Research And Application Based On Gaussian Bayesian Networks

As one kind of probability graph models combining probability statistics with graph theory,Bayesian network possesses unique advantages in uncertainty knowledge presentation and inference,therefore has been successfully applied in machine learning,financial analysis,medical diagnosis and many other fields.However,the traditional Bayesian network has strict discrete constraints on data attributes,which makes it tough to model ubiquitous continuous data in real life.In order to break through the limitation,this paper proposed an incremental structure learning strategy and a path inference algorithm for Gaussian Bayesian networks,then integrate them into a big data analysis platform named DODO.The specific work is as follows.For the structure learning problem of Gaussian Bayesian networks,we proposed an incremental updating strategy based on SBN algorithm.This strategy uses sliding window to locally update the established relational coefficient matrix,and sets change threshold as well as relational threshold to determine the final structure,which can improve convergence rate of the objective function and enable SBN algorithm to process the stream data.Contrast experiments on benchmark datasets verified the effectiveness of the proposed strategy.Experimental results showed that it can significantly reduce the computational complexity without loss of accuracy.In addition,compared with classical structure learning algorithms such as PC-stable,GS,HC and MMHC,this strategy often provided higher precision and recall rates.For the inference problem of Gaussian Bayesian networks,we proposed a path inference algorithm considering parent nodes impact.This algorithm simplifies the joint probability solvings based on observed data and the independent properties in Markov blanket,turning global inferences into local inferences.Given evidence node,its estimates with the maximum conditional probability are calculated using decomposition of conditional probability as well as dichotomy.Then the most probable cause is determined by comparing estimations with the true value.Through layer by layer analysis,paths for evidence occurred in the network can be finally confirmed.The variance difference of parent nodes and prior professional knowledge were used to evaluate the performance.Experimental results on benchmark datasets and cement raw mill datasets have demonstrated that this algorithm is able to capture causal relationships more effectively than variable elimination algorithm,at the same time more consistent with professional prior knowledge.Besides,this paper designed the correlation analysis module and fault diagnosis module for an industrial big data analysis platform called DODO.In these modules,Gaussian Bayesian networks methods including the above structure learning and path inference algorithms are integrated,which enable DODO deal with the tasks of correlation analysis and fault diagnosis in actual industrial production.The application in large electric heating units data is studied to prove the availability of DODO and exhibit the overall process of structure learning,parameter learning and path inference in Gaussian Bayesian networks.