Research On Nuclear System Fault Diagnosis Technology For Open Set Scenarios

Nuclear power plants are a very complex system that often malfunctions but cannot be detected in a timely manner.Fault diagnosis technology can identify faults by monitoring data from equipment and instruments,and issue alerts to remind people to handle them in a timely manner.However,current fault diagnosis techniques mainly focus on situations where the fault category is known.In response to the problem of unknown fault identification,this paper proposes a fault open set recognition model,which can identify as many unknown classes as possible without reducing the accuracy of known class recognition.The model consists of three main modules,as follows:First,an improved joint loss function is proposed,which is used to train the convolutional neural network and form a feature extraction module based on the convolutional neural network.The feature extracted by this module has smaller intra class distance and larger inter class distance,reducing the risk of unknown class samples falling into known class regions and helping to improve recognition accuracy in open set scenarios.Secondly,this article designs a statistical classifier based on extreme value theory,which is used to construct a meta recognition system and form a feature correction module.This module establishes a Weibull distribution model based on distance information in the feature space,outputs the probability that the sample belongs to each known class,and multiplies the probability value by weight by the feature output from the last fully connected layer of the network,thereby achieving feature correction.Further improve the accuracy of unknown fault identification.Finally,this paper proposes an open set fault diagnosis module based on hypothesis testing.This module takes the corrected features as the test statistics of each class,finds the quantile of each known class according to the significance level,and determines the rejection domain of each class.It judges whether the maximum value of the features falls into the rejection domain,and gives the diagnosis results at a certain confidence level.The model proposed in this article achieves open set recognition of nuclear system faults.In order to evaluate the effectiveness of the model,comparative experiments were conducted on handwritten digit sets and two sets of simulated kernel system fault datasets.The experimental results showed that the proposed model has good fault diagnosis performance,stable recognition performance on different datasets,and strong generalization ability.