Bogotá - Ingeniería - Doctorado en Ingeniería - Ingeniería Civil · 2023
Bayesian network methodology for decision support in forensic geotechnical engineering
Recent advances in engineering have increased the community’s expectation for civil engineering works to operate safely. Occasionally some of these works fail because of human errors or the unpredictable behavior of materials. Forensic engineering is the branch of forensic science in charge of investigating those engineering failures. Scientific methods used in forensic engineering guarantee that conclusions regarding the causes of an engineering failure come from reliable investigation processes. However, in the case of geotechnical failures, the inherent uncertainty of soil/rock materials, difficulties in evidence collection, and multiplicity of failure scenarios (hypotheses) pose a challenge in identifying the actual causes of failure. Therefore, conclusions about the causes of geotechnical failures sometimes seem arbitrary and biased because they are mainly based on expert judgment. Bayesian probabilistic tools can support decision-making about the causes of geotechnical failures. This thesis presents a Bayesian methodology for decision support in forensic geotechnical engineering based on two probabilistic techniques: Bayesian inference via posterior odds ratio and Bayesian Networks. The methodology compares probabilistically the hypotheses formulated as causes of failure and evaluates the influence of the amount of information (evidence) included in the analysis. Two benchmark problems and a case study were used to validate the applicability of the methodology. The results show that the Bayesian methodology identifies the most likely cause of a geotechnical failure, even when the amount of evidence is sparse. The use of the proposed methodology improves decision-making processes related to the causes of geotechnical failures. (Texto tomado de la fuente)
Full text 236 pages with text out of 240
Read the full thesis Record in the repository
Contents
- Introductionp. 21
- Backgroundp. 21
- Problem Statementp. 22
- Aim of the Thesisp. 23
- Outline of structurep. 23
- Introduction to Bayesian Inference and Bayesian Networksp. 25
- Bayesian Statisticsp. 25
- 2.1.1 Uncertainty and probabilityp. 25
- 2.1.2 Probability: Frequentists vs. Bayesiansp. 27
- 2.1.3 Bayes’ Theoremp. 28
- 2.1.4 Hypotheses Comparison: Bayes Factor and Posterior Oddsp. 30
- Bayesian Networksp. 25
- 2.2.1 Bayes’ Theorem and Bayesian Networksp. 33
- 2.2.2 Basic definitionsp. 33
- 2.2.3 Connection typesp. 34
- 2.2.4 D-separation and Markov Blanketsp. 37
- 2.2.5 From Causal Graphs to Bayesian Networksp. 39
- 2.2.6 The Chain Rulep. 40
- 2.2.7 Causal Graph Constructionp. 40
- 2.2.8 Eliciting Conditional Probability Distributions – CPDp. 46
- 2.2.9 Inference and Belief Updatingp. 50
- 2.2.10 Most Probable Explanationp. 53
- Bayesian Networks in Civil Engineering Applicationsp. 54
- Summaryp. 56
- Forensic Science and Forensic Engineeringp. 58
- Introduction and definitionsp. 58
- Characteristics for the development of a forensic engineering studyp. 59
