project of National [14] Ministry of Communication, Standard test methods of bitumen and
Natural Science Foundation of China (E080703) and the bituminous mixtures for highway engineering (JTG E20–2011), China
project of Heilongjiang Traffic and Transportation Communication Press, Beijing, China, 2011.
Department. [15] Xudong Zha, Temperature correction of modulus back-calculation in
asphalt pavement, Highway 6 (2006) 51–53.
[16] Xin Qiu, Qing Yang, Qinglong You, Back analysis on modulus of
References asphalt pavements based on FWD dynamic deflection basin param-
eter, J. Chang’an Univ. 33 (6) (2013) 42–48.
[1] Abubeker W. Ahmed, Sigurdur Erlingsson, Characterisation of heavy [17] Zhihong Xu, Mengyu Yang, Report for Study on Design Parameter
traffic axle load spectra for mechanistic-empirical pavement design of Pavement Materials. Institute of Road and Transportation, Tongji
applications, Int. J. Pavement Eng. 16 (6) (2015) 488–501. University, Shanghai, China, 1994.
[2] Tim Charles Martin, Experimental estimation of the relative deteri- [18] Yuanxun Zheng, Study on Temperature Correction of Dynamic
oration of flexible pavements under increased axle loads, Int. J. Deflection And Modulus Back-Calculation In Asphalt Pavement
Pavement Eng. 12 (1) (2011) 37–45. (Dissertation), Dalian Institute of Technology, Dalian, China, 2009.
[3] Sara I.R. Amorim, Jorge C. Pais, Aline C. Vale, Manuel J.C.
Minhoto, A model for equivalent axle load factors, Int. J. Pavement
Eng. 16 (10) (2015) 881–893.
Available online at www.sciencedirect.com
ScienceDirect
International Journal of Pavement Research and Technology 11 (2018) 427–434
www.elsevier.com/locate/IJPRT
Received 25 November 2016; received in revised form 1 September 2017; accepted 18 November 2017
Available online 23 November 2017
Abstract
The infrastructure system has experienced an unprecedented development in the last fifty years. Most of the road pavements have
been used for over twenty years and today require increasing maintenance to keep acceptable their levels of performance. In addition,
the use of virgin and raw materials has been strongly restricted by raising price and eco-friendly policies. As a result, researchers and
companies have focused their attention on recycling techniques with waste and second-hand materials that could lead to the construction
and restoration of civil infrastructures without negatively affecting their mechanical properties. The study presented in this paper pursues
this trend of research, by the evaluation of the performance of two different 100% Recycled Cement Bound Mixture (RCBM) for base
layers. Data were always compared to those obtained for a traditional Cement Bound Mixtures (CBM) entirely made of virgin aggre-
gates. The experimental programme was divided into two different and consecutive phases: a laboratory study and a trial field phase. In
the first, the aim was the mix design and the physical and mechanical characterization of three mixtures. In the second phase, a full-scale
trial field was realized: in situ tests and sampling were planned in four dates, corresponding to 0, 60, 180 and 365 days of trafficking.
Based on results, the replacement of virgin aggregates with waste materials in CBMs, according to these specific mix designs seems to
be a practicable solution for the construction of base layers with suitable performance.
Ó 2017 Chinese Society of Pavement Engineering. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND
license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
https://doi.org/10.1016/j.ijprt.2017.11.005
1996-6814/Ó 2017 Chinese Society of Pavement Engineering. Production and hosting by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
428 P. Tataranni et al. / International Journal of Pavement Research and Technology 11 (2018) 427–434
Pavement, RAP), crushed concrete, tiles, ceramic, bricks 2.1. Laboratory phase
and wood [10–12]. Recycling of these waste materials is
becoming more and more important as a means to improve Once the right amount of water to reach the maximum
the efficient and conscious usage of resources and to avoid dry density (EN 13286-2) was defined, 24 specimens for
the negative impacts connected to the use of raw materials. each mixture were prepared for the static mechanical char-
The target proposed by EU for 2020, is to increase the recy- acterization. Twelve of them were compacted according to
cling of C&D waste to 70% (Directive 2008/98/EC). In the AASHTO Mod Proctor compaction (85 blows, 5 layers)
last decade, a wide number of researches have been carried and 12 by the means of gyratory compaction (ASTM
out on the use of C&D waste, and many studies clearly D6925), according to a constant compaction pressure of
show the feasible reuse of these discarded materials within 600 kPa, 180 gyrations and an external angle of 1.25°.
bituminous or hydraulically bound mixtures for road pave- The specimens were kept under constant moisture and tem-
ments [13–15]. If we consider base and subbase layers, dif- perature curing conditions.
ferent studies have shown the good results achieved by the Tests for the determination of the Unconfined Compres-
partial substitution of virgin aggregates with RAP within sive Strength (EN 13286-41) and the Indirect Tensile
bound and unbound mixtures [16,17]. The technological Strength (EN 13286-42) were carried out after 3 and 7 days
evolution and the growing interest of researcher in recy- of curing. To provide statistical significance to the results, 3
cling have gradually allowed to increase the quantities of specimens for each mixture realized with both the com-
waste materials combined within recycled mixtures, with- paction processes were tested. The mechanical characteri-
out detrimental effects on the mechanical properties. Fol- zation was supported by the analysis of the Stiffness
lowing this trend, the present research was focused on the Modulus according to EN 12697-26 standard.
analysis of two different fully Recycled Cement Bound
Mixtures (RCBMs) for base layers. The primary goals of 2.2. Trial field phase
this study were:
On the basis of the correct mix designs and once
The analysis of the mechanical and physical properties defined the basic physical and mechanical properties of
conferred to the mixture by the total substitution of vir- the experimental recycled mixtures, these were laid for
gin aggregate with C&D wastes. the construction of a full-scale test truck. In order to
The evaluation of the feasibility of the entire process, assess the development of the mechanical properties of
from in plant production to the paving operations. the mixtures, in situ tests and collection of samples were
To achieve both targets, the experimental programme planned in four dates, corresponding to 0, 60, 180 and
was divided into a laboratory and a trial field phase, in 365 days of cumulative traffic. For each mixture, three
which the properties of the two experimental RCBMs were samples were prepared by the means of the gyratory com-
compared to a traditional Cement Bound Mixture (CBM) paction (ASTM D6925), with the material collected from
entirely made with virgin aggregates. the paving screed. The gyratory setting and the number of
revolution were the same used during the laboratory
2. Experimental programme, test methods and materials phase. The indirect tensile strength was also evaluated
(EN 13286-42). The samples cored and collected from
The research programme was divided into two: the trial field were tested according to EN 12697-26 to
evaluate the development of the mechanical properties
Laboratory: in this phase the optimal mix design was under traffic load. The change and development in stiff-
defined and the physical and mechanical characteristics ness was also assessed using a Light Weight Deflectometer
of the mixtures were evaluated. Two RCBM mixtures (LWD) (ASTM E2583).
were analysed, both made with more than 40% of
RAP and other waste aggregates: one (RCBMt) with 2.3. Materials
tiles waste, the second (RCBMc) with crushed concrete
waste. Their mechanical properties were compared with The experimental work here presented, involved the mix
those obtained from a fully virgin CBM, labelled CBMv. designs and the laboratory and in situ characterization of
Trial field: according to the experimental mix designs three different mixtures: a traditional CBM and two differ-
defined after the mechanical characterization, the three ent RCBMs.
mixtures were laid in a full-scale trial field. In this step While the traditional CBM (labelled CBMv) was made
all the processes, from the in plant production to the with virgin aggregate, the other two experimental mixtures
paving operations were assessed and in situ tests and were entirely produced with C&D wastes: one, named
collection of samples were planned at regular intervals. RCBMt designed with RAP and discarded tiles, the other,
The objective was the analysis of the development of named RCBMc with RAP and crushed concrete. For both
the mechanical properties of the mixtures under traffic CBMv and RCBMt a traditional Portland cement 32.5 was
loads. used as binder, while RCBMc was designed with an
P. Tataranni et al. / International Journal of Pavement Research and Technology 11 (2018) 427–434 429
Fig. 4. RCBMt, RCBMc and CBMv ITS average results after 3 and 7 days of curing (gyratory compaction).
Fig. 6. RCBMt, RCBMc and CBMv ITSM Moduli and average results
Table 2
after 7 days of curing.
ITS average results after 7 days of curing.
Mixture ITS lab (MPa) ITS t.f. (MPa) Variation (%)
specified in the most common Italian technical specifica- RCBMt 0.58 0.72 +24
tions for CBM mixtures (2.152 g/cm3 for RCBMt, 2.159 RCBMc 0.54 0.61 +13
CBMv 0.62 0.75 +21
g/cm3 for RCBMc and 2.165 g/cm3 CBMv). ITS tests were
432 P. Tataranni et al. / International Journal of Pavement Research and Technology 11 (2018) 427–434
Fig. 7. ITSM test average results at 60, 180 and 365 days of cumulative
traffic.
Table 3
ITSM test results for laboratory and trial field phase.
