Business-oriented fault localization based on probabilistic neural networks

Analyzed here is a business-oriented fault localization algorithm based on transitive closure fault propagation model and probability neural networks (PNN). Business-oriented fault localization is to construct a fault propagation model for each large complex software business. This strategy focuses on availability of key business other than scattered fault information. Because of the complex dependent relations between software, hardware, and middleware, fault of one component may propagate into correlated components and produce multiple alarms (symptoms). Transitive closure is possible symptoms domain of faults. Fault diagnosis can be transformed into a classification problem. In practice, PNN is often an excellent pattern classifier, outperforming other classifiers including back propagation (BP). It trains quickly since the training is done in one pass of each training vector, rather than several iterations. In our fault localization algorithm FLPNN, conditional probability of symptom is used as weight of hidden layer, and probability of fault is used as weight of output layer. Input of FLPNN is binary vector which represents the occurrence of symptom or not. In order to adapt the change of fault pattern, incremental learning algorithm of DFLPNN is also investigated. The simulation results show the validity and efficiency of FLPNN compared with MCA+ under lost and spurious symptom circumstances.