Some Research On Mechanical Fault Diagnosis With Bayesian Network

There are many uncertain factors in detction of system ,especially in fault detection prob-lem of modern mechanical system . we have to face such uncertainty due to the complexity ofdiagnostic object, the limitations of testing method, the ignorance of the system knowledge.Bayesian network is a visualization network model based on probability and graph theory,main advantages of Bayesian network lie in its strong ability of autonomous learning andconcise and visual expression.Bayesian net as a tool to deal with problems of uncertainty ,have extensive applicationin the diagnosis of complex mechanical system ,how ever in the application of bayesian netwe still need to consider many problems such as how to do approximate reasoning effectively, what we need to do when measurement errors exsit, there are another important problem,that's low efficiency caused by too many nodes in bayesian net, for example when Liu andzhang coped with fault diagnosis of the shuttle by the tool of bayesian network ,they put asmany as 150 censors to detect the message of the system ,inevitably too many bottom nodesgreatly increase costs of reasoning.This paper mainly focuses on these problems in application of bayesian network in me-chanical system detection , firstly we do some research about approximate reasoning in bayesiannetwork application , compares the two kinds of random simulation algorithm, and points outthe pros and cons for better implement in reality .Secondly, measurement messege error in mechanical system fault diagnosis is common,but the traditional methods are almost entirely about the denoising of continuous informa-tion, this paper introduced Gibbs sampling method to denoise discrete information on nodesof bayesian net to eliminate measurement error in practical application,and expect it to bepromoted.Finally, we consider to optimize the observation nodes of bayesian network constructedin system fault detection. in practical application, lack of experience and little known aboutsystem make us plant too much sensors to detect system information , sometimes these sensorshave much redundant system information, and too much observe nodes would cause increaseof study and reasoning cost . We take the turbine failure detection for example, discuss theapplication of bayesian network and the optimization of the observational nodes in BNT. we combine the traditional dimension-reduction methods, the principal component analysis andfactor analysis into it , extract fault symptoms , simplify observation nodes, reduce the cost ofresoning, and have achieved certain effect.