Mixture Indirect tensile stiffness modulus (MPa) at 20 °C
Lab analysis 60 days 180 days 365 days
RCBMt_1 7286 4695 6357 8952
RCBMt_2 6658 4284 6099 8152
RCBMt_3 6598 4476 5995 9192
RCBMt_4 7089
Avg. RCBMt 6908 4485 6150 8765
RCBMc_1 12,596 6218 8807 13,007
RCBMc_2 11,598 6691 8450 12,100
RCBMc_3 11,897 5985 8669 12,509
RCBMc_4 12,009
Avg. RCBMc 12,025 6298 8642 12,539
CBMv_1 10,895 7839 8018 10,210
CBMv_2 9663 7323 8299 10,089
Fig. 9. ELWD average results after 0, 60, 180 and 365 days of cumulative
CBMv_3 10,586 7596 8457 9235
traffic (mainly quarry dumpers).
CBMv_4 9012
Avg. CBMv 10,039 7586 8258 9845
f ð1 m2 Þr0 a
ELWD ¼ ð1Þ
d0
stiffer mixture (12,539 MPa) while RCBMt has the lower
values (8765 MPa). These data validate and prove the
where f is the plate rigidity factor, a is the radius of the
ITSM results on specimens prepared in laboratory. Never-
plate, r0 is the maximum value of the applied stress and d0
theless, even if the experimental mixtures are made with
is the measured deflection. Both values are measured under
100% recycled aggregates, values are in line with those
the centre of the plate. The LWD setting for the measure-
related to traditional CBM. Furthermore, considering the
ment was 150 mm (plate diameter) and 20 kg (falling mass).
heavy loads due to dumpers transit and since the road sec-
As recommended by literature, it was possible to obtain
tion was not designed for these traffic loads, results show
under the centre of the plate a measured deflection always
that there is no evidence of decay in the mechanical prop-
higher than 100 mm according to this configuration. LWD
erties of CBMs.
tests (ASTM E2583) were repeated after 0, 60, 180 and 365
days to assess the development of ELWD. Nine test points
4.2. Light Weight Deflectometer (LWD) analysis were fixed and tested for each section and for each repeti-
tion (Fig. 8).
The Light Weight Deflectometer (LWD) is a hand por- The measurement of the surface moduli is influenced by
table device, whose operating principle is a circular plate the wearing course and foundation stiffness. Nevertheless,
loaded by a falling mass. The surface modulus can be cal- given the LWD measuring depth and the layer thicknesses,
culated according to the Boussinesq theory, as a function authors consider that the results are reliable and represen-
of the applied stress and measured deflection, on the basis tative of the base stiffness. This is mainly because initial
of the Eq. (1): in situ test recorded a uniform foundation stiffness along
P. Tataranni et al. / International Journal of Pavement Research and Technology 11 (2018) 427–434 433
the whole trial field, and constant are the mechanical and ity and compactability characteristics of the experimen-
physical properties of the HMA surface layer. tal mixtures.
The development of ELWD average values is shown in The development of the stiffness moduli for RCBMc and
Fig. 9. RCBMt follows the same trend of the reference mixture.
In absolute terms, taking as reference literature ELWD val- However, the recycled mixture with crushed concrete
ues on cement treated mixtures, both experimental mixes are waste presents the highest stiffness even if compared to
able to provide sufficient stiffness properties to the road the CBMv. The high presence of RAP and C&D wastes
pavement. It is also worth noting that LWD tests have been does not determine a reduction in stiffness after 365 days
carried out in different climatic conditions: days 0 and 365 of traffic.
correspond to summer climate, day 60 to autumn, while tests LWD tests confirm the high performance of RCBMc as
at day 180 were performed in relatively cold weather. Thus, verified during the dynamic characterization on core
the surface bituminous layer’s temperature must be consid- samples. RCBMt presents lower stiffness that is compa-
ered for a correct analysis of the results. Stiffness moduli reg- rable to the virgin mixture.
istered at days 60 and 180 are generally higher because of the
lower temperature of the wearing course. However, from the According to the results here presented, the similarity of
comparison between the values measured at days 0 and 365, data obtained in the laboratory characterization and in the
having very similar temperature conditions, ELWD results test track phase is evident. The high mechanical properties
confirm the data obtained from ITSM tests on cores. achieved, confirm the quality of the experimental mix
RCBMc appears to be the stiffest mixture, while RCBMt designs and the validity of in plant production and paving
presents lower values but comparable to the reference mix- processes. However, the performances of the mixtures are
ture. Moreover, as for ITSM results, data show an increase strictly related to the type and quality of recycled materials
in stiffness from day 0 to day 365. Finally, the order of mod- used.
uli in the two days is similar for both testing methods.
References
5. Conclusions
[1] I.Z. Bribian, A.V. Capilla, A.A. Uson, ‘‘Life cycle assessment of
The objective of the present research was the laboratory building materials: comparative analysis of energy and environmental
impacts and evaluation of the eco efficiency improvement potential,
and in situ evaluation of the physical and mechanical prop- Build. Environ. 46 (2010) 1133–1140.
erties of 100% Recycled Cement Bound Mixtures for base [2] C. Sangiorgi, P. Tataranni, A. Simone, V. Vignali, C. Lantieri, G.
layers. Two experimental mixtures were analysed both con- Dondi, Waste bleaching clays as fillers in hot bituminous mixtures,
taining RAP and each a specific C&D waste: tiles waste or Constr. Build. Mater. 73 (2014) 320–325.
crushed concrete waste. The experimental programme con- [3] Dondi, G., Sangiorgi, C., Lantieri, C., Simone, A., Vignali, V. and
Lamperti, R. (2014). ‘‘Performance evaluation of Construction and
sidered a laboratory and a test track phase: test results have Demolition, and other waste materials”, in: 3rd International
been compared with those registered for a traditional CBM Conference on Transportation Infrastructure, ICTI 2014, Pisa, Italy.
and with common Italian technical specifications. [4] B. Gomex-Meijide, I. Perez, A.R. Pasandı̀n, Recycled construction
According to the presented experimental results, the fol- and demolition waste in Cold Asphalt Mixtures: evolutionary
properties, J. Cleaner Prod. 112 (2016) 588–598.
lowing conclusion can be deduced:
[5] C. Sangiorgi, P. Tataranni, A. Simone, V. Vignali, C. Lantieri, G.
Dondi, Assessment of waste bleaching clay as alternative filler for the
ITS and UCS after 3 and 7 days of curing show suitable production of porous asphalts, Constr. Build. Mater. 109 (2016) 1–7.
mechanical properties for both the experimental mix- [6] M. Bocci, F. Canestrari, A. Grilli, E. Pasquini, D. Lioi, Recycling
tures, even if RCBMt does not reach ITS limit after 3 techniques and environmental issues relating to the widening of an
days of curing. Results are lower but comparable to high traffic volume Italian motorway, Int. J. Pavement Res. Technol.
3 (2010) 171–177.
the reference CBM and in accordance with the limits [7] G. Dondi, P. Tataranni, M. Pettinari, C. Sangiorgi, A. Simone, V.
imposed by the technical specifications taken as refer- Vignali, Crumb rubber in cold recycled mixes: comparison between
ence. As for ITSM test, both recycled mixtures present traditional crumb rubber and cryogenic crumb rubber, Constr. Build.
suitable performance and RCBMc shows the highest Mater. 68 (2014) 370–375.
stiffness even if compared to the virgin mixture. This [8] M. Dinis-Almeida, J. Castro-Gomes, C. Sangiorgi, S.E. Zoorob, M.
L. Afonso, Performance of warm mix recycled asphalt containing up
may be related to the type of stabilizing agent. As indi- to 100% RAP, Constr. Build. Mater. 112 (2016) 1–6.
rect result of the mechanical characterization, the pres- [9] Mazzotta, F., Sangiorgi, C., Vignali, V., Lantieri, C. & Dondi, G.
ence of RAP and C&D wastes do not seem to prevent (2015), Rheological characterization of bituminous mastics added
the compactability properties and the achievement of with waste bleaching clays, in: Proceedings of 8th RILEM Sympo-
the higher layer densities. sium on testing and characterization of sustainable and innovative
bituminous materials, October 7-9, Ancona, Italy.
From ITS test on specimens prepared with the material [10] C.T. Chiu, T.H. Hsu, W.F. Yang, Life cycle assessment on using
taken from the paving screed, the good mechanical recycled materials for rehabilitating asphalt pavements, Resource.
properties evaluated during the laboratory phase are Conserv. Recycl. 52 (2008) 545–556.
confirmed. The in situ tests carried out after the laying [11] G. Dondi, C. Sangiorgi, C. Lantieri, Applying geostatistics to
(day 0), validate the mix designs and the good workabil- continuous compaction control of construction and demolition
434 P. Tataranni et al. / International Journal of Pavement Research and Technology 11 (2018) 427–434
materials for road embankments, J. Geotech. Geoenviron. Eng. 140 [15] S. Fatemi, R. Imaninasab, Performance evaluation of recycled
(3) (2014). asphalt mixtures by construction and demolition waste materials,
[12] C. Sangiorgi, C. Lantieri, G. Dondi, Construction and demolition Constr. Build. Mater. 120 (2016) 450–456.
waste recycling: an application for road construction, Int. J. Pave- [16] D. Deniz, E. Tutumler, J.S. Popovics, Evaluation of expansive
ment Eng. 16 (2015) 530–537. characteristics of reclaimed asphalt pavement and virgin aggregate
[13] Dondi, G., Mazzotta, F., Sangiorgi, C., Pettinari, M., Simone, A., used as base materials, Transport. Res. Rec. J. Transport. Res.
