Conditional Evidence Fusion Method And Its Application In Fault Diagnosis

As an Information fusion method, Dempster-Shafer evidence theory taking the advantages of representing, measuring and combining the uncertain information, has been widely used in fault diagnosis field. But, the existing evidence combination methods don't appropriately deal with the relationship among evidences, and have not a good way to represent and measure conditional information, such as if a then b. So, it is difficult to process the conditional information existing in fault diagnosis.This paper analyzes and researches conditional information appeared in fault diagnosis for information fusion, proposes some effective methods for representing, dealing with, and making full use of conditional information. The main works in the thesis are introduced as follows:(1) Give a general overview of fault diagnosis frame using evidence theory; summarize the existing method for obtaining the basic probability assignment and decision-making criterion. Point out some drawbacks of evidence theory and its extensions used in fault diagnosis and conditional information processing.(2) Fault diagnosis method based on the conditional evidence theory. In order to deal with the problems existed in fault diagnosis based on evidence theory, this paper analysis the importance of the priori information in diagnosis decision-making, apply conditional evidence theory to integrate evidence body and prior knowledge, and give a way of generating basic probability assignment (BPA). Finally, an example illustrates the effectiveness of the proposed method.(3) Information method of fault diagnosis based on fuzzy rule reasoning and evidence theory. In the fuzzy-reasoning expert system of fault diagnosis, the fuzzy rule base is often incomplete and uncertainty. It will lead to unreliable and uncertain outputs for some possible inputs. A fusion fault algorithm is presented based on fuzzy rule reasoning and evidence theory. First, by random set formalism of evidence and extension principle of ransom set, fuzzy inputs are mapped to outputs and outputs are integrated to generate basic probability assignments (BPA) measuring the uncertainties of outputs caused by incompletion of system and fuzziness of inputs. Second, Dempster's combination rule is used to fuse BPA coming from different incomplete rule bases and faults can be decided by fusion results. Finally, fault diagnosis tests on machine rotor show that proposed method can effectively improves diagnostic rates. (4) A new evidence updating rule based on conditional event. Updating or conditioning a body of evidence modeled within the DS framework plays an important role in most of artificial intelligence (AI) applications. Now, most of uncertain reasoning models represent the belief of rule by conditional probability, but, it has many limitations when standard conditional probability is used to measure the belief of expert system rule. In this paper, if…then…information is modeled by conditional event, and the belief of rule is measured by conditional event probability, then a new evidence updating method is presented using random conditional event. Two examples are given to illustrate the effeteness of the proposed method.