slope failure prediction using a decision tree: a case of...

Slope failure prediction using a decision tree: A case of engineered slopes in South Korea

Ivy F. Guevarra1 *, SangGi Hwang1, ByungOk Yu2

1Department of Civil, Environmental and Railroad Engineering Paichai University, Daejon, South Korea

2Geotechnical Research Group, Highway Research Center, Korea Highway Corporation

UP-NIGS AVR, 15 JUNE 2012

Presentation outline

Introduction

Objective

Methodology

Analytical result

Discussion

Conclusion

Slope Failure prediction using a decision tree

Problem: How useful are massive slope databases?

Annual slope survey

Massive data

Costly yet not fully utilized

How should we utilize it?

How useful is the database?


Objective

Massive Database

0

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

Igneous Metamorphic Sedimentaray Total

Lithology and Average

Failu

re n

um

ber

per

1 km

Total Area KyungGi OgChon YoungNam KyungSang

0

30

60

90

0 30 60 90

Slope Angle

Slo

pe H

eig

ht


1. Find informative slope prediction rules from the massive data

2. Quantify success rate 3. Find the most relevant slope

factors influencing slope failures in the Korean slope database

Statistical Analyses

Korean Engineered Slope Database

South Korea Highway slopes Byways slopes

Data source Korea highway corporation (KHC) Korea Institute of Geological, Mining and Materials( KIGAM ) Korea Meteorological Association (KMA)

13,036 slope entries 23 slope factors


Additional slope factors

Tectonic Domain

Tectonic Domain

Lithology

Precipitation

Selected Slope factors


METHODOLOGY: Data Mining

A young interdisciplinary field of computer science

Extracts useful information from an existing data set and transforms it into an understandable structure for further use

Does classification,anomaly detection, association, clustering, regression, sequence analysis

Classification generates prediction rules

Best for slope prediction studies

Methods Neural network, decision trees

Nonparametric: does not require the common

normal distribution assumption for input data

Internal structure of the Decision trees can be easily reviewed as compared to the

“black box” structure of the neural network.

Classification using Decision Trees

Decision Tree Classification


Classification Rules

MINED RULES 1. If Slope Dip > 50°, then slope Fails 2. If Slope Dip < 50°, Precipitation > 500mm,

Then slope Fails 3. If Slope Dip < 50°, Precipitation > 500mm,

Then slope Safe

Classifying a given slope

Classification Modeling and Validation Flow

WEKA 3.5 Software

J48 Decision Tree Technique

C4.5 Algorithm

Dept. of Computer Science

Waikato University

New Zealand

Database is divided into 10 sets 1 held out

9 used to generate rules

Rules applied to the hold out set

Error rate is generated

10 iterations 10 Error rates

Average Error rate as overall error

Least error rules chosen

Extreme case 1: Too much rules

Extreme case 2: Too few rules

Avoiding over and under fitting


Methodology: Ranking the slope factors


Attribute Ranking ALGORITHMS

Ranking based on relevant criteria

Uses 5 different rankers

Unique equation for relevance

independent mathematical, information and statistical functions

Weka 3.5 software

Attribute evaluators

Result:

Decision tree 174 rules

Represented in the diagram

Ranking Average ranking

Important attributes

Statistical Measure Overall success rate

Success rate of fail and safe slopes


Decision tree rules

If seepage is present, slope fails

1


Rules: Group 1

If seepage is present Upper, then slope fails

Middle , then slope fails

Lower

Rainfall is evaluated

Classified 829 (12%) observation

651 out of 829 failed

Caused by the presence of seepage

Commonly known rules

1

Decision tree rules

If seepage is present, slope fails If rainfall>267 mm,

slope fails

1

2


Rules: Group 2

If seepage is absent, slope dip >61°

Rainfall

Fracture Orientation

Weathering

Classified 21% observation

1016 out of 1413 failed

Characterized by high slope and high precipitation >267mm rainfall, slopes could fail, unless low angle

>501mm, fails except low angles and least weathered

2

Decision tree rules


Rules: Group 3 3a

If seepage is absent Slope dip is <61°

Known Fracture Orientation:3a

Unknown fracture orientation:3b

Group 3a

Antislope fails if rainfall>412mm

Low angles are safe unless notches are present

Most syn slopes fail

Safe if notches are absent

Safe in some tectonic domains (Kyungsang, Ogchon, Youngnam)

Decision tree rules

3bRules are more complicated 3b: Banded gneiss fail in high amount of rainfall (>454 mm).

