Introduction

I.1. Problem statement

Since the beginning of the 20th Century, our perception of technological artifacts has continued to evolve and ranges from complicated systems to complex systems combining technical, human, organizational and environmental components. The first major accidents, such as the Flixborough blast [DEP 75], the Three Mile Island accident [KEM 79] or the toxic cloud at Seveso [SEV 82] that occurred at the end of the 1980s, have highlighted the role played by humans and their organizations in the failure of technical systems. The analysis of these accidents helps us to understand that the components are not independent. Thus, they have to be considered jointly to assess the true risk presented by our industrial systems.

Nowadays, the technological objects used are taken in their environment and are defined as complex socio-technical systems. The increase in complexity is due to the complexity emerging from the interdependence between technical, human, organizational and environmental components. A system is a set of elements interacting with each other with shared dependencies.

Faced with regulations, industrial systems are now required to have a high level of risk management. This level should be continuously demonstrated and proved [DE 12]. It is also necessary to handle socio-technical systems from a global point of view. To cope with this, dependability analysis and decision-making methods should be improved. Dependability analysis is primarily focused on technical aspects to assess the intrinsic safety of a system and should take into account human, organizational and environmental factors. Companies need good assessment tools to determine their requirements. They also need to anticipate future trends/developments to better manage or optimize the consequences of their activity on people, goods and the environment, as well as their social and societal impact.

To manage the industrial system, the engineers produced models during the lifecycle of a plant to predict its future functioning states. These models are extracted from our knowledge of the systems and our objective. There are several aims in dependability analysis. For instance, the following items are particular areas of interest in the Research Center for Automatic Control of Nancy (CRAN) in the University of Lorraine1:

• - models that can assess the impacts of maintenance activities on the ability to maintain operational conditions and to aid decisions during maintenance;
• - models that can assess the impacts of control and pilotage activities on wear, degradations (faults) or failures of all or part of the system. Thus, they satisfy the main goals, i.e. service quality, low-risk situations for users, staff, environment, etc.;
• - models that help assess the efficiency of means to warrant an acceptable level of risk, whatever the operational constraints and environmental perturbations.

There are several domains of application of dependability. These domains can be associated with different aspects of system functioning such as management modes, governance, human factors, extreme events or rare events and their consequences on society, maintenance, control and supervision or risk reductions of socio-technical systems, etc.

Nevertheless, most engineers have neither the tools nor the methods to effectively understand the whole information set (knowledge and evidence) according to the operational constraints and disturbances that condition the functioning of socio-technical systems. This is the main paradigm regarding the management of socio-technical systems. The phenomena encountered are so complex, as a result of their heterogeneity and the number of nested mechanisms of different natures, that it is quite difficult to continuously meet the required objectives or levels of performance. Moreover, there are no exact analytical models that can describe all the phenomena encountered. It is also impossible to know all of the system states and to know and observe all of the component states at each point in time in order to determine the optimal decision. Engineers should bear in mind that all models are biased and partial. As a result, engineers need new methods to solve these modeling problems.

It is necessary to model systems and their components with a finite but unbounded set of states or performance levels, i.e. with multiple state systems. In addition, the component behaviors are conditioned by the operational constraints and environmental disturbances of the system. In such cases, dependability assessment becomes difficult because it should take into account the combining effects of dependent failures due to constraints, disturbances and the intrinsic multi-state nature of system components. This results in an increasing quantity of scenarios to model. It is cumbersome for the analyst and enforces bias and partiality.

Quantitative assessments are necessary to warrant the viability of systems and their performance regarding risk and dependability. It is thus necessary to handle an uncertain representation of the system to describe its functioning and dysfunctioning. This imperfect perception naturally leads to a probabilistic view of system states. The main difficulties are the integration of a huge amount of information to model industrial or socio-technical systems subject to a large set of interactions with its environment. To contribute to the solution of this modeling problem, this book shows the application of graph theoretic and probabilistic approaches using Bayesian networks (BN) in maintenance, risk analysis and management, as well as in control theory. Why is this choice made? In 2004, the Massachusetts Institute of Technology (MIT) published the rank of the next ten main revolutions in the industrial area. The use of BN was in fourth position.

In this book, the authors aim to formalize probabilistic graph model approaches like BN to solve different kinds of problems associated with dependability, maintenance and risk in complex systems. The book is oriented towards applications and the transfer of modeling methodologies to the industry and engineers. It does not focus on the algorithmic point of view, but on ways in which to build models for the dependability analysis of industrial systems. The book is inspired by some industrial problems the authors helped to solve with BN in several sectors.

The approach followed in the book is to link mathematical formalisms and their uses in industrial cases, industrial needs and the possibility of models. To do this, it is necessary to:

• - generalize and homogenize the scientific paradigms from several industrial problems in an inductive way and then to specify solutions to applications in risk, reliability, maintenance and control;
• - formalize and adapt methods to solve new industrial needs;
• - promote these methods beyond the specialist community and adopt the standardization processes to integrate these methods.

In addition, two classes of problems need to be solved:

• - the modeling of risk management, maintenance and reliability for socio-technical systems;
• - the integration of knowledge of reliability in the control and diagnosis of automated systems.

The objectives of the book are to contribute to:

• - modeling of complex systems to aid decisions in an uncertain context by proposing an efficient modeling method for the new challenges as socio-technical considerations;
• - taking into account the propagation of uncertainties in complex system models, especially the uncertainty due to unknown future operational conditions;
• - joint assessment of multi-sectorial risks by integrating organizational and human risks in the analyses;
• - managing the aging of components, by anticipating the maintenance and safety actions with respect to aging;
• - increasing the operational time with the objective of ensuring safety, risk management and quality even if faults or failures occur;
• - allocating control efficiently to reduce the risk according to component criticality, health state and operational conditions.

To address these key points, the probabilistic framework and, particularly, BN formalism are used. BN are not completely accepted by the industry because the formalism does not handle a dependability-oriented semantic. In addition, the BN formalism is not proposed in usual standard modeling methods such as fault tree analysis2, Markov chain3, reliability block diagram4, event tree analysis5 and petri nets6.

Currently, engineers mainly ask for the proof of correctness of models and results. The proof of computing correctness has been described in [PEA 88]. Thus, the remaining question concerns the correctness of the model built by the analyst. As BN offer a generic modeling framework, the subsequent question is how to attach a semantic to the model through a modeling methodology which is independent of the application. Moreover, the models obtained should achieve at least the same results and possibilities as well-recognized approaches. So, one objective of the book is to show and promote BN as a reference method.

I.2. Book structure

The scientific goal of this book is to formalize the probabilistic models of system functioning and dysfunctioning. The models built serve to assess the satisfaction of operational requirements and performances and the safety requirements through reliability and risk analyses. For this purpose, a system-centered model has to take into account the...