Multisensor Data Fusion And Its Application In Induction Motor Fault Diagnosis

Multisensor data fusion is an emerging interdisciplinary technology originated from military field and has received significant attention in various nonmilitary applications. Till now the most developed sections of data fusion are military related, but there is no unified theory framework. As data fusion technology become popular in civilian applications, the need for a general unified concept and theoretical basis is definitely urgent. This dissertation presents an idea to handle uncertainties in induction motor fault diagnosis with multisensor data fusion technology.Based on a detail synthesis of existing results, this dissertation constructs a general framework of data fusion subject. The subject framework includes concepts, models, architectures, and related algorithms. This framework presents a clear view of data fusion for nonmilitary researchers. As uncertainty theories are fundamental in data fusion technology, this dissertation has a detail review of data fusion related uncertainty theories, uncertainty measures and their relations.Various uncertainties exist in traditional induction motor diagnosis methods, the solutions always rely on validity and precision of sensors, accuracy of signal processing methods. Traditional methods have an inherent shortcoming because they mostly base on single sensor information. To reduce or eliminate this kind of uncertainty, multi source information shall be considered. Multisensor data fusion is an efficient method to treat this kind of uncertainty. This dissertation presents a fusion diagnosis system (FDS) structure based on the general data fusion framework. An analysis of FDS design process with system engineering method is provided. Preliminary studying of FDS with formal methods is also given.Fusion applications in motor fault diagnosis are mostly decision fusion. A diagnosis problem may be treated as a decision problem or a classification problem with multi source uncertainties. For a unit motor system, decision fusion may use centralized architecture. The data sources of FDS are fault features, uncertainty information are obtained with feature association. If pattern classification is used for diagnosis, uncertainty information in classification will be represented with proper uncertainty measure, and fusion these uncertainty information gives the diagnosis results.Algorithm varies from applications in data fusion. Performance and complexity of algorithms have much effect on implementation of data fusion algorithms. The dissertation studied the computation issue of evidence theory: the computation of basic probability assignment and Dempster combination rule. Theoretically, it's easier to handle the basic probability assignments, but the computation of Dempster combination rule is #P-complete, a variety of approximate and simplified methods are presented.Method of fault diagnosis with the support vector machine is also studied, the results are promising. Uncertainty information of classification for fault feature is represented by probabilities, and feature is associated for decision fusion.