- Use of the Scientific Methodp. 60
- Inductive, Deductive, and Abductive Processesp. 61
- A review of some general methods used in forensic engineeringp. 63
- Forensic Geotechnical Engineeringp. 22
- 3.6.1 Stages of the forensic geotechnical processp. 66
- 3.6.2 Additional aspects to considerp. 69
- discussionp. 185
- Summaryp. 56
- Proposed Bayesian Methodology for Decision Supportp. 79
- 4.1.1 Stage 1. Preliminary stepsp. 80
- 4.1.2 Stage 2. Constructing the Probabilistic Failure Modelp. 85
- 4.1.3 Stage 3. Probabilistic Hypothesis Comparisonp. 89
- The ERTC7 Benchmark Exercisep. 95
- 4.2.1 General description of the ERTC 7p. 95
- failure scenario)p. 99
- 4.2.3 Stage 2. Constructing the Probabilistic Failure Modelp. 85
- 4.2.4 Stage 3(a). Probabilistic Hypotheses Comparison via POR Techniquep. 105
- 4.2.5 Stage 3(b). Probabilistic Hypotheses Comparison Using BN Techniquep. 112
- The Breitenhagen Levee Failurep. 116
- 4.3.1 Geometry, Geotechnical Conditions, and Previous Forensic Studiesp. 116
- 4.3.2 Stage 1: Preliminary Steps. Failure hypotheses and Collected Evidencep. 118
- 4.3.3 Stage 2: Probabilistic Failure Modelp. 120
- 4.3.4 Stage 3 (a). Probabilistic Hypotheses Comparison via POR Techniquep. 122
- 4.3.5 Stage 3 (b). Probabilistic Hypotheses Comparison via Bayesian Networksp. 126
- Summaryp. 56
- Application of Bayesian Methodologyp. 138
- General Description and Geotechnical Conditionsp. 138
- 5.1.1 General descriptionp. 138
- 5.1.2 Geotechnical Characterization of the Sitep. 139
- 5.1.3 Structure, Foundation, and Construction Processp. 144
- 5.1.4 Monitoring System and Failure Descriptionp. 145
- Application of the Bayesian Methodologyp. 147
- 5.2.1 Stage 1: Collected Evidence and Hypotheses Formulationp. 147
- 5.2.2 Stage 2: Random Variables and Probabilistic Failure Modelp. 150
- 5.2.3 Stage 3. Bayesian Hypotheses Comparisonp. 156
- 5.2.4 Comparison of POR and BN Techniquesp. 181
- Summaryp. 56
- Discussionp. 185
- Key Findings and Implicationsp. 185
- 6.1.1 Improvement in Estimating Failure Causesp. 185
- 6.1.2 Beyond Traditional Back Analysisp. 186
- 6.1.3 Two Bayesian Techniques for Hypotheses Comparisonp. 187
- 6.1.4 Use of Standard Knowledge and Geotechnical Jargonp. 188
- Limitationsp. 189
- 6.2.1 General Limitationsp. 189
- 6.2.2 Posterior Odds Ratio (POR) Techniquep. 190
- 6.2.3 Bayesian Networks (BN) Techniquep. 191
- Conclusions and Suggestions for Future Workp. 195
- Conclusionsp. 195
- 7.1.1 Diagnosis of Standard Practices in Forensic Geotechnical Engineeringp. 195
- Odds Ratio and Bayesian Networksp. 196
- Suggestions for Future Workp. 199
- Referencesp. 201
- Modified from Baecher & Christian (2003)p. 26
- Figure 2-2.: Graphical representation for interpreting the BF and posterior odds valuesp. 32
- Figure 2-3.: Example of a simple causal graph (DAG)p. 34
- i). (a) Node A is instantiated, (b) Node C is instantiated, and (c) Node B is instantiatedp. 36
- instantiatedp. 36
- instantiatedp. 36
- Figure 2-7.: Markov blanket for node 𝐼p. 38
- connectionsp. 39
- examplep. 42