Vignali, V. and Tataranni, P. (2014), ‘‘Influence of cement and Board, 2167 (2010), Transportation Research Board, Washington, D.
limestone filler on the rheological properties of mastic in cold C., U.S.A..
bituminous recycled mixtures”, in: Proceedings of 3rd International [17] A. Arulrajah, M.M. Disfani, S. Horpibulsuk, C. Suksiripattanapong,
Conference on Transportation Infrastructures - ICTI 2014, April 22- N. Prongmanee, Physical properties and shear strength responses of
25, Pisa, Italy. recycled construction and demolition materials in unbound pavement
[14] D.X. Xuan, A.A.A. Molenaar, L.J.M. Houben, Evaluation of cement base/subbase applications, Constr. Build. Mater. 25 (2014) 245–257.
treatment of reclaimed construction and demolition waste as road
bases, J. Cleaner Prod. 100 (2015) 77–83.
International Journal of Pavement Engineering
To cite this article: Limon Barua, Bo Zou, Mohamadhossein Noruzoliaee & Sybil Derrible (2020):
A gradient boosting approach to understanding airport runway and taxiway pavement deterioration,
International Journal of Pavement Engineering, DOI: 10.1080/10298436.2020.1714616
Rubinfeld 1998). Identifying the (most) correct model specifi- relationship between IRI and pavement distress with the pur-
cation can be non-trivial and time consuming. Furthermore, pose to evaluate the possibility of using IRI as a critical rep-
when included input variables are highly correlated, as often resentation index for pavement performance.
is in practice (e.g. variation of temperature and the number Apart from BPNN, other extensions of ANN were proposed
of freeze–thaw cycles), the estimated coefficients for those vari- in pavement condition studies. Chou et al. (1994) applied a
ables can become insignificant, which compromises the ability fuzzy enhancement algorithm and ANN to classify pavement
of the model to infer the input variable effects on pavement cracks of flexible highway pavements based on image recog-
condition. nition. Shekharan (2000) developed a model using genetic
adaptive ANN training to forecast PCR for flexible, rigid, and
composite highway pavements. The input variables were pave-
Machine learning in pavement condition study
ment structure, age, traffic volume, and classification according
In contrast to statistical modelling, machine learning does not to the federal aid system. More recently, Bianchini and Bandini
entail assumptions on the functional relationship, data, or error (2010) designed a model using neuro-fuzzy reasoning to pre-
terms in model development. Model development is data dri- dict changes in Present Serviceability Index (PSI) based on
ven and thus more capable of uncovering the underlying data pavement surface curvature index, property of subgrade, rut
and model structure even when the structure is not apparent depth, cracking area, asphalt layer thickness, and incremental
(Vlahogianni et al. 2014). Compared to statistical modelling, traffic for flexible roadway pavements. Recurrent ANN was
machine learning can efficiently handle datasets of large sizes combined with Support Vector Classifier (SVC) in Tabatabaee
with any degree of complexity (Zhang and Haghani 2015, et al. (2013) to predict pavement performance. In this study,
Gorripaty et al. 2017, Lee et al. 2018, Barua et al. 2019a). SVC was first used to classify road sections. The classification
Among the machine learning methods used for pavement results then fed into the ANN to predict PSI. Shahnazari
conditions study, Artificial Neural Networks (ANN) is most et al. (2012) compared ANN and genetic programming in
popular method. Schwartz (1993) was probably the first one PCI prediction using various distress types, severity, and quan-
using ANN to model airfield PCI with pavement age, traffic, tity to forecast highway PCI. The authors found that ANN was
and subgrade strength as input variables. George and She- able to yield more accurate prediction than genetic program-
kharan (1998) developed an ANN model to characterise Pave- ming. Table 1 provides further information about the above-
ment Condition Rating (PCR), which is the combination of reviewed studies.
distress and roughness condition rating of flexible, rigid, and
composite highway pavements. Van der Gryp et al. (1998) eval-
Choice of gradient boosting machine
uated the capabilities of feed-forward ANN in determining
Visual Condition Index using various pavement distress types The literature review in the previous subsection clearly shows
and severities. Kargah-Ostadi et al. (2010) employed ANN pat- that ANN and its extensions are the dominant machine learn-
tern recognition to predict International Roughness Index (IRI) ing methods for modelling pavement conditions. Several other
of flexible highway pavements based on initial pavement machine learning methods also exist but their performance
roughness, pavement age, traffic load, climate condition, struc- remains largely underexplored. Among these methods, ensem-
tural property, subgrade properties, drainage type and con- ble-based methods, which build a bucket of relatively simple
dition, and M&R history. models (also termed “base learners”) to obtain a strong ensem-
As a variant of ANN, Backpropagation Neural Networks ble prediction, have reached a celebrity status in solving predic-
(BPNN) has been widely employed in pavement condition tion problems with great success such as winning the $1 million
forecast. Attoh-Okine (1994) developed a BPNN-based pave- Netflix Prize competition (Koren 2009, Zhou 2012). Decision
ment roughness progression model for flexible highway pave- tree-based ensemble methods that strategically combine mul-
ments using synthetically generated roughness. Pavement tiple single tree models instead of fitting the best single decision
structural deformation, incremental traffic loadings, the extent tree model are especially popular. By drawing insights and
of cracking and the thickness of surface layer, the incremental methods from both statistical and machine learning methods,
variation of rut depth, surface defects, and environmental and decision tree-based ensemble methods not only achieve
road age were used as input variables to determine pavement superior predictive performance than ANN models in some
roughness. Another BPNN model was developed by Eldin contexts (Ma et al. 2017, Ding et al. 2018, Lee and Derrible
and Senouci (1995) to determine condition rating considering 2019), but they can also identify and interpret the influence
15 distress types of rigid pavements. BPNN model was con- of individual input variables and their interactive effects,
sidered in Owusu-Ababio (1998) to model pavement distress which is not necessarily possible with ANN (Elith et al. 2008,
index of flexible highway pavements with pavement surface Zhang and Haghani 2015).
thickness, age, traffic level, base thickness, and roadway con- There are two types of decision tree-based ensemble
dition as input variables. Lou et al. (1999) used BPNN with methods: bagging and boosting. The most common bagging
crack index time series, pavement types, and age as input vari- method is Random Forest (RF) and the most common boosting
ables to model pavement crack condition. The work was method is Gradient Boosting Machine (GBM). RF relies on
extended by Yang (2004) who added ride and rut index as out- finding a consensus across multiple imperfect trees as opposed
puts. BPNN was also used to forecast serviceability rating for to using one single perfect tree which is prone to overfitting
flexible highway pavements (Choi et al. 2004). Finally, BPNN (Farid et al. 2014). Each tree is trained individually with a sub-
models were employed by Lin et al. (2003) to analyse the set of the training data. The final prediction comes from the
INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 3
Table 1. A summary of the machine learning models used for pavement condition prediction.
Method Response
category Authors (year) variable Input variables Pavement type Method used
ANN Schwartz (1993) Distress density Pavement age, traffic, and subgrade strength Flexible asphalt ANN
concrete airfield
pavements
George and PCR Different distress types Flexible, rigid, and ANN
Shekharan composite road
(1998) pavements
Van der Gryp Visual Condition Various distress types and severities Flexible road Feed-forward ANN
et al. (1998) Index pavements
Kargah-Ostadi IRI Initial roughness, pavement age, traffic volume, climatic Flexible pavements ANN pattern
et al. (2010) condition, structural property, subgrade properties, recognition
drainage type and conditions, maintenance and
rehabilitation treatment
BPNN Attoh-Okine Roughness Structural deformation, incremental traffic loadings, Flexible highway BPNN
(1994) progression cracking and thickness of surface layer, rut depth, pavements
surface defects, environmental variables, road age
Eldin and PCI 15 distress types Rigid road pavements BPNN
Senouci (1995)
Owusu-Ababio Pavement Surface thickness, pavement age, traffic level, base Flexible highway BPNN
(1998) Distress Index thickness, roadbed condition pavements
Lou et al. (1999) Pavement Crack Crack index time series, pavement type, and age Flexible and rigid BPNN
Index highway pavements
Yang (2004) PCR Crack index time series, pavement type, pavement age, Flexible and rigid BPNN
pavement cycle highway pavements
Lin et al. (2003) IRI Different distress types Highway and county BPNN
road pavements
Choi et al. (2004) IRI Stiffness, asphalt layer thickness, temperature, material Asphalt concrete road BPNN
types, air void, viscosity, traffic load, pavement age pavements
Other Chou et al. (1994) Classification of Image recognition Highway pavements Fuzzy enhancement
extensions of cracks algorithm and ANN
ANN Shekharan (2000) PCR Pavement structure, pavement age, traffic volume, and Flexible, composite, Genetic adoptive ANN
classification and rigid highway
pavements
Bianchini and Changes in PSI Surface curvature index, property of subgrade, rut depth, Flexible road Neuro-fuzzy
Bandini (2010) area of cracking, asphalt layer thickness, incremental pavements reasoning
traffic
Tabatabaee et al. PSI Pavement age, traffic load, season, maintenance, Highway pavements Combination of SVC
(2013) treatment history, current treatment, previous PSI and BPNN
Shahnazari et al. PCI Various types of distresses Road pavements ANN
(2012)
average of all individual tree-based predictions. In contrast, eliminating the need to retrain the entire model from scratch
GBM is based on the idea of adding new trees to the ensemble again. This saves computation time making GBM attractive
sequentially (Natekin and Knoll 2013). At each iteration, GBM for online applications.
