Risk Analysis, Vol. 34, No. 10, 2014

DOI: 10.1111/risa.12213

Using Prior Risk-Related Knowledge to Support Risk Management Decisions: Lessons Learnt from a Tunneling Project 1,∗ ´ Ibsen Chivata´ Cardenas, Saad S. H. Al-Jibouri,1 Johannes I. M. Halman,1 Wim van de Linde,2 and Frank Kaalberg3

The authors of this article have developed six probabilistic causal models for critical risks in tunnel works. The details of the models’ development and evaluation were reported in two earlier publications of this journal. Accordingly, as a remaining step, this article is focused on the investigation into the use of these models in a real case study project. The use of the models is challenging given the need to provide information on risks that usually are both project and context dependent. The latter is of particular concern in underground construction projects. Tunnel risks are the consequences of interactions between site- and projectspecific factors. Large variations and uncertainties in ground conditions as well as project singularities give rise to particular risk factors with very specific impacts. These circumstances mean that existing risk information, gathered from previous projects, is extremely difficult to use in other projects. This article considers these issues and addresses the extent to which prior risk-related knowledge, in the form of causal models, as the models developed for the investigation, can be used to provide useful risk information for the case study project. The identification and characterization of the causes and conditions that lead to failures and their interactions as well as their associated probabilistic information is assumed to be risk-related knowledge in this article. It is shown that, irrespective of existing constraints on using information and knowledge from past experiences, construction risk-related knowledge can be transferred and used from project to project in the form of comprehensive models based on probabilistic-causal relationships. The article also shows that the developed models provide guidance as to the use of specific remedial measures by means of the identification of critical risk factors, and therefore they support risk management decisions. Similarly, a number of limitations of the models are discussed. KEY WORDS: Bayesian networks; epistemic uncertainty; knowledge engineering; modeling risks; riskrelated knowledge reuse

1. INTRODUCTION Failures associated with collapse or excessive deformation of significant parts of works that lead to damage or injury are an ongoing concern in underground construction projects. According to the International Association of Engineering Insurers, more than €570 million of economic losses were incurred in 18 tunneling projects worldwide between 1994 and 2005 due to such failure events.(1) In geotechnical projects, abundant evidence has highlighted that a

1 Department of Construction Management and Engineering, Uni-

versity of Twente, P.O. Box 217, 7500 AE, Enschede, The Netherlands. 2 B.V. Kanaalkruising Sluiskil, Koegorsstraat 1C, 4538 PK, Terneuzen, The Netherlands. 3 Witteveen + Bos Raadgevende ingenieurs B.V., P.O. Box 233, 7400 AE, Deventer, The Netherlands. ∗ Address correspondence to Ibsen Chivata ´ Cardenas, ´ Department of Construction Management and Engineering, University of Twente, P.O. Box 217, 7500 AE, Enschede, The Netherlands; ibsen [email protected].

1923

C 2014 Society for Risk Analysis 0272-4332/14/0100-1923$22.00/1 

1924 significant proportion of these failures are a result of flaws in accessing and using the available knowledge and information, rather than a consequence of unknown factors or unexpected events such as unforeseen ground conditions.(2–6) Risk management (RM) can play a significant role in reducing the occurrence of these kinds of failures; for example, in the form of a tool to identify critical project risks and to then determine appropriate measures to manage them. Nevertheless, RM still has tremendous challenges to face if it is to prove its effectiveness in considerably reducing failures. For instance, Choi and Mahadevan(7) have indicated that risk assessment remains difficult for practicing engineers due to the requirement for data on a large number of input variables. This is especially true with underground construction projects since these usually involve many tasks and people and, consequently, the risks are also numerous. Paradoxically, risk information related to construction is often scarce, seldom documented, and in many cases unavailable.(8,9) Another factor that exacerbates these circumstances is that risks associated with underground construction failures are caused by factors that are highly interactive and might be little known or cannot be known with certainty.(10) For example, some ground conditions such as the exact nature of possible water inflows and the positions of important ground interfaces can be unpredictable.(3) Such constraints make RM in this type of work a very intractable process, and this might well have discouraged practitioners from using complicated risk analysis techniques(11,12) and instead led them to adopt simplistic analyses that rely largely on uncorroborated personal judgment to derive risk measures, which can be misleading when it comes to decision making. The conventional approach has largely been based on quantitative risk analysis to determine construction strategies in terms of project time and costs.(10) These traditional techniques do not stimulate practitioners to develop indepth understanding of the mechanisms of failures. Furthermore, these standard techniques have possibly hampered the systematic gathering of information on risks, remedial measures, and lessons learned from previous projects that could be reused when developing new projects.(13) To address the latter point, Tah and Carr(13) introduced the concept of a knowledge-driven RM process to the construction field. This entails the use of existing knowledge gained from past projects to support RM in new ones. The idea of using risk-related knowledge is seen as a trigger for using additional en-

