Agent-Based Diagnostic Reasoning for Root Cause Analysis
|has title::Agent-Based Diagnostic Reasoning for Root Cause Analysis|
|Master:||project within::Computational Intelligence and Selforganisation|
|Student name:||student name::B.W. Knopper|
|Second reader:||has second reader::Andy van der Mee|
|Company:||has company::CAMS/Force Vision (Royal Navy)|
There is a wide variety of systems aboard the ships of the Koninklijke Marine (Royal Navy). The weapon systems as well as the platform systems are growing more and more complex. If a failure of such a system (or part of such a system) were to occur then the cause of that system failure is hard to detect, simply because of the complexity of these systems. Currently, the diagnosis of a problem on board of a ship is done by mechanics, which means that the quality of diagnosis is dependent on the knowledge of the mechanic. Reduce manning is also becoming an issue. Not only does the domain complexity increase, but there is also a reduction of mechanics available because of cost reductions. In the future however, the Koninklijke Marine would like to have a semi- autonomous system that diagnoses and handles these failures, with the help of information and actions from the mechanics. Also interesting, is that the system may choose to ask certain components for example their status. Based on the combination of information from the mechanic and the components themselves the system can come up with a root cause. This diagnostic system would save the mechanics a significant amount of time and a performance increase can be reached since a diagnostic system might find the root cause to a certain problem faster than a mechanic could. On top of that there is always a “guarantee” of quality, in contrast with mechanics and their varying knowledge and experience (they might be on the ship for the first time!).
The goal of this Graduation Project is to make a prototype of such a diagnostic system. Since this project is done for the Koninklijke Marine there is a classification issue (classifies information). That combined with wanting to scope the project so that it can be concluded within the given time of around 6 months, the diagnostic system will first be developed for a computer network. This computer network consists of: l LAN-cables, m computers, n switches, etc. The advantage of developing the system for a computer network is that there are several components in the network and multiple failures that could be caused by various root causes. The network is also easily simulated, either physical or simulated by a computer. The goal is to develop a system that can find these root causes based on a certain set of failures (since there might be multiple failures when there is only one root cause). The computer network is a good substitute domain since it has overlapping characteristics with the (yet) unclassified application domain of the systems on the ship. As stated before, the computer network has quite a few different components, which can be seen as components from the system of the ship. If the prototype is able to diagnose the computer network for root causes it can be assumed that (if the causal network can be made) the prototype can also diagnose the ships’ systems.
The technique (and architecture) on which the prototype will be based is that of an intelligent agent using reasoning. The agent is going to use a causal network of the computer network to reason with and without assumptions and in the end produce the (possible) root cause(s). This will be the basic goal of the project. To make it more interesting there will be additional goals that can be reached next to the previous mentioned basic goal for this research project. One of them is the use of experience. The idea here is to let the agent use experience to reason even better (faster), using probabilities (to be able to choose between hypotheses better) that were built up in the past. The difference in reasoning here would be that the agent can use probabilities and has the possibility to reason about them. Another additional goal is the automatic generation of the causal network. The way to do this is to use a generic causal network per type component and link those using a design of the computer network. Then you would get an instantiated causal network of linked, through the design, generic causal networks. This would save the Koninklijke Marine a significant amount of time, since they don’t have to adjust the whole causal network per change in the design. The same goes for the components: when a component changes, they don’t have to change all instances of that type of component that occurs in the causal network. Only the specific components causal network has to be changed and the causal network has to be generated again. A third additional goal will be a “focused” causal network. You can imagine this focused causal network as a causal network that is built according to need. It has several aggregation levels (with similar cause and effect rules) and these separate levels will only be built if there is an assumption that the root cause will lie either in that level or in a level even beyond. The agent can build for example the first two layers/levels and decide if it needs the next level to be able to find the root cause. Using a focused causal network would save a significant amount of time in computational effort: first for not having to build the entire causal network, and second for not having to search through the entire causal network. To summarize: after (approximately) six months this project will have produced a diagnostic system, based on an intelligent agent that reasons with and about assumptions, that can do fault detection and in this case does that specifically on a computer network. How much functionality it has will depend on the amount of additional goals that are reached.