Inference Algorithms And Application To Expert Systems Based On Dynamic Causality Diagram

Dealing with uncertainty knowledge is a core problem of artificial intelligence (AI). So far, there are many approaches addressed. They may be classified into tow categories: probabilistic and non-probabilistic. The probabilistic approaches include the belief network, the markov network, the approaches used in PROSPECTOR, etc. The non-probabilistic approaches include the Dempster-Shafer theory, Fuzzy Set Logic, certainty factor theory (MYCIN), etc. Among the approaches above, Pearl's belief network is the most representative, due to its rigorousness and consistence in theory, the efficient local computation mechanism and intuitive graphical expression of knowledge. Dynamic causality diagram (DCD for short) as a new probabilistic methodology developed from belief network, was first proposed by Professor Zhang Qin in 1994. It introduces some new concepts in its knowledge representation, such as basic event, node event, logical gate and linkage event between different events. The linkage event could be independently represented so that it is more convenient for domain experts to express their knowledge than the belief network. The dynamic and online information could be included in the logic model of the application, because it introduces the symbol logical operation. Thus it is very helpful for the solution of multi-connectedness and cyclic causality in fault diagnoses of complex engineering system. One of the more mature and successful applications of AI is the Expert System. Expert systems are computer programs embodying knowledge about a narrow domain for solving problems related to that domain. The research on DCD based expert system is a key for Causality Diagram methodology to go into practical application from theoretic research. With focus on four features of DCD which are knowledge representation and inference algorithms, analytic algorithms, iteration algorithms and application of expert system, the main contents in the paper are given as follows: In terms of knowledge representaion and inference algorithms, firstly it introduces knowledge representaion style of DCD and inference algorithms of single-value DCD, multi-value DCD and a mixture model of a discrete DCD and a continuous DCD, secondly it points out that the inference algorithm of single-value DCD is a kind of NP-hard problem which have exponential computation complexity, finally it points out that multi-value DCD has higher computational complexity than single-value DCD, so it's necessary to find more efficiency approximate algorithms. In terms of analytic algorithms, with pointing out that there is a critical question in reasoning process of the Multi-value DCD that it does not meet the expectation of consistency and mutex in probabilistic reasoning process if we adopt the original reasoning algorithm of the Single-value DCD in reasoning process of the Multi-value DCD, it presents a reasoning algorithm which contains three new concepts, they are (1) Assuming that cause nodes do not affect result node directly, they only contribute an intensity to the probability distribution of result node, (2) Introducing the concept of unitizing coefficient to solve the problem of unitizing capability during reasoning process. (3) Declaring that there is a property that we can obtain the same inference consequence supposing that they are mutex among all states of all linkage events pointing to the same node in the multi-value causality diagram under the assumption above. The property can ensure all states of any node are mutex each other during reasoning process and hence the reasoning algorithm has been simplified. In terms of iteration algorithms, it proposes two kinds of iteration algorithms which get reasoning results by iterating methodology. One is iteration algorithm based on belief propagation of which the main idea is converting DCD into corresponding belief network firstly and then reasoning based on belief propagation for belief network. The other is iteration algorithm based on genetic algorithm for DCD's most probable explanation (MPE for short) problem, of which the main idea is searching the most probable explanation in possible status combination space of DCD under some evidences by the heuristic searching capability of genetic algorithm. In terms of application of expert system, it mainly introduces the architecture and the developing method of the developing platforms which based on two software developed by our researching group, a diagnosing platform named as "Intelligent Diagnosing Platform Based on Dynamical Causality Diagram DCDA1.0"and a forward reasoning platform named as "Forward Reasoning Platform Based on Dynamical Causality Diagram DCDB1.0". In developing processing we apply the OO methodology and the component techniques to our software. And the multi-layer architecture is adopted to construct an opening framework, so that the software can be used as an easy-to-use developing tool to build expert systems. Using our developing platform, according to the full-size 950MW nuclear power simulator at Beijing Nuclear Power Simulating Training Central in Tsinghua University, the representative fault diagnosis model of the 2nd loop has been build. Through the test of this model, experimentssuccessfully discover the 15 kinds of faults from simulator, and diagnose rapidity, the result is satisfied. In the end this paper summarizes the research work and points out further research focus.