accounts for the error of the previously ensembled trees and . GBM can capture and visualise the interactive effects among
tries to recover the error while preforming prediction in the input variables, e.g. the fact that a change in the response
next tree. As such, the error keeps decreasing in the subsequent variable with respect to a change in an input variable may
tree ensemble. Because of this difference, GBM often leads to depend on the values of other input variables.
greater prediction accuracy than RF (Hastie et al. 2009, Barua
et al. 2019b). The data size required by GBM can also be In light of the above discussions, we choose GBM as the pre-
much smaller than RF. Indeed, previous research found that ferred model to investigate the influence of individual input
GBM can handle complex datasets yielding smaller prediction variables and their interactive effects. The hypothesis of the
errors than RF (Freund and Schapire 1996). GBM has several study is that GBM can yield better performance than other
additional advantages (Friedman 2001, Regue and Recker machine learning methods in fitting airport pavement con-
2014, Ding et al. 2018), which are listed below: ditions and explaining the influence of input variables.
It is worth noting that the GBM approach and the genetic
. GBM works well with high-dimensional mixed types of algorithm (GA) bear some similarity in that both are iterative
numerical and categorical input variables. The performance and rely on multiple elements: in the GBM, the elements are
of GBM is invariant to transformations of the input variables decision trees; in GA, the elements are solutions. However,
and insensitive to outliers. GBM is also fundamentally different from GA. As mentioned
. The selection of input variables in GBM is internalised in the earlier, GBM produces a fitting model based on adding new
decision tree, making the modelling robust to irrelevant decision trees iteratively to the existing ensemble of decision
input variables. Imputing missing values is not needed trees. At each iteration, adding a new decision tree is to recover
since the decision tree can automatically handle them. the error of the previously ensembled trees. GA does not
. When new data become available, the boosting procedure in involve a procedure that is similar to ensembling. Instead, in
GBM can continue from the previous trained model, thereby searching for a good solution GA creates a sequence of solution
4 L. BARUA ET AL.
populations. Each solution population consists of elite, cross- Following the generally accepted rule (Ng 2018), a 60-20-20
over, and mutation child solutions that are generated based split of the data is adopted. Figure 1 shows the procedure for
on the previous (parent) solution population. Once the child the GBM model development.
solutions are created, the parent solutions will be removed In what follows, we detail the three steps. For the model
from consideration, which is different from GBM in which a training step, the description draws heavily from Friedman
decision tree, once created, will always be used throughout (2001) and Natekin and Knoll (2013). Thus interested readers
the model development process. may find further details in the two references.
The remainder of the paper is organised as follows. The next
section discusses model development including the overarching
picture and details of model training, validation, and testing. Model training
Then the pavement data collected from Chicago O’Hare Inter- Function estimation
national Airport is presented. The results of applying GBM to Let us use {yi , xi }N1 to denote the training data of known
airport pavement conditions are subsequently reported, includ- (y, x)-values, where yi and xi = (xi1 , xi2 , . . . , xid ) refer respect-
ing comparing its goodness-of-fit with other methods. The ively to the response variable and input variables of the ith
findings of this paper are summarised and future research observation. d input variables are considered. The goal of
directions are suggested in the end. model training is to reconstruct the unknown mapping
y = F ∗ (x) with an estimate F̂(x). F ∗ (x) is the one that mini-
mizes the expected value of some specified loss function
L(y, F(x)) over the joint distribution of all (y, x)-values:
Model development
F ∗ = arg min Ey,x L(y, F) (1)
Overarching picture of gradient boosting machine F
model development
where L(y, F) is the loss function associated with y and F,
In this study, GBM model development involves splitting the measuring the deviation of the predicted value from the true
data into three subsets and performing the following three value of the response variable.
steps: The response variable y may come from different distri-
butions. In machine learning theory, different distributions
. Step 1 (model training): Use the first subset of the data to naturally lead to different specifications for the loss function
train a set of GBM models each corresponding to one com- L(y, F). Given that our response variable is PCI, which is a con-
bination of model hyperparameter values. tinuous variable, the L2 square loss function
. Step 2 (model validation): Use the second subset of the data L(y, F)L2 = (1/2)(y − F)2 and the robust regression Huber
for model validation. This step pertains to: (1) selecting a loss function L(y, F)Huber,d are typical candidates (Natekin
trained model that yields the best prediction accuracy, and and Knoll 2013). Here we choose the Huber loss function
(2) cross-validation to ensure no overfitting. L(y, F)Huber,d as shown in Equation (2), which captures not
. Step 3 (model testing): Use the third subset of data to further only L2 square loss but also mean absolute error L1.
test prediction accuracy of the selected GBM model. L(y, F)Huber,d becomes L(y, F)L2 when the absolute error of
Figure 2. Shape of L2 (A), L1 (B), and Huber (C) loss functions (source: Natekin and Knoll 2013).
6 L. BARUA ET AL.
can follow standard procedures: fit, a small value of v (which means “shrinking” the model
2 fit) requires a large M. For the maximum depth of a tree, the
N
∂L(y, F(xi )) greater the depth, the larger the possible number of terminal
am = arg min − − bh(xi ; a)
a,b
i=1
∂F(xi ) F(x)=Fm−1 (x)
nodes in a tree, which allows for capturing higher-order inter-
action effects among the input variables. Nevertheless, captur-
(9)
ing more than the actual interaction effects can lead to
The obtained h(x; am ) is then used to replace overfitting (Hastie et al. 2009). Once the maximum depth of
∂L(y, F(xi )) a tree is fixed, the actual tree depth may be determined using
− in steepest descent. Specifically,
∂F(xi ) F(x)=Fm−1 (x)
cost-complexity pruning. Finally, requiring the minimum
the new line search can be expressed as: number of observations under a node is reasonable: we will
not want too few observations after the node is split.
rm = arg min L(y, Fm−1 (x) + vrh(x; am )) (10) In this paper, we use grid search to enumerate all possible
r
value combinations for the four hyperparameters and identify
where v [ (0, 1] is a hyperparameter in GBM called the “learn- the best combination. As mentioned before, each combination
ing rate”. rm is used to update F(x): results in a GBM model using the training data. We select the
GBM model having the highest prediction accuracy without
Fm (x) = Fm−1 (x) + vrm h(x; am ) (11)
overfitting. In doing so, R2 is used to measure prediction accu-
As can be seen in (11), the learning rate v controls the con- racy when applying each trained model to the training dataset
tribution of each base model h(x; am ) to F(x). The purpose of (Equation (12)).
introducing v is to “shrink” the update of F(x), so as to prevent N
the trained model from overfitting the data (Friedman 2001). (yi − ŷi )2
R2 = 1 − i=1N (12)
Summarising, the GBM algorithm can be represented as i=1 (yi −
y)2
follows:
where yi denotes the observed value of the ith observation, ŷi is
GBM Algorithm: the corresponding predicted value, y is the mean of the
1.
N 1 N
Initialization: F0 (x) = arg minr L(yi , r) observed values: y = y . Ranging from 0 to 1, an R2
i=1 i
2. i=1 n
For m = 1 to M do: value of 0 indicates that the trained model does not fit the
2
3.
N ∂L(y, F(x i ))
am = arg min − − bh(x i ; a) data at all. An R2 value of 1 signifies that the model fully cap-
4. a,b i=1 ∂F(x i ) F(x)=Fm−1 (x) tures the variability of the response variable values around its
rm = arg min L(y, Fm−1 (x) + v rh(x; am )) mean value.
r
5.
Fm (x) = Fm−1 (x) + v rm h(x; am ) We calculate R2 for each of the trained GBM models and
6. End for
sort the models in descending order based on R2 . These models
are then evaluated one by one starting from the one with the
highest R2 , as follows. We apply a trained model to the vali-
Model validation dation dataset to generate predicted values and calculate R2 .
