In this thesis a fault diagnostic method has been developed using BNs. The method uses FTs for the purpose of building the networks. In this way, the advantages of both techniques are employed: systematic construction with FTs and the ability to introduce evidence in the evaluation with BNs. The aim of the research was to establish if BNs could be used effectively for system diagnostics using a model-based approach. BNs have shown with this application they are ideal in the field of fault diagnostics when used with a model-based approach. The diagnostics methodology was applied to two systems, the water tank system and the fuel rig system, and the results were validated with two simulation codes in C++. This chapter summarises the achievements of the research and suggests further work that can be continued in the same direction.
6.2
Summary
6.2.1 Achievements
The main achievement of the research was the creation of a general approach that uses BNs for a model-based fault diagnostic system. Literature on BNs is dominated by case- based diagnostic reasoning, where the probabilities of the networks are obtained by training them using historical data known about previous faults. The method developed gives a gen- eral, straightforward and structured way to build a class of BNs and to evaluate them. The aim of the strategy is detecting when the system is faulty and diagnosing the cause or causes.
The diagnostic system is built in two stages: the modelling and preparation stage and the FTs and BNs development. The modelling and preparation stage and the development of the FTs follows the approach taken from a previously developed diagnostic system that uses FTA. FTs are then converted into BNs and these are manipulated to form a BN that models the entire process. When the observed sensor readings deviate from those expected, the BN posterior probability is calculated and a list of potential causes is obtained. Introducing BNs has several advantages with respect to the FTs that can be grouped in two aspects: the graphical representation and the probability evaluation. The graphical representation of the BNs is more concise because repeated events can be avoided and because BNs allow more modelling solutions such as the introduction of nodes with several states that represent the system components. Regarding the evaluation, that is the calculation of the causes of a fault, with the FTA approach, a different FT is built for every faulty scenario of the system. The prime implicants are then calculated for this FTs and finally the importance measured are considered. Using BNs enables to evaluate the same model for every scenario of the system by simply introducing evidence on the nodes that represent the symptoms. The posterior probability gives a measure of the likelihood for a component failure to be the cause of a fault. The calculation of the updated probability is almost instant with the software used.
6.2.2 Characteristics of the method
One of the aims of the research was to be able to model dynamic behaviours in the system. This was done introducing sensor patterns in the analysis and creating nodes in the BNs whose states represent the sensor patterns. The patterns symbolises dynamic trends that are observable in the sensors, therefore they show how a monitored variable changes over time. This is an important dynamic aspect in the system because considering static sen- sor readings can give a limitation to the number of scenarios and causes that can be analysed. Another relevant aspect of the diagnosis is the fact that is a general approach. In principle, any system for which FTs can be built can be diagnosed using this strategy. The way the component failures cause a system fault must be understood and the failure probabilities of the components must be available.
6.2.3 The water tank system
The diagnostic method is applied to the water tank system with two approaches. First, the system is studied in steady state and the FTs and BNs are built without considering the dynamics of the system. This approach has shown to provide accurate results however it
was unable to take into account all faulty scenarios. For this reason, the method has been improved introducing dynamic sensor patterns. The scale of the problem was relatively small and a more complex system had to be studied to prove the method applicability.
6.2.4 The fuel rig system
The fuel rig system is considerably larger than the water tank system, in terms of number of components and failure modes considered. The method was improved considering a division into sub-systems and modifying it introducing a two stage “modularization” strategy. In the first stage, a BN is evaluated to understand which sub-system contains the cause of a fault while in the second stage, the faulty sub-system or sub-systems are considered one at the time to identify the component failures that represent potential causes.
6.2.5 Validation of the method
In order to validate the results given by the diagnostic system, one should know all the ways a system can fail, that is, the scenarios, and, for each of them, the component failures that cause them. In this way, for each scenario, one can check if the method correctly identifies the actual causes. For both the water tank system and the fuel rig system, a simulation code in C++ was implemented. These codes produce the system scenarios when up to 3 failures occur and they also create the list of potential causes for each scenario. Comparing the results from the simulation and the ones from the diagnostic methods allows the accuracy of the method to be assessed.