´ Cardenas et al. gineering knowledge. (In this article, knowledge on the causes and conditions that lead to failures and their interactions as well as their associated probabilistic information is assumed to be risk-related knowledge.) In principle, it is possible to distinguish two basic and complementary strategies for accessing and using existing knowledge: a codification strategy and a personalization strategy. The first strategy is about codifying the knowledge and storing it in, for instance, databases. The personalization strategy incorporates the sharing of knowledge through personal interaction. Databases and repositories function as bodies of knowledge, or organizational memories, where experiences about risks and responses are continuously recorded.(13) Repository use has been particularly encouraging in addressing the issue of incomplete information on construction risks, but repositories have so far failed to be fully useful in RM. According to Dikmen et al.,(14) these standard instruments currently provide captured knowledge from past projects in terms of single facts, but without informing on the causes, conditions (risk factors), and possible interactions related to the failure events. Information captured, as it now is, is consequently difficult to reuse. This is also noticed by Marle and Vidal(15) and Marle et al.,(16) who indicated that risks in projects are generally managed thanks to the use of single tools that do not correctly show interactions among risks. (Note that throughout this article, risks are assumed to be unfavorable events influenced by linked factors and their interactions. Accordingly, the term “risk” when used alone refers to a potential failure event and the factors influencing it are called here “risk factors.”) These circumstances, among others, have led to traditional databases or repositories not being successfully applied and, consequently, personalized approaches for sharing knowledge are more commonly preferred in the construction industry. In addition, the limitations of the risk-related knowledge captured in databases or repositories pose enormous constraints if it is used, for instance, in underground construction projects. Such knowledge needs to provide useful information in situations where a project, its environment, and therefore its risks are continuously evolving. As such, the project risks need to be continuously assessed. New risks have to be continually identified and analyzed. The assessment of future risk requires a continuous update of data compiled from immediate past experiences.(3) A systematic updating process for risk

Using Prior Risk-Related Knowledge to Support Risk Management Decisions information has already been addressed by Choi and Mahadevan,(7) who developed a risk assessment methodology for construction projects that combined large quantities of existing data with project-specific information through updating processes. The personalization strategy to reuse knowledge is, however, not without its own shortcomings and difficulties. A major problem with this strategy concerns the enormous amount of information that can be required to support RM. Under these circumstances, and because of the need to gather and analyze large amounts of information, it is likely that some relevant risk issues and scenarios will be overlooked and not taken into account.(6) The use of risk models that comprehensively integrate risk-related knowledge can prevent failure scenarios not being taken into account. (“Risk models,” in this article, are assumed to be composed of a set of known possible risk factors as input variables linked to the potential failure events [risks] being the output variables.) Further, to enable decision making about risks, risk models can facilitate the analysis of identifying relevant failure modes and disclosing possible opportunities to avoid or mitigate the occurrence of risks. Insofar as risk models inform about interactions among risk factors, conditions, and potential critical failure events, they seem to be an alternative to the limited functionality of the databases or repositories usually employed in the construction industry. Such models might provide relevant information to support RM by enabling professionals to share their knowledge and allowing others to make use of it. Unfortunately, risk models that include these types of features are not yet applied for tunnel works. Moreover, models that comprehensively make the interrelationships between risk factors and failure events explicit are absent from the literature on tunnel works. Recent advances in knowledge engineering are promising in aiding the capture, permanent update, and reuse of risk-related knowledge gained from past experiences.(17–22) Their contributions lie in the use of Bayesian belief networks (BBNs) in conjunction with expert judgment to model knowledge. BBNs organize the body of knowledge in any given area by mapping out cause-and-effect relationships among key variables and encoding them with numbers that represent the extent to which one variable is likely to affect another.(23) The BBN approach is essentially a framework for representing the relationships between variables and for capturing the uncertainty in the dependencies between these variables using

