Agent-based Simulation of Maintenance for System Optimisation

The thorough planning and implementation of maintenance strategies save time and cost. However, the integration of maintenance actions into reliability analysis often leads to system models which are regarded as unmanageable and therefore refused at operational level. Common aggravating factors in modelling are state-explosion and restrictive model assumptions. From the authors’ point of view, agentbased modelling (ABM) of technical and organisational systems is promising to overcome some of the problems mentioned above. But ABM is uncommon in reliability analysis and its feasibility in this area has still to be shown. In doing so, ABM is compared with Markov modelling principles at first. The modelling of complex systems is implemented by the transition from a basic configuration to n uniform units. Although ABM as well Markov models add up in the same numerical results the former already highlights the potentiality of an improved system state handling. Subsequently, ABM extensions outreach Markov capabilities, e.g. operators are modelled as “agents” featuring their location (x; y) and availability (0; 1) among others. The planning of operator schedules is discussed. In addition, the ABM approach enables the implementation of various parameters (e.g., cost). As the Monte Carlo method is the general designation for stochastic simulation using random numbers, the affinity of ABM with Monte Carlo models (MCM) is also discussed. Identity is the property of an object that distinguishes it from other objects (e.g., safety valve). State describes the data stored in the object (e.g., open, closed) and behaviour specifies the methods (functions) in the object's interface by which the object can be used (e.g., valve opens when a specified threshold pressure is exceeded). According to (Jennings 2000) the general definition of the term agent provided by (Wooldridge 1997) becomes increasingly accepted. It considers an agent as an element which has a set of design objectives and is capable to autonomously perform flexible actions in a dynamic and unpredictable environment Ω in order to meet its goals: Def 2: Agents: objects with extended capabilities. These capabilities embrace rules of behaviour, autonomy, cooperation, mobility, memory, learning abilities, among others. Cooperation, which is considered as a core capability of an agent, comprises, e.g., perception, action and communication. 3 ABM IN RISK RESEARCH Where ABM is a well established concept in various fields (e.g. conflict research, urban planning, traffic modelling), it is not wide spread in risk assessment. An extended literature search gives information about the state of the art of ABM in the area of risk, reliability and hazard analysis as defined, e.g., in (ISO/IEC-Guide73 2002) and (DIN-40041 1990). The literature is mainly extracted from (WebSPIRS 2005; ISI 2006; ScienceDirect 2006), covering the core literature of the engineering field. Filter criteria are (a) publication year from 2000, (b) the methods described are clearly associated with risk assessment, (c) the methods described are allocatable to the principles of Definitions 1 and 2. However, the differentiation of the terms agent-based and objectoriented is often ambiguous in papers. It is not transparent to the reader whether the phrase “objectoriented” is used in strict terms of software development or it simply stands for modularisation. Three definitely object oriented and agent-based approaches to risk assessment are given in a) (Wyss and Duran 2001; Wyss and Duran 2002; Wyss, Duran et al. 2004), b) (Diergardt 2006; Kaegi, Mock et al. 2006) and c) (Weidl, Madsen et al. 2005). The latter methodology “integrates decision-theoretic troubleshooting within risk assessment for industrial process control” and advice on sequence of repair by probability-cost estimation is given. This methodology uses generic object-oriented Bayesian networks. The methodology and tool introduced in (Kavcic and Juricic 2001) is a computer-aided prototype which “automatically derives different representations of fault propagation: fault trees, fault event trees and a diagnostic rules table” for industrial processes. The methodology expressly follows an “object-oriented modelling paradigm”. The term agent is not used. Another object-oriented approach gives (Teoh and Case 2004) in order to automatically create an FMEA (Failure Modes and Effects Analysis) diagram. There, entities (objects in accordance with Definition 1) are “able to act or respond to the system through its distinctive ‘memory’ ” which follows Definition 2. An older paper from 1996 (Leone 1996) follows the same paradigm to construct a knowledge based system for Hazard and Operability Studies (HAZOP). Although not explicitly mentioned, the specified interplays among compound objects (chemical compounds, energy sources, environmental conditions) likely follow the definition of an agent. There are a few papers dealing with ABM of maintenance management and scheduling, e.g., (Zhou, Fox et al. 2004; Cerrada, Cardillo et al. 2006). However, the results rather comprise the analysis and optimisation of maintenance processes and do not provide any probabilities and other risk figures. A number of U.S. organisations (both commercial and governmental) are involved in the development of new models for analysing critical infrastructure interdependencies also related to risk assessment aspects. An extended list of such projects is shown in (Pederson, Dudenhoeffer et al. 2006). Some of these approaches comprise ABM but astonishingly, there is almost no scientific publication visible about their achievements. Furthermore, these developments mainly result in the design of new and very complex simulation tools which makes it difficult to verify their results and conclusions. Above all, some researchers in the field of financial risk assessment have discovered the promising potential of ABM to model and forecast the behaviour financial markets (Krawiecki, Holyst et al. 2002; Brown, Beyeler et al. 2004; Alkemper and Mango 2005; Krause, Andersson et al. 2005). Another example of an achievement related to safety analysis is (Helbing, Farkas et al. 2000). The author investigated a situation of escape panic in an overcrowded area with limited exits. Using an agent based simulation model of human behaviour, it was found that placing an obstacle in front of an emergency exit can increase the efficiency of such an exit due to unexpected self-organisation. 4 FROM MARKOV TO ABM A field of using the ABM approach in risk assessment is the availability computation of complex systems. In order to illustrate the extended capabilities of ABM, a system of n identical and independently operated units will be analyzed by means of Markov chains, Monte Carlo simulation and ABM. The unit states are Boolean (operating/non-operating). All simulations in this paper are implemented with AnyLogic (Borshchev, Karpov et al. 2002).