Optimal Troubleshooting for Electro-mechanical Systems

When a complex electromechanical system fails, the troubleshooting procedure adopted is often complex and tedious. No standard methods currently exist to optimize the sequence of steps in a troubleshooting process. The ad hoc methods generally followed are less than optimal methods and can result in high maintenance costs. This paper describes the use of behavioral models and multistage decision-making models in Bayesian networks for representing the troubleshooting process. It discusses advantages in using these methods and the difficulties in implementing them. An approximate method to obtain optimal decision sequence for a troubleshooting process on a complex electromechanical system is also described. 1. INTRODUCTION When a complex electromechanical system fails, the troubleshooting procedure that is adopted is often complex and tedious. Diagnosis is done in a series of tests and component replacement actions until enough information is obtained to isolate the fault and to bring the system back into operating condition. As the number of components increases, the costs incurred and the time spent fixing the system also increases. In troubleshooting a complex electromechanical system, it would be helpful to have a tool that suggests the most appropriate test to be carried out or the component to be replaced at each stage of the diagnostic process. This research uses the bleed air control system of the Boeing 737NG airplane as an example application and investigates the possibility of developing such a tool using Bayesian Networks. The bleed air control system has a history of high repair costs and is considered to be fairly complex, having several critical components that may have to be tested or replaced in the event of a failure. Most often the primary objective will be to repair the device, not just to determine what has gone wrong. At each stage of this process there may be many possible observations, tests and repairs that can be performed. In addition we may also have the option of calling a service: promoting the problem to a higher level of expertise that is guaranteed to be able to repair the device. Because these operations are expensive in terms of time and/or money, we wish to generate a sequence of actions that minimizes costs and results in a functioning device. This is known as an optimal troubleshooting plan. If there exists a methodology or tool that can suggest the optimal trouble shooting decision sequence, the costs involved can be reduced considerably. The need for such a tool has been the motivation for our research. 2. BACKGROUND AND MOTIVATION [Raiffa, 1968] describes the use of decision theory for solving problems involving decision making under uncertainty. Decision theory can be difficult to apply for complex diagnosis problems due to computational complexity. The application of Bayesian networks to diagnostic modeling can be more practicable than decision theory because of inherent assumptions about conditional independence. [Breese, et al. 1992] developed an expert system using probabilistic causal model for diagnosis of efficiency problems in a gas turbine. [D'Ambrosio, 1992] explained the application of Bayesian Networks for real-time decision-making. [Heckerman, et al.1995, 1996] introduced decision theoretic troubleshooting for making cost effective decisions. They used Bayesian networks for belief updating and diagnosis, generated a large set of problem instances, and developed a Monte-Carlo 1 Copyright © 2003 by ASME technique for estimating the troubleshooting costs for a given planer and domain. [Jensen.1996, Cowell, R.G., et al. 1999] describe the use of influence diagrams and Bayesian networks for optimal decision-making. 3. DESCRIPTION OF THE 737NG BLEED AIR CONTROL SYSTEM The 737NG bleed air control system (BACS) provides high-pressure air for use in cabin air conditioning, engine starting, lower cargo compartment heating and anti-icing systems. It bleeds air from the eighth and fourteenth stages of compressor on each side of the 737’s two jet engines. This air is routed through a heat exchanger called the pre-cooler where the bleed air is cooled with air from the engine’s fan. From the precooler, the air continues to the pneumatic manifold. The bleed air must be delivered to the manifold within specific temperature and pressure ranges. If the air was allowed to flow unrestricted, the manifold could be overheated and/or over pressurized. It is also important to be able to isolate the BACS on one engine from the other side if that system fails. A number of valves are used to regulate the air temperature and pressure and to insure that pressure is not lost through the BACS. The High Pressure Shut Off Valve (HPSOV) restricts flow from the fourteenth compressor stage when the pressure in the eighth stage is adequate. There is a low pressure Check valve (check) to prevent airflow into the eighth compressor stage. The pressure Relief Valve (PRV) located before the precooler, vents air to ambient if the system becomes over pressurized. The Fan Air Modulating Valve (FAMV) controls the rate of cooling airflow through the pre-cooler (Pclr.). The pressure Reducing and Shut Off Valve (PRSOV), which is after the pre-cooler, limits the air pressure supplied to the pneumatic manifold (man.). The PRSOV also provides over temperature protection for the manifold by reducing flow if the bleed air temperature is too high, and provides a checking function to prevent manifold pressure loss through the BACS. These components are interconnected with a series of ducts. The BACS has several sensors that are used to diagnose system failures. These sensors include analog readings of temperature and pressure at the pneumatic manifold. There are also switches that indicate when the PRSOV is closed, when the PRV is open, and when the HPSOV is open. 4. BAYESIAN NETWORKS A Bayesian network is a compact, expressive representation of uncertain relationships among parameters in a domain. It is a graphical model for probabilistic relationships among a set of variables. A Bayesian network consists of a set of variables called nodes and a set of directed arcs connecting them. The variables are connected based on their causal relationships. Bayesian networks can be used to obtain the information about some variables given the information on others. Each variable has a finite set of mutually exclusive states. The variables together with the directed arcs, forms a directed acyclic graph (A directed graph is acyclic if there is no directed path A1→ A2→… An , such that A1= An.). Some of the variables are identified as the cause nodes i.e., those whose state can affect the state of other nodes. They are also called parent nodes and often their state cannot be directly known. Other nodes represent the end result produced because of the state of the cause nodes; they are called the effect nodes and normally can be readily observed and help in obtaining some kind of information regarding the status of the cause nodes, so they are also called information nodes. These information or effect nodes do not have any child nodes connected to them. The directed arcs indicated the causal influence from the parent node to the child node. For each variable v with parents p1,…,pn there is specified a conditional probability table P (v/ p1, …, pn). This conditional probability table gives the information or decides what state the variable v takes for a given set of states taken by it's parent nodes. Figure 1 shows a very simple Bayesian network. The node A is called the Parent or the cause node and the node B is called the child or the effect node. The directed arc connecting them represents the causal relationship between them. The node A has two states; True/False and B have three states above normal, below normal, Normal. The associated probability tables look like those shown in Fig 2. Figure 2(a) shows the associated prior probabilities of variable A in states true /false. Figure 2(b) shows the table associated with variable B.