If the difference between this R2 and the R2 associated with
Grid search to select the best model the training dataset is less than a threshold (0.1 in this
In Step 1, many GBM models are trained, each with a distinct study), then the model is selected as the best model. Otherwise,
set of hyperparameter values. Because of the different hyper- the difference in R2 suggests presence of overfitting. In this case,
parameter values used, the resulting trained models will have the model is discarded and the next model for evaluation is
different prediction performance. In this paper, four hyper- studied. In the end, the best combination of hyperparameter
parameters are of interest: the number of regression trees values for GBM by accounting for both prediction accuracy
(M), the learning rate (v), the maximum depth of a tree, and and no overfitting is identified.
the minimum sample leaf of a tree (i.e. the minimum number
of observations a node needs to have in order to be considered
for splitting). A question that naturally arises is: what are the Cross-validation
best combined values for the hyperparameters to yield the To further assure that the selected hyperparameter values lead
best prediction performance without overfitting? to a good GBM model, k-fold cross-validation is also performed.
To answer the question, let us first have some intuitive Specifically, the training and validation datasets are merged and
understanding on why these hyperparameters are important. randomly divided into k subsets. Then, k − 1 subsets are
For M, having too few trees (small M) will obviously not pro- selected for training a GBM model using the selected hyperpara-
duce a good model fit. When more trees are added (larger M), meter values. The trained model is then used for prediction
the GBM model will become more complex and have improved using the remaining subset and the R2 ’s of the training subset
fit with the training data. However, fitting the training data too and the testing subset are calculated. This process is repeated
closely can be counterproductive as it often leads to poor gen- k times. If the average R2 associated with model validation is
eralisation ability – i.e. overfitting. Therefore, an optimal M much lower than with model training, then the model is dis-
needs to be determined. In doing so, a tradeoff should be recog- carded, and the next best combination of hyperparameter values
nised between M and v since both parameters affect the model (from the previous subsection) is evaluated. Otherwise, the
prediction error (Friedman 2001): to achieve the same model selected hyperparameters and associated GBM model are kept.
INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 7
Model testing Accordingly, two input variables are included for each inspec-
tion record. The first variable counts the number of minor
The model testing step is to provide another layer of under-
M&R activities (AC patching, AC crack sealing, PCC patching,
standing and confidence on how accurate the model prediction
PCC joint sealing, PCC crack sealing, PCC slab replacement,
is on new data. More specifically, the 20% of the data not used
subbase stabilisation, bituminous surface treatment, and sub-
in the previous two steps are used to check if the selected GBM
base permanent asphalt treatment) since the last inspection.
model can yield good prediction accuracy if implemented on
The second variable is a 0–1 indicator of whether a major
other data. Root Mean Square Error (RMSE) is used in addition
action has taken place since the last inspection. As already men-
R2 to measure prediction accuracy. As shown in Equation (13),
tioned, such an action will upgrade PCI of any value to 100.
RMSE is defined as the square root of the averaged squared
As traffic loading naturally affects pavement deterioration, it
difference between predicted and observed values (i.e. ŷi and
would be ideal to have traffic loading as an input variable. How-
yi ) over all observations in the training dataset (N ′ denotes
ever, full information about traffic loading is not available to us.
the testing dataset size). A lower RMSE value means a smaller
As a surrogate, for runway pavements a 0–1 variable is intro-
average difference, and thus a better fit of the model.
duced to indicate whether a section is located at the centre,
i.e. keel (taking value 1), or the side (taking value 0) of a run-
1 N′
RMSE = ′ (ŷi − yi )2 (13) way. Keel, which means courage way, is the area between the
N i main landing gears of aircraft outer wheels. Keel sections are
subject to much greater air traffic loading than at side locations
due to the lateral channelisation of aircraft load (Gerardi et al.
2007). As such, keel section PCI could decrease faster. In
Data
addition, the time elapsed since the last inspection, which is
The data used in this study come from Chicago O’Hare Inter- included in the models, may partly capture the cumulative
national Airport. The pavements considered consist of 163 run- effect of traffic loading over time, since cumulative traffic and
way sections and 377 taxiway sections. Each pavement section time can have high correlation.
exhibits fairly homogenous characteristics including pavement Apart from the pavement-specific information, weather
age, material type, and traffic load. A pavement section can information at each time of inspection is also collected from
have multiple records each corresponding to an inspection. the National Oceanic and Atmospheric Administration
The inspections span from the construction date (the beginning (NOAA), including the amount of rainfall in mm and the num-
of the pavement service life) to the latest inspection in 2017. ber of freeze-thaw cycles since the last inspection. Following
The inspections have been conducted once every three years Baker and Ruschy (1995), a freeze-thaw cycle occurs when a
since 2000. Prior to 2000, the time between two inspections day has a minimum temperature of 25.7°F or lower and a maxi-
was more irregular, but most of the time intervals were less mum temperature of 43.3°F or higher.
than 6 years. For each inspection, information on the PCI To summarise, each data point used in the GBM model
value, inspection date, construction date, pavement material development corresponds to an inspection record. The input
type, and past M&R activities is collected. In total, the dataset variables in the GBM models are: pavement age, time elapsed
has 1,091 and 2,090 records for runways and taxiways, since the last inspection, PCI value from the last inspection,
respectively. material type dummies, the number of minor M&R activities,
For runways, three pavement material types are considered: whether a major action has taken place, the amount of rainfall,
Portland Cement Concrete (PCC), Asphalt Overlay on PCC and the number of freeze-thaw cycles since the last inspection.
(APC), and Asphalt Overlay on Asphalt Concrete (AAC). For runway pavement sections, a 0–1 variable indicating the
The three material types are also considered for taxiways keel or side location of a pavement section is further included.
with the addition of Asphalt Concrete (AC). Each pavement
type is associated with a dummy input variable. Accordingly,
Results
three dummy variables are introduced for runways and four
for taxiways to indicate the material type of each pavement sec- This section presents model development results. We first
tion at each inspection. report the performance of the developed GBM models in
At each inspection, pavement age is determined by the terms of R2 and RMSE. Then, we compare the results from
difference between the inspection date and the construction the GBM models and other methods, including linear
date. Note that if one or more major actions (i.e. reconstruction regression, nonlinear regression, ANN, and RF. Finally, as the
and rehabilitation, where the latter refers to AC mill and over- focus of the paper, we investigate the contributions of the
lay and PCC overlay) were taken, the pavement age and PCI different input variables and their interactions to PCI using
would be reset to 0 and 100 at the time of reconstruction. relative importance and partial dependence plots.
Also note that the pavement material type of a section at
each inspection depends on the implemented M&R action(s)
Performance of the GBM models
since the previous inspection. For example, if AC mill and over-
lay was implemented on a PCC pavement, the material type will Two separate GBM models are developed for runways and taxi-
change to APC in the next inspection record. way pavements. Table 2 shows the R2 values of the selected
M&R activities can be classified as major and minor (to be GBM models after model training, validation, and testing.
exact, minor maintenance and major rehabilitation actions). The generally high R2 values indicate good fit of both models.
8 L. BARUA ET AL.
Table 2. R2 of the selected GBM models applied to the runway and taxiway The predicted and actual PCI values of runway and taxiway
datasets.
sections in different datasets are plotted in Figures 3 and 4. As
Training Validation Cross- Testing
dataset dataset validation dataset
can be seen in the figures, most data points in each graph are
close to the 45-degree line. Consistent with the R2 and RMSE
Runway 0.97 0.90 0.90 ± 0.09 0.91
model results above, the greater concentration of data points along
Taxiway 0.92 0.85 0.86 ± 0.07 0.86 the 45-degree lines for runway pavements suggests better
model model fit of the runway GBM model.
To further illustrate the good fit of the two GBM models,
Figures 5 and 6 plot the predicted and actual PCI curves start-
Between runways and taxiways, the runway model has higher ing from the construction dates for three runway pavement sec-
R2 ’s, which could be due to the use of an extra input variable tions and three taxiway pavement sections that are randomly
indicating the lateral location (keel vs. side) of a pavement sec- selected. Note that the first point on a predicted PCI curve
tion to partly capture the traffic loading impact. We observe has the same PCI value as the first point on the corresponding
small differences between the mean R2 values in cross-vali- actual PCI curve, as no prediction can be made for that point.
dation and the R2 values in model training. The R2 values for The selected runway sections are with flexible pavement and
the testing datasets are also high, suggesting a high level of pre- the taxiway sections are with rigid pavement. Figures 5–6
diction accuracy when the selected models are applied to new show that the predicted and actual curves match quite well. It
datasets. is clear that the pavement deterioration is not linear nor mono-
The RMSE values presented in Table 3 tell a similar story. tonic over time, with jumps caused by M&R interventions. The
Smaller RMSE values are obtained for runways than for taxi- effects of the M&R interventions on PCI values can be correctly
ways. Considering that PCI ranges from 0 to 100, an RMSE captured by the GBM models.
of 4 for runways and 8 for taxiways using the testing dataset Finally, given that information on traffic loading is available
corroborate the good performance of the selected GBM models for a subset of the data, two GBM models, one for runways and
when applied to a new dataset. one for taxiways, are further developed with the addition of a
traffic loading variable. The subset of the data covers 26 runway
sections and 31 taxiway sections for 1998–2013, providing in
Table 3. RMSE for the selected GBM models.
total 473 records for runways and 924 for taxiways. In this sub-
Dataset Training Validation Testing
set, the aircraft type of each departure / arrival flight using a
Runway 2.4 3.7 4.0
Taxiway 5.1 8.0 8.1 runway / taxiway pavement section is recorded. To approxi-
mate the traffic loading on a pavement section, the maximum
Figure 5. Predicted and actual deterioration curves for three selected runway pavement sections.