However, if weathering is less intense, it is safe

Least important slope factors


Rules: Group 3

If seepage is absent Slope dip is <61°

Known Fracture Orientation:3a

Unknown fracture orientation:3b

Notches, rainfall, slope dip,

lithology,tectonic domain

Group 3b

Widely branching, not easy to describe

Granitic rock are strong, except for slope dip>41° and weathering is high

If rainfall is lower than 455mm but dipping >43.2° and lithology are shale, schist, diorite, layered conglomerate and sandstone, then it could fail.

Natural slope, height, length were least used

3b

Ranking Results


Ranking Results Ranking

1.Fracture

2.Lithology

3.Seepage

4.Rainfall

5.Slope dip

6.Notches

7.Domain

8.Topography

9.Weathering

10.Natural slope

11.Heigh

12.Length

Selectors:

1.Information Gain 2.Gain Ratio

3.Symmetrical Uncertainty 4.ReliefF 5. Chi-square


Result: Statistical Validity


Number of instances 100% 6,828

Correctly predicted instances

71.98% 4,915

Incorrectly predicted instances

26.02% 1,916

Success rate

Discussion: Ranking

Observed discrepancy in ranking between selector and decision tree

Higher ranks supports the higher branch

Seepage(3rd), slope dip (5th), rainfall (4th) and fracture orientation (1st) control the tree

Not all highly ranked control the structure of the tree: lithology (2nd)

Discussion: Ranking

Attribute selector credits

homogeneity of the divided data

higher amount divided

ability to make a unique classification

Height, length, natural slope dip

homogeneous, but the divided records are minimal

Contributes little in making the classification unique

Should not be part of sequential ranks

Discussion: Slope Factors

Slope Height: A higher rock slope is likely to generate unstable block geometry,

DT rules did not show this

But engineered slopes are designed and constructed such that height is a critical factors in reinforcement selection.

Detection of common slope stability phenomenon Failure when seepage is noted

Lower slope angle, raises safety factor. Further branching were created using rainfall and lithology rendered safe

rules

BUT has high failure rate 72% in higher slope dip, 48% in lower

Discussion: Slope Factors

Rainfall: Heavy and continuous rainfall could trigger failure

Detected by DT

Limitation: no failure date, cant correlate with the exact amount of rainfall

Limitation: Rainfall data should be improved

Dip direction of fractures Fracture dipping towards slope face, fails slope

Fracture dipping away, safe slope

Discussion: Complex Rules

Slope failure is governed by not just one but a combination of many 3b: If lithology is schist:

slopes with notches fail

Slopes without are safe

If lithology is banded gneiss

Higher rainfall fails Presence of notches, high intensity of weathering

Lower rainfall is safe

Conclusion

Limitations of the study

Safety of the engineered slope has to be evaluated by detailed site specific information

Depending on the trained rule, not all slope factors were used in classification

However, this DB simplifies such information extensively

Generated174 slope rules

to be used to preliminary studies

Conclusion

Commonly known rules If seepage is present, then slope fails

If rainfall has greater amount, then slope fails

If slope dip is higher, then slopes fails

Complex rules If rainfall is high, slopes with banded gneiss tend to

fail.

If rainfall is less, banded gneiss is safe

If the intensity of weathering is high, slope with banded gneiss fail but slopes with lesser intensity are safe.

Least important slope factors


Conclusion

Despite limitations, the rules predicted 61% of safe and 80% of failed slope, and predicted 72% of combined safe and failed slope

Gathering massive slope data is expensive, requires a lot of effort, suffers from inconsistencies of survey, such people become skeptical for failure analysis studies

72% is encouraging

Rules to be tested in new observation

Conclusion

Important slope attributes are ranked orientation of major fractures, lithology, seepage,

rainfall, slope dip, notches, domain, lithology, topography, weathering

Least important attribute Natural slope, height, length


Thank you all!

Thank you too!

Department of Civil, Environmental and Railroad Engineering, Paichai University, Doma 2-dong, Seo-Gu, Dajeon, South Korea Geotechnical Research Group, Highway and Transportation Technology Institute HwaSeong-Ri, GyeongGi-Do, South Korea Korea highway corporation (KHC) Seongnam-Si, GyeongGi-Do, South Korea Korea Institute of Geological, Mining and Materials( KIGAM ) Gajeong-dong, Yuseong, Daejon South Korea Korea Meteorological Administration (KMA) Dongjak-Gu, Seoul, South Korea WEKA 3.2 Department of Computer Science, University of Waikato Knighton Road, Hamilton, New Zealand

kMining.com (http://www.kmining.com)

slope failure prediction using a decision tree: a case of...

Documents