1925

conditional probabilities.(24) In this approach, both experimental data and expert judgments are used as source data for modeling knowledge. Judgments made by experts can be viewed as preliminary propositions about variables whose exact values are uncertain but that are known partially from undocumented experiences of the experts or inferred from other sources of knowledge.(25) In a previous stage of this research, the use of the BBN approach in conjunction with expert judgments was shown to be suitable for handling the problem of incomplete information and to cope with the uncertainty and complexity of construction risks.(26,27) Expert judgment has been deployed not only to augment the available information but also to reduce the complexity of risk characterization. Experts helped to identify a manageable number of critical variables relevant to failure events occurring. The expert judgments were collected using a structured elicitation method, and combined with available information using BBNs. The BBNs played two additional roles. First, they provided probabilistic consistency to both the judgments and the information available. Second, they facilitated risk analysis in developing further risk-related knowledge. In this research, such risk analysis not only contributed to delivering a comprehensive understanding of the risks involved, it also yields information on opportunities to control specific risks. These two characteristics have made the BBNs approach a desirable tool to model and transfer knowledge in comparison to competitive approaches such as fuzzy set modeling. In our investigation, BBNs have successfully been used to develop six models for critical risks in tunnel works. The details of the development of the models were reported in two articles published in this journal.(26,27) These two antecedent articles showed how constraints such as incomplete information on construction risks, as well as the epistemic uncertainty of risks, were addressed. Issues such as expert knowledge aggregation, propagation, and critical risk identification were thoroughly analyzed in these papers. This article, therefore, is focused on describing the results that were obtained in applying the developed models in a real ongoing tunneling project that was the remaining step in the investigation. To this end, the article provides evidence that construction risk-related knowledge can be transferred and used from project to project in the form of comprehensive models based on probabilistic-causal relationships rather than through statistics, or even probability

´ Cardenas et al.

1926 values, attached to isolated variables that are usually encountered in databases or repositories. Nonetheless, some relevant information as to the models’ development was included in this article to provide necessary background. The remainder of this article is divided into eight sections. An overview of Bayesian networks is given in Section 2. Section 3 provides an overview of the developed models for six important risks in tunneling projects. In Section 4, it is explained how the developed models were customized to be used in specific projects. Details of the case study are provided in Section 5 and following these is a discussion of the results (Section 6). In addition, the contributions and limitations of the proposed approach and the research are elaborated in Section 7. The last section describes conclusions of the research. 2. BAYESIAN BELIEF NETWORKS OVERVIEW BBNs organize the body of knowledge in any given area by mapping out cause-and-effect relationships among key variables and encoding them with numbers that represent the extent to which one variable is likely to affect another.(23) The BBNs approach is a suitable way of representing complex and uncertain relationships among many factors that contribute to the occurrence of risks.(28,29) A thorough analysis on the advantages and drawbacks of BBNs is reported by Liu et al.(30) The main reservation posed by Liu et al.(30) consists of the use of the axioms of probability theory to deal with ambiguous and uncertain information that is often encountered in many risk analysis problems. Fang and Marle(31) reported the limitations of BBNs in modeling cyclic causal interactions between risk factors, which is seen as a potentially misleading factor that could affect decision making. BBNs might be used to construct models composed of scenarios based on a set of known possible risk factors associated with the risks being analyzed. These possible scenarios can be structured as a set of mutually exclusive and collectively exhaustive elements to which a probability distribution can be attributed. In a BBN, the causal interrelationships between risk factors are represented by distributions of conditional probabilities.(19) The mechanism of inference used throughout a BBN is the Bayesian theorem. With this mechanism, it is possible to propagate the information available on a variable and to consistently estimate changes in the other network vari-

A

P(B|A)

B

Fig. 1. Probabilistic relationship between the variables A and B.

ables. With two directly related variables, the probabilities can be computed as follows:(32) P[effect] = [P[effect|cause]P[cause]]/ (P[cause|effect]),

(1)

where P[cause] is probability that the cause occurs, P[effect] is probability that the effect occurs, P[effect|cause] is the conditional probability of the effect, given the cause, P[cause|effect] is the conditional probability of the cause, given the effect. The posterior probability of the cause given the effect can similarly be derived as: P[cause|effect] = [P[effect|cause]P[cause]]/ P[effect].

(2)

The graphical representation for two directly related variables is depicted in Fig. 1. BBNs are expressed graphically in the form of diagrams. Risk factors are represented by nodes. Nodes that have interdependencies are connected graphically by arcs, whereas independent nodes are not connected. The direction of the arcs reflects the direction of causal influence, which might be scientifically known or else based on expert judgment. A BBN model produced in this research is shown in Fig. 2. The information on conditional probabilities reflecting the relationships between the risk factors in this model is not indicated on the diagram but is internally stored and accessible to any user. 3. OVERVIEW OF THE DEVELOPED MODELS In an earlier stage of this research, as described ´ in Cardenas et al.,(26,27) six risk models were developed for the following critical risks in tunnel works: (1) Face instability of bored tunnels in soft soils when using slurry shields. (2) Face instability of bored tunnels in soft soils when using earth pressure balance shields. (3) Excessive ground deformation leading to surface settlement in bored tunnels in soft soils.