Figure 6. Predicted and actual deterioration curves for three selected taxiway pavement sections.
take-off weights (MTOW) / maximum landing weights (MLW) relationship between the input variables and the response vari-
of all departure / arrival flights between the current and the able (Hastie et al. 2009). On the other hand, too many neurons
immediate previous inspections are summed for each runway can cause overfitting and/or be slow to train. Some general rec-
/ taxiway pavement section. The MTOW and MLW infor- ommendations for selecting the number of neurons are given in
mation for each aircraft type is collected from the manufac- Heaton (2008), as follows:
turers’ websites.
With the new traffic loading variable, the two trained GBM . The number of neurons should be between the size of the
models using the subset of the original data yield R2 ’s of 0.78 input layer (i.e. the number of input variables) and the
and 0.74 for runways and taxiways respectively. These R2 ’s size of the output layer (i.e. the number of response
are better than R2 ’s using the same data without the traffic load- variables).
ing variable (0.65 and 0.62 for runways and taxiways), confi- . The number of neurons should be 2/3 the size of the input
rming the importance of traffic loading in affecting PCIs. On layer plus the size of the output layer.
the other hand, the R2 ’s are lower than without the traffic load- . The number of neurons should be less than twice the size of
ing variable but using the original, full dataset (which has R2 ’s the input layer.
of 0.97 and 0.92 for runways and taxiways). The reason is that
having larger data helps GBM learn the relationship between In our study, the number of input variables is 11 for both
the response and input variables, resulting in better model fit runways and taxiways. There is only one response variable,
(Brain and Webb 1999). i.e. PCI. Following the above recommendations, we develop a
number of models for both runways and taxiways with the
number of neurons ranging from 1 to 11. The models are eval-
Comparing GBM with other methods
uated based on R2 associated with the testing data. The highest
We further compare the performance of the developed GBM R2 is found to occur when 10 neurons are used, for both taxi-
models with linear regression, nonlinear regression, ANN, ways and runways.
and RF models. For ANN, a feed forward model with one hid- In developing the RF model, three hyperparameters need to
den layer along with one input layer and one output layer is be tuned: the number of trees, the maximum number of fea-
considered. While multiple hidden layers can be considered tures used for splitting a tree, and the minimum sample leaf
in ANN, in practice there is no reason to use more than one size (i.e. the minimum number of samples required for a leaf
hidden layer (Heaton 2008). A hidden layer consists of multiple node). To this end, a large number of combinations of hyper-
hidden units called neurons. Choosing an appropriate number parameter values are tested. Specifically, we set the number of
of neurons is necessary for selecting the best model. Using too trees between 1 and 1000, the maximum number of features
few neurons may not predict the response variable properly equal to the number of input variables, and the minimum
because they lack the flexibility to capture the nonlinear sample leaf size between 1 and 20. We find that RF models
10 L. BARUA ET AL.
Table 4. R2 and RMSE for different ML methods. variable (Friedman 2001). In other words, the relative impor-
R2 RMSE tance of an input variable is intimately associated with the
Runways Taxiways Runways Taxiways change in the model fit with and without the variable.
Linear regression 0.68 0.67 12.2 12.2 In a rough sense, one could view the relative importance of a
Quadratic regression 0.76 0.74 6.9 11.9 variable as analogous to the t-statistic of the variable’s coeffi-
ANN 0.79 0.76 6.7 10.1
RF 0.86 0.81 5.3 9.2 cient in linear regression, which indicates the coefficient’s stat-
GBM 0.91 0.86 4.0 8.1 istical significance, as suggested in Darlington (1968), Breiman
(2017), and Greenwell et al. (2018). Let us use an example to
illustrate. Say a linear regression model has two input variables
with 300 trees and a minimum sample leaf size of 1 yield the
A and B. Variable A’s t-statistic is 10. Variable B’s t-statistic is
highest R2 for both runways and taxiway using the testing
2. As both variables are statistically significant from 0 at 0.05
dataset.
level, often we are content with the result and do not care
Table 4 compares the R2 and RMSE results of linear
further about the difference in the t-statistics. However, if we
regression, quadratic regression (which includes as input vari-
develop a GBM model, then the difference in the t-statistics
ables all possible first- and second-order terms), ANN, RF, and
would likely manifest itself in a higher relative importance of
GBM using the testing data. Results indicate the order of per-
variable A than variable B.
formance: linear regression < quadratic regression < ANN
Based on the above analogy, it is important to clarify that the
< RF < GBM, which substantiates the superiority of GBM in
relative importance of an input variable is a concept more com-
this study and thus our modelling choice.
parable to “the statistical significance of a variable’s coefficient”,
rather than “the marginal effect of the variable (i.e. the absolute
Contributions of input variables to pavement value of the variable’s coefficient)” in a linear regression
deterioration context.
To facilitate comparison, the relative importance is further
As already mentioned, the main purpose of developing the scaled, such that the relative importance of the most influential
GBM models is to gain an understanding of the influence of input variable is 100. Because the relative importance is
the input variables and their interactions on PCI. Fortunately, measured in terms of the improvement in the squared error,
this is possible with GBM. In this subsection, we use the relative it is always positive.
importance of input variables and the partial dependence plots Figures 7 and 8 plot the relative importance of the input
to interpret the underlying relationship between different input variables in our GBM models for runways and taxiways
variables and PCI. respectively. For runways, pavement age has the highest relative
importance, followed by the time elapsed since the last inspec-
Relative importance tion. PCI of the previous inspection and weather conditions,
The relative importance of each input variable on the response including both the number of freeze-thaw cycles and the
variable is computed following Friedman (2001) and Hastie amount of rainfall, have non-trivial relative importance as
et al. (2009). Specifically, for an input variable, its relative well. The keel location and the number of minor M&R actions
importance in each decision tree in a GBM model is the sum have less contributions. The occurrence of a major M&R action
of improvements in the squared error from each split that has even smaller influence. The influence of different pavement
involves the input variable. Such relative importance is then material types is limited.
averaged over all the decision trees in a GBM model. In plainer One reason for the limited importance of keel location,
language, the relative importance of an input variable describes number of minor M&R actions, occurrence of a major M&R
the influence of the variable on the variation of the response action, and pavement material types stems from the definition
of the relative importance measure and the fact that these vari- are conditioned for x\l , then F̂(x), the estimate of F ∗ (x), can
ables can take only limited values. Indeed, all of these variables, be considered as a function only of xl : F̂l (xl ) = F̂(xl |x\l ).
except for the number of minor M&R actions (which is an inte- Under the assumption that the dependence of F̂l (xl ) on x\l is
ger), are binary variables and can take only 0 or 1 value. Con- not too strong, the average function of F̂l (xl ), i.e. F l (xl ), can
sequently, these variables provide limited possibilities for tree be estimated from the training data:
splitting compared to other input variables that take continu-
ous values. As each split improves model squared error, a vari- 1 N
able involved in fewer splits would likely lead to less F l (xl ) = F̂(xl , xi,\l ) (14)
N i=1
improvement in the model squared error and therefore have
a smaller relative importance. Equation (14) means that we account for the (empirical)
For taxiways, the relative importance of the input variables is average influence of the input variables not in xl while comput-
similar. However, there exist the main differences: (1) the rela- ing F l (xl ). We first consider xl to have only a single input vari-
tive importance of the number of freeze-thaw cycles is able. Using Equation (14), PDPs are constructed for six input
decreased, from being comparable to only about half the rela- variables which have the highest relative importance. The
tive importance of the amount of rainfall and PCI of previous results are shown in Figures 9 and 10. The plots are not strictly
inspection; (2) the relative importance of the time elapsed smooth, which is a consequence of using tree-based models.
since the last inspection is also decreased, especially compared Decision trees produce discontinuous piecewise constant
to that of pavement age. A possible explanation for the former models. This carries over to sums of trees (Hastie et al. 2009).
could be that taxiway pavements are subject to less speedy and Also, for some observations the time elapsed since the last
intense aircraft traffic loading than runway pavements, leading inspection were over 40 years. Consequently, very large num-
to less cracking. Because freeze-thaw cycles influence PCI bers of freeze-thaw cycles and large amounts of rainfall since
mainly through cracking, the relative importance of the num- the last inspection exist in the data.
