Research And Application Of Combat System Failure Prediction Model Based On Small Samples

Combat system is a combination system of hardware and software.The current research on the field of failure prediction for combat system focus on the failure of hardware or devices mainly,ignoring the failure of software,which restricted the overall performance of combat system to a degree.Therefore,taking the submarine combat system for example,and using the metrics-based software defect prediction method,then the small sample problems in combat systems was solved through the method of integration and transfer learning,the combat system failure prediction model based on small samples was put forward.Considering the differences on the performance of multiple operating systems in submarine combat system,a metric-selection method facing operating system stability based on the expert authority scoring for the stability of each system was presented.For the importance of network communication in the integration network,being the center of submarine combat system,a metric-selection method for the stability of communication was presented to make sure the Logistic function between the way of communication and the probability of communication failure through fitting its data distribution.The results show better predictions on both operating system and network communication.With the properties of small samples and high requirement for the reliability of submarine combat system,a cost-sensitive system failure prediction model based on Boosting method was presented,which takes the false positives and false negatives cost into account.The problem of deleting wrongly when selecting attributes was solved by deleting its subset one by one.And cost sensitivity was introduced into the selection of attribute subset,the updating of weight and the selection of adaptive threshold,The results show that false negative rate is reduced most while the false positive rate is increased a little and its overall performance is better than that of previous integrated K-NN model.As for small sample characteristics of few references in one software of combat system but lots of references in others,a system failure prediction method with multi-source transfer learning while considering KL divergence was presented and the model was established.Based on the previous multi-source transfer learning model which focused on prediction error or cost only,the KL divergence was exploited to measure the similarity between source data sets and target data set as a new effect factor to build the model.The results show that the Accuracy and F-Measure value increased great after considering the KL divergence as a factor.The results based on the data sets of submarine combat system show that the model presented is suitable for its failure prediction,and improves the prediction performance of the combat system effectively.