- earthquake, (c) Inward movement of a diaphragm wall, and (d) dam failurep. 44
- Figure 2-12.: Example of a BN constructed from a definitional idiomp. 46
- Figure 2-13.: Example of a BN constructed from an induction idiomp. 46
- Figure 2-14.: Example of a BN constructed from a reconciliation idiomp. 46
- values (Kjærulff & Madsen, 2013)p. 49
- publications by year. (b) Cumulative publications by yearp. 54
- Figure 3-1.: Abductive process (modified from Campos et al., 2001)p. 63
- Figure 3-2.: The Investigation Pyramid (Modified from Noon, 2009)p. 64
- Figure 3-3.: The investigative process proposed by Bell (2000)p. 65
- Figure 3-4.: Typical steps of a forensic investigation suggested in ASCE (2018)p. 65
- Greenspan et al. (1989)p. 75
- Engineeringp. 30
- Figure 4-2. Typical convergence plot for 𝑃(𝑓) and standard deviation of 𝑃(𝑓)p. 89
- Figure 4-3. The ERTC7 benchmark exercise (Modified from Schweiger, 2006)p. 96
- modelingp. 98
- Figure 4-5. The ERTC7 hypothetical failure scenariop. 99
- mass functions (PMF) for continuous and discrete random variablesp. 103
- Factor of Safety (FoS)p. 104
- Figure 4-8. ERTC7 Benchmark exercise: Convergence plot for 𝑃(𝑓)p. 104
- Figure 4-10. Prior and posterior odds values for 𝐻1 vs 𝐻0. Evidence 𝑑1 and 𝑑2p. 108
- 𝐻3 vs 𝐻1, (e) 𝐻3 vs 𝐻2, (f) Comparison summary of H𝑖 𝑣𝑠 H𝑗 for 𝑂𝐻ij = 1.0p. 111
- excavation failurep. 106
- and evidence nodesp. 114
- stratigraphy (Modified from Kool et al., 2019)p. 117
- Mohr-Coulomb, (b) Undrained model SHANSEPp. 122
- 1.0p. 124
- failurep. 18
- and SGp. 127
- circles represent evidence and hypothesis nodes, respectivelyp. 129
- and a Mohr-Coulomb soil constitutive modelp. 131
- conditions and SHANSEP constitutive modelp. 132
- Drained condition (Mohr-Coulomb), (b) Undrained condition (SHANSEP)p. 133
- (Modified from Google Earth, 2022)p. 139
- developed during the forensic investigation (Modified from Google Earth, 2022)p. 140
- 2012)p. 19
- (Unal, 2012)p. 142
- Figure 5-5. Stratigraphy and simplified scheme of the Green Office Projectp. 144
- construction stage (Modified from Google Earth 2022)p. 146
- elevationp. 151
- Projectp. 153
- Figure 5-9. Green Office Project: Histograms of the Factor of Safety (FoS)p. 155
- Figure 5-10. Convergence plots for FoS and standard deviation (FoS)p. 155
- Figure 5-11. Flow chart used to construc the Python script (Green office FEM)p. 156
- (1:1)p. 160
- Simple DAG, (b) Final version of the DAG fragment with the FoS as a mediating nodep. 162
- measured at point Bp. 163
- simple cause-effect idiom. (b) Expanded DAG fragment with indicator nodesp. 164
- and green nodes represent evidence, hypothesis, and mediating nodes, respectivelyp. 165
- conditional probability tables (CPTs)p. 168
- (i.e., evidence not included in the nodes)p. 169
- evidence: 𝑃𝑧= 0.0 𝑚 𝑙𝑒𝑐𝑡𝑢𝑟𝑒, 𝐼𝑛= 𝑁𝑜 𝑑𝑎𝑚𝑎𝑔𝑒p. 171
- BN for evidence: 𝑃𝑧= 0.0 𝑚 𝑙𝑒𝑐𝑡𝑢𝑟𝑒, 𝐼𝑛= 𝑁𝑜 𝑑𝑎𝑚𝑎𝑔𝑒, 𝐸𝑜= 𝑜𝑏𝑠𝑒𝑟𝑣𝑒 𝑓𝑎𝑖𝑙𝑢𝑟𝑒p. 172
- 0.3 𝑚, 𝐸𝑜= 𝑜𝑏𝑠𝑒𝑟𝑣𝑒 𝑓𝑎𝑖𝑙𝑢𝑟𝑒p. 174
- Figure 5-22. Monthly rainfall from 2007 to 2011 in the project area (Unal, 2012)p. 175
- 2012)p. 19
- on 11th streetp. 176
- Figure 5-26. Updated BN after the inclusion of actual evidence:p. 179