ber of freeze-thaw cycles is expected to become less. In a similar The partial dependence of PCI on the input variables can all
vein, as the time elapsed since the last inspection partly capture be intuitively interpreted. For runways, PCI decreases almost
the traffic effect (the longer the time, the greater the amount of exponentially with pavement age. As the time elapsed since
cumulated traffic), the less speedy and intense traffic loading on the last inspection increases, the PCI value decreases with a
taxiways than on runways suggests less damage on the taxi- sudden drop when the time elapsed reaches 7,000 days
ways, or smaller influence of the time elapsed variable on the (about 19 years). Such a large time interval, as noted above,
variations of taxiway PCI values. only occurred for the very old pavements (e.g. those con-
structed in the 1950s). PCI value increases almost linearly
Partial dependence plot with the PCI of the previous inspection, implying that the cur-
The nature of the dependence of the response variable on a sub- rent condition has a natural intimate correlation with historic
set of input variables can be alternatively identified using Partial condition. As the number of freeze-thaw cycles increases, pave-
Dependence Plot (PDP), which provides a graphic rendering of ment conditions deteriorate more, resulting in lower PCI
the effect of one or two input variables on the response variable values. It is interesting to observe that PCI tends to increase
after accounting for the average effects of the remaining input with the increase of rainfall at the beginning, suggesting that
variables (Friedman 2001). moderate rainfall would actually be beneficial for pavement
The construction of PDP is as follows. Let xl denote the health. This can be explained by the fact that less rainfall,
input variables x of interest, and x\l denote the remaining which is often associated with more sunshine, causes pavement
input variables. For graphic display, a low cardinality of xl , oxidation and consequently deterioration (Farris 2017). How-
i.e. with one or two variables, is most useful. If specific values ever, when rainfall becomes excessive, such a positive effect
12 L. BARUA ET AL.
Figure 9. Partial dependence plots of six input variables with the highest relative importance for runways.
of rainfall on PCI will cease to exist, possibly due to the adverse for different x2 values, with the difference only in the absolute
effect from water puddling and infiltration into the pavement. PCI value as PCI is also a function of x2 , then the interaction
Finally, given that the keel location variable has only two values between x1 and x2 is weak (the same is true by exchanging x1
0 and 1, its PDP is also discrete. The result shows that, as and x2 ). This appears to the case for pavement age – the time
expected, PCI will be less in keel than in side locations. elapsed since the last inspection, and pavement age – the num-
The PDPs for taxiways show very similar relationships. The ber of freeze-thaw cycles for runways; and for pavement age –
only new plot, which is the partial dependence of PCI on the the time elapsed since the last inspection, and pavement age –
number of minor M&R actions, indicates that as we increase the amount of rainfall for taxiways.
the number of minor M&R actions, the PCI value will increase. In contrast to the above, the interaction between pavement
This is again consistent with expectation. age and PCI value from the last inspection is more apparent for
Figures 11 and 12 show the partial dependence on joint both runways and taxiways. The effect of PCI value from the
values of several two-variable combinations, for runways and last inspection is small when a runway pavement section is rela-
taxiways respectively. The combinations are pavement age tively new until about 5–6 years (2,000 days). As the pavement
(the variable with the highest relative importance) with the continues to age, the effect becomes more significant. The
input variables of the second, third, and fourth highest relative dependence of the pavement age effect on the PCI value from
importance. Note that the vertical axis has a range covering the last inspection can also be discerned. Similar conclusions
positive and negative values, with zero as the benchmark corre- can be drawn for taxiways. For example, the dependence of
sponding to the average PCI value among all observations. taxiway PCI on the PCI value from the previous inspection fol-
Overall, the non-flat surfaces in the PDPs confirm the lows generally two trends that are delimited by a pavement age
dependence of PCI on each of the input variables. What is of about 11 years (4,000 day). The trends are not difficult to
more interesting is the interactions between two input variables interpret: as a pavement section ages, its PCI would likely be
in a given pair. Specifically, given two variables x1 and x2 , if a more closely correlated with its recent condition, as measured
changing trend of PCI against x1 remains largely the same by the PCI value from the last inspection.
INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 13
Figure 10. Partial dependence plots of six input variables with the highest relative importance for taxiways.
Conclusion Funding
In this paper, we develop Gradient Boosting Machine (GBM) This work was supported by Chicago Department of Aviation: [grant
models to investigate the influence of a variety of input vari- number C8199].
ables and their interactions on airfield pavement deterioration.
By including input variables that capture characteristics such as ORCID
pavement age, previous condition, material, maintenance & Bo Zou http://orcid.org/0000-0003-4485-5548
rehabilitation (M&R) history, weather conditions, and traffic
loading, GBM models are developed following a systematic
procedure composed of model training, validation, and testing. References
The models are applied to runways and taxiways at Chicago Attoh-Okine, N.O., 1994. Predicting roughness progression in flexible
O’Hare International Airport. It is found that pavement age, pavements using artificial neural networks. In: Proc., 3rd int. conf. on
time elapse since the last inspection, PCI value from the last managing pavements, transportation research board. San Antonio,
TX, 55–62.
inspection, and weather conditions are the most influential
Baker, D.G. and Ruschy, D.L., 1995. Calculated and measured air and soil
input variables. The results show that GBM yields better freeze-thaw frequencies. Journal of Applied Meteorology, 34 (10), 2197–
model goodness-of-fit than linear regression, nonlinear 2205.
regression, artificial neural networks, and random forest, Barua, L., et al., 2019b. Predicting airport runway and taxiway pavement
provided that the input variable values can be accurately conditions: a gradient boosting approach. In: Proceedings of the 98th
annual meeting of the transportation research board. Washington, DC.
obtained.
Barua, L., Zou, B., and Zhou, Y., 2019a. Machine learning for international
The developed models and the associated results will help freight transportation management: a comprehensive review. Under review.
airport agencies and operators understand and confirm the Bianchini, A. and Bandini, P., 2010. Prediction of pavement performance
effects of a variety of influencing factors and their interactions through neuro-fuzzy reasoning. Computer-Aided Civil and
on airport pavement deterioration. In particular, airport Infrastructure Engineering, 25 (1), 39–54.
Brain, D. and Webb, G., 1999. On the effect of data set size on bias and
agencies and operators will gain further knowledge in assessing
variance in classification learning. In: R. Debbie, ed. Proceedings of
the effectiveness of historic M&R actions, which in turn the fourth Australian knowledge acquisition workshop. Sydney:
informs future pavement maintenance management. In University of New South Wales, 117–128.
addition, the gained knowledge about the relative importance Breiman, L., 2017. Classification and regression trees. New York: Routledge.
of different input variables will help airport agencies and oper- Carnahan, J.V., et al., 1987. Optimal maintenance decisions for
pavement management. Journal of Transportation Engineering, 113
ators identify influential factors for which data collection efforts
(5), 554–572.
should be preserved / enhanced (e.g. frequency of data collec- Chan, P., Oppermann, M., and Wu, S.S., 1997. North Carolina’s experience
tion and quality requirement of the collected data), which is in development of pavement performance prediction and modeling.
especially important if data collection efforts are cost-sensitive. Transportation Research Record: Journal of the Transportation
For future research, an important direction is to collect more Research Board, 1592, 80–88.
Choi, J.H., Adams, T.M., and Bahia, H.U., 2004. Pavement roughness
data on traffic loading, which is only available for some of the
modeling using back-propagation neural networks. Computer-Aided
years in the dataset. Although as a remedy we include in the Civil and Infrastructure Engineering, 19 (4), 295–303.
runway model a 0–1 variable indicating the keel/side location Chou, J., O’Neill, W.A., and Cheng, H.D., 1994. Pavement distress classifi-
of a pavement section to partly capture the traffic load effect, cation using neural networks. In: Proceedings of the IEEE international
the binary nature of the variable suggests that it is not a perfect conference on systems, man and cybernetics society. San Antonio, TX:
IEEE, 397–401.
surrogate. Another direction is to explore a hybrid approach
Darlington, R.B., 1968. Multiple regression in psychological research and
that combines GBM and other machine learning methods practice. Psychological Bulletin, 69 (3), 161.
(e.g. using a support vector classifier to first classify pavements) Ding, C., Cao, X.J., and Næss, P., 2018. Applying gradient boosting
to harness the strengths of each method in model development decision trees to examine non-linear effects of the built environment
which may further enhance the interpretability of the influence on driving distance in Oslo. Transportation Research Part A: Policy
and Practice, 110, 107–117.
of different factors and their interactions on pavement
Eldin, N.N. and Senouci, A.B., 1995. Condition rating of rigid pavements
deterioration. by neural networks. Canadian Journal of Civil Engineering, 22 (5), 861–
870.
Elith, J., Leathwick, J.R., and Hastie, T., 2008. A working guide to boosted
Acknowledgement regression trees. Journal of Animal Ecology, 77 (4), 802–813.
Farid, D.M., et al., 2014. Hybrid decision tree and naïve Bayes classifiers for
This research was funded by the Chicago Department of Aviation multi-class classification tasks. Expert Systems with Applications, 41 (4),
(CDA) and the O’Hare BPC/CarePlus programme through the Center 1937–1946.
of Excellence for Airport Technology (CEAT) at the University of Illi- Farris, D., 2017. How does weather effect asphalt pavement? Available
nois at Urbana-Champaign. We are very grateful for the enthusiastic from: http://bitroads.com/About-Us/Blog/entryid/6/how-does-weather-
support from Professor David Lange, Ross Anderson, Jim Chilton, effect-asphalt-pavement [Accessed 23 August 2018].