Using Prior Risk-Related Knowledge to Support Risk Management Decisions

Excessive overburden pressure

Ground conditions

Large variations of permeability in the ground

Design factors

Overcutting in curved alignments

Insufficient soil cover depth

Excessive overcutting

Slow learning in driving the TBM from a shaft

Lower than specified grouting pressure

Grout hardening faster than expected

Grouting pipelines blockage

Failed cleaning of the grouting pipes

Grout pumps failure

Large ground strength/stiffness variation

Suboptimal structural design of the shield

Poor groutability of the ground

Excessive TBM steering movements

Sensitive ground disturbance of by stress changes

Excessive pore ground water pressure

Excessive Ko

Suboptimal shield geometrical specifications

Bentonite flow over TBM shield (face-tail)

Excessive deformations in the shield tail

Excessive tapering of the shield

Excessive ground shear by roughness cutting wheel Excessive volume loss Unplanned TBM stoppages

Incomplete or retarded filling of the tail gap

Excessive internal support pressure variation Insufficient nominal grout stiffness

Grouting process delays or disturbance Insufficient grout supply

Construction-related factors

1927

Inap. arrangement of the grouting injection system

Misleading monit. of surface ground movement Miscommunication operator – monit. office

Wrong estimation of grouting pressures

Lining-grouting operator mistake

Grout hardening slower than expected

Incomplete countermeasures

Response factors Reaction to a retarded or incomplete filling

Machine operator mistake

Fig. 2. Risk model for excessive volume loss leading to settlement in tunnels bored in soft soils.

(4) Excessive deformation, damage, or leakage in concrete linings. (5) Collapse, excessive deformation, and water inflow into shaft excavations. (6) Collapse, excessive deformation, and water inflow when excavating cross-passages in soft soils. These six models contain the following aspects:

r The r r

relevant risk factors that can lead to the failure events being considered. The plausible relationships among these risk factors. The strength of these relationships measured in terms of conditional probabilities.

The information used to develop the models was initially collected from risk databases, reports of failure events, and specialized literature on tunnel works. This information was complemented by experts’ judgments on relevant risk factors, plausible relationships among them, and estimates of

probability and conditional probabilities. Thirty-one experts involved in ongoing or past underground construction projects, such as bored tunnels and deep shaft excavations, in the Netherlands participated in this investigation. In the models, the relevant factors are assumed to be those identified as such by the experts consulted. The risk factors are viewed as deviations from the requirements of a tunnel project (assumptions, expectations, specifications, tolerances, and thresholds) that could cause risks to occur or exacerbate their impact. Since information on impacts related to construction risks, such as injuries or loss of life, damage to third parties, additional costs, delays in project completion, or failure to meet quality requirements, is very project specific, it is deliberately excluded from the risk models. As such, information on risk impacts has to be incorporated into the models on a case-by-case basis, as and when required. Face-to-face interviews formed the main method used to elicit judgments from experts. The elicitation

1928 process involved three rounds of interviews. The first round allowed the most important risks to be identified. A second round of interviews provided data on the relevant risk factors that could lead to the main risks under study. The plausible relationships among the risk factors were also initially identified in this round. The third and final round rendered information on the strengths of the relationships identified, assessed in terms of conditional probabilities. The final round of interviews was also designed to review divergences in offered information and to internally validate the data. Seven specialists took part in the first round of interviews, 24 and 11 in the second and third rounds, respectively. This means that a number of experts were able to participate in all of the three rounds. All the participating experts originated from the Netherlands or Germany. The experts all had a minimum of 10 years of tunneling experience. A further explanation of the experts’ elicitation method used in this research is available ´ from Cardenas et al.(26) Particular details of the models are provided in the following. Each variable in the models is regarded as an event or condition that represents a fault event, a state of failure, or an unfavorable condition. Fault events or states of failure associated with a variable can be events in which a risk factor exceeds a predetermined threshold. Accordingly, most of the variables have two possible states: “absent” or “present.” The “present” status was further discretized into five probability categories. In line with this, the identified experts were asked to provide estimates of conditional probabilities in terms of qualitative probabilities using a five-category scale consisting of categories labeled: “frequent,” “probable,” “occasional,” “remote,” and “improbable”; representing the probability intervals: >30%, 10–30%, 5–10%, 1–5%, and

Using prior risk-related knowledge to support risk management decisions: lessons learnt from a tunneling project.

The authors of this article have developed six probabilistic causal models for critical risks in tunnel works. The details of the models' development ...
827KB Sizes 0 Downloads 4 Views