Tim Morgan, Zachary Bergman, Adam Hardy, Brad McMullen, Michael Freund, Y. and Schapire, R.E., 1996. Experiments with a new boosting
Vonic, Donald McGady, Shawn Gould, Matthew Yohn, and Carlos algorithm. In: Machine learning: proceedings of the thirteenth inter-
Mendez of this effort. national conference, 96, 148–156.
Friedman, J.H., 2001. Greedy function approximation: a gradient boosting
machine. Annals of Statistics, 29 (5), 1189–1232.
Disclosure statement Gendreau, M. and Soriano, P., 1998. Airport pavement management sys-
tems: an appraisal of existing methodologies. Transportation Research
No potential conflict of interest was reported by the authors. Part A: Policy and Practice, 32 (3), 197–214.
INTERNATIONAL JOURNAL OF PAVEMENT ENGINEERING 15
George, K. and Shekharan, A., 1998. Updates of pavement performance Nazari, F., et al., 2017. Optimal facility-specific inspection and mainten-
modeling. In: 3rd International conference on road & airfield pavement ance decisions under measurement uncertainty: Unifying framework.
technology, proceedings. Beijing: Huajie International Publishing Co Journal of Infrastructure Systems, 23 (4), 04017036.
Ltd, 402–410. Ng, A., 2018. coursera.org. Available from: https://www.coursera.org/
Gerardi, T., et al., 2007. Airfield concrete pavement smoothness: a hand- learn/machine-learning [Accessed 25 June 2018].
book. Rep. IPRF, 2002–4. Noruzoliaee, M. and Zou, B., 2015. Protecting the airfield assets at O’Hare
Golabi, K., Kulkarni, R.B., and Way, G.B., 1982. A statewide pavement International Airport: capability enhancement for predicting future
management system. Interfaces, 12 (6), 5–21. runway pavement performance. Technical report. Prepared for the
Gorripaty, S., et al., 2017. Identifying similar days for air traffic manage- Chicago Department of Aviation. doi:10.13140/RG.2.2.19882.16326.
ment. Journal of Air Transport Management, 65, 144–155. Noruzoliaee, M. and Zou, B., 2017. Protecting the airfield assets at O’Hare
Greenwell, B.M., Boehmke, B.C., and McCarthy, A.J., 2018. A simple and International Airport: developing prediction capability for taxiway
effective model-based variable importance measure. arXiv preprint pavement conditions and review of federal guidelines and state-of-
arXiv:1805.04755. the-practice in airfield assets condition rating. Technical report.
Hastie, T., Tibshirani, R., and Friedman, J., 2009. The elements of statistical Prepared for the Chicago Department of Aviation. doi:10.13140/RG.2.
learning. New York, NY: Springer. 2.16526.72007.
Heaton, J., 2008. Introduction to neural networks with Java. St. Louis: Noruzoliaee, M. and Zou, B., 2019. Airfield infrastructure management
Heaton Research, Inc. using network-level optimization and stochastic duration modeling.
Hong, F. and Prozzi, J.A., 2006. Estimation of pavement performance Infrastructures, 4 (1), 2.
deterioration using Bayesian approach. Journal of Infrastructure Ortiz-García, J.J., Costello, S., and Snaith, M.S., 2006. Derivation of tran-
Systems, 12 (2), 77–86. sition probability matrices for pavement deterioration modeling.
Jiang, Y., Saito, M., and Sinha, K.C., 1988. Bridge performance prediction Journal of Transportation Engineering, 132 (2), 141–161.
model using the Markov chain. Transportation Research Record, 1180, Owusu-Ababio, S., 1998. Application of neural networks to modeling thick
31–39. asphalt pavement performance. In: A.A. Balkema, ed. Artificial intelli-
Kargah-Ostadi, N., Stoffels, S., and Tabatabaee, N., 2010. Network-level gence and mathematical methods in pavement and geomechanical sys-
pavement roughness prediction model for rehabilitation recommen- tems. Rotterdam: CRC Press, 23–30.
dations. Transportation Research Record: Journal of the Pindyck, R.S. and Rubinfeld, D.L., 1998. Econometric models and economic
Transportation Research Board, 2155, 124–133. forecasts. New York: McGraw-Hill.
Koren, Y., 2009. The bellkor solution to the netflix grand prize. Netflix Prozzi, J. and Madanat, S, 2000. Using duration models to analyze exper-
Prize Documentation, 81, 1–10. imental pavement failure data. Transportation Research Record: Journal
Lee, D. and Derrible, S., 2019. Predicting residential water consumption: of the Transportation Research Board, 1699, 87–94.
modeling techniques and data perspectives. ASCE Journal of Water Regue, R. and Recker, W., 2014. Proactive vehicle routing with inferred
Resources Planning and Management, in press. demand to solve the bikesharing rebalancing problem. Transportation
Lee, D., Derrible, S., and Pereira, F.C., 2018. Comparison of four types of Research Part E: Logistics and Transportation Review, 72, 192–209.
artificial neural network and a multinomial logit model for travel mode Schwartz, C.W., 1993. Infrastructure condition forecasting using neural
choice modeling. Transportation Research Record: Journal of the networks. Proceedings of Infrastructure Planning and Management,
Transportation Research Board, 2672 (49), 101–112. 282–284. American Society of Civil Engineers.
Lin, J.D., Yau, J.T., and Hsiao, L.H., 2003. Correlation analysis between Shahnazari, H., et al., 2012. Application of soft computing for prediction of
international roughness index (IRI) and pavement distress by neural pavement condition index. Journal of Transportation Engineering, 138
network. In: 82th Transportation research board annual meeting, (12), 1495–1506.
(CD-ROM). Washington, DC. Shekharan, A.R., 2000. Pavement performance prediction by artificial
Lou, Z., Lu, J.J., and Gunaratne, M., 1999. Road surface crack condition neural networks. In: A.A. Balkema, ed. Computational intelligence
forecasting using neural network models. Final report for Florida applications in pavement and geomechanical systems. Rotterdam: CRC
State department of transportation. Press, 89–98.
Ma, X., et al., 2017. Prioritizing influential factors for freeway incident Straub, D., 2009. Stochastic modeling of deterioration processes through
clearance time prediction using the gradient boosting decision trees dynamic Bayesian networks. Journal of Engineering Mechanics, 135
method. IEEE Transactions on Intelligent Transportation Systems, 18 (10), 1089–1099.
(9), 2303–2310. Tabatabaee, N., Ziyadi, M., and Shafahi, Y., 2013. Two-stage support
Madanat, S., Bulusu, S., and Mahmoud, A., 1995b. Estimation of infra- vector classifier and recurrent neural network predictor for pave-
structure distress initiation and progression models. Journal of ment performance modeling. Journal of Infrastructure Systems,
Infrastructure Systems, 1 (3), 146–150. 19 (3), 266–274.
Madanat, S. and Ibrahim, W.H.W., 1995. Poisson regression models of Van der Gryp, A., et al., 1998. Determining the visual condition index of
infrastructure transition probabilities. Journal of Transportation flexible pavements using artificial neural networks. In: 4th
Engineering, 121 (3), 267–272. International conference on managing pavements, 115–129.
Madanat, S.M., Karlaftis, M.G., and McCarthy, P.S., 1997. Probabilistic Vlahogianni, E.I., Karlaftis, M.G., and Golias, J.C., 2014. Short-term traffic
infrastructure deterioration models with panel data. Journal of forecasting: where we are and where we’re going. Transportation
Infrastructure Systems, 3 (1), 4–9. Research Part C: Emerging Technologies, 43, 3–19.
Madanat, S., Mishalani, R., and Ibrahim, W.H.W., 1995a. Estimation of Walgama Wellalage, N., Zhang, T., and Dwight, R., 2015. Calibrating Markov
infrastructure transition probabilities from condition rating data. chain-based deterioration models for predicting future conditions of rail-
Journal of Infrastructure Systems, 1 (2), 120–125. way bridge elements. Journal of Bridge Engineering, 20 (2), 1–13.
Mauch, M. and Madanat, S., 2001. Semiparametric hazard rate models of Yang, J., 2004. Road crack condition performance modeling using recurrent
reinforced concrete bridge deck deterioration. Journal of Infrastructure Markov chains and artificial neural networks. Dissertation (PhD),
Systems, 7 (2), 49–57. University of South Florida.
Mishalani, R.G. and Madanat, S.M., 2002. Computation of infrastructure Zhang, Y. and Haghani, A., 2015. A gradient boosting method to improve
transition probabilities using stochastic duration models. Journal of travel time prediction. Transportation Research Part C: Emerging
Infrastructure Systems, 8 (4), 139–148. Technologies, 58, 308–324.
Nakat, Z.S. and Madanat, S.M., 2008. Stochastic duration modeling of pave- Zhou, Z.H., 2012. Ensemble methods: foundations and algorithms. Boca
ment overlay crack initiation. Journal of Infrastructure Systems, 14 (3), Raton, FL: Chapman and Hall/CRC.
185–192. Zou, B. and Madanat, S, 2012. Incorporating delay effects into airport run-
Natekin, A. and Knoll, A., 2013. Gradient boosting machines, a tutorial. way pavement management Systems. Journal of Infrastructure Systems,
Frontiers in Neurorobotics, 7, 21. 18 (3), 183–193.